NumPy Reference

Release 2.0.0.dev8464

Written by the NumPy community

June 20, 2010

 CONTENTS

1 Array objects 3
1.1 The N-dimensional array (ndarray) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Scalars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
1.3 Data type objects (dtype) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
1.4 Indexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
1.5 Standard array subclasses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
1.6 Masked arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
1.7 The Array Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366

2 Universal functions (ufunc) 371

2.1 Broadcasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
2.2 Output type determination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
2.3 Use of internal buffers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
2.4 Error handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
2.5 Casting Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
2.6 ufunc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
2.7 Available ufuncs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385

3 Routines 389
3.1 Array creation routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
3.2 Array manipulation routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
3.3 Indexing routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
3.4 Data type routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481
3.5 Input and output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 494
3.6 Discrete Fourier Transform (numpy.fft) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
3.7 Linear algebra (numpy.linalg) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
3.8 Random sampling (numpy.random) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
3.9 Sorting and searching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 655
3.10 Logic functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 668
3.11 Binary operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684
3.12 Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 692
3.13 Mathematical functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 712
3.14 Functional programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 772
3.15 Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 777
3.16 Financial functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 790
3.17 Set routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 799
3.18 Window functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803
3.19 Floating point error handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815
3.20 Masked array operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 822

i
 3.21 Numpy-specific help functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 943
3.22 Miscellaneous routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 945
3.23 Test Support (numpy.testing) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 946
3.24 Asserts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 947
3.25 Mathematical functions with automatic domain (numpy.emath) . . . . . . . . . . . . . . . . . . . 957
3.26 Matrix library (numpy.matlib) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 957
3.27 Optionally Scipy-accelerated routines (numpy.dual) . . . . . . . . . . . . . . . . . . . . . . . . . 957
3.28 Numarray compatibility (numpy.numarray) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 958
3.29 Old Numeric compatibility (numpy.oldnumeric) . . . . . . . . . . . . . . . . . . . . . . . . . . 958
3.30 C-Types Foreign Function Interface (numpy.ctypeslib) . . . . . . . . . . . . . . . . . . . . . . 958
3.31 String operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 960

4 Packaging (numpy.distutils) 999

4.1 Modules in numpy.distutils . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 999
4.2 Building Installable C libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1010
4.3 Conversion of .src files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1011

5 Numpy C-API 1013

5.1 Python Types and C-Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1013
5.2 System configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1027
5.3 Data Type API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1029
5.4 Array API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1032
5.5 UFunc API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1064
5.6 Generalized Universal Function API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1070
5.7 Numpy core libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1072

6 Numpy internals 1075

6.1 Numpy C Code Explanations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1075
6.2 Internal organization of numpy arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1082
6.3 Multidimensional Array Indexing Order Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1083

7 Acknowledgements 1085

Bibliography 1087

Index 1097

ii
 NumPy Reference, Release 2.0.0.dev8464

Release
2.0.dev8464
Date
June 20, 2010
This reference manual details functions, modules, and objects included in Numpy, describing what they are and what
they do. For learning how to use NumPy, see also NumPy User Guide (in NumPy User Guide).

CONTENTS 1
NumPy Reference, Release 2.0.0.dev8464

2 CONTENTS
 CHAPTER

ONE

ARRAY OBJECTS

NumPy provides an N-dimensional array type, the ndarray, which describes a collection of items of the same type.
The items can be indexed using for example N integers.
All ndarrays are homogenous: every item takes up the same size block of memory, and all blocks are interpreted in
exactly the same way. How each item in the array is to be interpreted is specified by a separate data-type object, one
of which is associated with every array. In addition to basic types (integers, floats, etc.), the data type objects can also
represent data structures.
An item extracted from an array, e.g., by indexing, is represented by a Python object whose type is one of the array
scalar types built in Numpy. The array scalars allow easy manipulation of also more complicated arrangements of
data.

Figure 1.1: Figure Conceptual diagram showing the relationship between the three fundamental objects used to de-
scribe the data in an array: 1) the ndarray itself, 2) the data-type object that describes the layout of a single fixed-size
element of the array, 3) the array-scalar Python object that is returned when a single element of the array is accessed.

1.1 The N-dimensional array (ndarray)

An ndarray is a (usually fixed-size) multidimensional container of items of the same type and size. The number
of dimensions and items in an array is defined by its shape, which is a tuple of N positive integers that specify
the sizes of each dimension. The type of items in the array is specified by a separate data-type object (dtype), one of
which is associated with each ndarray.
As with other container objects in Python, the contents of an ndarray can be accessed and modified by indexing
or slicing the array (using, for example, N integers), and via the methods and attributes of the ndarray. Different

3
NumPy Reference, Release 2.0.0.dev8464

ndarrays can share the same data, so that changes made in one ndarray may be visible in another. That is, an
ndarray can be a view to another ndarray, and the data it is referring to is taken care of by the base ndarray.
ndarrays can also be views to memory owned by Python strings or objects implementing the buffer or array
interfaces.
Example
A 2-dimensional array of size 2 x 3, composed of 4-byte integer elements:

>>> x = np.array([[1, 2, 3], [4, 5, 6]], np.int32)

>>> type(x)
<type numpy.ndarray>
>>> x.shape
(2, 3)
>>> x.dtype
dtype(int32)

The array can be indexed using Python container-like syntax:

>>> x[1,2] # i.e., the element of x in the *second* row, *third*

column, namely, 6.

For example slicing can produce views of the array:

>>> y = x[:,1]
>>> y
array([2, 5])
>>> y[0] = 9 # this also changes the corresponding element in x
>>> y
array([9, 5])
>>> x
array([[1, 9, 3],
[4, 5, 6]])

1.1.1 Constructing arrays

New arrays can be constructed using the routines detailed in Array creation routines, and also by using the low-level
ndarray constructor:
ndarray An array object represents a multidimensional, homogeneous array of fixed-size items.
class ndarray()
An array object represents a multidimensional, homogeneous array of fixed-size items. An associated data-type
object describes the format of each element in the array (its byte-order, how many bytes it occupies in memory,
whether it is an integer, a floating point number, or something else, etc.)
Arrays should be constructed using array, zeros or empty (refer to the See Also section below). The parameters
given here refer to a low-level method (ndarray(...)) for instantiating an array.
For more information, refer to the numpy module and examine the the methods and attributes of an array.
Parameters
(for the __new__ method; see Notes below) :
shape : tuple of ints
Shape of created array.
dtype : data-type, optional

4 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Any object that can be interpreted as a numpy data type.

buffer : object exposing buffer interface, optional
Used to fill the array with data.
offset : int, optional
Offset of array data in buffer.
strides : tuple of ints, optional
Strides of data in memory.
order : {C, F}, optional
Row-major or column-major order.
See Also:

array
Construct an array.
zeros
Create an array, each element of which is zero.
empty
Create an array, but leave its allocated memory unchanged (i.e., it contains garbage).
dtype
Create a data-type.

Notes
There are two modes of creating an array using __new__:
1.If buffer is None, then only shape, dtype, and order are used.
2.If buffer is an object exposing the buffer interface, then all keywords are interpreted.
No __init__ method is needed because the array is fully initialized after the __new__ method.

Examples
These examples illustrate the low-level ndarray constructor. Refer to the See Also section above for easier ways
of constructing an ndarray.
First mode, buffer is None:

>>> np.ndarray(shape=(2,2), dtype=float, order=F)

array([[ -1.13698227e+002, 4.25087011e-303],
[ 2.88528414e-306, 3.27025015e-309]]) #random

Second mode:

>>> np.ndarray((2,), buffer=np.array([1,2,3]),

... offset=np.int_().itemsize,
... dtype=int) # offset = 1*itemsize, i.e. skip first element
array([2, 3])

1.1. The N-dimensional array (ndarray) 5

NumPy Reference, Release 2.0.0.dev8464

Attributes
T Same as self.transpose(), except that self is returned if self.ndim < 2.
data Python buffer object pointing to the start of the arrays data.
dtype Data-type of the arrays elements.
flags Information about the memory layout of the array.
flat A 1-D iterator over the array.
imag The imaginary part of the array.
real The real part of the array.
size Number of elements in the array.
itemsize Length of one array element in bytes.
nbytes Total bytes consumed by the elements of the array.
ndim Number of array dimensions.
shape Tuple of array dimensions.
strides Tuple of bytes to step in each dimension when traversing an array.
ctypes An object to simplify the interaction of the array with the ctypes module.
base Base object if memory is from some other object.
T
Same as self.transpose(), except that self is returned if self.ndim < 2.

Examples

>>> x = np.array([[1.,2.],[3.,4.]])
>>> x
array([[ 1., 2.],
[ 3., 4.]])
>>> x.T
array([[ 1., 3.],
[ 2., 4.]])
>>> x = np.array([1.,2.,3.,4.])
>>> x
array([ 1., 2., 3., 4.])
>>> x.T
array([ 1., 2., 3., 4.])

data
Python buffer object pointing to the start of the arrays data.
dtype
Data-type of the arrays elements.
Parameters
None :
Returns
d : numpy dtype object
See Also:
numpy.dtype

Examples

>>> x
array([[0, 1],
[2, 3]])
>>> x.dtype
dtype(int32)

6 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> type(x.dtype)
<type numpy.dtype>

flags
Information about the memory layout of the array.

Notes
The flags object can be accessed dictionary-like (as in a.flags[WRITEABLE]), or by using lower-
cased attribute names (as in a.flags.writeable). Short flag names are only supported in dictionary
access.
Only the UPDATEIFCOPY, WRITEABLE, and ALIGNED flags can be changed by the user, via direct
assignment to the attribute or dictionary entry, or by calling ndarray.setflags.
The array flags cannot be set arbitrarily:
UPDATEIFCOPY can only be set False.
ALIGNED can only be set True if the data is truly aligned.
WRITEABLE can only be set True if the array owns its own memory or the ultimate owner of the
memory exposes a writeable buffer interface or is a string.

Attributes
flat
A 1-D iterator over the array.
This is a numpy.flatiter instance, which acts similarly to, but is not a subclass of, Pythons built-in
iterator object.
See Also:

flatten
Return a copy of the array collapsed into one dimension.

flatiter

Examples

>>> x = np.arange(1, 7).reshape(2, 3)

>>> x
array([[1, 2, 3],
[4, 5, 6]])
>>> x.flat[3]
4
>>> x.T
array([[1, 4],
[2, 5],
[3, 6]])
>>> x.T.flat[3]
5
>>> type(x.flat)
<type numpy.flatiter>

An assignment example:

1.1. The N-dimensional array (ndarray) 7

NumPy Reference, Release 2.0.0.dev8464

>>> x.flat = 3; x
array([[3, 3, 3],
[3, 3, 3]])
>>> x.flat[[1,4]] = 1; x
array([[3, 1, 3],
[3, 1, 3]])

imag
The imaginary part of the array.

Examples

>>> x = np.sqrt([1+0j, 0+1j])

>>> x.imag
array([ 0. , 0.70710678])
>>> x.imag.dtype
dtype(float64)

real
The real part of the array.
See Also:

numpy.real
equivalent function

Examples

>>> x = np.sqrt([1+0j, 0+1j])

>>> x.real
array([ 1. , 0.70710678])
>>> x.real.dtype
dtype(float64)

size
Number of elements in the array.
Equivalent to np.prod(a.shape), i.e., the product of the arrays dimensions.

Examples

>>> x = np.zeros((3, 5, 2), dtype=np.complex128)

>>> x.size
30
>>> np.prod(x.shape)
30

itemsize
Length of one array element in bytes.

Examples

>>> x = np.array([1,2,3], dtype=np.float64)

>>> x.itemsize
8
>>> x = np.array([1,2,3], dtype=np.complex128)

8 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> x.itemsize
16

nbytes
Total bytes consumed by the elements of the array.

Notes
Does not include memory consumed by non-element attributes of the array object.

Examples

>>> x = np.zeros((3,5,2), dtype=np.complex128)

>>> x.nbytes
480
>>> np.prod(x.shape) * x.itemsize
480

ndim
Number of array dimensions.

Examples

>>> x = np.array([1, 2, 3])

>>> x.ndim
1
>>> y = np.zeros((2, 3, 4))
>>> y.ndim
3

shape
Tuple of array dimensions.

Notes
May be used to reshape the array, as long as this would not require a change in the total number of
elements

Examples

>>> x = np.array([1, 2, 3, 4])

>>> x.shape
(4,)
>>> y = np.zeros((2, 3, 4))
>>> y.shape
(2, 3, 4)
>>> y.shape = (3, 8)
>>> y
array([[ 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0.]])
>>> y.shape = (3, 6)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: total size of new array must be unchanged

strides
Tuple of bytes to step in each dimension when traversing an array.

1.1. The N-dimensional array (ndarray) 9

NumPy Reference, Release 2.0.0.dev8464

The byte offset of element (i[0], i[1], ..., i[n]) in an array a is:

offset = sum(np.array(i) * a.strides)

A more detailed explanation of strides can be found in the ndarray.rst file in the NumPy reference guide.
See Also:
numpy.lib.stride_tricks.as_strided

Notes
Imagine an array of 32-bit integers (each 4 bytes):

x = np.array([[0, 1, 2, 3, 4],
[5, 6, 7, 8, 9]], dtype=np.int32)

This array is stored in memory as 40 bytes, one after the other (known as a contiguous block of memory).
The strides of an array tell us how many bytes we have to skip in memory to move to the next position
along a certain axis. For example, we have to skip 4 bytes (1 value) to move to the next column, but 20
bytes (5 values) to get to the same position in the next row. As such, the strides for the array x will be
(20, 4).

Examples

>>> y = np.reshape(np.arange(2*3*4), (2,3,4))

>>> y
array([[[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]],
[[12, 13, 14, 15],
[16, 17, 18, 19],
[20, 21, 22, 23]]])
>>> y.strides
(48, 16, 4)
>>> y[1,1,1]
17
>>> offset=sum(y.strides * np.array((1,1,1)))
>>> offset/y.itemsize
17

>>> x = np.reshape(np.arange(5*6*7*8), (5,6,7,8)).transpose(2,3,1,0)

>>> x.strides
(32, 4, 224, 1344)
>>> i = np.array([3,5,2,2])
>>> offset = sum(i * x.strides)
>>> x[3,5,2,2]
813
>>> offset / x.itemsize
813

ctypes
An object to simplify the interaction of the array with the ctypes module.
This attribute creates an object that makes it easier to use arrays when calling shared libraries with the
ctypes module. The returned object has, among others, data, shape, and strides attributes (see Notes
below) which themselves return ctypes objects that can be used as arguments to a shared library.

10 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Parameters
None :
Returns
c : Python object
Possessing attributes data, shape, strides, etc.
See Also:
numpy.ctypeslib

Notes
Below are the public attributes of this object which were documented in Guide to NumPy (we have
omitted undocumented public attributes, as well as documented private attributes):
data: A pointer to the memory area of the array as a Python integer. This memory area may contain
data that is not aligned, or not in correct byte-order. The memory area may not even be writeable.
The array flags and data-type of this array should be respected when passing this attribute to arbitrary
C-code to avoid trouble that can include Python crashing. User Beware! The value of this attribute is
exactly the same as self._array_interface_[data][0].
shape (c_intp*self.ndim): A ctypes array of length self.ndim where the basetype is the C-integer
corresponding to dtype(p) on this platform. This base-type could be c_int, c_long, or c_longlong
depending on the platform. The c_intp type is defined accordingly in numpy.ctypeslib. The ctypes
array contains the shape of the underlying array.
strides (c_intp*self.ndim): A ctypes array of length self.ndim where the basetype is the same as for
the shape attribute. This ctypes array contains the strides information from the underlying array.
This strides information is important for showing how many bytes must be jumped to get to the next
element in the array.
data_as(obj): Return the data pointer cast to a particular c-types object. For example, calling
self._as_parameter_ is equivalent to self.data_as(ctypes.c_void_p). Perhaps you want to use the data
as a pointer to a ctypes array of floating-point data: self.data_as(ctypes.POINTER(ctypes.c_double)).
shape_as(obj): Return the shape tuple as an array of some other c-types type. For example:
self.shape_as(ctypes.c_short).
strides_as(obj): Return the strides tuple as an array of some other c-types type. For example:
self.strides_as(ctypes.c_longlong).
Be careful using the ctypes attribute - especially on temporary arrays or arrays constructed on the fly. For
example, calling (a+b).ctypes.data_as(ctypes.c_void_p) returns a pointer to memory that
is invalid because the array created as (a+b) is deallocated before the next Python statement. You can avoid
this problem using either c=a+b or ct=(a+b).ctypes. In the latter case, ct will hold a reference to
the array until ct is deleted or re-assigned.
If the ctypes module is not available, then the ctypes attribute of array objects still returns something useful,
but ctypes objects are not returned and errors may be raised instead. In particular, the object will still have
the as parameter attribute which will return an integer equal to the data attribute.

Examples

>>> import ctypes

>>> x
array([[0, 1],
[2, 3]])
>>> x.ctypes.data
30439712

1.1. The N-dimensional array (ndarray) 11

NumPy Reference, Release 2.0.0.dev8464

>>> x.ctypes.data_as(ctypes.POINTER(ctypes.c_long))
<ctypes.LP_c_long object at 0x01F01300>
>>> x.ctypes.data_as(ctypes.POINTER(ctypes.c_long)).contents
c_long(0)
>>> x.ctypes.data_as(ctypes.POINTER(ctypes.c_longlong)).contents
c_longlong(4294967296L)
>>> x.ctypes.shape
<numpy.core._internal.c_long_Array_2 object at 0x01FFD580>
>>> x.ctypes.shape_as(ctypes.c_long)
<numpy.core._internal.c_long_Array_2 object at 0x01FCE620>
>>> x.ctypes.strides
<numpy.core._internal.c_long_Array_2 object at 0x01FCE620>
>>> x.ctypes.strides_as(ctypes.c_longlong)
<numpy.core._internal.c_longlong_Array_2 object at 0x01F01300>

base
Base object if memory is from some other object.

Examples
The base of an array that owns its memory is None:

>>> x = np.array([1,2,3,4])
>>> x.base is None
True

Slicing creates a view, whose memory is shared with x:

>>> y = x[2:]
>>> y.base is x
True

Methods

all([axis, out]) Returns True if all elements evaluate to True.

any([axis, out]) Returns True if any of the elements of a evaluate to True.
argmax([axis, out]) Return indices of the maximum values along the given axis.
argmin([axis, out]) Return indices of the minimum values along the given axis of a.
argsort([axis, kind, order]) Returns the indices that would sort this array.
astype(t) Copy of the array, cast to a specified type.
byteswap(inplace) Swap the bytes of the array elements
choose(choices[, out, mode]) Use an index array to construct a new array from a set of choices.
clip(a_min, a_max[, out]) Return an array whose values are limited to [a_min, a_max].
compress(condition[, axis, out]) Return selected slices of this array along given axis.
conj() Complex-conjugate all elements.
conjugate() Return the complex conjugate, element-wise.
copy([order]) Return a copy of the array.
cumprod([axis, dtype, out]) Return the cumulative product of the elements along the given axis.
cumsum([axis, dtype, out]) Return the cumulative sum of the elements along the given axis.
diagonal([offset, axis1, axis2]) Return specified diagonals.
dot
dump(file) Dump a pickle of the array to the specified file.
dumps() Returns the pickle of the array as a string.
fill(value) Fill the array with a scalar value.
flatten([order]) Return a copy of the array collapsed into one dimension.
Continued on next page

12 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Table 1.1 continued from previous page

getfield(dtype, offset) Returns a field of the given array as a certain type.
item(*args) Copy an element of an array to a standard Python scalar and return it.
itemset
max([axis, out]) Return the maximum along a given axis.
mean([axis, dtype, out]) Returns the average of the array elements along given axis.
min([axis, out]) Return the minimum along a given axis.
newbyteorder([new_order]) Return the array with the same data viewed with a different byte order.
nonzero() Return the indices of the elements that are non-zero.
prod([axis, dtype, out]) Return the product of the array elements over the given axis
ptp([axis, out]) Peak to peak (maximum - minimum) value along a given axis.
put(indices, values[, mode]) Set a.flat[n] = values[n] for all n in indices.
ravel() Return a flattened array.
repeat(repeats[, axis]) Repeat elements of an array.
reshape(shape[, order]) Returns an array containing the same data with a new shape.
resize(new_shape[, refcheck]) Change shape and size of array in-place.
round([decimals, out]) Return an array rounded a to the given number of decimals.
searchsorted(v[, side]) Find indices where elements of v should be inserted in a to maintain order.
setfield(val, dtype[, offset]) Put a value into a specified place in a field defined by a data-type.
setflags([write, align, uic]) Set array flags WRITEABLE, ALIGNED, and UPDATEIFCOPY, respectively.
sort([axis, kind, order]) Sort an array, in-place.
squeeze() Remove single-dimensional entries from the shape of a.
std([axis, dtype, out, ddof]) Returns the standard deviation of the array elements along given axis.
sum([axis, dtype, out]) Return the sum of the array elements over the given axis.
swapaxes(axis1, axis2) Return a view of the array with axis1 and axis2 interchanged.
take(indices[, axis, out, mode]) Return an array formed from the elements of a at the given indices.
tofile(fid[, sep, format]) Write array to a file as text or binary (default).
tolist() Return the array as a (possibly nested) list.
tostring([order]) Construct a Python string containing the raw data bytes in the array.
trace([offset, axis1, axis2, dtype, out]) Return the sum along diagonals of the array.
transpose(*axes) Returns a view of the array with axes transposed.
var([axis, dtype, out, ddof]) Returns the variance of the array elements, along given axis.
view([dtype, type]) New view of array with the same data.

all(axis=None, out=None)
Returns True if all elements evaluate to True.
Refer to numpy.all for full documentation.
See Also:

numpy.all
equivalent function

any(axis=None, out=None)
Returns True if any of the elements of a evaluate to True.
Refer to numpy.any for full documentation.
See Also:

numpy.any
equivalent function

1.1. The N-dimensional array (ndarray) 13

NumPy Reference, Release 2.0.0.dev8464

argmax(axis=None, out=None)
Return indices of the maximum values along the given axis.
Refer to numpy.argmax for full documentation.
See Also:

numpy.argmax
equivalent function

argmin(axis=None, out=None)
Return indices of the minimum values along the given axis of a.
Refer to numpy.argmin for detailed documentation.
See Also:

numpy.argmin
equivalent function

argsort(axis=-1, kind=quicksort, order=None)

Returns the indices that would sort this array.
Refer to numpy.argsort for full documentation.
See Also:

numpy.argsort
equivalent function

astype(t)
Copy of the array, cast to a specified type.
Parameters
t : string or dtype
Typecode or data-type to which the array is cast.

Examples

>>> x = np.array([1, 2, 2.5])

>>> x
array([ 1. , 2. , 2.5])

>>> x.astype(int)
array([1, 2, 2])

byteswap(inplace)
Swap the bytes of the array elements
Toggle between low-endian and big-endian data representation by returning a byteswapped array, option-
ally swapped in-place.
Parameters
inplace: bool, optional :
If True, swap bytes in-place, default is False.
Returns
out: ndarray :

14 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

The byteswapped array. If inplace is True, this is a view to self.

Examples

>>> A = np.array([1, 256, 8755], dtype=np.int16)

>>> map(hex, A)
[0x1, 0x100, 0x2233]
>>> A.byteswap(True)
array([ 256, 1, 13090], dtype=int16)
>>> map(hex, A)
[0x100, 0x1, 0x3322]

Arrays of strings are not swapped

>>> A = np.array([ceg, fac])

>>> A.byteswap()
array([ceg, fac],
dtype=|S3)

choose(choices, out=None, mode=raise)

Use an index array to construct a new array from a set of choices.
Refer to numpy.choose for full documentation.
See Also:

numpy.choose
equivalent function

clip(a_min, a_max, out=None)

Return an array whose values are limited to [a_min, a_max].
Refer to numpy.clip for full documentation.
See Also:

numpy.clip
equivalent function

compress(condition, axis=None, out=None)

Return selected slices of this array along given axis.
Refer to numpy.compress for full documentation.
See Also:

numpy.compress
equivalent function

conj()
Complex-conjugate all elements.
Refer to numpy.conjugate for full documentation.
See Also:

numpy.conjugate
equivalent function

1.1. The N-dimensional array (ndarray) 15

NumPy Reference, Release 2.0.0.dev8464

conjugate()
Return the complex conjugate, element-wise.
Refer to numpy.conjugate for full documentation.
See Also:

numpy.conjugate
equivalent function

copy(order=C)
Return a copy of the array.
Parameters
order : {C, F, A}, optional
By default, the result is stored in C-contiguous (row-major) order in memory. If order is
F, the result has Fortran (column-major) order. If order is A (Any), then the result
has the same order as the input.

Examples

>>> x = np.array([[1,2,3],[4,5,6]], order=F)

>>> y = x.copy()

>>> x.fill(0)

>>> x
array([[0, 0, 0],
[0, 0, 0]])

>>> y
array([[1, 2, 3],
[4, 5, 6]])

>>> y.flags[C_CONTIGUOUS]
True

cumprod(axis=None, dtype=None, out=None)

Return the cumulative product of the elements along the given axis.
Refer to numpy.cumprod for full documentation.
See Also:

numpy.cumprod
equivalent function

cumsum(axis=None, dtype=None, out=None)

Return the cumulative sum of the elements along the given axis.
Refer to numpy.cumsum for full documentation.
See Also:

16 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

numpy.cumsum
equivalent function

diagonal(offset=0, axis1=0, axis2=1)

Return specified diagonals.
Refer to numpy.diagonal for full documentation.
See Also:

numpy.diagonal
equivalent function

dot()

dump(file)
Dump a pickle of the array to the specified file. The array can be read back with pickle.load or numpy.load.
Parameters
file : str
A string naming the dump file.
dumps()
Returns the pickle of the array as a string. pickle.loads or numpy.loads will convert the string back to an
array.
Parameters
None :
fill(value)
Fill the array with a scalar value.
Parameters
value : scalar
All elements of a will be assigned this value.

Examples

>>> a = np.array([1, 2])

>>> a.fill(0)
>>> a
array([0, 0])
>>> a = np.empty(2)
>>> a.fill(1)
>>> a
array([ 1., 1.])

flatten(order=C)
Return a copy of the array collapsed into one dimension.
Parameters
order : {C, F}, optional
Whether to flatten in C (row-major) or Fortran (column-major) order. The default is C.
Returns
y : ndarray
A copy of the input array, flattened to one dimension.

1.1. The N-dimensional array (ndarray) 17

NumPy Reference, Release 2.0.0.dev8464

See Also:

ravel
Return a flattened array.
flat
A 1-D flat iterator over the array.

Examples

>>> a = np.array([[1,2], [3,4]])

>>> a.flatten()
array([1, 2, 3, 4])
>>> a.flatten(F)
array([1, 3, 2, 4])

getfield(dtype, offset)
Returns a field of the given array as a certain type.
A field is a view of the array data with each itemsize determined by the given type and the offset into the
current array, i.e. from offset * dtype.itemsize to (offset+1) * dtype.itemsize.
Parameters
dtype : str
String denoting the data type of the field.
offset : int
Number of dtype.itemsizes to skip before beginning the element view.

Examples

>>> x = np.diag([1.+1.j]*2)
>>> x
array([[ 1.+1.j, 0.+0.j],
[ 0.+0.j, 1.+1.j]])
>>> x.dtype
dtype(complex128)

>>> x.getfield(complex64, 0) # Note how this != x

array([[ 0.+1.875j, 0.+0.j ],
[ 0.+0.j , 0.+1.875j]], dtype=complex64)

>>> x.getfield(complex64,1) # Note how different this is than x

array([[ 0. +5.87173204e-39j, 0. +0.00000000e+00j],
[ 0. +0.00000000e+00j, 0. +5.87173204e-39j]], dtype=complex64)

>>> x.getfield(complex128, 0) # == x
array([[ 1.+1.j, 0.+0.j],
[ 0.+0.j, 1.+1.j]])

If the argument dtype is the same as x.dtype, then offset != 0 raises a ValueError:

>>> x.getfield(complex128, 1)
Traceback (most recent call last):

18 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

File "<stdin>", line 1, in <module>

ValueError: Need 0 <= offset <= 0 for requested type but received offset = 1

>>> x.getfield(float64, 0)
array([[ 1., 0.],
[ 0., 1.]])

>>> x.getfield(float64, 1)
array([[ 1.77658241e-307, 0.00000000e+000],
[ 0.00000000e+000, 1.77658241e-307]])

item(*args)
Copy an element of an array to a standard Python scalar and return it.
Parameters
*args : Arguments (variable number and type)
none: in this case, the method only works for arrays with one element (a.size == 1), which
element is copied into a standard Python scalar object and returned.
int_type: this argument is interpreted as a flat index into the array, specifying which ele-
ment to copy and return.
tuple of int_types: functions as does a single int_type argument, except that the argument
is interpreted as an nd-index into the array.
Returns
z : Standard Python scalar object
A copy of the specified element of the array as a suitable Python scalar

Notes
When the data type of a is longdouble or clongdouble, item() returns a scalar array object because there is
no available Python scalar that would not lose information. Void arrays return a buffer object for item(),
unless fields are defined, in which case a tuple is returned.
item is very similar to a[args], except, instead of an array scalar, a standard Python scalar is returned. This
can be useful for speeding up access to elements of the array and doing arithmetic on elements of the array
using Pythons optimized math.

Examples

>>> x = np.random.randint(9, size=(3, 3))

>>> x
array([[3, 1, 7],
[2, 8, 3],
[8, 5, 3]])
>>> x.item(3)
2
>>> x.item(7)
5
>>> x.item((0, 1))
1
>>> x.item((2, 2))
3

itemset()

1.1. The N-dimensional array (ndarray) 19

NumPy Reference, Release 2.0.0.dev8464

max(axis=None, out=None)
Return the maximum along a given axis.
Refer to numpy.amax for full documentation.
See Also:

numpy.amax
equivalent function

mean(axis=None, dtype=None, out=None)

Returns the average of the array elements along given axis.
Refer to numpy.mean for full documentation.
See Also:

numpy.mean
equivalent function

min(axis=None, out=None)
Return the minimum along a given axis.
Refer to numpy.amin for full documentation.
See Also:

numpy.amin
equivalent function

newbyteorder(new_order=S)
Return the array with the same data viewed with a different byte order.
Equivalent to:

arr.view(arr.dtype.newbytorder(new_order))

Changes are also made in all fields and sub-arrays of the array data type.
Parameters
new_order : string, optional
Byte order to force; a value from the byte order specifications above. new_order codes
can be any of:

* S - swap dtype from current to opposite endian

* {<, L} - little endian
* {>, B} - big endian
* {=, N} - native order
* {|, I} - ignore (no change to byte order)

The default value (S) results in swapping the current byte order. The code does a
case-insensitive check on the first letter of new_order for the alternatives above. For
example, any of B or b or biggish are valid to specify big-endian.
Returns
new_arr : array
New array object with the dtype reflecting given change to the byte order.

20 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

nonzero()
Return the indices of the elements that are non-zero.
Refer to numpy.nonzero for full documentation.
See Also:

numpy.nonzero
equivalent function

prod(axis=None, dtype=None, out=None)

Return the product of the array elements over the given axis
Refer to numpy.prod for full documentation.
See Also:

numpy.prod
equivalent function

ptp(axis=None, out=None)
Peak to peak (maximum - minimum) value along a given axis.
Refer to numpy.ptp for full documentation.
See Also:

numpy.ptp
equivalent function

put(indices, values, mode=raise)

Set a.flat[n] = values[n] for all n in indices.
Refer to numpy.put for full documentation.
See Also:

numpy.put
equivalent function

ravel([order])
Return a flattened array.
Refer to numpy.ravel for full documentation.
See Also:

numpy.ravel
equivalent function
ndarray.flat
a flat iterator on the array.

repeat(repeats, axis=None)
Repeat elements of an array.
Refer to numpy.repeat for full documentation.
See Also:

1.1. The N-dimensional array (ndarray) 21

NumPy Reference, Release 2.0.0.dev8464

numpy.repeat
equivalent function

reshape(shape, order=C)
Returns an array containing the same data with a new shape.
Refer to numpy.reshape for full documentation.
See Also:

numpy.reshape
equivalent function

resize(new_shape, refcheck=True)
Change shape and size of array in-place.
Parameters
new_shape : tuple of ints, or n ints
Shape of resized array.
refcheck : bool, optional
If False, reference count will not be checked. Default is True.
Returns
None :
Raises
ValueError :
If a does not own its own data or references or views to it exist, and the data memory
must be changed.
SystemError :
If the order keyword argument is specified. This behaviour is a bug in NumPy.
See Also:

resize
Return a new array with the specified shape.

Notes
This reallocates space for the data area if necessary.
Only contiguous arrays (data elements consecutive in memory) can be resized.
The purpose of the reference count check is to make sure you do not use this array as a buffer for another
Python object and then reallocate the memory. However, reference counts can increase in other ways so if
you are sure that you have not shared the memory for this array with another Python object, then you may
safely set refcheck to False.

Examples
Shrinking an array: array is flattened (in the order that the data are stored in memory), resized, and re-
shaped:

22 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> a = np.array([[0, 1], [2, 3]], order=C)

>>> a.resize((2, 1))
>>> a
array([[0],
[1]])

>>> a = np.array([[0, 1], [2, 3]], order=F)

>>> a.resize((2, 1))
>>> a
array([[0],
[2]])

Enlarging an array: as above, but missing entries are filled with zeros:

>>> b = np.array([[0, 1], [2, 3]])

>>> b.resize(2, 3) # new_shape parameter doesnt have to be a tuple
>>> b
array([[0, 1, 2],
[3, 0, 0]])

Referencing an array prevents resizing...

>>> c = a
>>> a.resize((1, 1))
...
ValueError: cannot resize an array that has been referenced ...

Unless refcheck is False:

>>> a.resize((1, 1), refcheck=False)

>>> a
array([[0]])
>>> c
array([[0]])

round(decimals=0, out=None)
Return an array rounded a to the given number of decimals.
Refer to numpy.around for full documentation.
See Also:

numpy.around
equivalent function

searchsorted(v, side=left)
Find indices where elements of v should be inserted in a to maintain order.
For full documentation, see numpy.searchsorted
See Also:

numpy.searchsorted
equivalent function

1.1. The N-dimensional array (ndarray) 23

NumPy Reference, Release 2.0.0.dev8464

setfield(val, dtype, offset=0)

Put a value into a specified place in a field defined by a data-type.
Place val into as field defined by dtype and beginning offset bytes into the field.
Parameters
val : object
Value to be placed in field.
dtype : dtype object
Data-type of the field in which to place val.
offset : int, optional
The number of bytes into the field at which to place val.
Returns
None :
See Also:
getfield

Examples

>>> x = np.eye(3)
>>> x.getfield(np.float64)
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])
>>> x.setfield(3, np.int32)
>>> x.getfield(np.int32)
array([[3, 3, 3],
[3, 3, 3],
[3, 3, 3]])
>>> x
array([[ 1.00000000e+000, 1.48219694e-323, 1.48219694e-323],
[ 1.48219694e-323, 1.00000000e+000, 1.48219694e-323],
[ 1.48219694e-323, 1.48219694e-323, 1.00000000e+000]])
>>> x.setfield(np.eye(3), np.int32)
>>> x
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])

setflags(write=None, align=None, uic=None)

Set array flags WRITEABLE, ALIGNED, and UPDATEIFCOPY, respectively.
These Boolean-valued flags affect how numpy interprets the memory area used by a (see Notes below).
The ALIGNED flag can only be set to True if the data is actually aligned according to the type. The
UPDATEIFCOPY flag can never be set to True. The flag WRITEABLE can only be set to True if the array
owns its own memory, or the ultimate owner of the memory exposes a writeable buffer interface, or is a
string. (The exception for string is made so that unpickling can be done without copying memory.)
Parameters
write : bool, optional
Describes whether or not a can be written to.
align : bool, optional

24 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Describes whether or not a is aligned properly for its type.

uic : bool, optional
Describes whether or not a is a copy of another base array.

Notes
Array flags provide information about how the memory area used for the array is to be interpreted. There
are 6 Boolean flags in use, only three of which can be changed by the user: UPDATEIFCOPY, WRITE-
ABLE, and ALIGNED.
WRITEABLE (W) the data area can be written to;
ALIGNED (A) the data and strides are aligned appropriately for the hardware (as determined by the com-
piler);
UPDATEIFCOPY (U) this array is a copy of some other array (referenced by .base). When this array is
deallocated, the base array will be updated with the contents of this array.
All flags can be accessed using their first (upper case) letter as well as the full name.

Examples

>>> y
array([[3, 1, 7],
[2, 0, 0],
[8, 5, 9]])
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : True
WRITEABLE : True
ALIGNED : True
UPDATEIFCOPY : False
>>> y.setflags(write=0, align=0)
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : True
WRITEABLE : False
ALIGNED : False
UPDATEIFCOPY : False
>>> y.setflags(uic=1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: cannot set UPDATEIFCOPY flag to True

sort(axis=-1, kind=quicksort, order=None)

Sort an array, in-place.
Parameters
axis : int, optional
Axis along which to sort. Default is -1, which means sort along the last axis.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm. Default is quicksort.
order : list, optional

1.1. The N-dimensional array (ndarray) 25

NumPy Reference, Release 2.0.0.dev8464

When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
See Also:

numpy.sort
Return a sorted copy of an array.
argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in sorted array.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = np.array([[1,4], [3,1]])

>>> a.sort(axis=1)
>>> a
array([[1, 4],
[1, 3]])
>>> a.sort(axis=0)
>>> a
array([[1, 3],
[1, 4]])

Use the order keyword to specify a field to use when sorting a structured array:

>>> a = np.array([(a, 2), (c, 1)], dtype=[(x, S1), (y, int)])

>>> a.sort(order=y)
>>> a
array([(c, 1), (a, 2)],
dtype=[(x, |S1), (y, <i4)])

squeeze()
Remove single-dimensional entries from the shape of a.
Refer to numpy.squeeze for full documentation.
See Also:

numpy.squeeze
equivalent function

std(axis=None, dtype=None, out=None, ddof=0)

Returns the standard deviation of the array elements along given axis.
Refer to numpy.std for full documentation.
See Also:

numpy.std
equivalent function

26 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

sum(axis=None, dtype=None, out=None)

Return the sum of the array elements over the given axis.
Refer to numpy.sum for full documentation.
See Also:

numpy.sum
equivalent function

swapaxes(axis1, axis2)
Return a view of the array with axis1 and axis2 interchanged.
Refer to numpy.swapaxes for full documentation.
See Also:

numpy.swapaxes
equivalent function

take(indices, axis=None, out=None, mode=raise)

Return an array formed from the elements of a at the given indices.
Refer to numpy.take for full documentation.
See Also:

numpy.take
equivalent function

tofile(fid, sep="", format="%s")

Write array to a file as text or binary (default).
Data is always written in C order, independent of the order of a. The data produced by this method can
be recovered using the function fromfile().
Parameters
fid : file or str
An open file object, or a string containing a filename.
sep : str
Separator between array items for text output. If (empty), a binary file is written,
equivalent to file.write(a.tostring()).
format : str
Format string for text file output. Each entry in the array is formatted to text by first
converting it to the closest Python type, and then using format % item.

Notes
This is a convenience function for quick storage of array data. Information on endianness and precision
is lost, so this method is not a good choice for files intended to archive data or transport data between
machines with different endianness. Some of these problems can be overcome by outputting the data as
text files, at the expense of speed and file size.
tolist()
Return the array as a (possibly nested) list.

1.1. The N-dimensional array (ndarray) 27

NumPy Reference, Release 2.0.0.dev8464

Return a copy of the array data as a (nested) Python list. Data items are converted to the nearest compatible
Python type.
Parameters
none :
Returns
y : list
The possibly nested list of array elements.

Notes
The array may be recreated, a = np.array(a.tolist()).

Examples

>>> a = np.array([1, 2])

>>> a.tolist()
[1, 2]
>>> a = np.array([[1, 2], [3, 4]])
>>> list(a)
[array([1, 2]), array([3, 4])]
>>> a.tolist()
[[1, 2], [3, 4]]

tostring(order=C)
Construct a Python string containing the raw data bytes in the array.
Constructs a Python string showing a copy of the raw contents of data memory. The string can be produced
in either C or Fortran, or Any order (the default is C-order). Any order means C-order unless the
F_CONTIGUOUS flag in the array is set, in which case it means Fortran order.
Parameters
order : {C, F, None}, optional
Order of the data for multidimensional arrays: C, Fortran, or the same as for the original
array.
Returns
s : str
A Python string exhibiting a copy of as raw data.

Examples

>>> x = np.array([[0, 1], [2, 3]])

>>> x.tostring()
\x00\x00\x00\x00\x01\x00\x00\x00\x02\x00\x00\x00\x03\x00\x00\x00
>>> x.tostring(C) == x.tostring()
True
>>> x.tostring(F)
\x00\x00\x00\x00\x02\x00\x00\x00\x01\x00\x00\x00\x03\x00\x00\x00

trace(offset=0, axis1=0, axis2=1, dtype=None, out=None)

Return the sum along diagonals of the array.
Refer to numpy.trace for full documentation.
See Also:

28 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

numpy.trace
equivalent function

transpose(*axes)
Returns a view of the array with axes transposed.
For a 1-D array, this has no effect. (To change between column and row vectors, first cast the 1-D ar-
ray into a matrix object.) For a 2-D array, this is the usual matrix transpose. For an n-D array, if axes
are given, their order indicates how the axes are permuted (see Examples). If axes are not provided
and a.shape = (i[0], i[1], ... i[n-2], i[n-1]), then a.transpose().shape
= (i[n-1], i[n-2], ... i[1], i[0]).
Parameters
axes : None, tuple of ints, or n ints
None or no argument: reverses the order of the axes.
tuple of ints: i in the j-th place in the tuple means as i-th axis becomes a.transpose()s
j-th axis.
n ints: same as an n-tuple of the same ints (this form is intended simply as a convenience
alternative to the tuple form)
Returns
out : ndarray
View of a, with axes suitably permuted.
See Also:

ndarray.T
Array property returning the array transposed.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> a
array([[1, 2],
[3, 4]])
>>> a.transpose()
array([[1, 3],
[2, 4]])
>>> a.transpose((1, 0))
array([[1, 3],
[2, 4]])
>>> a.transpose(1, 0)
array([[1, 3],
[2, 4]])

var(axis=None, dtype=None, out=None, ddof=0)

Returns the variance of the array elements, along given axis.
Refer to numpy.var for full documentation.
See Also:

numpy.var
equivalent function

1.1. The N-dimensional array (ndarray) 29

NumPy Reference, Release 2.0.0.dev8464

view(dtype=None, type=None)
New view of array with the same data.
Parameters
dtype : data-type
Data-type descriptor of the returned view, e.g. float32 or int16.
type : python type
Type of the returned view, e.g. ndarray or matrix.

Notes
a.view() is used two different ways.
a.view(some_dtype) or a.view(dtype=some_dtype) constructs a view of the arrays memory with a different
dtype. This can cause a reinterpretation of the bytes of memory.
a.view(ndarray_subclass), or a.view(type=ndarray_subclass), just returns an instance of ndarray_subclass
that looks at the same array (same shape, dtype, etc.). This does not cause a reinterpretation of the memory.

Examples

>>> x = np.array([(1, 2)], dtype=[(a, np.int8), (b, np.int8)])

Viewing array data using a different type and dtype:

>>> y = x.view(dtype=np.int16, type=np.matrix)

>>> y
matrix([[513]], dtype=int16)
>>> print type(y)
<class numpy.matrixlib.defmatrix.matrix>

Creating a view on a structured array so it can be used in calculations

>>> x = np.array([(1, 2),(3,4)], dtype=[(a, np.int8), (b, np.int8)])

>>> xv = x.view(dtype=np.int8).reshape(-1,2)
>>> xv
array([[1, 2],
[3, 4]], dtype=int8)
>>> xv.mean(0)
array([ 2., 3.])

Making changes to the view changes the underlying array

>>> xv[0,1] = 20
>>> print x
[(1, 20) (3, 4)]

Using a view to convert an array to a record array:

>>> z = x.view(np.recarray)
>>> z.a
array([1], dtype=int8)

Views share data:

30 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> x[0] = (9, 10)

>>> z[0]
(9, 10)

1.1.2 Indexing arrays

Arrays can be indexed using an extended Python slicing syntax, array[selection]. Similar syntax is also used
for accessing fields in a record array.
See Also:
Array Indexing.

1.1.3 Internal memory layout of an ndarray

An instance of class ndarray consists of a contiguous one-dimensional segment of computer memory (owned by
the array, or by some other object), combined with an indexing scheme that maps N integers into the location of an
item in the block. The ranges in which the indices can vary is specified by the shape of the array. How many
bytes each item takes and how the bytes are interpreted is defined by the data-type object associated with the array.
A segment of memory is inherently 1-dimensional, and there are many different schemes for arranging the items of
an N-dimensional array in a 1-dimensional block. Numpy is flexible, and ndarray objects can accommodate any
strided indexing scheme. In a strided scheme, the N-dimensional index (n0 , n1 , ..., nN 1 ) corresponds to the offset
(in bytes):
N
X 1
noffset = sk nk
k=0

from the beginning of the memory block associated with the array. Here, sk are integers which specify the strides
of the array. The column-major order (used, for example, in the Fortran language and in Matlab) and row-major order
(used in C) schemes are just specific kinds of strided scheme, and correspond to the strides:
k1
Y N
Y 1
scolumn
k = dj , srow
k = dj .
j=0 j=k+1

where dj = self.itemsize * self.shape[j].

Both the C and Fortran orders are contiguous, i.e., single-segment, memory layouts, in which every part of the memory
block can be accessed by some combination of the indices.
Data in new ndarrays is in the row-major (C) order, unless otherwise specified, but, for example, basic array slicing
often produces views in a different scheme.
Note: Several algorithms in NumPy work on arbitrarily strided arrays. However, some algorithms require single-
segment arrays. When an irregularly strided array is passed in to such algorithms, a copy is automatically made.

1.1.4 Array attributes

Array attributes reflect information that is intrinsic to the array itself. Generally, accessing an array through its at-
tributes allows you to get and sometimes set intrinsic properties of the array without creating a new array. The exposed
attributes are the core parts of an array and only some of them can be reset meaningfully without creating a new array.
Information on each attribute is given below.

1.1. The N-dimensional array (ndarray) 31

NumPy Reference, Release 2.0.0.dev8464

Memory layout

The following attributes contain information about the memory layout of the array:
ndarray.flags Information about the memory layout of the array.
ndarray.shape Tuple of array dimensions.
ndarray.strides Tuple of bytes to step in each dimension when traversing an array.
ndarray.ndim Number of array dimensions.
ndarray.data Python buffer object pointing to the start of the arrays data.
ndarray.size Number of elements in the array.
ndarray.itemsize Length of one array element in bytes.
ndarray.nbytes Total bytes consumed by the elements of the array.
ndarray.base Base object if memory is from some other object.
flags
Information about the memory layout of the array.

Notes
The flags object can be accessed dictionary-like (as in a.flags[WRITEABLE]), or by using lowercased
attribute names (as in a.flags.writeable). Short flag names are only supported in dictionary access.
Only the UPDATEIFCOPY, WRITEABLE, and ALIGNED flags can be changed by the user, via direct assign-
ment to the attribute or dictionary entry, or by calling ndarray.setflags.
The array flags cannot be set arbitrarily:
UPDATEIFCOPY can only be set False.
ALIGNED can only be set True if the data is truly aligned.
WRITEABLE can only be set True if the array owns its own memory or the ultimate owner of the memory
exposes a writeable buffer interface or is a string.

Attributes
shape
Tuple of array dimensions.

Notes
May be used to reshape the array, as long as this would not require a change in the total number of elements

Examples

>>> x = np.array([1, 2, 3, 4])

>>> x.shape
(4,)
>>> y = np.zeros((2, 3, 4))
>>> y.shape
(2, 3, 4)
>>> y.shape = (3, 8)
>>> y
array([[ 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0.]])
>>> y.shape = (3, 6)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: total size of new array must be unchanged

32 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

strides
Tuple of bytes to step in each dimension when traversing an array.
The byte offset of element (i[0], i[1], ..., i[n]) in an array a is:

offset = sum(np.array(i) * a.strides)

A more detailed explanation of strides can be found in the ndarray.rst file in the NumPy reference guide.
See Also:
numpy.lib.stride_tricks.as_strided

Notes
Imagine an array of 32-bit integers (each 4 bytes):

x = np.array([[0, 1, 2, 3, 4],
[5, 6, 7, 8, 9]], dtype=np.int32)

This array is stored in memory as 40 bytes, one after the other (known as a contiguous block of memory). The
strides of an array tell us how many bytes we have to skip in memory to move to the next position along a certain
axis. For example, we have to skip 4 bytes (1 value) to move to the next column, but 20 bytes (5 values) to get
to the same position in the next row. As such, the strides for the array x will be (20, 4).

Examples

>>> y = np.reshape(np.arange(2*3*4), (2,3,4))

>>> y
array([[[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]],
[[12, 13, 14, 15],
[16, 17, 18, 19],
[20, 21, 22, 23]]])
>>> y.strides
(48, 16, 4)
>>> y[1,1,1]
17
>>> offset=sum(y.strides * np.array((1,1,1)))
>>> offset/y.itemsize
17

>>> x = np.reshape(np.arange(5*6*7*8), (5,6,7,8)).transpose(2,3,1,0)

>>> x.strides
(32, 4, 224, 1344)
>>> i = np.array([3,5,2,2])
>>> offset = sum(i * x.strides)
>>> x[3,5,2,2]
813
>>> offset / x.itemsize
813

ndim
Number of array dimensions.

1.1. The N-dimensional array (ndarray) 33

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.array([1, 2, 3])

>>> x.ndim
1
>>> y = np.zeros((2, 3, 4))
>>> y.ndim
3

data
Python buffer object pointing to the start of the arrays data.
size
Number of elements in the array.
Equivalent to np.prod(a.shape), i.e., the product of the arrays dimensions.

Examples

>>> x = np.zeros((3, 5, 2), dtype=np.complex128)

>>> x.size
30
>>> np.prod(x.shape)
30

itemsize
Length of one array element in bytes.

Examples

>>> x = np.array([1,2,3], dtype=np.float64)

>>> x.itemsize
8
>>> x = np.array([1,2,3], dtype=np.complex128)
>>> x.itemsize
16

nbytes
Total bytes consumed by the elements of the array.

Notes
Does not include memory consumed by non-element attributes of the array object.

Examples

>>> x = np.zeros((3,5,2), dtype=np.complex128)

>>> x.nbytes
480
>>> np.prod(x.shape) * x.itemsize
480

base
Base object if memory is from some other object.

Examples
The base of an array that owns its memory is None:

34 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> x = np.array([1,2,3,4])
>>> x.base is None
True

Slicing creates a view, whose memory is shared with x:

>>> y = x[2:]
>>> y.base is x
True

Data type

See Also:
Data type objects
The data type object associated with the array can be found in the dtype attribute:
ndarray.dtype Data-type of the arrays elements.
dtype
Data-type of the arrays elements.
Parameters
None :
Returns
d : numpy dtype object
See Also:
numpy.dtype

Examples

>>> x
array([[0, 1],
[2, 3]])
>>> x.dtype
dtype(int32)
>>> type(x.dtype)
<type numpy.dtype>

Other attributes

ndarray.T Same as self.transpose(), except that self is returned if self.ndim < 2.

ndarray.real The real part of the array.
ndarray.imag The imaginary part of the array.
ndarray.flat A 1-D iterator over the array.
ndarray.ctypes An object to simplify the interaction of the array with the ctypes module.
__array_priority__
T
Same as self.transpose(), except that self is returned if self.ndim < 2.

1.1. The N-dimensional array (ndarray) 35

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.array([[1.,2.],[3.,4.]])
>>> x
array([[ 1., 2.],
[ 3., 4.]])
>>> x.T
array([[ 1., 3.],
[ 2., 4.]])
>>> x = np.array([1.,2.,3.,4.])
>>> x
array([ 1., 2., 3., 4.])
>>> x.T
array([ 1., 2., 3., 4.])

real
The real part of the array.
See Also:

numpy.real
equivalent function

Examples

>>> x = np.sqrt([1+0j, 0+1j])

>>> x.real
array([ 1. , 0.70710678])
>>> x.real.dtype
dtype(float64)

imag
The imaginary part of the array.

Examples

>>> x = np.sqrt([1+0j, 0+1j])

>>> x.imag
array([ 0. , 0.70710678])
>>> x.imag.dtype
dtype(float64)

flat
A 1-D iterator over the array.
This is a numpy.flatiter instance, which acts similarly to, but is not a subclass of, Pythons built-in iterator
object.
See Also:

flatten
Return a copy of the array collapsed into one dimension.

flatiter

36 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.arange(1, 7).reshape(2, 3)

>>> x
array([[1, 2, 3],
[4, 5, 6]])
>>> x.flat[3]
4
>>> x.T
array([[1, 4],
[2, 5],
[3, 6]])
>>> x.T.flat[3]
5
>>> type(x.flat)
<type numpy.flatiter>

An assignment example:

>>> x.flat = 3; x
array([[3, 3, 3],
[3, 3, 3]])
>>> x.flat[[1,4]] = 1; x
array([[3, 1, 3],
[3, 1, 3]])

ctypes
An object to simplify the interaction of the array with the ctypes module.
This attribute creates an object that makes it easier to use arrays when calling shared libraries with the ctypes
module. The returned object has, among others, data, shape, and strides attributes (see Notes below) which
themselves return ctypes objects that can be used as arguments to a shared library.
Parameters
None :
Returns
c : Python object
Possessing attributes data, shape, strides, etc.
See Also:
numpy.ctypeslib

Notes
Below are the public attributes of this object which were documented in Guide to NumPy (we have omitted
undocumented public attributes, as well as documented private attributes):
data: A pointer to the memory area of the array as a Python integer. This memory area may contain data
that is not aligned, or not in correct byte-order. The memory area may not even be writeable. The array
flags and data-type of this array should be respected when passing this attribute to arbitrary C-code to avoid
trouble that can include Python crashing. User Beware! The value of this attribute is exactly the same as
self._array_interface_[data][0].
shape (c_intp*self.ndim): A ctypes array of length self.ndim where the basetype is the C-integer corre-
sponding to dtype(p) on this platform. This base-type could be c_int, c_long, or c_longlong depending
on the platform. The c_intp type is defined accordingly in numpy.ctypeslib. The ctypes array contains the
shape of the underlying array.

1.1. The N-dimensional array (ndarray) 37

NumPy Reference, Release 2.0.0.dev8464

strides (c_intp*self.ndim): A ctypes array of length self.ndim where the basetype is the same as for the
shape attribute. This ctypes array contains the strides information from the underlying array. This strides
information is important for showing how many bytes must be jumped to get to the next element in the
array.
data_as(obj): Return the data pointer cast to a particular c-types object. For example, calling
self._as_parameter_ is equivalent to self.data_as(ctypes.c_void_p). Perhaps you want to use the data as
a pointer to a ctypes array of floating-point data: self.data_as(ctypes.POINTER(ctypes.c_double)).
shape_as(obj): Return the shape tuple as an array of some other c-types type. For example:
self.shape_as(ctypes.c_short).
strides_as(obj): Return the strides tuple as an array of some other c-types type. For example:
self.strides_as(ctypes.c_longlong).
Be careful using the ctypes attribute - especially on temporary arrays or arrays constructed on the fly. For
example, calling (a+b).ctypes.data_as(ctypes.c_void_p) returns a pointer to memory that is
invalid because the array created as (a+b) is deallocated before the next Python statement. You can avoid this
problem using either c=a+b or ct=(a+b).ctypes. In the latter case, ct will hold a reference to the array
until ct is deleted or re-assigned.
If the ctypes module is not available, then the ctypes attribute of array objects still returns something useful, but
ctypes objects are not returned and errors may be raised instead. In particular, the object will still have the as
parameter attribute which will return an integer equal to the data attribute.

Examples

>>> import ctypes

>>> x
array([[0, 1],
[2, 3]])
>>> x.ctypes.data
30439712
>>> x.ctypes.data_as(ctypes.POINTER(ctypes.c_long))
<ctypes.LP_c_long object at 0x01F01300>
>>> x.ctypes.data_as(ctypes.POINTER(ctypes.c_long)).contents
c_long(0)
>>> x.ctypes.data_as(ctypes.POINTER(ctypes.c_longlong)).contents
c_longlong(4294967296L)
>>> x.ctypes.shape
<numpy.core._internal.c_long_Array_2 object at 0x01FFD580>
>>> x.ctypes.shape_as(ctypes.c_long)
<numpy.core._internal.c_long_Array_2 object at 0x01FCE620>
>>> x.ctypes.strides
<numpy.core._internal.c_long_Array_2 object at 0x01FCE620>
>>> x.ctypes.strides_as(ctypes.c_longlong)
<numpy.core._internal.c_longlong_Array_2 object at 0x01F01300>

Array interface

See Also:
The Array Interface.
__array_interface__ Python-side of the array interface
__array_struct__ C-side of the array interface

38 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

ctypes foreign function interface

ndarray.ctypes An object to simplify the interaction of the array with the ctypes module.
ctypes
An object to simplify the interaction of the array with the ctypes module.
This attribute creates an object that makes it easier to use arrays when calling shared libraries with the ctypes
module. The returned object has, among others, data, shape, and strides attributes (see Notes below) which
themselves return ctypes objects that can be used as arguments to a shared library.
Parameters
None :
Returns
c : Python object
Possessing attributes data, shape, strides, etc.
See Also:
numpy.ctypeslib

Notes
Below are the public attributes of this object which were documented in Guide to NumPy (we have omitted
undocumented public attributes, as well as documented private attributes):
data: A pointer to the memory area of the array as a Python integer. This memory area may contain data
that is not aligned, or not in correct byte-order. The memory area may not even be writeable. The array
flags and data-type of this array should be respected when passing this attribute to arbitrary C-code to avoid
trouble that can include Python crashing. User Beware! The value of this attribute is exactly the same as
self._array_interface_[data][0].
shape (c_intp*self.ndim): A ctypes array of length self.ndim where the basetype is the C-integer corre-
sponding to dtype(p) on this platform. This base-type could be c_int, c_long, or c_longlong depending
on the platform. The c_intp type is defined accordingly in numpy.ctypeslib. The ctypes array contains the
shape of the underlying array.
strides (c_intp*self.ndim): A ctypes array of length self.ndim where the basetype is the same as for the
shape attribute. This ctypes array contains the strides information from the underlying array. This strides
information is important for showing how many bytes must be jumped to get to the next element in the
array.
data_as(obj): Return the data pointer cast to a particular c-types object. For example, calling
self._as_parameter_ is equivalent to self.data_as(ctypes.c_void_p). Perhaps you want to use the data as
a pointer to a ctypes array of floating-point data: self.data_as(ctypes.POINTER(ctypes.c_double)).
shape_as(obj): Return the shape tuple as an array of some other c-types type. For example:
self.shape_as(ctypes.c_short).
strides_as(obj): Return the strides tuple as an array of some other c-types type. For example:
self.strides_as(ctypes.c_longlong).
Be careful using the ctypes attribute - especially on temporary arrays or arrays constructed on the fly. For
example, calling (a+b).ctypes.data_as(ctypes.c_void_p) returns a pointer to memory that is
invalid because the array created as (a+b) is deallocated before the next Python statement. You can avoid this
problem using either c=a+b or ct=(a+b).ctypes. In the latter case, ct will hold a reference to the array
until ct is deleted or re-assigned.

1.1. The N-dimensional array (ndarray) 39

NumPy Reference, Release 2.0.0.dev8464

If the ctypes module is not available, then the ctypes attribute of array objects still returns something useful, but
ctypes objects are not returned and errors may be raised instead. In particular, the object will still have the as
parameter attribute which will return an integer equal to the data attribute.

Examples

>>> import ctypes

>>> x
array([[0, 1],
[2, 3]])
>>> x.ctypes.data
30439712
>>> x.ctypes.data_as(ctypes.POINTER(ctypes.c_long))
<ctypes.LP_c_long object at 0x01F01300>
>>> x.ctypes.data_as(ctypes.POINTER(ctypes.c_long)).contents
c_long(0)
>>> x.ctypes.data_as(ctypes.POINTER(ctypes.c_longlong)).contents
c_longlong(4294967296L)
>>> x.ctypes.shape
<numpy.core._internal.c_long_Array_2 object at 0x01FFD580>
>>> x.ctypes.shape_as(ctypes.c_long)
<numpy.core._internal.c_long_Array_2 object at 0x01FCE620>
>>> x.ctypes.strides
<numpy.core._internal.c_long_Array_2 object at 0x01FCE620>
>>> x.ctypes.strides_as(ctypes.c_longlong)
<numpy.core._internal.c_longlong_Array_2 object at 0x01F01300>

1.1.5 Array methods

An ndarray object has many methods which operate on or with the array in some fashion, typically returning an
array result. These methods are briefly explained below. (Each methods docstring has a more complete description.)
For the following methods there are also corresponding functions in numpy: all, any, argmax, argmin,
argsort, choose, clip, compress, copy, cumprod, cumsum, diagonal, imag, max, mean, min,
nonzero, prod, ptp, put, ravel, real, repeat, reshape, round, searchsorted, sort, squeeze,
std, sum, swapaxes, take, trace, transpose, var.

40 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Array conversion

ndarray.item(*args) Copy an element of an array to a standard Python scalar and return it.
ndarray.tolist() Return the array as a (possibly nested) list.
ndarray.itemset
ndarray.tostring([order]) Construct a Python string containing the raw data bytes in the array.
ndarray.tofile(fid[, sep, Write array to a file as text or binary (default).
format])
ndarray.dump(file) Dump a pickle of the array to the specified file.
ndarray.dumps() Returns the pickle of the array as a string.
ndarray.astype(t) Copy of the array, cast to a specified type.
ndarray.byteswap(inplace) Swap the bytes of the array elements
ndarray.copy([order]) Return a copy of the array.
ndarray.view([dtype, type]) New view of array with the same data.
ndarray.getfield(dtype, offset) Returns a field of the given array as a certain type.
ndarray.setflags([write, align, Set array flags WRITEABLE, ALIGNED, and UPDATEIFCOPY,
uic]) respectively.
ndarray.fill(value) Fill the array with a scalar value.
item(*args)
Copy an element of an array to a standard Python scalar and return it.
Parameters
*args : Arguments (variable number and type)
none: in this case, the method only works for arrays with one element (a.size == 1), which
element is copied into a standard Python scalar object and returned.
int_type: this argument is interpreted as a flat index into the array, specifying which element
to copy and return.
tuple of int_types: functions as does a single int_type argument, except that the argument is
interpreted as an nd-index into the array.
Returns
z : Standard Python scalar object
A copy of the specified element of the array as a suitable Python scalar

Notes
When the data type of a is longdouble or clongdouble, item() returns a scalar array object because there is no
available Python scalar that would not lose information. Void arrays return a buffer object for item(), unless
fields are defined, in which case a tuple is returned.
item is very similar to a[args], except, instead of an array scalar, a standard Python scalar is returned. This can
be useful for speeding up access to elements of the array and doing arithmetic on elements of the array using
Pythons optimized math.

Examples

>>> x = np.random.randint(9, size=(3, 3))

>>> x
array([[3, 1, 7],
[2, 8, 3],
[8, 5, 3]])
>>> x.item(3)
2
>>> x.item(7)

1.1. The N-dimensional array (ndarray) 41

NumPy Reference, Release 2.0.0.dev8464

5
>>> x.item((0, 1))
1
>>> x.item((2, 2))
3

tolist()
Return the array as a (possibly nested) list.
Return a copy of the array data as a (nested) Python list. Data items are converted to the nearest compatible
Python type.
Parameters
none :
Returns
y : list
The possibly nested list of array elements.

Notes
The array may be recreated, a = np.array(a.tolist()).

Examples

>>> a = np.array([1, 2])

>>> a.tolist()
[1, 2]
>>> a = np.array([[1, 2], [3, 4]])
>>> list(a)
[array([1, 2]), array([3, 4])]
>>> a.tolist()
[[1, 2], [3, 4]]

itemset()

tostring(order=C)
Construct a Python string containing the raw data bytes in the array.
Constructs a Python string showing a copy of the raw contents of data memory. The string can be produced
in either C or Fortran, or Any order (the default is C-order). Any order means C-order unless the
F_CONTIGUOUS flag in the array is set, in which case it means Fortran order.
Parameters
order : {C, F, None}, optional
Order of the data for multidimensional arrays: C, Fortran, or the same as for the original
array.
Returns
s : str
A Python string exhibiting a copy of as raw data.

Examples

>>> x = np.array([[0, 1], [2, 3]])

>>> x.tostring()

42 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

\x00\x00\x00\x00\x01\x00\x00\x00\x02\x00\x00\x00\x03\x00\x00\x00
>>> x.tostring(C) == x.tostring()
True
>>> x.tostring(F)
\x00\x00\x00\x00\x02\x00\x00\x00\x01\x00\x00\x00\x03\x00\x00\x00

tofile(fid, sep="", format="%s")

Write array to a file as text or binary (default).
Data is always written in C order, independent of the order of a. The data produced by this method can be
recovered using the function fromfile().
Parameters
fid : file or str
An open file object, or a string containing a filename.
sep : str
Separator between array items for text output. If (empty), a binary file is written,
equivalent to file.write(a.tostring()).
format : str
Format string for text file output. Each entry in the array is formatted to text by first
converting it to the closest Python type, and then using format % item.

Notes
This is a convenience function for quick storage of array data. Information on endianness and precision is lost,
so this method is not a good choice for files intended to archive data or transport data between machines with
different endianness. Some of these problems can be overcome by outputting the data as text files, at the expense
of speed and file size.
dump(file)
Dump a pickle of the array to the specified file. The array can be read back with pickle.load or numpy.load.
Parameters
file : str
A string naming the dump file.
dumps()
Returns the pickle of the array as a string. pickle.loads or numpy.loads will convert the string back to an array.
Parameters
None :
astype(t)
Copy of the array, cast to a specified type.
Parameters
t : string or dtype
Typecode or data-type to which the array is cast.

Examples

>>> x = np.array([1, 2, 2.5])

>>> x
array([ 1. , 2. , 2.5])

1.1. The N-dimensional array (ndarray) 43

NumPy Reference, Release 2.0.0.dev8464

>>> x.astype(int)
array([1, 2, 2])

byteswap(inplace)
Swap the bytes of the array elements
Toggle between low-endian and big-endian data representation by returning a byteswapped array, optionally
swapped in-place.
Parameters
inplace: bool, optional :
If True, swap bytes in-place, default is False.
Returns
out: ndarray :
The byteswapped array. If inplace is True, this is a view to self.

Examples

>>> A = np.array([1, 256, 8755], dtype=np.int16)

>>> map(hex, A)
[0x1, 0x100, 0x2233]
>>> A.byteswap(True)
array([ 256, 1, 13090], dtype=int16)
>>> map(hex, A)
[0x100, 0x1, 0x3322]

Arrays of strings are not swapped

>>> A = np.array([ceg, fac])

>>> A.byteswap()
array([ceg, fac],
dtype=|S3)

copy(order=C)
Return a copy of the array.
Parameters
order : {C, F, A}, optional
By default, the result is stored in C-contiguous (row-major) order in memory. If order is
F, the result has Fortran (column-major) order. If order is A (Any), then the result
has the same order as the input.

Examples

>>> x = np.array([[1,2,3],[4,5,6]], order=F)

>>> y = x.copy()

>>> x.fill(0)

44 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> x
array([[0, 0, 0],
[0, 0, 0]])

>>> y
array([[1, 2, 3],
[4, 5, 6]])

>>> y.flags[C_CONTIGUOUS]
True

view(dtype=None, type=None)
New view of array with the same data.
Parameters
dtype : data-type
Data-type descriptor of the returned view, e.g. float32 or int16.
type : python type
Type of the returned view, e.g. ndarray or matrix.

Notes
a.view() is used two different ways.
a.view(some_dtype) or a.view(dtype=some_dtype) constructs a view of the arrays memory with a different
dtype. This can cause a reinterpretation of the bytes of memory.
a.view(ndarray_subclass), or a.view(type=ndarray_subclass), just returns an instance of ndarray_subclass that
looks at the same array (same shape, dtype, etc.). This does not cause a reinterpretation of the memory.

Examples

>>> x = np.array([(1, 2)], dtype=[(a, np.int8), (b, np.int8)])

Viewing array data using a different type and dtype:

>>> y = x.view(dtype=np.int16, type=np.matrix)

>>> y
matrix([[513]], dtype=int16)
>>> print type(y)
<class numpy.matrixlib.defmatrix.matrix>

Creating a view on a structured array so it can be used in calculations

>>> x = np.array([(1, 2),(3,4)], dtype=[(a, np.int8), (b, np.int8)])

>>> xv = x.view(dtype=np.int8).reshape(-1,2)
>>> xv
array([[1, 2],
[3, 4]], dtype=int8)
>>> xv.mean(0)
array([ 2., 3.])

Making changes to the view changes the underlying array

1.1. The N-dimensional array (ndarray) 45

NumPy Reference, Release 2.0.0.dev8464

>>> xv[0,1] = 20
>>> print x
[(1, 20) (3, 4)]

Using a view to convert an array to a record array:

>>> z = x.view(np.recarray)
>>> z.a
array([1], dtype=int8)

Views share data:

>>> x[0] = (9, 10)

>>> z[0]
(9, 10)

getfield(dtype, offset)
Returns a field of the given array as a certain type.
A field is a view of the array data with each itemsize determined by the given type and the offset into the current
array, i.e. from offset * dtype.itemsize to (offset+1) * dtype.itemsize.
Parameters
dtype : str
String denoting the data type of the field.
offset : int
Number of dtype.itemsizes to skip before beginning the element view.

Examples

>>> x = np.diag([1.+1.j]*2)
>>> x
array([[ 1.+1.j, 0.+0.j],
[ 0.+0.j, 1.+1.j]])
>>> x.dtype
dtype(complex128)

>>> x.getfield(complex64, 0) # Note how this != x

array([[ 0.+1.875j, 0.+0.j ],
[ 0.+0.j , 0.+1.875j]], dtype=complex64)

>>> x.getfield(complex64,1) # Note how different this is than x

array([[ 0. +5.87173204e-39j, 0. +0.00000000e+00j],
[ 0. +0.00000000e+00j, 0. +5.87173204e-39j]], dtype=complex64)

>>> x.getfield(complex128, 0) # == x
array([[ 1.+1.j, 0.+0.j],
[ 0.+0.j, 1.+1.j]])

If the argument dtype is the same as x.dtype, then offset != 0 raises a ValueError:

46 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> x.getfield(complex128, 1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: Need 0 <= offset <= 0 for requested type but received offset = 1

>>> x.getfield(float64, 0)
array([[ 1., 0.],
[ 0., 1.]])

>>> x.getfield(float64, 1)
array([[ 1.77658241e-307, 0.00000000e+000],
[ 0.00000000e+000, 1.77658241e-307]])

setflags(write=None, align=None, uic=None)

Set array flags WRITEABLE, ALIGNED, and UPDATEIFCOPY, respectively.
These Boolean-valued flags affect how numpy interprets the memory area used by a (see Notes below). The
ALIGNED flag can only be set to True if the data is actually aligned according to the type. The UPDATEIF-
COPY flag can never be set to True. The flag WRITEABLE can only be set to True if the array owns its own
memory, or the ultimate owner of the memory exposes a writeable buffer interface, or is a string. (The exception
for string is made so that unpickling can be done without copying memory.)
Parameters
write : bool, optional
Describes whether or not a can be written to.
align : bool, optional
Describes whether or not a is aligned properly for its type.
uic : bool, optional
Describes whether or not a is a copy of another base array.

Notes
Array flags provide information about how the memory area used for the array is to be interpreted. There are
6 Boolean flags in use, only three of which can be changed by the user: UPDATEIFCOPY, WRITEABLE, and
ALIGNED.
WRITEABLE (W) the data area can be written to;
ALIGNED (A) the data and strides are aligned appropriately for the hardware (as determined by the compiler);
UPDATEIFCOPY (U) this array is a copy of some other array (referenced by .base). When this array is deallo-
cated, the base array will be updated with the contents of this array.
All flags can be accessed using their first (upper case) letter as well as the full name.

Examples

>>> y
array([[3, 1, 7],
[2, 0, 0],
[8, 5, 9]])
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : True

1.1. The N-dimensional array (ndarray) 47

NumPy Reference, Release 2.0.0.dev8464

WRITEABLE : True
ALIGNED : True
UPDATEIFCOPY : False
>>> y.setflags(write=0, align=0)
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : True
WRITEABLE : False
ALIGNED : False
UPDATEIFCOPY : False
>>> y.setflags(uic=1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: cannot set UPDATEIFCOPY flag to True

fill(value)
Fill the array with a scalar value.
Parameters
value : scalar
All elements of a will be assigned this value.

Examples

>>> a = np.array([1, 2])

>>> a.fill(0)
>>> a
array([0, 0])
>>> a = np.empty(2)
>>> a.fill(1)
>>> a
array([ 1., 1.])

Shape manipulation

For reshape, resize, and transpose, the single tuple argument may be replaced with n integers which will be interpreted
as an n-tuple.
ndarray.reshape(shape[, order]) Returns an array containing the same data with a new shape.
ndarray.resize(new_shape[, refcheck]) Change shape and size of array in-place.
ndarray.transpose(*axes) Returns a view of the array with axes transposed.
ndarray.swapaxes(axis1, axis2) Return a view of the array with axis1 and axis2 interchanged.
ndarray.flatten([order]) Return a copy of the array collapsed into one dimension.
ndarray.ravel() Return a flattened array.
ndarray.squeeze() Remove single-dimensional entries from the shape of a.
reshape(shape, order=C)
Returns an array containing the same data with a new shape.
Refer to numpy.reshape for full documentation.
See Also:

numpy.reshape
equivalent function

48 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

resize(new_shape, refcheck=True)
Change shape and size of array in-place.
Parameters
new_shape : tuple of ints, or n ints
Shape of resized array.
refcheck : bool, optional
If False, reference count will not be checked. Default is True.
Returns
None :
Raises
ValueError :
If a does not own its own data or references or views to it exist, and the data memory
must be changed.
SystemError :
If the order keyword argument is specified. This behaviour is a bug in NumPy.
See Also:

resize
Return a new array with the specified shape.

Notes
This reallocates space for the data area if necessary.
Only contiguous arrays (data elements consecutive in memory) can be resized.
The purpose of the reference count check is to make sure you do not use this array as a buffer for another Python
object and then reallocate the memory. However, reference counts can increase in other ways so if you are sure
that you have not shared the memory for this array with another Python object, then you may safely set refcheck
to False.

Examples
Shrinking an array: array is flattened (in the order that the data are stored in memory), resized, and reshaped:

>>> a = np.array([[0, 1], [2, 3]], order=C)

>>> a.resize((2, 1))
>>> a
array([[0],
[1]])

>>> a = np.array([[0, 1], [2, 3]], order=F)

>>> a.resize((2, 1))
>>> a
array([[0],
[2]])

Enlarging an array: as above, but missing entries are filled with zeros:

1.1. The N-dimensional array (ndarray) 49

NumPy Reference, Release 2.0.0.dev8464

>>> b = np.array([[0, 1], [2, 3]])

>>> b.resize(2, 3) # new_shape parameter doesnt have to be a tuple
>>> b
array([[0, 1, 2],
[3, 0, 0]])

Referencing an array prevents resizing...

>>> c = a
>>> a.resize((1, 1))
...
ValueError: cannot resize an array that has been referenced ...

Unless refcheck is False:

>>> a.resize((1, 1), refcheck=False)

>>> a
array([[0]])
>>> c
array([[0]])

transpose(*axes)
Returns a view of the array with axes transposed.
For a 1-D array, this has no effect. (To change between column and row vectors, first cast the 1-D array into a ma-
trix object.) For a 2-D array, this is the usual matrix transpose. For an n-D array, if axes are given, their order in-
dicates how the axes are permuted (see Examples). If axes are not provided and a.shape = (i[0], i[1],
... i[n-2], i[n-1]), then a.transpose().shape = (i[n-1], i[n-2], ... i[1],
i[0]).
Parameters
axes : None, tuple of ints, or n ints
None or no argument: reverses the order of the axes.
tuple of ints: i in the j-th place in the tuple means as i-th axis becomes a.transpose()s j-th
axis.
n ints: same as an n-tuple of the same ints (this form is intended simply as a convenience
alternative to the tuple form)
Returns
out : ndarray
View of a, with axes suitably permuted.
See Also:

ndarray.T
Array property returning the array transposed.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> a
array([[1, 2],
[3, 4]])
>>> a.transpose()
array([[1, 3],

50 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

[2, 4]])
>>> a.transpose((1, 0))
array([[1, 3],
[2, 4]])
>>> a.transpose(1, 0)
array([[1, 3],
[2, 4]])

swapaxes(axis1, axis2)
Return a view of the array with axis1 and axis2 interchanged.
Refer to numpy.swapaxes for full documentation.
See Also:

numpy.swapaxes
equivalent function

flatten(order=C)
Return a copy of the array collapsed into one dimension.
Parameters
order : {C, F}, optional
Whether to flatten in C (row-major) or Fortran (column-major) order. The default is
C.
Returns
y : ndarray
A copy of the input array, flattened to one dimension.
See Also:

ravel
Return a flattened array.
flat
A 1-D flat iterator over the array.

Examples

>>> a = np.array([[1,2], [3,4]])

>>> a.flatten()
array([1, 2, 3, 4])
>>> a.flatten(F)
array([1, 3, 2, 4])

ravel([order])
Return a flattened array.
Refer to numpy.ravel for full documentation.
See Also:

numpy.ravel
equivalent function
ndarray.flat
a flat iterator on the array.

1.1. The N-dimensional array (ndarray) 51

NumPy Reference, Release 2.0.0.dev8464

squeeze()
Remove single-dimensional entries from the shape of a.
Refer to numpy.squeeze for full documentation.
See Also:

numpy.squeeze
equivalent function

Item selection and manipulation

For array methods that take an axis keyword, it defaults to None. If axis is None, then the array is treated as a 1-D
array. Any other value for axis represents the dimension along which the operation should proceed.
ndarray.take(indices[, axis, out, Return an array formed from the elements of a at the given indices.
mode])
ndarray.put(indices, values[, mode]) Set a.flat[n] = values[n] for all n in indices.
ndarray.repeat(repeats[, axis]) Repeat elements of an array.
ndarray.choose(choices[, out, Use an index array to construct a new array from a set of choices.
mode])
ndarray.sort([axis, kind, order]) Sort an array, in-place.
ndarray.argsort([axis, kind, order]) Returns the indices that would sort this array.
ndarray.searchsorted(v[, side]) Find indices where elements of v should be inserted in a to maintain
order.
ndarray.nonzero() Return the indices of the elements that are non-zero.
ndarray.compress(condition[, axis, Return selected slices of this array along given axis.
out])
ndarray.diagonal([offset, axis1, Return specified diagonals.
axis2])
take(indices, axis=None, out=None, mode=raise)
Return an array formed from the elements of a at the given indices.
Refer to numpy.take for full documentation.
See Also:

numpy.take
equivalent function

put(indices, values, mode=raise)

Set a.flat[n] = values[n] for all n in indices.
Refer to numpy.put for full documentation.
See Also:

numpy.put
equivalent function

repeat(repeats, axis=None)
Repeat elements of an array.
Refer to numpy.repeat for full documentation.
See Also:

52 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

numpy.repeat
equivalent function

choose(choices, out=None, mode=raise)

Use an index array to construct a new array from a set of choices.
Refer to numpy.choose for full documentation.
See Also:

numpy.choose
equivalent function

sort(axis=-1, kind=quicksort, order=None)

Sort an array, in-place.
Parameters
axis : int, optional
Axis along which to sort. Default is -1, which means sort along the last axis.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm. Default is quicksort.
order : list, optional
When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
See Also:

numpy.sort
Return a sorted copy of an array.
argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in sorted array.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = np.array([[1,4], [3,1]])

>>> a.sort(axis=1)
>>> a
array([[1, 4],
[1, 3]])
>>> a.sort(axis=0)
>>> a
array([[1, 3],
[1, 4]])

Use the order keyword to specify a field to use when sorting a structured array:

1.1. The N-dimensional array (ndarray) 53

NumPy Reference, Release 2.0.0.dev8464

>>> a = np.array([(a, 2), (c, 1)], dtype=[(x, S1), (y, int)])

>>> a.sort(order=y)
>>> a
array([(c, 1), (a, 2)],
dtype=[(x, |S1), (y, <i4)])

argsort(axis=-1, kind=quicksort, order=None)

Returns the indices that would sort this array.
Refer to numpy.argsort for full documentation.
See Also:

numpy.argsort
equivalent function

searchsorted(v, side=left)
Find indices where elements of v should be inserted in a to maintain order.
For full documentation, see numpy.searchsorted
See Also:

numpy.searchsorted
equivalent function

nonzero()
Return the indices of the elements that are non-zero.
Refer to numpy.nonzero for full documentation.
See Also:

numpy.nonzero
equivalent function

compress(condition, axis=None, out=None)

Return selected slices of this array along given axis.
Refer to numpy.compress for full documentation.
See Also:

numpy.compress
equivalent function

diagonal(offset=0, axis1=0, axis2=1)

Return specified diagonals.
Refer to numpy.diagonal for full documentation.
See Also:

numpy.diagonal
equivalent function

54 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Calculation

Many of these methods take an argument named axis. In such cases,

If axis is None (the default), the array is treated as a 1-D array and the operation is performed over the entire
array. This behavior is also the default if self is a 0-dimensional array or array scalar. (An array scalar is
an instance of the types/classes float32, float64, etc., whereas a 0-dimensional array is an ndarray instance
containing precisely one array scalar.)
If axis is an integer, then the operation is done over the given axis (for each 1-D subarray that can be created
along the given axis).
Example of the axis argument
A 3-dimensional array of size 3 x 3 x 3, summed over each of its three axes

>>> x
array([[[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8]],
[[ 9, 10, 11],
[12, 13, 14],
[15, 16, 17]],
[[18, 19, 20],
[21, 22, 23],
[24, 25, 26]]])
>>> x.sum(axis=0)
array([[27, 30, 33],
[36, 39, 42],
[45, 48, 51]])
>>> # for sum, axis is the first keyword, so we may omit it,
>>> # specifying only its value
>>> x.sum(0), x.sum(1), x.sum(2)
(array([[27, 30, 33],
[36, 39, 42],
[45, 48, 51]]),
array([[ 9, 12, 15],
[36, 39, 42],
[63, 66, 69]]),
array([[ 3, 12, 21],
[30, 39, 48],
[57, 66, 75]]))

The parameter dtype specifies the data type over which a reduction operation (like summing) should take place. The
default reduce data type is the same as the data type of self. To avoid overflow, it can be useful to perform the reduction
using a larger data type.
For several methods, an optional out argument can also be provided and the result will be placed into the output array
given. The out argument must be an ndarray and have the same number of elements. It can have a different data
type in which case casting will be performed.

1.1. The N-dimensional array (ndarray) 55

NumPy Reference, Release 2.0.0.dev8464

ndarray.argmax([axis, out]) Return indices of the maximum values along the given axis.
ndarray.min([axis, out]) Return the minimum along a given axis.
ndarray.argmin([axis, out]) Return indices of the minimum values along the given axis of
a.
ndarray.ptp([axis, out]) Peak to peak (maximum - minimum) value along a given axis.
ndarray.clip(a_min, a_max[, out]) Return an array whose values are limited to [a_min,
a_max].
ndarray.conj() Complex-conjugate all elements.
ndarray.round([decimals, out]) Return an array rounded a to the given number of decimals.
ndarray.trace([offset, axis1, axis2, dtype, Return the sum along diagonals of the array.
out])
ndarray.sum([axis, dtype, out]) Return the sum of the array elements over the given axis.
ndarray.cumsum([axis, dtype, out]) Return the cumulative sum of the elements along the given
axis.
ndarray.mean([axis, dtype, out]) Returns the average of the array elements along given axis.
ndarray.var([axis, dtype, out, ddof]) Returns the variance of the array elements, along given axis.
ndarray.std([axis, dtype, out, ddof]) Returns the standard deviation of the array elements along
given axis.
ndarray.prod([axis, dtype, out]) Return the product of the array elements over the given axis
ndarray.cumprod([axis, dtype, out]) Return the cumulative product of the elements along the given
axis.
ndarray.all([axis, out]) Returns True if all elements evaluate to True.
ndarray.any([axis, out]) Returns True if any of the elements of a evaluate to True.
argmax(axis=None, out=None)
Return indices of the maximum values along the given axis.
Refer to numpy.argmax for full documentation.
See Also:

numpy.argmax
equivalent function

min(axis=None, out=None)
Return the minimum along a given axis.
Refer to numpy.amin for full documentation.
See Also:

numpy.amin
equivalent function

argmin(axis=None, out=None)
Return indices of the minimum values along the given axis of a.
Refer to numpy.argmin for detailed documentation.
See Also:

numpy.argmin
equivalent function

ptp(axis=None, out=None)
Peak to peak (maximum - minimum) value along a given axis.
Refer to numpy.ptp for full documentation.

56 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

See Also:

numpy.ptp
equivalent function

clip(a_min, a_max, out=None)

Return an array whose values are limited to [a_min, a_max].
Refer to numpy.clip for full documentation.
See Also:

numpy.clip
equivalent function

conj()
Complex-conjugate all elements.
Refer to numpy.conjugate for full documentation.
See Also:

numpy.conjugate
equivalent function

round(decimals=0, out=None)
Return an array rounded a to the given number of decimals.
Refer to numpy.around for full documentation.
See Also:

numpy.around
equivalent function

trace(offset=0, axis1=0, axis2=1, dtype=None, out=None)

Return the sum along diagonals of the array.
Refer to numpy.trace for full documentation.
See Also:

numpy.trace
equivalent function

sum(axis=None, dtype=None, out=None)

Return the sum of the array elements over the given axis.
Refer to numpy.sum for full documentation.
See Also:

numpy.sum
equivalent function

cumsum(axis=None, dtype=None, out=None)

Return the cumulative sum of the elements along the given axis.
Refer to numpy.cumsum for full documentation.

1.1. The N-dimensional array (ndarray) 57

NumPy Reference, Release 2.0.0.dev8464

See Also:

numpy.cumsum
equivalent function

mean(axis=None, dtype=None, out=None)

Returns the average of the array elements along given axis.
Refer to numpy.mean for full documentation.
See Also:

numpy.mean
equivalent function

var(axis=None, dtype=None, out=None, ddof=0)

Returns the variance of the array elements, along given axis.
Refer to numpy.var for full documentation.
See Also:

numpy.var
equivalent function

std(axis=None, dtype=None, out=None, ddof=0)

Returns the standard deviation of the array elements along given axis.
Refer to numpy.std for full documentation.
See Also:

numpy.std
equivalent function

prod(axis=None, dtype=None, out=None)

Return the product of the array elements over the given axis
Refer to numpy.prod for full documentation.
See Also:

numpy.prod
equivalent function

cumprod(axis=None, dtype=None, out=None)

Return the cumulative product of the elements along the given axis.
Refer to numpy.cumprod for full documentation.
See Also:

numpy.cumprod
equivalent function

all(axis=None, out=None)
Returns True if all elements evaluate to True.
Refer to numpy.all for full documentation.

58 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

See Also:

numpy.all
equivalent function

any(axis=None, out=None)
Returns True if any of the elements of a evaluate to True.
Refer to numpy.any for full documentation.
See Also:

numpy.any
equivalent function

1.1.6 Arithmetic and comparison operations

Arithmetic and comparison operations on ndarrays are defined as element-wise operations, and generally yield
ndarray objects as results.
Each of the arithmetic operations (+, -, *, /, //, %, divmod(), ** or pow(), <<, >>, &, ^, |, ~) and the
comparisons (==, <, >, <=, >=, !=) is equivalent to the corresponding universal function (or ufunc for short) in
Numpy. For more information, see the section on Universal Functions.
Comparison operators:
ndarray.__lt__ x.__lt__(y) <==> x<y
ndarray.__le__ x.__le__(y) <==> x<=y
ndarray.__gt__ x.__gt__(y) <==> x>y
ndarray.__ge__ x.__ge__(y) <==> x>=y
ndarray.__eq__ x.__eq__(y) <==> x==y
ndarray.__ne__ x.__ne__(y) <==> x!=y
__lt__()
x.__lt__(y) <==> x<y
__le__()
x.__le__(y) <==> x<=y
__gt__()
x.__gt__(y) <==> x>y
__ge__()
x.__ge__(y) <==> x>=y
__eq__()
x.__eq__(y) <==> x==y
__ne__()
x.__ne__(y) <==> x!=y
Truth value of an array (bool):
ndarray.__nonzero__ x.__nonzero__() <==> x != 0
__nonzero__()
x.__nonzero__() <==> x != 0
Note: Truth-value testing of an array invokes ndarray.__nonzero__, which raises an error if the number of
elements in the the array is larger than 1, because the truth value of such arrays is ambiguous. Use .any() and

1.1. The N-dimensional array (ndarray) 59

NumPy Reference, Release 2.0.0.dev8464

.all() instead to be clear about what is meant in such cases. (If the number of elements is 0, the array evaluates to
False.)
Unary operations:
ndarray.__neg__ x.__neg__() <==> -x
ndarray.__pos__ x.__pos__() <==> +x
ndarray.__abs__(x)
ndarray.__invert__ x.__invert__() <==> ~x
__neg__()
x.__neg__() <==> -x
__pos__()
x.__pos__() <==> +x
__abs__() <==> abs(x)

__invert__()
x.__invert__() <==> ~x
Arithmetic:
ndarray.__add__ x.__add__(y) <==> x+y
ndarray.__sub__ x.__sub__(y) <==> x-y
ndarray.__mul__ x.__mul__(y) <==> x*y
ndarray.__div__ x.__div__(y) <==> x/y
ndarray.__truediv__ x.__truediv__(y) <==> x/y
ndarray.__floordiv__ x.__floordiv__(y) <==> x//y
ndarray.__mod__ x.__mod__(y) <==> x%y
ndarray.__divmod__(x, y)
ndarray.__pow__(x)
ndarray.__lshift__ x.__lshift__(y) <==> x<<y
ndarray.__rshift__ x.__rshift__(y) <==> x>>y
ndarray.__and__ x.__and__(y) <==> x&y
ndarray.__or__ x.__or__(y) <==> x|y
ndarray.__xor__ x.__xor__(y) <==> x^y
__add__()
x.__add__(y) <==> x+y
__sub__()
x.__sub__(y) <==> x-y
__mul__()
x.__mul__(y) <==> x*y
__div__()
x.__div__(y) <==> x/y
__truediv__()
x.__truediv__(y) <==> x/y
__floordiv__()
x.__floordiv__(y) <==> x//y
__mod__()
x.__mod__(y) <==> x%y
__divmod__(y) <==> divmod(x, y)

60 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

__pow__(y, [z], ) <==> pow(x, y, [z])

__lshift__()
x.__lshift__(y) <==> x<<y
__rshift__()
x.__rshift__(y) <==> x>>y
__and__()
x.__and__(y) <==> x&y
__or__()
x.__or__(y) <==> x|y
__xor__()
x.__xor__(y) <==> x^y
Note:
Any third argument to pow is silently ignored, as the underlying ufunc takes only two arguments.
The three division operators are all defined; div is active by default, truediv is active when __future__
division is in effect.
Because ndarray is a built-in type (written in C), the __r{op}__ special methods are not directly defined.
The functions called to implement many arithmetic special methods for arrays can be modified using
set_numeric_ops.
Arithmetic, in-place:
ndarray.__iadd__ x.__iadd__(y) <==> x+y
ndarray.__isub__ x.__isub__(y) <==> x-y
ndarray.__imul__ x.__imul__(y) <==> x*y
ndarray.__idiv__ x.__idiv__(y) <==> x/y
ndarray.__itruediv__ x.__itruediv__(y) <==> x/y
ndarray.__ifloordiv__ x.__ifloordiv__(y) <==> x//y
ndarray.__imod__ x.__imod__(y) <==> x%y
ndarray.__ipow__ x.__ipow__(y) <==> x**y
ndarray.__ilshift__ x.__ilshift__(y) <==> x<<y
ndarray.__irshift__ x.__irshift__(y) <==> x>>y
ndarray.__iand__ x.__iand__(y) <==> x&y
ndarray.__ior__ x.__ior__(y) <==> x|y
ndarray.__ixor__ x.__ixor__(y) <==> x^y
__iadd__()
x.__iadd__(y) <==> x+y
__isub__()
x.__isub__(y) <==> x-y
__imul__()
x.__imul__(y) <==> x*y
__idiv__()
x.__idiv__(y) <==> x/y
__itruediv__()
x.__itruediv__(y) <==> x/y
__ifloordiv__()
x.__ifloordiv__(y) <==> x//y

1.1. The N-dimensional array (ndarray) 61

NumPy Reference, Release 2.0.0.dev8464

__imod__()
x.__imod__(y) <==> x%y
__ipow__()
x.__ipow__(y) <==> x**y
__ilshift__()
x.__ilshift__(y) <==> x<<y
__irshift__()
x.__irshift__(y) <==> x>>y
__iand__()
x.__iand__(y) <==> x&y
__ior__()
x.__ior__(y) <==> x|y
__ixor__()
x.__ixor__(y) <==> x^y

Warning: In place operations will perform the calculation using the precision decided by the data type of the
two operands, but will silently downcast the result (if necessary) so it can fit back into the array. Therefore,
for mixed precision calculations, A {op}= B can be different than A = A {op} B. For example, suppose
a = ones((3,3)). Then, a += 3j is different than a = a + 3j: while they both perform the same
computation, a += 3 casts the result to fit back in a, whereas a = a + 3j re-binds the name a to the result.

1.1.7 Special methods

For standard library functions:

ndarray.__copy__() Return a copy of the array.
ndarray.__deepcopy__ a.__deepcopy__() -> Deep copy of array.
ndarray.__reduce__() For pickling.
ndarray.__setstate__(version, shape, dtype, ...) For unpickling.
__copy__([order])
Return a copy of the array.
Parameters
order : {C, F, A}, optional
If order is C (False) then the result is contiguous (default). If order is Fortran (True)
then the result has fortran order. If order is Any (None) then the result has fortran
order only if the array already is in fortran order.
__deepcopy__()
a.__deepcopy__() -> Deep copy of array.
Used if copy.deepcopy is called on an array.
__reduce__()
For pickling.
__setstate__(version, shape, dtype, isfortran, rawdata)
For unpickling.
Parameters
version : int

62 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

optional pickle version. If omitted defaults to 0.

shape : tuple
dtype : data-type
isFortran : bool
rawdata : string or list
a binary string with the data (or a list if a is an object array)
Basic customization:
ndarray.__new__
ndarray.__array__ a.__array__(|dtype) -> reference if type unchanged, copy otherwise.
ndarray.__array_wrap__ a.__array_wrap__(obj) -> Object of same type as ndarray object a.
__array__()
a.__array__(|dtype) -> reference if type unchanged, copy otherwise.
Returns either a new reference to self if dtype is not given or a new array of provided data type if dtype is
different from the current dtype of the array.
__array_wrap__()
a.__array_wrap__(obj) -> Object of same type as ndarray object a.
Container customization: (see Indexing)
ndarray.__len__(x)
ndarray.__getitem__ x.__getitem__(y) <==> x[y]
ndarray.__setitem__ x.__setitem__(i, y) <==> x[i]=y
ndarray.__getslice__ x.__getslice__(i, j) <==> x[i:j]
ndarray.__setslice__ x.__setslice__(i, j, y) <==> x[i:j]=y
ndarray.__contains__ x.__contains__(y) <==> y in x
__len__() <==> len(x)

__getitem__()
x.__getitem__(y) <==> x[y]
__setitem__()
x.__setitem__(i, y) <==> x[i]=y
__getslice__()
x.__getslice__(i, j) <==> x[i:j]
Use of negative indices is not supported.
__setslice__()
x.__setslice__(i, j, y) <==> x[i:j]=y
Use of negative indices is not supported.
__contains__()
x.__contains__(y) <==> y in x
Conversion; the operations complex, int, long, float, oct, and hex. They work only on arrays that have one
element in them and return the appropriate scalar.

1.1. The N-dimensional array (ndarray) 63

NumPy Reference, Release 2.0.0.dev8464

ndarray.__int__(x)
ndarray.__long__(x)
ndarray.__float__(x)
ndarray.__oct__(x)
ndarray.__hex__(x)
__int__() <==> int(x)

__long__() <==> long(x)

__float__() <==> float(x)

__oct__() <==> oct(x)

__hex__() <==> hex(x)

String representations:
ndarray.__str__(x)
ndarray.__repr__(x)
__str__() <==> str(x)

__repr__() <==> repr(x)

1.2 Scalars

Python defines only one type of a particular data class (there is only one integer type, one floating-point type, etc.).
This can be convenient in applications that dont need to be concerned with all the ways data can be represented in a
computer. For scientific computing, however, more control is often needed.
In NumPy, there are 21 new fundamental Python types to describe different types of scalars. These type descriptors
are mostly based on the types available in the C language that CPython is written in, with several additional types
compatible with Pythons types.
Array scalars have the same attributes and methods as ndarrays. 1 This allows one to treat items of an array partly
on the same footing as arrays, smoothing out rough edges that result when mixing scalar and array operations.
Array scalars live in a hierarchy (see the Figure below) of data types. They can be detected using the hierarchy:
For example, isinstance(val, np.generic) will return True if val is an array scalar object. Alternatively,
what kind of array scalar is present can be determined using other members of the data type hierarchy. Thus, for
example isinstance(val, np.complexfloating) will return True if val is a complex valued type, while
isinstance(val, np.flexible) will return true if val is one of the flexible itemsize array types (string,
unicode, void).

1.2.1 Built-in scalar types

The built-in scalar types are shown below. Along with their (mostly) C-derived names, the integer, float, and complex
data-types are also available using a bit-width convention so that an array of the right size can always be ensured (e.g.
1 However, array scalars are immutable, so none of the array scalar attributes are settable.

64 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Figure 1.2: Figure: Hierarchy of type objects representing the array data types. Not shown are the two integer types
intp and uintp which just point to the integer type that holds a pointer for the platform. All the number types can
be obtained using bit-width names as well.

1.2. Scalars 65
NumPy Reference, Release 2.0.0.dev8464

int8, float64, complex128). Two aliases (intp and uintp) pointing to the integer type that is sufficiently
large to hold a C pointer are also provided. The C-like names are associated with character codes, which are shown in
the table. Use of the character codes, however, is discouraged.
Five of the scalar types are essentially equivalent to fundamental Python types and therefore inherit from them as well
as from the generic array scalar type:
Array scalar type Related Python type
int_ IntType
float_ FloatType
complex_ ComplexType
str_ StringType
unicode_ UnicodeType
The bool_ data type is very similar to the Python BooleanType but does not inherit from it because Pythons
BooleanType does not allow itself to be inherited from, and on the C-level the size of the actual bool data is not the
same as a Python Boolean scalar.

Warning: The bool_ type is not a subclass of the int_ type (the bool_ is not even a number type). This is
different than Pythons default implementation of bool as a sub-class of int.

Tip: The default data type in Numpy is float_.

In the tables below, platform? means that the type may not be available on all platforms. Compatibility with
different C or Python types is indicated: two types are compatible if their data is of the same size and interpreted in
the same way.
Booleans:
Type Remarks Character code
bool_ compatible: Python bool ?
bool8 8 bits
Integers:
byte compatible: C char b
short compatible: C short h
intc compatible: C int i
int_ compatible: Python int l
longlong compatible: C long long q
intp large enough to fit a pointer p
int8 8 bits
int16 16 bits
int32 32 bits
int64 64 bits
Unsigned integers:
ubyte compatible: C unsigned char B
ushort compatible: C unsigned short H
uintc compatible: C unsigned int I
uint compatible: Python int L
ulonglong compatible: C long long Q
uintp large enough to fit a pointer P
uint8 8 bits
uint16 16 bits
uint32 32 bits
uint64 64 bits

66 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Floating-point numbers:
single compatible: C float f
double compatible: C double
float_ compatible: Python float d
longfloat compatible: C long float g
float32 32 bits
float64 64 bits
float96 96 bits, platform?
float128 128 bits, platform?
Complex floating-point numbers:
csingle F
complex_ compatible: Python complex D
clongfloat G
complex64 two 32-bit floats
complex128 two 64-bit floats
complex192 two 96-bit floats, platform?
complex256 two 128-bit floats, platform?
Any Python object:
object_ any Python object O
Note: The data actually stored in object arrays (i.e., arrays having dtype object_) are references to Python objects,
not the objects themselves. Hence, object arrays behave more like usual Python lists, in the sense that their contents
need not be of the same Python type.
The object type is also special because an array containing object_ items does not return an object_ object on
item access, but instead returns the actual object that the array item refers to.
The following data types are flexible. They have no predefined size: the data they describe can be of different length
in different arrays. (In the character codes # is an integer denoting how many elements the data type consists of.)
str_ compatible: Python str S#
unicode_ compatible: Python unicode U#
void V#

Warning: Numeric Compatibility: If you used old typecode characters in your Numeric code (which was never
recommended), you will need to change some of them to the new characters. In particular, the needed changes
are c -> S1, b -> B, 1 -> b, s -> h, w -> H, and u -> I. These changes make the type character
convention more consistent with other Python modules such as the struct module.

1.2.2 Attributes

The array scalar objects have an array priority of NPY_SCALAR_PRIORITY (-1,000,000.0). They also do
not (yet) have a ctypes attribute. Otherwise, they share the same attributes as arrays:

1.2. Scalars 67
NumPy Reference, Release 2.0.0.dev8464

generic.flags integer value of flags

generic.shape tuple of array dimensions
generic.strides tuple of bytes steps in each dimension
generic.ndim number of array dimensions
generic.data pointer to start of data
generic.size number of elements in the gentype
generic.itemsize length of one element in bytes
generic.base base object
generic.dtype get array data-descriptor
generic.real real part of scalar
generic.imag imaginary part of scalar
generic.flat a 1-d view of scalar
generic.T transpose
generic.__array_interface__ Array protocol: Python side
generic.__array_struct__ Array protocol: struct
generic.__array_priority__ Array priority.
generic.__array_wrap__ sc.__array_wrap__(obj) return scalar from array
flags
integer value of flags
shape
tuple of array dimensions
strides
tuple of bytes steps in each dimension
ndim
number of array dimensions
data
pointer to start of data
size
number of elements in the gentype
itemsize
length of one element in bytes
base
base object
dtype
get array data-descriptor
real
real part of scalar
imag
imaginary part of scalar
flat
a 1-d view of scalar
T
transpose
__array_interface__
Array protocol: Python side
__array_struct__
Array protocol: struct

68 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

__array_priority__
Array priority.
__array_wrap__()
sc.__array_wrap__(obj) return scalar from array

1.2.3 Indexing

See Also:
Indexing, Data type objects (dtype)
Array scalars can be indexed like 0-dimensional arrays: if x is an array scalar,
x[()] returns a 0-dimensional ndarray
x[field-name] returns the array scalar in the field field-name. (x can have fields, for example, when it
corresponds to a record data type.)

1.2.4 Methods

Array scalars have exactly the same methods as arrays. The default behavior of these methods is to internally convert
the scalar to an equivalent 0-dimensional array and to call the corresponding array method. In addition, math operations
on array scalars are defined so that the same hardware flags are set and used to interpret the results as for ufunc, so that
the error state used for ufuncs also carries over to the math on array scalars.
The exceptions to the above rules are given below:
generic Base class for numpy scalar types.
generic.__array__ sc.__array__(|type) return 0-dim array
generic.__array_wrap__ sc.__array_wrap__(obj) return scalar from array
generic.__squeeze__
generic.byteswap Not implemented (virtual attribute)
generic.__reduce__
generic.__setstate__
generic.setflags Not implemented (virtual attribute)
class generic()
Base class for numpy scalar types.
Class from which most (all?) numpy scalar types are derived. For consistency, exposes the same API as ndarray,
despite many consequent attributes being either get-only, or completely irrelevant. This is the class from which
it is strongly suggested users should derive custom scalar types.

Methods

all Not implemented (virtual attribute)

any Not implemented (virtual attribute)
argmax Not implemented (virtual attribute)
argmin Not implemented (virtual attribute)
argsort Not implemented (virtual attribute)
astype Not implemented (virtual attribute)
byteswap Not implemented (virtual attribute)
choose Not implemented (virtual attribute)
clip Not implemented (virtual attribute)
compress Not implemented (virtual attribute)
Continued on next page

1.2. Scalars 69
NumPy Reference, Release 2.0.0.dev8464

Table 1.2 continued from previous page

conj
conjugate Not implemented (virtual attribute)
copy Not implemented (virtual attribute)
cumprod Not implemented (virtual attribute)
cumsum Not implemented (virtual attribute)
diagonal Not implemented (virtual attribute)
dump Not implemented (virtual attribute)
dumps Not implemented (virtual attribute)
fill Not implemented (virtual attribute)
flatten Not implemented (virtual attribute)
getfield Not implemented (virtual attribute)
item Not implemented (virtual attribute)
itemset Not implemented (virtual attribute)
max Not implemented (virtual attribute)
mean Not implemented (virtual attribute)
min Not implemented (virtual attribute)
newbyteorder([new_order]) Return a new dtype with a different byte order.
nonzero Not implemented (virtual attribute)
prod Not implemented (virtual attribute)
ptp Not implemented (virtual attribute)
put Not implemented (virtual attribute)
ravel Not implemented (virtual attribute)
repeat Not implemented (virtual attribute)
reshape Not implemented (virtual attribute)
resize Not implemented (virtual attribute)
round Not implemented (virtual attribute)
searchsorted Not implemented (virtual attribute)
setfield Not implemented (virtual attribute)
setflags Not implemented (virtual attribute)
sort Not implemented (virtual attribute)
squeeze Not implemented (virtual attribute)
std Not implemented (virtual attribute)
sum Not implemented (virtual attribute)
swapaxes Not implemented (virtual attribute)
take Not implemented (virtual attribute)
tofile Not implemented (virtual attribute)
tolist Not implemented (virtual attribute)
tostring Not implemented (virtual attribute)
trace Not implemented (virtual attribute)
transpose Not implemented (virtual attribute)
var Not implemented (virtual attribute)
view Not implemented (virtual attribute)

all()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
any()

70 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Not implemented (virtual attribute)

Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
argmax()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
argmin()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
argsort()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
astype()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
byteswap()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
choose()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:

1.2. Scalars 71
NumPy Reference, Release 2.0.0.dev8464

The
clip()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
compress()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
conj()

conjugate()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
copy()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
cumprod()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
cumsum()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The

72 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

diagonal()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
dump()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
dumps()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
fill()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
flatten()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
getfield()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
item()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:

1.2. Scalars 73
NumPy Reference, Release 2.0.0.dev8464

The
itemset()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
max()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
mean()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
min()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
newbyteorder(new_order=S)
Return a new dtype with a different byte order.
Changes are also made in all fields and sub-arrays of the data type.
The new_order code can be any from the following:
{<, L} - little endian
{>, B} - big endian
{=, N} - native order
S - swap dtype from current to opposite endian
{|, I} - ignore (no change to byte order)

Parameters
new_order : str, optional
Byte order to force; a value from the byte order specifications above. The default value
(S) results in swapping the current byte order. The code does a case-insensitive check
on the first letter of new_order for the alternatives above. For example, any of B or b
or biggish are valid to specify big-endian.

74 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Returns
new_dtype : dtype
New dtype object with the given change to the byte order.

nonzero()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
prod()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
ptp()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
put()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
ravel()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
repeat()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The

1.2. Scalars 75
NumPy Reference, Release 2.0.0.dev8464

reshape()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
resize()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
round()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
searchsorted()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
setfield()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
setflags()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
sort()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:

76 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

The
squeeze()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
std()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
sum()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
swapaxes()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
take()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
tofile()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
tolist()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.

1.2. Scalars 77
NumPy Reference, Release 2.0.0.dev8464

See Also:
The
tostring()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
trace()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
transpose()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
var()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
view()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
__array__()
sc.__array__(|type) return 0-dim array
__array_wrap__()
sc.__array_wrap__(obj) return scalar from array
byteswap()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the attributes
of the ndarray class so as to provide a uniform API.
See Also:

78 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

The
__reduce__()

__setstate__()

setflags()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the attributes
of the ndarray class so as to provide a uniform API.
See Also:
The

1.2.5 Defining new types

There are two ways to effectively define a new array scalar type (apart from composing record dtypes from the built-in
scalar types): One way is to simply subclass the ndarray and overwrite the methods of interest. This will work to a
degree, but internally certain behaviors are fixed by the data type of the array. To fully customize the data type of an
array you need to define a new data-type, and register it with NumPy. Such new types can only be defined in C, using
the Numpy C-API.

1.3 Data type objects (dtype)

A data type object (an instance of numpy.dtype class) describes how the bytes in the fixed-size block of memory
corresponding to an array item should be interpreted. It describes the following aspects of the data:
1. Type of the data (integer, float, Python object, etc.)
2. Size of the data (how many bytes is in e.g. the integer)
3. Byte order of the data (little-endian or big-endian)
4. If the data type is a record, an aggregate of other data types, (e.g., describing an array item consisting of an
integer and a float),
(a) what are the names of the fields of the record, by which they can be accessed,
(b) what is the data-type of each field, and
(c) which part of the memory block each field takes.
5. If the data is a sub-array, what is its shape and data type.
To describe the type of scalar data, there are several built-in scalar types in Numpy for various precision of integers,
floating-point numbers, etc. An item extracted from an array, e.g., by indexing, will be a Python object whose type is
the scalar type associated with the data type of the array.
Note that the scalar types are not dtype objects, even though they can be used in place of one whenever a data type
specification is needed in Numpy. Record data types are formed by creating a data type whose fields contain other
data types. Each field has a name by which it can be accessed. The parent data type should be of sufficient size to
contain all its fields; the parent can for example be based on the void type which allows an arbitrary item size. Record
data types may also contain other record types and fixed-size sub-array data types in their fields. Finally, a data type
can describe items that are themselves arrays of items of another data type. These sub-arrays must, however, be of a
fixed size. If an array is created using a data-type describing a sub-array, the dimensions of the sub-array are appended

1.3. Data type objects (dtype) 79

NumPy Reference, Release 2.0.0.dev8464

to the shape of the array when the array is created. Sub-arrays in a field of a record behave differently, see Record
Access.
Example
A simple data type containing a 32-bit big-endian integer: (see Specifying and constructing data types for details on
construction)

>>> dt = np.dtype(>i4)
>>> dt.byteorder
>
>>> dt.itemsize
4
>>> dt.name
int32
>>> dt.type is np.int32
True

The corresponding array scalar type is int32.

Example
A record data type containing a 16-character string (in field name) and a sub-array of two 64-bit floating-point
number (in field grades):

>>> dt = np.dtype([(name, np.str_, 16), (grades, np.float64, (2,))])

>>> dt[name]
dtype(|S16)
>>> dt[grades]
dtype((float64,(2,)))

Items of an array of this data type are wrapped in an array scalar type that also has two fields:

>>> x = np.array([(Sarah, (8.0, 7.0)), (John, (6.0, 7.0))], dtype=dt)

>>> x[1]
(John, [6.0, 7.0])
>>> x[1][grades]
array([ 6., 7.])
>>> type(x[1])
<type numpy.void>
>>> type(x[1][grades])
<type numpy.ndarray>

1.3.1 Specifying and constructing data types

Whenever a data-type is required in a NumPy function or method, either a dtype object or something that can be
converted to one can be supplied. Such conversions are done by the dtype constructor:
dtype Create a data type object.
class dtype()
Create a data type object.
A numpy array is homogeneous, and contains elements described by a dtype object. A dtype object can be
constructed from different combinations of fundamental numeric types.
Parameters
obj :

80 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Object to be converted to a data type object.

align : bool, optional
Add padding to the fields to match what a C compiler would output for a similar C-
struct. Can be True only if obj is a dictionary or a comma-separated string.
copy : bool, optional
Make a new copy of the data-type object. If False, the result may just be a reference
to a built-in data-type object.

Examples
Using array-scalar type:

>>> np.dtype(np.int16)
dtype(int16)

Record, one field name f1, containing int16:

>>> np.dtype([(f1, np.int16)])

dtype([(f1, <i2)])

Record, one field named f1, in itself containing a record with one field:

>>> np.dtype([(f1, [(f1, np.int16)])])

dtype([(f1, [(f1, <i2)])])

Record, two fields: the first field contains an unsigned int, the second an int32:

>>> np.dtype([(f1, np.uint), (f2, np.int32)])

dtype([(f1, <u4), (f2, <i4)])

Using array-protocol type strings:

>>> np.dtype([(a,f8),(b,S10)])
dtype([(a, <f8), (b, |S10)])

Using comma-separated field formats. The shape is (2,3):

>>> np.dtype("i4, (2,3)f8")

dtype([(f0, <i4), (f1, <f8, (2, 3))])

Using tuples. int is a fixed type, 3 the fields shape. void is a flexible type, here of size 10:

>>> np.dtype([(hello,(np.int,3)),(world,np.void,10)])
dtype([(hello, <i4, 3), (world, |V10)])

Subdivide int16 into 2 int8s, called x and y. 0 and 1 are the offsets in bytes:

>>> np.dtype((np.int16, {x:(np.int8,0), y:(np.int8,1)}))

dtype((<i2, [(x, |i1), (y, |i1)]))

Using dictionaries. Two fields named gender and age:

1.3. Data type objects (dtype) 81

NumPy Reference, Release 2.0.0.dev8464

>>> np.dtype({names:[gender,age], formats:[S1,np.uint8]})

dtype([(gender, |S1), (age, |u1)])

Offsets in bytes, here 0 and 25:

>>> np.dtype({surname:(S25,0),age:(np.uint8,25)})
dtype([(surname, |S25), (age, |u1)])

Methods
newbyteorder([new_order]) Return a new dtype with a different byte order.
newbyteorder(new_order=S)
Return a new dtype with a different byte order.
Changes are also made in all fields and sub-arrays of the data type.
Parameters
new_order : string, optional
Byte order to force; a value from the byte order specifications below. The default value
(S) results in swapping the current byte order. new_order codes can be any of:

* S - swap dtype from current to opposite endian

* {<, L} - little endian
* {>, B} - big endian
* {=, N} - native order
* {|, I} - ignore (no change to byte order)

The code does a case-insensitive check on the first letter of new_order for these alterna-
tives. For example, any of > or B or b or brian are valid to specify big-endian.
Returns
new_dtype : dtype
New dtype object with the given change to the byte order.

Notes
Changes are also made in all fields and sub-arrays of the data type.

Examples

>>> import sys

>>> sys_is_le = sys.byteorder == little
>>> native_code = sys_is_le and < or >
>>> swapped_code = sys_is_le and > or <
>>> native_dt = np.dtype(native_code+i2)
>>> swapped_dt = np.dtype(swapped_code+i2)
>>> native_dt.newbyteorder(S) == swapped_dt
True
>>> native_dt.newbyteorder() == swapped_dt
True
>>> native_dt == swapped_dt.newbyteorder(S)
True
>>> native_dt == swapped_dt.newbyteorder(=)
True
>>> native_dt == swapped_dt.newbyteorder(N)
True

82 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> native_dt == native_dt.newbyteorder(|)

True
>>> np.dtype(<i2) == native_dt.newbyteorder(<)
True
>>> np.dtype(<i2) == native_dt.newbyteorder(L)
True
>>> np.dtype(>i2) == native_dt.newbyteorder(>)
True
>>> np.dtype(>i2) == native_dt.newbyteorder(B)
True

What can be converted to a data-type object is described below:

dtype object
Used as-is.
None
The default data type: float_.
Array-scalar types
The 21 built-in array scalar type objects all convert to an associated data-type object. This is true for their
sub-classes as well.
Note that not all data-type information can be supplied with a type-object: for example, flexible data-types
have a default itemsize of 0, and require an explicitly given size to be useful.
Example

>>> dt = np.dtype(np.int32) # 32-bit integer

>>> dt = np.dtype(np.complex128) # 128-bit complex floating-point number

Generic types
The generic hierarchical type objects convert to corresponding type objects according to the associations:
number, inexact, floating float
complexfloating cfloat
integer, signedinteger int_
unsignedinteger uint
character string
generic, flexible void
Built-in Python types
Several python types are equivalent to a corresponding array scalar when used to generate a dtype object:
int int_
bool bool_
float float_
complex cfloat
str string
unicode unicode_
buffer void
(all others) object_
Example

1.3. Data type objects (dtype) 83

NumPy Reference, Release 2.0.0.dev8464

>>> dt = np.dtype(float) # Python-compatible floating-point number

>>> dt = np.dtype(int) # Python-compatible integer
>>> dt = np.dtype(object) # Python object

Types with .dtype

Any type object with a dtype attribute: The attribute will be accessed and used directly. The attribute
must return something that is convertible into a dtype object.
Several kinds of strings can be converted. Recognized strings can be prepended with > (big-endian), < (little-
endian), or = (hardware-native, the default), to specify the byte order.
One-character strings
Each built-in data-type has a character code (the updated Numeric typecodes), that uniquely identifies it.
Example

>>> dt = np.dtype(b) # byte, native byte order

>>> dt = np.dtype(>H) # big-endian unsigned short
>>> dt = np.dtype(<f) # little-endian single-precision float
>>> dt = np.dtype(d) # double-precision floating-point number

Array-protocol type strings (see The Array Interface)

The first character specifies the kind of data and the remaining characters specify how many bytes of data.
The supported kinds are
b Boolean
i (signed) integer
u unsigned integer
f floating-point
c complex-floating point
S, a string
U unicode
V anything (void)
Example

>>> dt = np.dtype(i4) # 32-bit signed integer

>>> dt = np.dtype(f8) # 64-bit floating-point number
>>> dt = np.dtype(c16) # 128-bit complex floating-point number
>>> dt = np.dtype(a25) # 25-character string

String with comma-separated fields

Numarray introduced a short-hand notation for specifying the format of a record as a comma-separated
string of basic formats.
A basic format in this context is an optional shape specifier followed by an array-protocol type string.
Parenthesis are required on the shape if it is greater than 1-d. NumPy allows a modification on the format
in that any string that can uniquely identify the type can be used to specify the data-type in a field. The
generated data-type fields are named f0, f2, ..., f<N-1> where N (>1) is the number of comma-
separated basic formats in the string. If the optional shape specifier is provided, then the data-type for the
corresponding field describes a sub-array.
Example
field named f0 containing a 32-bit integer

84 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

field named f1 containing a 2 x 3 sub-array of 64-bit floating-point numbers

field named f2 containing a 32-bit floating-point number

>>> dt = np.dtype("i4, (2,3)f8, f4")

field named f0 containing a 3-character string

field named f1 containing a sub-array of shape (3,) containing 64-bit unsigned integers
field named f2 containing a 3 x 4 sub-array containing 10-character strings

>>> dt = np.dtype("a3, 3u8, (3,4)a10")

Type strings
Any string in numpy.sctypeDict.keys():
Example

>>> dt = np.dtype(uint32) # 32-bit unsigned integer

>>> dt = np.dtype(Float64) # 64-bit floating-point number

(flexible_dtype, itemsize)
The first argument must be an object that is converted to a flexible data-type object (one whose element
size is 0), the second argument is an integer providing the desired itemsize.
Example

>>> dt = np.dtype((void, 10)) # 10-byte wide data block

>>> dt = np.dtype((str, 35)) # 35-character string
>>> dt = np.dtype((U, 10)) # 10-character unicode string

(fixed_dtype, shape)
The first argument is any object that can be converted into a fixed-size data-type object. The second
argument is the desired shape of this type. If the shape parameter is 1, then the data-type object is
equivalent to fixed dtype. If shape is a tuple, then the new dtype defines a sub-array of the given shape.
Example

>>> dt = np.dtype((np.int32, (2,2))) # 2 x 2 integer sub-array

>>> dt = np.dtype((S10, 1)) # 10-character string
>>> dt = np.dtype((i4, (2,3)f8, f4, (2,3))) # 2 x 3 record sub-array

(base_dtype, new_dtype)
Both arguments must be convertible to data-type objects in this case. The base_dtype is the data-type
object that the new data-type builds on. This is how you could assign named fields to any built-in data-
type object.
Example
32-bit integer, whose first two bytes are interpreted as an integer via field real, and the following two
bytes via field imag.

>>> dt = np.dtype((np.int32,{real:(np.int16, 0),imag:(np.int16, 2)})

32-bit integer, which is interpreted as consisting of a sub-array of shape (4,) containing 8-bit integers:

1.3. Data type objects (dtype) 85

NumPy Reference, Release 2.0.0.dev8464

>>> dt = np.dtype((np.int32, (np.int8, 4)))

32-bit integer, containing fields r, g, b, a that interpret the 4 bytes in the integer as four unsigned integers:

>>> dt = np.dtype((i4, [(r,u1),(g,u1),(b,u1),(a,u1)]))

[(field_name, field_dtype, field_shape), ...]

obj should be a list of fields where each field is described by a tuple of length 2 or 3. (Equivalent to the
descr item in the __array_interface__ attribute.)
The first element, field_name, is the field name (if this is then a standard field name, f#, is assigned).
The field name may also be a 2-tuple of strings where the first string is either a title (which may be any
string or unicode string) or meta-data for the field which can be any object, and the second string is the
name which must be a valid Python identifier.
The second element, field_dtype, can be anything that can be interpreted as a data-type.
The optional third element field_shape contains the shape if this field represents an array of the data-type
in the second element. Note that a 3-tuple with a third argument equal to 1 is equivalent to a 2-tuple.
This style does not accept align in the dtype constructor as it is assumed that all of the memory is
accounted for by the array interface description.
Example
Data-type with fields big (big-endian 32-bit integer) and little (little-endian 32-bit integer):

>>> dt = np.dtype([(big, >i4), (little, <i4)])

Data-type with fields R, G, B, A, each being an unsigned 8-bit integer:

>>> dt = np.dtype([(R,u1), (G,u1), (B,u1), (A,u1)])

{names: ..., formats: ..., offsets: ..., titles: ...}

This style has two required and two optional keys. The names and formats keys are required. Their
respective values are equal-length lists with the field names and the field formats. The field names must
be strings and the field formats can be any object accepted by dtype constructor.
The optional keys in the dictionary are offsets and titles and their values must each be lists of the same
length as the names and formats lists. The offsets value is a list of byte offsets (integers) for each field,
while the titles value is a list of titles for each field (None can be used if no title is desired for that field).
The titles can be any string or unicode object and will add another entry to the fields dictionary
keyed by the title and referencing the same field tuple which will contain the title as an additional tuple
member.
Example
Data type with fields r, g, b, a, each being a 8-bit unsigned integer:

>>> dt = np.dtype({names: [r,g,b,a],

... formats: [uint8, uint8, uint8, uint8]})

Data type with fields r and b (with the given titles), both being 8-bit unsigned integers, the first at byte
position 0 from the start of the field and the second at position 2:

86 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> dt = np.dtype({names: [r,b], formats: [u1, u1],

... offsets: [0, 2],
... titles: [Red pixel, Blue pixel]})

{field1: ..., field2: ..., ...}

This style allows passing in the fields attribute of a data-type object.
obj should contain string or unicode keys that refer to (data-type, offset) or (data-type,
offset, title) tuples.
Example
Data type containing field col1 (10-character string at byte position 0), col2 (32-bit float at byte posi-
tion 10), and col3 (integers at byte position 14):

>>> dt = np.dtype({col1: (S10, 0), col2: (float32, 10),

col3: (int, 14)})

1.3.2 dtype

Numpy data type descriptions are instances of the dtype class.

Attributes

The type of the data is described by the following dtype attributes:

dtype.type The type object used to instantiate a scalar of this data-type.
dtype.kind A character code (one of biufcSUV) identifying the general kind of data.
dtype.char A unique character code for each of the 21 different built-in types.
dtype.num A unique number for each of the 21 different built-in types.
dtype.str The array-protocol typestring of this data-type object.
type
The type object used to instantiate a scalar of this data-type.
kind
A character code (one of biufcSUV) identifying the general kind of data.
char
A unique character code for each of the 21 different built-in types.
num
A unique number for each of the 21 different built-in types.
These are roughly ordered from least-to-most precision.
str
The array-protocol typestring of this data-type object.
Size of the data is in turn described by:
dtype.name A bit-width name for this data-type.
dtype.itemsize The element size of this data-type object.
name
A bit-width name for this data-type.
Un-sized flexible data-type objects do not have this attribute.

1.3. Data type objects (dtype) 87

NumPy Reference, Release 2.0.0.dev8464

itemsize
The element size of this data-type object.
For 18 of the 21 types this number is fixed by the data-type. For the flexible data-types, this number can be
anything.
Endianness of this data:
dtype.byteorder A character indicating the byte-order of this data-type object.
byteorder
A character indicating the byte-order of this data-type object.
One of:
= native
< little-endian
> big-endian
| not applicable
All built-in data-type objects have byteorder either = or |.

Examples

>>> dt = np.dtype(i2)
>>> dt.byteorder
=
>>> # endian is not relevant for 8 bit numbers
>>> np.dtype(i1).byteorder
|
>>> # or ASCII strings
>>> np.dtype(S2).byteorder
|
>>> # Even if specific code is given, and it is native
>>> # = is the byteorder
>>> import sys
>>> sys_is_le = sys.byteorder == little
>>> native_code = sys_is_le and < or >
>>> swapped_code = sys_is_le and > or <
>>> dt = np.dtype(native_code + i2)
>>> dt.byteorder
=
>>> # Swapped code shows up as itself
>>> dt = np.dtype(swapped_code + i2)
>>> dt.byteorder == swapped_code
True

Information about sub-data-types in a record:

dtype.fields Dictionary of named fields defined for this data type, or None.
dtype.names Ordered list of field names, or None if there are no fields.
fields
Dictionary of named fields defined for this data type, or None.
The dictionary is indexed by keys that are the names of the fields. Each entry in the dictionary is a tuple fully
describing the field:

(dtype, offset[, title])

88 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

If present, the optional title can be any object (if it is a string or unicode then it will also be a key in the fields
dictionary, otherwise its meta-data). Notice also that the first two elements of the tuple can be passed directly
as arguments to the ndarray.getfield and ndarray.setfield methods.
See Also:
ndarray.getfield, ndarray.setfield

Examples

>>> dt = np.dtype([(name, np.str_, 16), (grades, np.float64, (2,))])

>>> print dt.fields
{grades: (dtype((float64,(2,))), 16), name: (dtype(|S16), 0)}

names
Ordered list of field names, or None if there are no fields.
The names are ordered according to increasing byte offset. This can be used, for example, to walk through all
of the named fields in offset order.

Examples

>>> dt = np.dtype([(name, np.str_, 16), (grades, np.float64, (2,))])

>>> dt.names
(name, grades)

For data types that describe sub-arrays:

dtype.subdtype Tuple (item_dtype, shape) if this dtype describes a sub-array, and
dtype.shape Shape tuple of the sub-array if this data type describes a sub-array,
subdtype
Tuple (item_dtype, shape) if this dtype describes a sub-array, and None otherwise.
The shape is the fixed shape of the sub-array described by this data type, and item_dtype the data type of the
array.
If a field whose dtype object has this attribute is retrieved, then the extra dimensions implied by shape are tacked
on to the end of the retrieved array.
shape
Shape tuple of the sub-array if this data type describes a sub-array, and () otherwise.
Attributes providing additional information:
dtype.hasobjectBoolean indicating whether this dtype contains any reference-counted objects in any fields
or sub-dtypes.
dtype.flags Bit-flags describing how this data type is to be interpreted.
dtype.isbuiltinInteger indicating how this dtype relates to the built-in dtypes.
dtype.isnative Boolean indicating whether the byte order of this dtype is native
dtype.descr Array-interface compliant full description of the data-type.
dtype.alignmentThe required alignment (bytes) of this data-type according to the compiler.
hasobject
Boolean indicating whether this dtype contains any reference-counted objects in any fields or sub-dtypes.
Recall that what is actually in the ndarray memory representing the Python object is the memory address of that
object (a pointer). Special handling may be required, and this attribute is useful for distinguishing data types
that may contain arbitrary Python objects and data-types that wont.

1.3. Data type objects (dtype) 89

NumPy Reference, Release 2.0.0.dev8464

flags
Bit-flags describing how this data type is to be interpreted.
Bit-masks are in numpy.core.multiarray as the constants ITEM_HASOBJECT, LIST_PICKLE,
ITEM_IS_POINTER, NEEDS_INIT, NEEDS_PYAPI, USE_GETITEM, USE_SETITEM. A full explanation of
these flags is in C-API documentation; they are largely useful for user-defined data-types.
isbuiltin
Integer indicating how this dtype relates to the built-in dtypes.
Read-only.
0 if this is a structured array type, with fields
1 if this is a dtype compiled into numpy (such as ints, floats etc)
2 if the dtype is for a user-defined numpy type A user-defined type uses the numpy C-API machinery to extend
numpy to handle a new array type. See User-defined data-types (in NumPy User Guide) in the Numpy
manual.

Examples

>>> dt = np.dtype(i2)
>>> dt.isbuiltin
1
>>> dt = np.dtype(f8)
>>> dt.isbuiltin
1
>>> dt = np.dtype([(field1, f8)])
>>> dt.isbuiltin
0

isnative
Boolean indicating whether the byte order of this dtype is native to the platform.
descr
Array-interface compliant full description of the data-type.
The format is that required by the descr key in the __array_interface__ attribute.
alignment
The required alignment (bytes) of this data-type according to the compiler.
More information is available in the C-API section of the manual.

Methods

Data types have the following method for changing the byte order:
dtype.newbyteorder([new_order]) Return a new dtype with a different byte order.
newbyteorder(new_order=S)
Return a new dtype with a different byte order.
Changes are also made in all fields and sub-arrays of the data type.
Parameters
new_order : string, optional
Byte order to force; a value from the byte order specifications below. The default value
(S) results in swapping the current byte order. new_order codes can be any of:

90 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

* S - swap dtype from current to opposite endian

* {<, L} - little endian
* {>, B} - big endian
* {=, N} - native order
* {|, I} - ignore (no change to byte order)

The code does a case-insensitive check on the first letter of new_order for these alter-
natives. For example, any of > or B or b or brian are valid to specify big-endian.
Returns
new_dtype : dtype
New dtype object with the given change to the byte order.

Notes
Changes are also made in all fields and sub-arrays of the data type.

Examples

>>> import sys

>>> sys_is_le = sys.byteorder == little
>>> native_code = sys_is_le and < or >
>>> swapped_code = sys_is_le and > or <
>>> native_dt = np.dtype(native_code+i2)
>>> swapped_dt = np.dtype(swapped_code+i2)
>>> native_dt.newbyteorder(S) == swapped_dt
True
>>> native_dt.newbyteorder() == swapped_dt
True
>>> native_dt == swapped_dt.newbyteorder(S)
True
>>> native_dt == swapped_dt.newbyteorder(=)
True
>>> native_dt == swapped_dt.newbyteorder(N)
True
>>> native_dt == native_dt.newbyteorder(|)
True
>>> np.dtype(<i2) == native_dt.newbyteorder(<)
True
>>> np.dtype(<i2) == native_dt.newbyteorder(L)
True
>>> np.dtype(>i2) == native_dt.newbyteorder(>)
True
>>> np.dtype(>i2) == native_dt.newbyteorder(B)
True

The following methods implement the pickle protocol:

dtype.__reduce__
dtype.__setstate__
__reduce__()

__setstate__()

1.3. Data type objects (dtype) 91

NumPy Reference, Release 2.0.0.dev8464

1.4 Indexing

ndarrays can be indexed using the standard Python x[obj] syntax, where x is the array and obj the selection.
There are three kinds of indexing available: record access, basic slicing, advanced indexing. Which one occurs
depends on obj.
Note: In Python, x[(exp1, exp2, ..., expN)] is equivalent to x[exp1, exp2, ..., expN]; the
latter is just syntactic sugar for the former.

1.4.1 Basic Slicing

Basic slicing extends Pythons basic concept of slicing to N dimensions. Basic slicing occurs when obj is a slice
object (constructed by start:stop:step notation inside of brackets), an integer, or a tuple of slice objects and
integers. Ellipsis and newaxis objects can be interspersed with these as well. In order to remain backward
compatible with a common usage in Numeric, basic slicing is also initiated if the selection object is any sequence
(such as a list) containing slice objects, the Ellipsis object, or the newaxis object, but no integer arrays or
other embedded sequences. The simplest case of indexing with N integers returns an array scalar representing the
corresponding item. As in Python, all indices are zero-based: for the i-th index ni , the valid range is 0 ni < di
where di is the i-th element of the shape of the array. Negative indices are interpreted as counting from the end of the
array (i.e., if i < 0, it means ni + i).
All arrays generated by basic slicing are always views of the original array.
The standard rules of sequence slicing apply to basic slicing on a per-dimension basis (including using a step index).
Some useful concepts to remember include:
The basic slice syntax is i:j:k where i is the starting index, j is the stopping index, and k is the step (k 6= 0).
This selects the m elements (in the corresponding dimension) with index values i, i + k, ..., i + (m - 1) k where
m = q + (r 6= 0) and q and r are the quotient and remainder obtained by dividing j - i by k: j - i = q k + r, so
that i + (m - 1) k < j.
Example

>>> x = np.array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

>>> x[1:7:2]
array([1, 3, 5])

Negative i and j are interpreted as n + i and n + j where n is the number of elements in the corresponding
dimension. Negative k makes stepping go towards smaller indices.
Example

>>> x[-2:10]
array([8, 9])
>>> x[-3:3:-1]
array([7, 6, 5, 4])

Assume n is the number of elements in the dimension being sliced. Then, if i is not given it defaults to 0 for k >
0 and n for k < 0 . If j is not given it defaults to n for k > 0 and -1 for k < 0 . If k is not given it defaults to 1.
Note that :: is the same as : and means select all indices along this axis.
Example

>>> x[5:]
array([5, 6, 7, 8, 9])

92 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

If the number of objects in the selection tuple is less than N , then : is assumed for any subsequent dimensions.
Example

>>> x = np.array([[[1],[2],[3]], [[4],[5],[6]]])

>>> x.shape
(2, 3, 1)
>>> x[1:2]
array([[[4],
[5],
[6]]])

Ellipsis expand to the number of : objects needed to make a selection tuple of the same length as x.ndim.
Only the first ellipsis is expanded, any others are interpreted as :.
Example

>>> x[...,0]
array([[1, 2, 3],
[4, 5, 6]])

Each newaxis object in the selection tuple serves to expand the dimensions of the resulting selection by one
unit-length dimension. The added dimension is the position of the newaxis object in the selection tuple.
Example

>>> x[:,np.newaxis,:,:].shape
(2, 1, 3, 1)

An integer, i, returns the same values as i:i+1 except the dimensionality of the returned object is reduced by 1.
In particular, a selection tuple with the p-th element an integer (and all other entries :) returns the corresponding
sub-array with dimension N - 1. If N = 1 then the returned object is an array scalar. These objects are explained
in Scalars.
If the selection tuple has all entries : except the p-th entry which is a slice object i:j:k, then the returned
array has dimension N formed by concatenating the sub-arrays returned by integer indexing of elements i, i+k,
..., i + (m - 1) k < j,
Basic slicing with more than one non-: entry in the slicing tuple, acts like repeated application of slicing using
a single non-: entry, where the non-: entries are successively taken (with all other non-: entries replaced by
:). Thus, x[ind1,...,ind2,:] acts like x[ind1][...,ind2,:] under basic slicing.

Warning: The above is not true for advanced slicing.

You may use slicing to set values in the array, but (unlike lists) you can never grow the array. The size of the
value to be set in x[obj] = value must be (broadcastable) to the same shape as x[obj].
Note: Remember that a slicing tuple can always be constructed as obj and used in the x[obj] notation. Slice objects
can be used in the construction in place of the [start:stop:step] notation. For example, x[1:10:5,::-1]
can also be implemented as obj = (slice(1,10,5), slice(None,None,-1)); x[obj] . This can be
useful for constructing generic code that works on arrays of arbitrary dimension.
newaxis
The newaxis object can be used in the basic slicing syntax discussed above. None can also be used instead
of newaxis.

1.4. Indexing 93
NumPy Reference, Release 2.0.0.dev8464

1.4.2 Advanced indexing

Advanced indexing is triggered when the selection object, obj, is a non-tuple sequence object, an ndarray (of data
type integer or bool), or a tuple with at least one sequence object or ndarray (of data type integer or bool). There are
two types of advanced indexing: integer and Boolean.
Advanced indexing always returns a copy of the data (contrast with basic slicing that returns a view).

Integer

Integer indexing allows selection of arbitrary items in the array based on their N-dimensional index. This kind of
selection occurs when advanced indexing is triggered and the selection object is not an array of data type bool. For the
discussion below, when the selection object is not a tuple, it will be referred to as if it had been promoted to a 1-tuple,
which will be called the selection tuple. The rules of advanced integer-style indexing are:
If the length of the selection tuple is larger than N an error is raised.
All sequences and scalars in the selection tuple are converted to intp indexing arrays.
All selection tuple objects must be convertible to intp arrays, slice objects, or the Ellipsis object.
The first Ellipsis object will be expanded, and any other Ellipsis objects will be treated as full slice (:)
objects. The expanded Ellipsis object is replaced with as many full slice (:) objects as needed to make the
length of the selection tuple N .
If the selection tuple is smaller than N, then as many : objects as needed are added to the end of the selection
tuple so that the modified selection tuple has length N.
All the integer indexing arrays must be broadcastable to the same shape.
The shape of the output (or the needed shape of the object to be used for setting) is the broadcasted shape.
After expanding any ellipses and filling out any missing : objects in the selection tuple, then let Nt be the
number of indexing arrays, and let Ns = N Nt be the number of slice objects. Note that Nt > 0 (or we
wouldnt be doing advanced integer indexing).
If Ns = 0 then the M-dimensional result is constructed by varying the index tuple (i_1, ..., i_M) over
the range of the result shape and for each value of the index tuple (ind_1, ..., ind_M):

result[i_1, ..., i_M] == x[ind_1[i_1, ..., i_M], ind_2[i_1, ..., i_M],

..., ind_N[i_1, ..., i_M]]

Example
Suppose the shape of the broadcasted indexing arrays is 3-dimensional and N is 2. Then the result is found by
letting i, j, k run over the shape found by broadcasting ind_1 and ind_2, and each i, j, k yields:

result[i,j,k] = x[ind_1[i,j,k], ind_2[i,j,k]]

If Ns > 0, then partial indexing is done. This can be somewhat mind-boggling to understand, but if you think
in terms of the shapes of the arrays involved, it can be easier to grasp what happens. In simple cases (i.e.
one indexing array and N - 1 slice objects) it does exactly what you would expect (concatenation of repeated
application of basic slicing). The rule for partial indexing is that the shape of the result (or the interpreted shape
of the object to be used in setting) is the shape of x with the indexed subspace replaced with the broadcasted
indexing subspace. If the index subspaces are right next to each other, then the broadcasted indexing space
directly replaces all of the indexed subspaces in x. If the indexing subspaces are separated (by slice objects),
then the broadcasted indexing space is first, followed by the sliced subspace of x.
Example

94 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Suppose x.shape is (10,20,30) and ind is a (2,3,4)-shaped indexing intp array, then result =
x[...,ind,:] has shape (10,2,3,4,30) because the (20,)-shaped subspace has been replaced with a
(2,3,4)-shaped broadcasted indexing subspace. If we let i, j, k loop over the (2,3,4)-shaped subspace then
result[...,i,j,k,:] = x[...,ind[i,j,k],:]. This example produces the same result as
x.take(ind, axis=-2).
Example
Now let x.shape be (10,20,30,40,50) and suppose ind_1 and ind_2 are broadcastable to the shape (2,3,4).
Then x[:,ind_1,ind_2] has shape (10,2,3,4,40,50) because the (20,30)-shaped subspace from X has
been replaced with the (2,3,4) subspace from the indices. However, x[:,ind_1,:,ind_2] has shape
(2,3,4,10,30,50) because there is no unambiguous place to drop in the indexing subspace, thus it is tacked-
on to the beginning. It is always possible to use .transpose() to move the subspace anywhere desired.
(Note that this example cannot be replicated using take.)

Boolean

This advanced indexing occurs when obj is an array object of Boolean type (such as may be returned from com-
parison operators). It is always equivalent to (but faster than) x[obj.nonzero()] where, as described above,
obj.nonzero() returns a tuple (of length obj.ndim) of integer index arrays showing the True elements of obj.
The special case when obj.ndim == x.ndim is worth mentioning. In this case x[obj] returns a 1-dimensional
array filled with the elements of x corresponding to the True values of obj. The search order will be C-style (last
index varies the fastest). If obj has True values at entries that are outside of the bounds of x, then an index error will
be raised.
You can also use Boolean arrays as element of the selection tuple. In such instances, they will always be interpreted
as nonzero(obj) and the equivalent integer indexing will be done.

Warning: The definition of advanced indexing means that x[(1,2,3),] is fundamentally different than
x[(1,2,3)]. The latter is equivalent to x[1,2,3] which will trigger basic selection while the former will
trigger advanced indexing. Be sure to understand why this is occurs.
Also recognize that x[[1,2,3]] will trigger advanced indexing, whereas x[[1,2,slice(None)]] will
trigger basic slicing.

1.4.3 Record Access

See Also:
Data type objects (dtype), Scalars
If the ndarray object is a record array, i.e. its data type is a record data type, the fields of the array can be accessed
by indexing the array with strings, dictionary-like.
Indexing x[field-name] returns a new view to the array, which is of the same shape as x (except when the field
is a sub-array) but of data type x.dtype[field-name] and contains only the part of the data in the specified
field. Also record array scalars can be indexed this way.
If the accessed field is a sub-array, the dimensions of the sub-array are appended to the shape of the result.
Example

>>> x = np.zeros((2,2), dtype=[(a, np.int32), (b, np.float64, (3,3))])

>>> x[a].shape
(2, 2)
>>> x[a].dtype

1.4. Indexing 95
NumPy Reference, Release 2.0.0.dev8464

dtype(int32)
>>> x[b].shape
(2, 2, 3, 3)
>>> x[b].dtype
dtype(float64)

1.4.4 Flat Iterator indexing

x.flat returns an iterator that will iterate over the entire array (in C-contiguous style with the last index varying
the fastest). This iterator object can also be indexed using basic slicing or advanced indexing as long as the selection
object is not a tuple. This should be clear from the fact that x.flat is a 1-dimensional view. It can be used for integer
indexing with 1-dimensional C-style-flat indices. The shape of any returned array is therefore the shape of the integer
indexing object.

1.5 Standard array subclasses

The ndarray in NumPy is a new-style Python built-in-type. Therefore, it can be inherited from (in Python or in
C) if desired. Therefore, it can form a foundation for many useful classes. Often whether to sub-class the array object
or to simply use the core array component as an internal part of a new class is a difficult decision, and can be simply a
matter of choice. NumPy has several tools for simplifying how your new object interacts with other array objects, and
so the choice may not be significant in the end. One way to simplify the question is by asking yourself if the object
you are interested in can be replaced as a single array or does it really require two or more arrays at its core.
Note that asarray always returns the base-class ndarray. If you are confident that your use of the array object can
handle any subclass of an ndarray, then asanyarray can be used to allow subclasses to propagate more cleanly
through your subroutine. In principal a subclass could redefine any aspect of the array and therefore, under strict
guidelines, asanyarray would rarely be useful. However, most subclasses of the arrayobject will not redefine
certain aspects of the array object such as the buffer interface, or the attributes of the array. One important example,
however, of why your subroutine may not be able to handle an arbitrary subclass of an array is that matrices redefine
the * operator to be matrix-multiplication, rather than element-by-element multiplication.

1.5.1 Special attributes and methods

See Also:
Subclassing ndarray (in NumPy User Guide)
Numpy provides several hooks that subclasses of ndarray can customize:
__array_finalize__(self )
This method is called whenever the system internally allocates a new array from obj, where obj is a subclass
(subtype) of the ndarray. It can be used to change attributes of self after construction (so as to ensure
a 2-d matrix for example), or to update meta-information from the parent. Subclasses inherit a default
implementation of this method that does nothing.
__array_prepare__(array, context=None)
At the beginning of every ufunc, this method is called on the input object with the highest array priority, or the
output object if one was specified. The output array is passed in and whatever is returned is passed to the ufunc.
Subclasses inherit a default implementation of this method which simply returns the output array unmodified.
Subclasses may opt to use this method to transform the output array into an instance of the subclass and update
metadata before returning the array to the ufunc for computation.

96 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

__array_wrap__(array, context=None)
At the end of every ufunc, this method is called on the input object with the highest array priority, or the output
object if one was specified. The ufunc-computed array is passed in and whatever is returned is passed to the
user. Subclasses inherit a default implementation of this method, which transforms the array into a new instance
of the objects class. Subclasses may opt to use this method to transform the output array into an instance of the
subclass and update metadata before returning the array to the user.
__array_priority__
The value of this attribute is used to determine what type of object to return in situations where there is more
than one possibility for the Python type of the returned object. Subclasses inherit a default value of 1.0 for this
attribute.
__array__([dtype])
If a class having the __array__ method is used as the output object of an ufunc, results will be written to the
object returned by __array__.

1.5.2 Matrix objects

matrix objects inherit from the ndarray and therefore, they have the same attributes and methods of ndarrays. There
are six important differences of matrix objects, however, that may lead to unexpected results when you use matrices
but expect them to act like arrays:
1. Matrix objects can be created using a string notation to allow Matlab-style syntax where spaces separate columns
and semicolons (;) separate rows.
2. Matrix objects are always two-dimensional. This has far-reaching implications, in that m.ravel() is still two-
dimensional (with a 1 in the first dimension) and item selection returns two-dimensional objects so that sequence
behavior is fundamentally different than arrays.
3. Matrix objects over-ride multiplication to be matrix-multiplication. Make sure you understand this for func-
tions that you may want to receive matrices. Especially in light of the fact that asanyarray(m) returns a
matrix when m is a matrix.
4. Matrix objects over-ride power to be matrix raised to a power. The same warning about using power inside a
function that uses asanyarray(...) to get an array object holds for this fact.
5. The default __array_priority__ of matrix objects is 10.0, and therefore mixed operations with ndarrays always
produce matrices.
6. Matrices have special attributes which make calculations easier. These are
matrix.T transpose
matrix.H hermitian (conjugate) transpose
matrix.I inverse
matrix.A base array
T
transpose
H
hermitian (conjugate) transpose
I
inverse
A
base array

1.5. Standard array subclasses 97

NumPy Reference, Release 2.0.0.dev8464

Warning: Matrix objects over-ride multiplication, *, and power, **, to be matrix-multiplication and matrix
power, respectively. If your subroutine can accept sub-classes and you do not convert to base- class arrays, then
you must use the ufuncs multiply and power to be sure that you are performing the correct operation for all inputs.

The matrix class is a Python subclass of the ndarray and can be used as a reference for how to construct your own
subclass of the ndarray. Matrices can be created from other matrices, strings, and anything else that can be converted
to an ndarray . The name mat is an alias for matrix in NumPy.
matrix Returns a matrix from an array-like object, or from a string of data.
asmatrix(data[, dtype]) Interpret the input as a matrix.
bmat(obj[, ldict, gdict]) Build a matrix object from a string, nested sequence, or array.
class matrix()
Returns a matrix from an array-like object, or from a string of data. A matrix is a specialized 2-d array that
retains its 2-d nature through operations. It has certain special operators, such as * (matrix multiplication) and
** (matrix power).
Parameters
data : array_like or string
If data is a string, the string is interpreted as a matrix with commas or spaces separating
columns, and semicolons separating rows.
dtype : data-type
Data-type of the output matrix.
copy : bool
If data is already an ndarray, then this flag determines whether the data is copied, or
whether a view is constructed.
See Also:
array

Examples

>>> a = np.matrix(1 2; 3 4)
>>> print a
[[1 2]
[3 4]]

>>> np.matrix([[1, 2], [3, 4]])

matrix([[1, 2],
[3, 4]])

Methods

all([axis, out]) Test whether all matrix elements along a given axis evaluate to True.
any([axis, out]) Test whether any array element along a given axis evaluates to True.
argmax([axis, out]) Indices of the maximum values along an axis.
argmin([axis, out]) Return the indices of the minimum values along an axis.
argsort([axis, kind, order]) Returns the indices that would sort this array.
astype(t) Copy of the array, cast to a specified type.
byteswap(inplace) Swap the bytes of the array elements
choose(choices[, out, mode]) Use an index array to construct a new array from a set of choices.
Continued on next page

98 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Table 1.3 continued from previous page

clip(a_min, a_max[, out]) Return an array whose values are limited to [a_min, a_max].
compress(condition[, axis, out]) Return selected slices of this array along given axis.
conj() Complex-conjugate all elements.
conjugate() Return the complex conjugate, element-wise.
copy([order]) Return a copy of the array.
cumprod([axis, dtype, out]) Return the cumulative product of the elements along the given axis.
cumsum([axis, dtype, out]) Return the cumulative sum of the elements along the given axis.
diagonal([offset, axis1, axis2]) Return specified diagonals.
dot
dump(file) Dump a pickle of the array to the specified file.
dumps() Returns the pickle of the array as a string.
fill(value) Fill the array with a scalar value.
flatten([order]) Return a copy of the array collapsed into one dimension.
getA() Return self as an ndarray object.
getA1() Return self as a flattened ndarray.
getH() Returns the (complex) conjugate transpose of self.
getI() Returns the (multiplicative) inverse of invertible self.
getT() Returns the transpose of the matrix.
getfield(dtype, offset) Returns a field of the given array as a certain type.
item(*args) Copy an element of an array to a standard Python scalar and return it.
itemset
max([axis, out]) Return the maximum value along an axis.
mean([axis, dtype, out]) Returns the average of the matrix elements along the given axis.
min([axis, out]) Return the minimum value along an axis.
newbyteorder([new_order]) Return the array with the same data viewed with a different byte order.
nonzero() Return the indices of the elements that are non-zero.
prod([axis, dtype, out]) Return the product of the array elements over the given axis.
ptp([axis, out]) Peak-to-peak (maximum - minimum) value along the given axis.
put(indices, values[, mode]) Set a.flat[n] = values[n] for all n in indices.
ravel() Return a flattened array.
repeat(repeats[, axis]) Repeat elements of an array.
reshape(shape[, order]) Returns an array containing the same data with a new shape.
resize(new_shape[, refcheck]) Change shape and size of array in-place.
round([decimals, out]) Return an array rounded a to the given number of decimals.
searchsorted(v[, side]) Find indices where elements of v should be inserted in a to maintain order.
setfield(val, dtype[, offset]) Put a value into a specified place in a field defined by a data-type.
setflags([write, align, uic]) Set array flags WRITEABLE, ALIGNED, and UPDATEIFCOPY, respectively.
sort([axis, kind, order]) Sort an array, in-place.
squeeze() Remove single-dimensional entries from the shape of a.
std([axis, dtype, out, ddof]) Return the standard deviation of the array elements along the given axis.
sum([axis, dtype, out]) Returns the sum of the matrix elements, along the given axis.
swapaxes(axis1, axis2) Return a view of the array with axis1 and axis2 interchanged.
take(indices[, axis, out, mode]) Return an array formed from the elements of a at the given indices.
tofile(fid[, sep, format]) Write array to a file as text or binary (default).
tolist() Return the matrix as a (possibly nested) list.
tostring([order]) Construct a Python string containing the raw data bytes in the array.
trace([offset, axis1, axis2, dtype, out]) Return the sum along diagonals of the array.
transpose(*axes) Returns a view of the array with axes transposed.
var([axis, dtype, out, ddof]) Returns the variance of the matrix elements, along the given axis.
view([dtype, type]) New view of array with the same data.

1.5. Standard array subclasses 99

NumPy Reference, Release 2.0.0.dev8464

all(axis=None, out=None)
Test whether all matrix elements along a given axis evaluate to True.
Parameters
See numpy.all for complete descriptions :
See Also:
numpy.all

Notes
This is the same as ndarray.all, but it returns a matrix object.

Examples

>>> x = np.matrix(np.arange(12).reshape((3,4))); x
matrix([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> y = x[0]; y
matrix([[0, 1, 2, 3]])
>>> (x == y)
matrix([[ True, True, True, True],
[False, False, False, False],
[False, False, False, False]], dtype=bool)
>>> (x == y).all()
False
>>> (x == y).all(0)
matrix([[False, False, False, False]], dtype=bool)
>>> (x == y).all(1)
matrix([[ True],
[False],
[False]], dtype=bool)

any(axis=None, out=None)
Test whether any array element along a given axis evaluates to True.
Refer to numpy.any for full documentation.
Parameters
axis: int, optional :
Axis along which logical OR is performed
out: ndarray, optional :
Output to existing array instead of creating new one, must have same shape as expected
output
Returns
any : bool, ndarray
Returns a single bool if axis is None; otherwise, returns ndarray
argmax(axis=None, out=None)
Indices of the maximum values along an axis.
Parameters
See numpy.argmax for complete descriptions :
See Also:

100 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

numpy.argmax

Notes
This is the same as ndarray.argmax, but returns a matrix object where ndarray.argmax would return an
ndarray.

Examples

>>> x = np.matrix(np.arange(12).reshape((3,4))); x
matrix([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> x.argmax()
11
>>> x.argmax(0)
matrix([[2, 2, 2, 2]])
>>> x.argmax(1)
matrix([[3],
[3],
[3]])

argmin(axis=None, out=None)
Return the indices of the minimum values along an axis.
Parameters
See numpy.argmin for complete descriptions. :
See Also:
numpy.argmin

Notes
This is the same as ndarray.argmin, but returns a matrix object where ndarray.argmin would return an
ndarray.

Examples

>>> x = -np.matrix(np.arange(12).reshape((3,4))); x
matrix([[ 0, -1, -2, -3],
[ -4, -5, -6, -7],
[ -8, -9, -10, -11]])
>>> x.argmin()
11
>>> x.argmin(0)
matrix([[2, 2, 2, 2]])
>>> x.argmin(1)
matrix([[3],
[3],
[3]])

argsort(axis=-1, kind=quicksort, order=None)

Returns the indices that would sort this array.
Refer to numpy.argsort for full documentation.
See Also:

1.5. Standard array subclasses 101

NumPy Reference, Release 2.0.0.dev8464

numpy.argsort
equivalent function

astype(t)
Copy of the array, cast to a specified type.
Parameters
t : string or dtype
Typecode or data-type to which the array is cast.

Examples

>>> x = np.array([1, 2, 2.5])

>>> x
array([ 1. , 2. , 2.5])

>>> x.astype(int)
array([1, 2, 2])

byteswap(inplace)
Swap the bytes of the array elements
Toggle between low-endian and big-endian data representation by returning a byteswapped array, option-
ally swapped in-place.
Parameters
inplace: bool, optional :
If True, swap bytes in-place, default is False.
Returns
out: ndarray :
The byteswapped array. If inplace is True, this is a view to self.

Examples

>>> A = np.array([1, 256, 8755], dtype=np.int16)

>>> map(hex, A)
[0x1, 0x100, 0x2233]
>>> A.byteswap(True)
array([ 256, 1, 13090], dtype=int16)
>>> map(hex, A)
[0x100, 0x1, 0x3322]

Arrays of strings are not swapped

>>> A = np.array([ceg, fac])

>>> A.byteswap()
array([ceg, fac],
dtype=|S3)

choose(choices, out=None, mode=raise)

Use an index array to construct a new array from a set of choices.
Refer to numpy.choose for full documentation.
See Also:

102 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

numpy.choose
equivalent function

clip(a_min, a_max, out=None)

Return an array whose values are limited to [a_min, a_max].
Refer to numpy.clip for full documentation.
See Also:

numpy.clip
equivalent function

compress(condition, axis=None, out=None)

Return selected slices of this array along given axis.
Refer to numpy.compress for full documentation.
See Also:

numpy.compress
equivalent function

conj()
Complex-conjugate all elements.
Refer to numpy.conjugate for full documentation.
See Also:

numpy.conjugate
equivalent function

conjugate()
Return the complex conjugate, element-wise.
Refer to numpy.conjugate for full documentation.
See Also:

numpy.conjugate
equivalent function

copy(order=C)
Return a copy of the array.
Parameters
order : {C, F, A}, optional
By default, the result is stored in C-contiguous (row-major) order in memory. If order is
F, the result has Fortran (column-major) order. If order is A (Any), then the result
has the same order as the input.

Examples

>>> x = np.array([[1,2,3],[4,5,6]], order=F)

>>> y = x.copy()

1.5. Standard array subclasses 103

NumPy Reference, Release 2.0.0.dev8464

>>> x.fill(0)

>>> x
array([[0, 0, 0],
[0, 0, 0]])

>>> y
array([[1, 2, 3],
[4, 5, 6]])

>>> y.flags[C_CONTIGUOUS]
True

cumprod(axis=None, dtype=None, out=None)

Return the cumulative product of the elements along the given axis.
Refer to numpy.cumprod for full documentation.
See Also:

numpy.cumprod
equivalent function

cumsum(axis=None, dtype=None, out=None)

Return the cumulative sum of the elements along the given axis.
Refer to numpy.cumsum for full documentation.
See Also:

numpy.cumsum
equivalent function

diagonal(offset=0, axis1=0, axis2=1)

Return specified diagonals.
Refer to numpy.diagonal for full documentation.
See Also:

numpy.diagonal
equivalent function

dot()

dump(file)
Dump a pickle of the array to the specified file. The array can be read back with pickle.load or numpy.load.
Parameters
file : str
A string naming the dump file.
dumps()
Returns the pickle of the array as a string. pickle.loads or numpy.loads will convert the string back to an
array.

104 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Parameters
None :
fill(value)
Fill the array with a scalar value.
Parameters
value : scalar
All elements of a will be assigned this value.

Examples

>>> a = np.array([1, 2])

>>> a.fill(0)
>>> a
array([0, 0])
>>> a = np.empty(2)
>>> a.fill(1)
>>> a
array([ 1., 1.])

flatten(order=C)
Return a copy of the array collapsed into one dimension.
Parameters
order : {C, F}, optional
Whether to flatten in C (row-major) or Fortran (column-major) order. The default is C.
Returns
y : ndarray
A copy of the input array, flattened to one dimension.
See Also:

ravel
Return a flattened array.
flat
A 1-D flat iterator over the array.

Examples

>>> a = np.array([[1,2], [3,4]])

>>> a.flatten()
array([1, 2, 3, 4])
>>> a.flatten(F)
array([1, 3, 2, 4])

getA()
Return self as an ndarray object.
Equivalent to np.asarray(self).
Parameters
None :
Returns
ret : ndarray

1.5. Standard array subclasses 105

NumPy Reference, Release 2.0.0.dev8464

self as an ndarray

Examples

>>> x = np.matrix(np.arange(12).reshape((3,4))); x
matrix([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> x.getA()
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])

getA1()
Return self as a flattened ndarray.
Equivalent to np.asarray(x).ravel()
Parameters
None :
Returns
ret : ndarray
self, 1-D, as an ndarray

Examples

>>> x = np.matrix(np.arange(12).reshape((3,4))); x
matrix([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> x.getA1()
array([ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11])

getH()
Returns the (complex) conjugate transpose of self.
Equivalent to np.transpose(self) if self is real-valued.
Parameters
None :
Returns
ret : matrix object
complex conjugate transpose of self

Examples

>>> x = np.matrix(np.arange(12).reshape((3,4)))
>>> z = x - 1j*x; z
matrix([[ 0. +0.j, 1. -1.j, 2. -2.j, 3. -3.j],
[ 4. -4.j, 5. -5.j, 6. -6.j, 7. -7.j],
[ 8. -8.j, 9. -9.j, 10.-10.j, 11.-11.j]])
>>> z.getH()
matrix([[ 0. +0.j, 4. +4.j, 8. +8.j],
[ 1. +1.j, 5. +5.j, 9. +9.j],
[ 2. +2.j, 6. +6.j, 10.+10.j],
[ 3. +3.j, 7. +7.j, 11.+11.j]])

106 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

getI()
Returns the (multiplicative) inverse of invertible self.
Parameters
None :
Returns
ret : matrix object
If self is non-singular, ret is such that ret * self == self * ret ==
np.matrix(np.eye(self[0,:].size) all return True.
Raises
numpy.linalg.linalg.LinAlgError: Singular matrix :
If self is singular.
See Also:
linalg.inv

Examples

>>> m = np.matrix([1, 2; 3, 4]); m

matrix([[1, 2],
[3, 4]])
>>> m.getI()
matrix([[-2. , 1. ],
[ 1.5, -0.5]])
>>> m.getI() * m
matrix([[ 1., 0.],
[ 0., 1.]])

getT()
Returns the transpose of the matrix.
Does not conjugate! For the complex conjugate transpose, use getH.
Parameters
None :
Returns
ret : matrix object
The (non-conjugated) transpose of the matrix.
See Also:
transpose, getH

Examples

>>> m = np.matrix([1, 2; 3, 4])

>>> m
matrix([[1, 2],
[3, 4]])
>>> m.getT()
matrix([[1, 3],
[2, 4]])

getfield(dtype, offset)
Returns a field of the given array as a certain type.

1.5. Standard array subclasses 107

NumPy Reference, Release 2.0.0.dev8464

A field is a view of the array data with each itemsize determined by the given type and the offset into the
current array, i.e. from offset * dtype.itemsize to (offset+1) * dtype.itemsize.
Parameters
dtype : str
String denoting the data type of the field.
offset : int
Number of dtype.itemsizes to skip before beginning the element view.

Examples

>>> x = np.diag([1.+1.j]*2)
>>> x
array([[ 1.+1.j, 0.+0.j],
[ 0.+0.j, 1.+1.j]])
>>> x.dtype
dtype(complex128)

>>> x.getfield(complex64, 0) # Note how this != x

array([[ 0.+1.875j, 0.+0.j ],
[ 0.+0.j , 0.+1.875j]], dtype=complex64)

>>> x.getfield(complex64,1) # Note how different this is than x

array([[ 0. +5.87173204e-39j, 0. +0.00000000e+00j],
[ 0. +0.00000000e+00j, 0. +5.87173204e-39j]], dtype=complex64)

>>> x.getfield(complex128, 0) # == x
array([[ 1.+1.j, 0.+0.j],
[ 0.+0.j, 1.+1.j]])

If the argument dtype is the same as x.dtype, then offset != 0 raises a ValueError:

>>> x.getfield(complex128, 1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: Need 0 <= offset <= 0 for requested type but received offset = 1

>>> x.getfield(float64, 0)
array([[ 1., 0.],
[ 0., 1.]])

>>> x.getfield(float64, 1)
array([[ 1.77658241e-307, 0.00000000e+000],
[ 0.00000000e+000, 1.77658241e-307]])

item(*args)
Copy an element of an array to a standard Python scalar and return it.
Parameters
*args : Arguments (variable number and type)
none: in this case, the method only works for arrays with one element (a.size == 1), which
element is copied into a standard Python scalar object and returned.

108 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

int_type: this argument is interpreted as a flat index into the array, specifying which ele-
ment to copy and return.
tuple of int_types: functions as does a single int_type argument, except that the argument
is interpreted as an nd-index into the array.
Returns
z : Standard Python scalar object
A copy of the specified element of the array as a suitable Python scalar

Notes
When the data type of a is longdouble or clongdouble, item() returns a scalar array object because there is
no available Python scalar that would not lose information. Void arrays return a buffer object for item(),
unless fields are defined, in which case a tuple is returned.
item is very similar to a[args], except, instead of an array scalar, a standard Python scalar is returned. This
can be useful for speeding up access to elements of the array and doing arithmetic on elements of the array
using Pythons optimized math.

Examples

>>> x = np.random.randint(9, size=(3, 3))

>>> x
array([[3, 1, 7],
[2, 8, 3],
[8, 5, 3]])
>>> x.item(3)
2
>>> x.item(7)
5
>>> x.item((0, 1))
1
>>> x.item((2, 2))
3

itemset()

max(axis=None, out=None)
Return the maximum value along an axis.
Parameters
See amax for complete descriptions :
See Also:
amax, ndarray.max

Notes
This is the same as ndarray.max, but returns a matrix object where ndarray.max would return an ndarray.

Examples

>>> x = np.matrix(np.arange(12).reshape((3,4))); x
matrix([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> x.max()

1.5. Standard array subclasses 109

NumPy Reference, Release 2.0.0.dev8464

11
>>> x.max(0)
matrix([[ 8, 9, 10, 11]])
>>> x.max(1)
matrix([[ 3],
[ 7],
[11]])

mean(axis=None, dtype=None, out=None)

Returns the average of the matrix elements along the given axis.
Refer to numpy.mean for full documentation.
See Also:
numpy.mean

Notes
Same as ndarray.mean except that, where that returns an ndarray, this returns a matrix object.

Examples

>>> x = np.matrix(np.arange(12).reshape((3, 4)))

>>> x
matrix([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> x.mean()
5.5
>>> x.mean(0)
matrix([[ 4., 5., 6., 7.]])
>>> x.mean(1)
matrix([[ 1.5],
[ 5.5],
[ 9.5]])

min(axis=None, out=None)
Return the minimum value along an axis.
Parameters
See amin for complete descriptions. :
See Also:
amin, ndarray.min

Notes
This is the same as ndarray.min, but returns a matrix object where ndarray.min would return an ndarray.

Examples

>>> x = -np.matrix(np.arange(12).reshape((3,4))); x
matrix([[ 0, -1, -2, -3],
[ -4, -5, -6, -7],
[ -8, -9, -10, -11]])
>>> x.min()
-11
>>> x.min(0)

110 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

matrix([[ -8, -9, -10, -11]])

>>> x.min(1)
matrix([[ -3],
[ -7],
[-11]])

newbyteorder(new_order=S)
Return the array with the same data viewed with a different byte order.
Equivalent to:

arr.view(arr.dtype.newbytorder(new_order))

Changes are also made in all fields and sub-arrays of the array data type.
Parameters
new_order : string, optional
Byte order to force; a value from the byte order specifications above. new_order codes
can be any of:

* S - swap dtype from current to opposite endian

* {<, L} - little endian
* {>, B} - big endian
* {=, N} - native order
* {|, I} - ignore (no change to byte order)

The default value (S) results in swapping the current byte order. The code does a
case-insensitive check on the first letter of new_order for the alternatives above. For
example, any of B or b or biggish are valid to specify big-endian.
Returns
new_arr : array
New array object with the dtype reflecting given change to the byte order.
nonzero()
Return the indices of the elements that are non-zero.
Refer to numpy.nonzero for full documentation.
See Also:

numpy.nonzero
equivalent function

prod(axis=None, dtype=None, out=None)

Return the product of the array elements over the given axis.
Refer to prod for full documentation.
See Also:
prod, ndarray.prod

Notes
Same as ndarray.prod, except, where that returns an ndarray, this returns a matrix object instead.

1.5. Standard array subclasses 111

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.matrix(np.arange(12).reshape((3,4))); x
matrix([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> x.prod()
0
>>> x.prod(0)
matrix([[ 0, 45, 120, 231]])
>>> x.prod(1)
matrix([[ 0],
[ 840],
[7920]])

ptp(axis=None, out=None)
Peak-to-peak (maximum - minimum) value along the given axis.
Refer to numpy.ptp for full documentation.
See Also:
numpy.ptp

Notes
Same as ndarray.ptp, except, where that would return an ndarray object, this returns a matrix object.

Examples

>>> x = np.matrix(np.arange(12).reshape((3,4))); x
matrix([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> x.ptp()
11
>>> x.ptp(0)
matrix([[8, 8, 8, 8]])
>>> x.ptp(1)
matrix([[3],
[3],
[3]])

put(indices, values, mode=raise)

Set a.flat[n] = values[n] for all n in indices.
Refer to numpy.put for full documentation.
See Also:

numpy.put
equivalent function

ravel([order])
Return a flattened array.
Refer to numpy.ravel for full documentation.
See Also:

112 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

numpy.ravel
equivalent function
ndarray.flat
a flat iterator on the array.

repeat(repeats, axis=None)
Repeat elements of an array.
Refer to numpy.repeat for full documentation.
See Also:

numpy.repeat
equivalent function

reshape(shape, order=C)
Returns an array containing the same data with a new shape.
Refer to numpy.reshape for full documentation.
See Also:

numpy.reshape
equivalent function

resize(new_shape, refcheck=True)
Change shape and size of array in-place.
Parameters
new_shape : tuple of ints, or n ints
Shape of resized array.
refcheck : bool, optional
If False, reference count will not be checked. Default is True.
Returns
None :
Raises
ValueError :
If a does not own its own data or references or views to it exist, and the data memory
must be changed.
SystemError :
If the order keyword argument is specified. This behaviour is a bug in NumPy.
See Also:

resize
Return a new array with the specified shape.

Notes
This reallocates space for the data area if necessary.
Only contiguous arrays (data elements consecutive in memory) can be resized.

1.5. Standard array subclasses 113

NumPy Reference, Release 2.0.0.dev8464

The purpose of the reference count check is to make sure you do not use this array as a buffer for another
Python object and then reallocate the memory. However, reference counts can increase in other ways so if
you are sure that you have not shared the memory for this array with another Python object, then you may
safely set refcheck to False.

Examples
Shrinking an array: array is flattened (in the order that the data are stored in memory), resized, and re-
shaped:

>>> a = np.array([[0, 1], [2, 3]], order=C)

>>> a.resize((2, 1))
>>> a
array([[0],
[1]])

>>> a = np.array([[0, 1], [2, 3]], order=F)

>>> a.resize((2, 1))
>>> a
array([[0],
[2]])

Enlarging an array: as above, but missing entries are filled with zeros:

>>> b = np.array([[0, 1], [2, 3]])

>>> b.resize(2, 3) # new_shape parameter doesnt have to be a tuple
>>> b
array([[0, 1, 2],
[3, 0, 0]])

Referencing an array prevents resizing...

>>> c = a
>>> a.resize((1, 1))
...
ValueError: cannot resize an array that has been referenced ...

Unless refcheck is False:

>>> a.resize((1, 1), refcheck=False)

>>> a
array([[0]])
>>> c
array([[0]])

round(decimals=0, out=None)
Return an array rounded a to the given number of decimals.
Refer to numpy.around for full documentation.
See Also:

numpy.around
equivalent function

114 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

searchsorted(v, side=left)
Find indices where elements of v should be inserted in a to maintain order.
For full documentation, see numpy.searchsorted
See Also:

numpy.searchsorted
equivalent function

setfield(val, dtype, offset=0)

Put a value into a specified place in a field defined by a data-type.
Place val into as field defined by dtype and beginning offset bytes into the field.
Parameters
val : object
Value to be placed in field.
dtype : dtype object
Data-type of the field in which to place val.
offset : int, optional
The number of bytes into the field at which to place val.
Returns
None :
See Also:
getfield

Examples

>>> x = np.eye(3)
>>> x.getfield(np.float64)
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])
>>> x.setfield(3, np.int32)
>>> x.getfield(np.int32)
array([[3, 3, 3],
[3, 3, 3],
[3, 3, 3]])
>>> x
array([[ 1.00000000e+000, 1.48219694e-323, 1.48219694e-323],
[ 1.48219694e-323, 1.00000000e+000, 1.48219694e-323],
[ 1.48219694e-323, 1.48219694e-323, 1.00000000e+000]])
>>> x.setfield(np.eye(3), np.int32)
>>> x
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])

setflags(write=None, align=None, uic=None)

Set array flags WRITEABLE, ALIGNED, and UPDATEIFCOPY, respectively.
These Boolean-valued flags affect how numpy interprets the memory area used by a (see Notes below).
The ALIGNED flag can only be set to True if the data is actually aligned according to the type. The

1.5. Standard array subclasses 115

NumPy Reference, Release 2.0.0.dev8464

UPDATEIFCOPY flag can never be set to True. The flag WRITEABLE can only be set to True if the array
owns its own memory, or the ultimate owner of the memory exposes a writeable buffer interface, or is a
string. (The exception for string is made so that unpickling can be done without copying memory.)
Parameters
write : bool, optional
Describes whether or not a can be written to.
align : bool, optional
Describes whether or not a is aligned properly for its type.
uic : bool, optional
Describes whether or not a is a copy of another base array.

Notes
Array flags provide information about how the memory area used for the array is to be interpreted. There
are 6 Boolean flags in use, only three of which can be changed by the user: UPDATEIFCOPY, WRITE-
ABLE, and ALIGNED.
WRITEABLE (W) the data area can be written to;
ALIGNED (A) the data and strides are aligned appropriately for the hardware (as determined by the com-
piler);
UPDATEIFCOPY (U) this array is a copy of some other array (referenced by .base). When this array is
deallocated, the base array will be updated with the contents of this array.
All flags can be accessed using their first (upper case) letter as well as the full name.

Examples

>>> y
array([[3, 1, 7],
[2, 0, 0],
[8, 5, 9]])
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : True
WRITEABLE : True
ALIGNED : True
UPDATEIFCOPY : False
>>> y.setflags(write=0, align=0)
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : True
WRITEABLE : False
ALIGNED : False
UPDATEIFCOPY : False
>>> y.setflags(uic=1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: cannot set UPDATEIFCOPY flag to True

sort(axis=-1, kind=quicksort, order=None)

Sort an array, in-place.

116 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Parameters
axis : int, optional
Axis along which to sort. Default is -1, which means sort along the last axis.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm. Default is quicksort.
order : list, optional
When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
See Also:

numpy.sort
Return a sorted copy of an array.
argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in sorted array.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = np.array([[1,4], [3,1]])

>>> a.sort(axis=1)
>>> a
array([[1, 4],
[1, 3]])
>>> a.sort(axis=0)
>>> a
array([[1, 3],
[1, 4]])

Use the order keyword to specify a field to use when sorting a structured array:

>>> a = np.array([(a, 2), (c, 1)], dtype=[(x, S1), (y, int)])

>>> a.sort(order=y)
>>> a
array([(c, 1), (a, 2)],
dtype=[(x, |S1), (y, <i4)])

squeeze()
Remove single-dimensional entries from the shape of a.
Refer to numpy.squeeze for full documentation.
See Also:

numpy.squeeze
equivalent function

1.5. Standard array subclasses 117

NumPy Reference, Release 2.0.0.dev8464

std(axis=None, dtype=None, out=None, ddof=0)

Return the standard deviation of the array elements along the given axis.
Refer to numpy.std for full documentation.
See Also:
numpy.std

Notes
This is the same as ndarray.std, except that where an ndarray would be returned, a matrix object is returned
instead.

Examples

>>> x = np.matrix(np.arange(12).reshape((3, 4)))

>>> x
matrix([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> x.std()
3.4520525295346629
>>> x.std(0)
matrix([[ 3.26598632, 3.26598632, 3.26598632, 3.26598632]])
>>> x.std(1)
matrix([[ 1.11803399],
[ 1.11803399],
[ 1.11803399]])

sum(axis=None, dtype=None, out=None)

Returns the sum of the matrix elements, along the given axis.
Refer to numpy.sum for full documentation.
See Also:
numpy.sum

Notes
This is the same as ndarray.sum, except that where an ndarray would be returned, a matrix object is
returned instead.

Examples

>>> x = np.matrix([[1, 2], [4, 3]])

>>> x.sum()
10
>>> x.sum(axis=1)
matrix([[3],
[7]])
>>> x.sum(axis=1, dtype=float)
matrix([[ 3.],
[ 7.]])
>>> out = np.zeros((1, 2), dtype=float)
>>> x.sum(axis=1, dtype=float, out=out)
matrix([[ 3.],
[ 7.]])

118 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

swapaxes(axis1, axis2)
Return a view of the array with axis1 and axis2 interchanged.
Refer to numpy.swapaxes for full documentation.
See Also:

numpy.swapaxes
equivalent function

take(indices, axis=None, out=None, mode=raise)

Return an array formed from the elements of a at the given indices.
Refer to numpy.take for full documentation.
See Also:

numpy.take
equivalent function

tofile(fid, sep="", format="%s")

Write array to a file as text or binary (default).
Data is always written in C order, independent of the order of a. The data produced by this method can
be recovered using the function fromfile().
Parameters
fid : file or str
An open file object, or a string containing a filename.
sep : str
Separator between array items for text output. If (empty), a binary file is written,
equivalent to file.write(a.tostring()).
format : str
Format string for text file output. Each entry in the array is formatted to text by first
converting it to the closest Python type, and then using format % item.

Notes
This is a convenience function for quick storage of array data. Information on endianness and precision
is lost, so this method is not a good choice for files intended to archive data or transport data between
machines with different endianness. Some of these problems can be overcome by outputting the data as
text files, at the expense of speed and file size.
tolist()
Return the matrix as a (possibly nested) list.
See ndarray.tolist for full documentation.
See Also:
ndarray.tolist

Examples

>>> x = np.matrix(np.arange(12).reshape((3,4))); x
matrix([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],

1.5. Standard array subclasses 119

NumPy Reference, Release 2.0.0.dev8464

[ 8, 9, 10, 11]])
>>> x.tolist()
[[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11]]

tostring(order=C)
Construct a Python string containing the raw data bytes in the array.
Constructs a Python string showing a copy of the raw contents of data memory. The string can be produced
in either C or Fortran, or Any order (the default is C-order). Any order means C-order unless the
F_CONTIGUOUS flag in the array is set, in which case it means Fortran order.
Parameters
order : {C, F, None}, optional
Order of the data for multidimensional arrays: C, Fortran, or the same as for the original
array.
Returns
s : str
A Python string exhibiting a copy of as raw data.

Examples

>>> x = np.array([[0, 1], [2, 3]])

>>> x.tostring()
\x00\x00\x00\x00\x01\x00\x00\x00\x02\x00\x00\x00\x03\x00\x00\x00
>>> x.tostring(C) == x.tostring()
True
>>> x.tostring(F)
\x00\x00\x00\x00\x02\x00\x00\x00\x01\x00\x00\x00\x03\x00\x00\x00

trace(offset=0, axis1=0, axis2=1, dtype=None, out=None)

Return the sum along diagonals of the array.
Refer to numpy.trace for full documentation.
See Also:

numpy.trace
equivalent function

transpose(*axes)
Returns a view of the array with axes transposed.
For a 1-D array, this has no effect. (To change between column and row vectors, first cast the 1-D ar-
ray into a matrix object.) For a 2-D array, this is the usual matrix transpose. For an n-D array, if axes
are given, their order indicates how the axes are permuted (see Examples). If axes are not provided
and a.shape = (i[0], i[1], ... i[n-2], i[n-1]), then a.transpose().shape
= (i[n-1], i[n-2], ... i[1], i[0]).
Parameters
axes : None, tuple of ints, or n ints
None or no argument: reverses the order of the axes.
tuple of ints: i in the j-th place in the tuple means as i-th axis becomes a.transpose()s
j-th axis.

120 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

n ints: same as an n-tuple of the same ints (this form is intended simply as a convenience
alternative to the tuple form)
Returns
out : ndarray
View of a, with axes suitably permuted.
See Also:

ndarray.T
Array property returning the array transposed.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> a
array([[1, 2],
[3, 4]])
>>> a.transpose()
array([[1, 3],
[2, 4]])
>>> a.transpose((1, 0))
array([[1, 3],
[2, 4]])
>>> a.transpose(1, 0)
array([[1, 3],
[2, 4]])

var(axis=None, dtype=None, out=None, ddof=0)

Returns the variance of the matrix elements, along the given axis.
Refer to numpy.var for full documentation.
See Also:
numpy.var

Notes
This is the same as ndarray.var, except that where an ndarray would be returned, a matrix object is returned
instead.

Examples

>>> x = np.matrix(np.arange(12).reshape((3, 4)))

>>> x
matrix([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> x.var()
11.916666666666666
>>> x.var(0)
matrix([[ 10.66666667, 10.66666667, 10.66666667, 10.66666667]])
>>> x.var(1)
matrix([[ 1.25],
[ 1.25],
[ 1.25]])

1.5. Standard array subclasses 121

NumPy Reference, Release 2.0.0.dev8464

view(dtype=None, type=None)
New view of array with the same data.
Parameters
dtype : data-type
Data-type descriptor of the returned view, e.g. float32 or int16.
type : python type
Type of the returned view, e.g. ndarray or matrix.

Notes
a.view() is used two different ways.
a.view(some_dtype) or a.view(dtype=some_dtype) constructs a view of the arrays memory with a different
dtype. This can cause a reinterpretation of the bytes of memory.
a.view(ndarray_subclass), or a.view(type=ndarray_subclass), just returns an instance of ndarray_subclass
that looks at the same array (same shape, dtype, etc.). This does not cause a reinterpretation of the memory.

Examples

>>> x = np.array([(1, 2)], dtype=[(a, np.int8), (b, np.int8)])

Viewing array data using a different type and dtype:

>>> y = x.view(dtype=np.int16, type=np.matrix)

>>> y
matrix([[513]], dtype=int16)
>>> print type(y)
<class numpy.matrixlib.defmatrix.matrix>

Creating a view on a structured array so it can be used in calculations

>>> x = np.array([(1, 2),(3,4)], dtype=[(a, np.int8), (b, np.int8)])

>>> xv = x.view(dtype=np.int8).reshape(-1,2)
>>> xv
array([[1, 2],
[3, 4]], dtype=int8)
>>> xv.mean(0)
array([ 2., 3.])

Making changes to the view changes the underlying array

>>> xv[0,1] = 20
>>> print x
[(1, 20) (3, 4)]

Using a view to convert an array to a record array:

>>> z = x.view(np.recarray)
>>> z.a
array([1], dtype=int8)

Views share data:

122 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> x[0] = (9, 10)

>>> z[0]
(9, 10)

asmatrix(data, dtype=None)
Interpret the input as a matrix.
Unlike matrix, asmatrix does not make a copy if the input is already a matrix or an ndarray. Equivalent to
matrix(data, copy=False).
Parameters
data : array_like
Input data.
Returns
mat : matrix
data interpreted as a matrix.

Examples

>>> x = np.array([[1, 2], [3, 4]])

>>> m = np.asmatrix(x)

>>> x[0,0] = 5

>>> m
matrix([[5, 2],
[3, 4]])

bmat(obj, ldict=None, gdict=None)

Build a matrix object from a string, nested sequence, or array.
Parameters
obj : string, sequence or array
Input data. Variables names in the current scope may be referenced, even if obj is a
string.
Returns
out : matrix
Returns a matrix object, which is a specialized 2-D array.
See Also:
matrix

Examples

>>> A = np.mat(1 1; 1 1)
>>> B = np.mat(2 2; 2 2)
>>> C = np.mat(3 4; 5 6)
>>> D = np.mat(7 8; 9 0)

All the following expressions construct the same block matrix:

1.5. Standard array subclasses 123

NumPy Reference, Release 2.0.0.dev8464

>>> np.bmat([[A, B], [C, D]])

matrix([[1, 1, 2, 2],
[1, 1, 2, 2],
[3, 4, 7, 8],
[5, 6, 9, 0]])
>>> np.bmat(np.r_[np.c_[A, B], np.c_[C, D]])
matrix([[1, 1, 2, 2],
[1, 1, 2, 2],
[3, 4, 7, 8],
[5, 6, 9, 0]])
>>> np.bmat(A,B; C,D)
matrix([[1, 1, 2, 2],
[1, 1, 2, 2],
[3, 4, 7, 8],
[5, 6, 9, 0]])

Example 1: Matrix creation from a string

>>> a=mat(1 2 3; 4 5 3)
>>> print (a*a.T).I
[[ 0.2924 -0.1345]
[-0.1345 0.0819]]

Example 2: Matrix creation from nested sequence

>>> mat([[1,5,10],[1.0,3,4j]])
matrix([[ 1.+0.j, 5.+0.j, 10.+0.j],
[ 1.+0.j, 3.+0.j, 0.+4.j]])

Example 3: Matrix creation from an array

>>> mat(random.rand(3,3)).T
matrix([[ 0.7699, 0.7922, 0.3294],
[ 0.2792, 0.0101, 0.9219],
[ 0.3398, 0.7571, 0.8197]])

1.5.3 Memory-mapped file arrays

Memory-mapped files are useful for reading and/or modifying small segments of a large file with regular layout,
without reading the entire file into memory. A simple subclass of the ndarray uses a memory-mapped file for the data
buffer of the array. For small files, the over-head of reading the entire file into memory is typically not significant,
however for large files using memory mapping can save considerable resources.
Memory-mapped-file arrays have one additional method (besides those they inherit from the ndarray): .flush()
which must be called manually by the user to ensure that any changes to the array actually get written to disk.
Note: Memory-mapped arrays use the the Python memory-map object which (prior to Python 2.5) does not allow
files to be larger than a certain size depending on the platform. This size is always < 2GB even on 64-bit systems.
memmap Create a memory-map to an array stored in a binary file on disk.
memmap.flush() Write any changes in the array to the file on disk.
class memmap()
Create a memory-map to an array stored in a binary file on disk.

124 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Memory-mapped files are used for accessing small segments of large files on disk, without reading the entire
file into memory. Numpys memmaps are array-like objects. This differs from Pythons mmap module, which
uses file-like objects.
Parameters
filename : str or file-like object
The file name or file object to be used as the array data buffer.
dtype : data-type, optional
The data-type used to interpret the file contents. Default is uint8.
mode : {r+, r, w+, c}, optional
The file is opened in this mode:
r Open existing file for reading only.
r+ Open existing file for reading and writing.
w+ Create or overwrite existing file for reading and writing.
c Copy-on-write: assignments affect data in memory, but changes are not saved to disk. The file on disk is
read-only.
Default is r+.
offset : int, optional
In the file, array data starts at this offset. Since offset is measured in bytes, it should be
a multiple of the byte-size of dtype. Requires shape=None. The default is 0.
shape : tuple, optional
The desired shape of the array. By default, the returned array will be 1-D with the
number of elements determined by file size and data-type.
order : {C, F}, optional
Specify the order of the ndarray memory layout: C (row-major) or Fortran (column-
major). This only has an effect if the shape is greater than 1-D. The default order is
C.

Notes
The memmap object can be used anywhere an ndarray is accepted. Given a memmap fp, isinstance(fp,
numpy.ndarray) returns True.
Memory-mapped arrays use the Python memory-map object which (prior to Python 2.5) does not allow files to
be larger than a certain size depending on the platform. This size is always < 2GB even on 64-bit systems.

Examples

>>> data = np.arange(12, dtype=float32)

>>> data.resize((3,4))

This example uses a temporary file so that doctest doesnt write files to your directory. You would use a normal
filename.

>>> from tempfile import mkdtemp

>>> import os.path as path
>>> filename = path.join(mkdtemp(), newfile.dat)

Create a memmap with dtype and shape that matches our data:

1.5. Standard array subclasses 125

NumPy Reference, Release 2.0.0.dev8464

>>> fp = np.memmap(filename, dtype=float32, mode=w+, shape=(3,4))

>>> fp
memmap([[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.]], dtype=float32)

Write data to memmap array:

>>> fp[:] = data[:]

>>> fp
memmap([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]], dtype=float32)

>>> fp.filename == path.abspath(filename)

True

Deletion flushes memory changes to disk before removing the object:

>>> del fp

Load the memmap and verify data was stored:

>>> newfp = np.memmap(filename, dtype=float32, mode=r, shape=(3,4))

>>> newfp
memmap([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]], dtype=float32)

Read-only memmap:

>>> fpr = np.memmap(filename, dtype=float32, mode=r, shape=(3,4))

>>> fpr.flags.writeable
False

Copy-on-write memmap:

>>> fpc = np.memmap(filename, dtype=float32, mode=c, shape=(3,4))

>>> fpc.flags.writeable
True

Its possible to assign to copy-on-write array, but values are only written into the memory copy of the array, and
not written to disk:

>>> fpc
memmap([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]], dtype=float32)
>>> fpc[0,:] = 0
>>> fpc
memmap([[ 0., 0., 0., 0.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]], dtype=float32)

File on disk is unchanged:

126 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> fpr
memmap([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]], dtype=float32)

Offset into a memmap:

>>> fpo = np.memmap(filename, dtype=float32, mode=r, offset=16)

>>> fpo
memmap([ 4., 5., 6., 7., 8., 9., 10., 11.], dtype=float32)

Attributes
filename str Path to the mapped file.
offset int Offset position in the file.
mode str File mode.

Methods
close() Close the memmap file.
flush() Write any changes in the array to the file on disk.
close()
Close the memmap file. Does nothing.
flush()
Write any changes in the array to the file on disk.
For further information, see memmap.
Parameters
None :
See Also:
memmap
flush()
Write any changes in the array to the file on disk.
For further information, see memmap.
Parameters
None :
See Also:
memmap
Example:

>>> a = memmap(newfile.dat, dtype=float, mode=w+, shape=1000)

>>> a[10] = 10.0
>>> a[30] = 30.0
>>> del a
>>> b = fromfile(newfile.dat, dtype=float)
>>> print b[10], b[30]
10.0 30.0
>>> a = memmap(newfile.dat, dtype=float)

1.5. Standard array subclasses 127

NumPy Reference, Release 2.0.0.dev8464

>>> print a[10], a[30]

10.0 30.0

1.5.4 Character arrays (numpy.char)

See Also:
Creating character arrays (numpy.char)
Note: The chararray class exists for backwards compatibility with Numarray, it is not recommended for new devel-
opment. Starting from numpy 1.4, if one needs arrays of strings, it is recommended to use arrays of dtype object_,
string_ or unicode_, and use the free functions in the numpy.char module for fast vectorized string operations.
These are enhanced arrays of either string_ type or unicode_ type. These arrays inherit from the ndarray,
but specially-define the operations +, *, and % on a (broadcasting) element-by-element basis. These operations are
not available on the standard ndarray of character type. In addition, the chararray has all of the standard
string (and unicode) methods, executing them on an element-by-element basis. Perhaps the easiest way to create
a chararray is to use self.view(chararray) where self is an ndarray of str or unicode data-type. However, a
chararray can also be created using the numpy.chararray constructor, or via the numpy.char.array function:
chararray Provides a convenient view on arrays of string and unicode
values.
core.defchararray.array(obj[, itemsize, Create a chararray.
...])
class chararray()
Provides a convenient view on arrays of string and unicode values.
Note: The chararray class exists for backwards compatibility with Numarray, it is not recommended for new
development. Starting from numpy 1.4, if one needs arrays of strings, it is recommended to use arrays of dtype
object_, string_ or unicode_, and use the free functions in the numpy.char module for fast vectorized string
operations.
Versus a regular Numpy array of type str or unicode, this class adds the following functionality:
1.values automatically have whitespace removed from the end when indexed
2.comparison operators automatically remove whitespace from the end when comparing values
3.vectorized string operations are provided as methods (e.g. endswith) and infix operators (e.g. "+", "*",
"%")
chararrays should be created using numpy.char.array or numpy.char.asarray, rather than this con-
structor directly.
This constructor creates the array, using buffer (with offset and strides) if it is not None. If buffer is None, then
constructs a new array with strides in C order, unless both len(shape) >= 2 and order=Fortran,
in which case strides is in Fortran order.
Parameters
shape : tuple
Shape of the array.
itemsize : int, optional
Length of each array element, in number of characters. Default is 1.
unicode : bool, optional
Are the array elements of type unicode (True) or string (False). Default is False.

128 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

buffer : int, optional

Memory address of the start of the array data. Default is None, in which case a new
array is created.
offset : int, optional
Fixed stride displacement from the beginning of an axis? Default is 0. Needs to be >=0.
strides : array_like of ints, optional
Strides for the array (see ndarray.strides for full description). Default is None.
order : {C, F}, optional
The order in which the array data is stored in memory: C -> row major order (the
default), F -> column major (Fortran) order.

Examples

>>> charar = np.chararray((3, 3))

>>> charar[:] = a
>>> charar
chararray([[a, a, a],
[a, a, a],
[a, a, a]],
dtype=|S1)

>>> charar = np.chararray(charar.shape, itemsize=5)

>>> charar[:] = abc
>>> charar
chararray([[abc, abc, abc],
[abc, abc, abc],
[abc, abc, abc]],
dtype=|S5)

Methods

astype(t) Copy of the array, cast

argsort([axis, kind, order])
copy([order]) Return a copy of
count(sub[, start, end]) Returns an array with the number of non-overlapping
decode([encoding, errors]) Calls str.decode elem
dump(file) Dump a pickle of the ar
dumps() Returns the pickle of the
encode([encoding, errors]) Calls str.encode elem
endswith(suffix[, start, end]) Returns a boolean array which
expandtabs([tabsize]) Return a copy of each string element where all
fill(value) Fill the array with
find(sub[, start, end]) For each element, return the lowest index
flatten([order]) Return a copy of the array
getfield(dtype, offset) Returns a field of the given
index(sub[, start, end]) Like find, but raises ValueError wh
isalnum() Returns true for each element if all characters in the string
isalpha() Returns true for each element if all characters in the strin
isdecimal() For each element in self,
isdigit() Returns true for each element if all characters in the st

1.5. Standard array subclasses 129

NumPy Reference, Release 2.0.0.dev8464

Table 1.4 continued from

islower() Returns true for each element if all cased characters in the st
isnumeric() For each element in self,
isspace() Returns true for each element if there are only whitespace ch
istitle() Returns true for each element if the element is a tit
isupper() Returns true for each element if all cased characters in the
item(*args) Copy an element of an array to
join(seq) Return a string which is the concatenation
ljust(width[, fillchar]) Return an array with the elements of se
lower() Return an array with the elements
lstrip([chars]) For each element in self, return a
nonzero() Return the indices of the
put(indices, values[, mode]) Set a.flat[n] = values[n] for
ravel() Return a flattened
repeat(repeats[, axis]) Repeat elements of
replace(old, new[, count]) For each element in self, return a copy of the
reshape(shape[, order]) Returns an array containing the sam
resize(new_shape[, refcheck]) Change shape and size
rfind(sub[, start, end]) For each element in self, return the highest index in the string
rindex(sub[, start, end]) Like rfind, but raises ValueEr
rjust(width[, fillchar]) Return an array with the elements of se
rsplit([sep, maxsplit]) For each element in self, return a list of the
rstrip([chars]) For each element in self, return a
searchsorted(v[, side]) Find indices where elements of v should
setfield(val, dtype[, offset]) Put a value into a specified place
setflags([write, align, uic]) Set array flags WRITEABLE, ALIGNE
sort([axis, kind, order]) Sort an array,
split([sep, maxsplit]) For each element in self, return a list of the
splitlines([keepends]) For each element in self, return a list of
squeeze() Remove single-dimensional entries from
startswith(prefix[, start, end]) Returns a boolean array which
strip([chars]) For each element in self, return a copy
swapaxes(axis1, axis2) Return a view of the array
swapcase() For each element in self, return a copy of the stri
take(indices[, axis, out, mode]) Return an array formed from the el
title() For each element in self, return a titlecased version of the string:
tofile(fid[, sep, format]) Write array to a file as
tolist() Return the array as a
tostring([order]) Construct a Python string containing the
translate(table[, deletechars]) For each element in self, return a copy of the string where all characters occurring in the option
transpose(*axes) Returns a view of the
upper() Return an array with the elements
view([dtype, type]) New view of array with
zfill(width) Return the numeric string left-filled with

astype(t)
Copy of the array, cast to a specified type.
Parameters
t : string or dtype
Typecode or data-type to which the array is cast.

130 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.array([1, 2, 2.5])

>>> x
array([ 1. , 2. , 2.5])

>>> x.astype(int)
array([1, 2, 2])

argsort(axis=-1, kind=quicksort, order=None)

copy(order=C)
Return a copy of the array.
Parameters
order : {C, F, A}, optional
By default, the result is stored in C-contiguous (row-major) order in memory. If order is
F, the result has Fortran (column-major) order. If order is A (Any), then the result
has the same order as the input.

Examples

>>> x = np.array([[1,2,3],[4,5,6]], order=F)

>>> y = x.copy()

>>> x.fill(0)

>>> x
array([[0, 0, 0],
[0, 0, 0]])

>>> y
array([[1, 2, 3],
[4, 5, 6]])

>>> y.flags[C_CONTIGUOUS]
True

count(sub, start=0, end=None)

Returns an array with the number of non-overlapping occurrences of substring sub in the range [start,
end].
See Also:
char.count
decode(encoding=None, errors=None)
Calls str.decode element-wise.
See Also:
char.decode

1.5. Standard array subclasses 131

NumPy Reference, Release 2.0.0.dev8464

dump(file)
Dump a pickle of the array to the specified file. The array can be read back with pickle.load or numpy.load.
Parameters
file : str
A string naming the dump file.
dumps()
Returns the pickle of the array as a string. pickle.loads or numpy.loads will convert the string back to an
array.
Parameters
None :
encode(encoding=None, errors=None)
Calls str.encode element-wise.
See Also:
char.encode
endswith(suffix, start=0, end=None)
Returns a boolean array which is True where the string element in self ends with suffix, otherwise False.
See Also:
char.endswith
expandtabs(tabsize=8)
Return a copy of each string element where all tab characters are replaced by one or more spaces.
See Also:
char.expandtabs
fill(value)
Fill the array with a scalar value.
Parameters
value : scalar
All elements of a will be assigned this value.

Examples

>>> a = np.array([1, 2])

>>> a.fill(0)
>>> a
array([0, 0])
>>> a = np.empty(2)
>>> a.fill(1)
>>> a
array([ 1., 1.])

find(sub, start=0, end=None)

For each element, return the lowest index in the string where substring sub is found.
See Also:
char.find
flatten(order=C)
Return a copy of the array collapsed into one dimension.

132 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Parameters
order : {C, F}, optional
Whether to flatten in C (row-major) or Fortran (column-major) order. The default is C.
Returns
y : ndarray
A copy of the input array, flattened to one dimension.
See Also:

ravel
Return a flattened array.
flat
A 1-D flat iterator over the array.

Examples

>>> a = np.array([[1,2], [3,4]])

>>> a.flatten()
array([1, 2, 3, 4])
>>> a.flatten(F)
array([1, 3, 2, 4])

getfield(dtype, offset)
Returns a field of the given array as a certain type.
A field is a view of the array data with each itemsize determined by the given type and the offset into the
current array, i.e. from offset * dtype.itemsize to (offset+1) * dtype.itemsize.
Parameters
dtype : str
String denoting the data type of the field.
offset : int
Number of dtype.itemsizes to skip before beginning the element view.

Examples

>>> x = np.diag([1.+1.j]*2)
>>> x
array([[ 1.+1.j, 0.+0.j],
[ 0.+0.j, 1.+1.j]])
>>> x.dtype
dtype(complex128)

>>> x.getfield(complex64, 0) # Note how this != x

array([[ 0.+1.875j, 0.+0.j ],
[ 0.+0.j , 0.+1.875j]], dtype=complex64)

>>> x.getfield(complex64,1) # Note how different this is than x

array([[ 0. +5.87173204e-39j, 0. +0.00000000e+00j],
[ 0. +0.00000000e+00j, 0. +5.87173204e-39j]], dtype=complex64)

1.5. Standard array subclasses 133

NumPy Reference, Release 2.0.0.dev8464

>>> x.getfield(complex128, 0) # == x
array([[ 1.+1.j, 0.+0.j],
[ 0.+0.j, 1.+1.j]])

If the argument dtype is the same as x.dtype, then offset != 0 raises a ValueError:

>>> x.getfield(complex128, 1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: Need 0 <= offset <= 0 for requested type but received offset = 1

>>> x.getfield(float64, 0)
array([[ 1., 0.],
[ 0., 1.]])

>>> x.getfield(float64, 1)
array([[ 1.77658241e-307, 0.00000000e+000],
[ 0.00000000e+000, 1.77658241e-307]])

index(sub, start=0, end=None)

Like find, but raises ValueError when the substring is not found.
See Also:
char.index
isalnum()
Returns true for each element if all characters in the string are alphanumeric and there is at least one
character, false otherwise.
See Also:
char.isalnum
isalpha()
Returns true for each element if all characters in the string are alphabetic and there is at least one character,
false otherwise.
See Also:
char.isalpha
isdecimal()
For each element in self, return True if there are only decimal characters in the element.
See Also:
char.isdecimal
isdigit()
Returns true for each element if all characters in the string are digits and there is at least one character,
false otherwise.
See Also:
char.isdigit
islower()
Returns true for each element if all cased characters in the string are lowercase and there is at least one
cased character, false otherwise.

134 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

See Also:
char.islower
isnumeric()
For each element in self, return True if there are only numeric characters in the element.
See Also:
char.isnumeric
isspace()
Returns true for each element if there are only whitespace characters in the string and there is at least one
character, false otherwise.
See Also:
char.isspace
istitle()
Returns true for each element if the element is a titlecased string and there is at least one character, false
otherwise.
See Also:
char.istitle
isupper()
Returns true for each element if all cased characters in the string are uppercase and there is at least one
character, false otherwise.
See Also:
char.isupper
item(*args)
Copy an element of an array to a standard Python scalar and return it.
Parameters
*args : Arguments (variable number and type)
none: in this case, the method only works for arrays with one element (a.size == 1), which
element is copied into a standard Python scalar object and returned.
int_type: this argument is interpreted as a flat index into the array, specifying which ele-
ment to copy and return.
tuple of int_types: functions as does a single int_type argument, except that the argument
is interpreted as an nd-index into the array.
Returns
z : Standard Python scalar object
A copy of the specified element of the array as a suitable Python scalar

Notes
When the data type of a is longdouble or clongdouble, item() returns a scalar array object because there is
no available Python scalar that would not lose information. Void arrays return a buffer object for item(),
unless fields are defined, in which case a tuple is returned.
item is very similar to a[args], except, instead of an array scalar, a standard Python scalar is returned. This
can be useful for speeding up access to elements of the array and doing arithmetic on elements of the array
using Pythons optimized math.

1.5. Standard array subclasses 135

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.random.randint(9, size=(3, 3))

>>> x
array([[3, 1, 7],
[2, 8, 3],
[8, 5, 3]])
>>> x.item(3)
2
>>> x.item(7)
5
>>> x.item((0, 1))
1
>>> x.item((2, 2))
3

join(seq)
Return a string which is the concatenation of the strings in the sequence seq.
See Also:
char.join
ljust(width, fillchar= )
Return an array with the elements of self left-justified in a string of length width.
See Also:
char.ljust
lower()
Return an array with the elements of self converted to lowercase.
See Also:
char.lower
lstrip(chars=None)
For each element in self, return a copy with the leading characters removed.
See Also:
char.lstrip
nonzero()
Return the indices of the elements that are non-zero.
Refer to numpy.nonzero for full documentation.
See Also:

numpy.nonzero
equivalent function

put(indices, values, mode=raise)

Set a.flat[n] = values[n] for all n in indices.
Refer to numpy.put for full documentation.
See Also:

numpy.put
equivalent function

136 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

ravel([order])
Return a flattened array.
Refer to numpy.ravel for full documentation.
See Also:

numpy.ravel
equivalent function
ndarray.flat
a flat iterator on the array.

repeat(repeats, axis=None)
Repeat elements of an array.
Refer to numpy.repeat for full documentation.
See Also:

numpy.repeat
equivalent function

replace(old, new, count=None)

For each element in self, return a copy of the string with all occurrences of substring old replaced by new.
See Also:
char.replace
reshape(shape, order=C)
Returns an array containing the same data with a new shape.
Refer to numpy.reshape for full documentation.
See Also:

numpy.reshape
equivalent function

resize(new_shape, refcheck=True)
Change shape and size of array in-place.
Parameters
new_shape : tuple of ints, or n ints
Shape of resized array.
refcheck : bool, optional
If False, reference count will not be checked. Default is True.
Returns
None :
Raises
ValueError :
If a does not own its own data or references or views to it exist, and the data memory
must be changed.
SystemError :

1.5. Standard array subclasses 137

NumPy Reference, Release 2.0.0.dev8464

If the order keyword argument is specified. This behaviour is a bug in NumPy.

See Also:

resize
Return a new array with the specified shape.

Notes
This reallocates space for the data area if necessary.
Only contiguous arrays (data elements consecutive in memory) can be resized.
The purpose of the reference count check is to make sure you do not use this array as a buffer for another
Python object and then reallocate the memory. However, reference counts can increase in other ways so if
you are sure that you have not shared the memory for this array with another Python object, then you may
safely set refcheck to False.

Examples
Shrinking an array: array is flattened (in the order that the data are stored in memory), resized, and re-
shaped:

>>> a = np.array([[0, 1], [2, 3]], order=C)

>>> a.resize((2, 1))
>>> a
array([[0],
[1]])

>>> a = np.array([[0, 1], [2, 3]], order=F)

>>> a.resize((2, 1))
>>> a
array([[0],
[2]])

Enlarging an array: as above, but missing entries are filled with zeros:

>>> b = np.array([[0, 1], [2, 3]])

>>> b.resize(2, 3) # new_shape parameter doesnt have to be a tuple
>>> b
array([[0, 1, 2],
[3, 0, 0]])

Referencing an array prevents resizing...

>>> c = a
>>> a.resize((1, 1))
...
ValueError: cannot resize an array that has been referenced ...

Unless refcheck is False:

>>> a.resize((1, 1), refcheck=False)

>>> a
array([[0]])
>>> c
array([[0]])

138 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

rfind(sub, start=0, end=None)

For each element in self, return the highest index in the string where substring sub is found, such that sub
is contained within [start, end].
See Also:
char.rfind
rindex(sub, start=0, end=None)
Like rfind, but raises ValueError when the substring sub is not found.
See Also:
char.rindex
rjust(width, fillchar= )
Return an array with the elements of self right-justified in a string of length width.
See Also:
char.rjust
rsplit(sep=None, maxsplit=None)
For each element in self, return a list of the words in the string, using sep as the delimiter string.
See Also:
char.rsplit
rstrip(chars=None)
For each element in self, return a copy with the trailing characters removed.
See Also:
char.rstrip
searchsorted(v, side=left)
Find indices where elements of v should be inserted in a to maintain order.
For full documentation, see numpy.searchsorted
See Also:

numpy.searchsorted
equivalent function

setfield(val, dtype, offset=0)

Put a value into a specified place in a field defined by a data-type.
Place val into as field defined by dtype and beginning offset bytes into the field.
Parameters
val : object
Value to be placed in field.
dtype : dtype object
Data-type of the field in which to place val.
offset : int, optional
The number of bytes into the field at which to place val.
Returns
None :

1.5. Standard array subclasses 139

NumPy Reference, Release 2.0.0.dev8464

See Also:
getfield

Examples

>>> x = np.eye(3)
>>> x.getfield(np.float64)
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])
>>> x.setfield(3, np.int32)
>>> x.getfield(np.int32)
array([[3, 3, 3],
[3, 3, 3],
[3, 3, 3]])
>>> x
array([[ 1.00000000e+000, 1.48219694e-323, 1.48219694e-323],
[ 1.48219694e-323, 1.00000000e+000, 1.48219694e-323],
[ 1.48219694e-323, 1.48219694e-323, 1.00000000e+000]])
>>> x.setfield(np.eye(3), np.int32)
>>> x
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])

setflags(write=None, align=None, uic=None)

Set array flags WRITEABLE, ALIGNED, and UPDATEIFCOPY, respectively.
These Boolean-valued flags affect how numpy interprets the memory area used by a (see Notes below).
The ALIGNED flag can only be set to True if the data is actually aligned according to the type. The
UPDATEIFCOPY flag can never be set to True. The flag WRITEABLE can only be set to True if the array
owns its own memory, or the ultimate owner of the memory exposes a writeable buffer interface, or is a
string. (The exception for string is made so that unpickling can be done without copying memory.)
Parameters
write : bool, optional
Describes whether or not a can be written to.
align : bool, optional
Describes whether or not a is aligned properly for its type.
uic : bool, optional
Describes whether or not a is a copy of another base array.

Notes
Array flags provide information about how the memory area used for the array is to be interpreted. There
are 6 Boolean flags in use, only three of which can be changed by the user: UPDATEIFCOPY, WRITE-
ABLE, and ALIGNED.
WRITEABLE (W) the data area can be written to;
ALIGNED (A) the data and strides are aligned appropriately for the hardware (as determined by the com-
piler);
UPDATEIFCOPY (U) this array is a copy of some other array (referenced by .base). When this array is
deallocated, the base array will be updated with the contents of this array.

140 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

All flags can be accessed using their first (upper case) letter as well as the full name.

Examples

>>> y
array([[3, 1, 7],
[2, 0, 0],
[8, 5, 9]])
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : True
WRITEABLE : True
ALIGNED : True
UPDATEIFCOPY : False
>>> y.setflags(write=0, align=0)
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : True
WRITEABLE : False
ALIGNED : False
UPDATEIFCOPY : False
>>> y.setflags(uic=1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: cannot set UPDATEIFCOPY flag to True

sort(axis=-1, kind=quicksort, order=None)

Sort an array, in-place.
Parameters
axis : int, optional
Axis along which to sort. Default is -1, which means sort along the last axis.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm. Default is quicksort.
order : list, optional
When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
See Also:

numpy.sort
Return a sorted copy of an array.
argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in sorted array.

Notes
See sort for notes on the different sorting algorithms.

1.5. Standard array subclasses 141

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.array([[1,4], [3,1]])

>>> a.sort(axis=1)
>>> a
array([[1, 4],
[1, 3]])
>>> a.sort(axis=0)
>>> a
array([[1, 3],
[1, 4]])

Use the order keyword to specify a field to use when sorting a structured array:

>>> a = np.array([(a, 2), (c, 1)], dtype=[(x, S1), (y, int)])

>>> a.sort(order=y)
>>> a
array([(c, 1), (a, 2)],
dtype=[(x, |S1), (y, <i4)])

split(sep=None, maxsplit=None)
For each element in self, return a list of the words in the string, using sep as the delimiter string.
See Also:
char.split
splitlines(keepends=None)
For each element in self, return a list of the lines in the element, breaking at line boundaries.
See Also:
char.splitlines
squeeze()
Remove single-dimensional entries from the shape of a.
Refer to numpy.squeeze for full documentation.
See Also:

numpy.squeeze
equivalent function

startswith(prefix, start=0, end=None)

Returns a boolean array which is True where the string element in self starts with prefix, otherwise False.
See Also:
char.startswith
strip(chars=None)
For each element in self, return a copy with the leading and trailing characters removed.
See Also:
char.strip
swapaxes(axis1, axis2)
Return a view of the array with axis1 and axis2 interchanged.
Refer to numpy.swapaxes for full documentation.

142 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

See Also:

numpy.swapaxes
equivalent function

swapcase()
For each element in self, return a copy of the string with uppercase characters converted to lowercase and
vice versa.
See Also:
char.swapcase
take(indices, axis=None, out=None, mode=raise)
Return an array formed from the elements of a at the given indices.
Refer to numpy.take for full documentation.
See Also:

numpy.take
equivalent function

title()
For each element in self, return a titlecased version of the string: words start with uppercase characters, all
remaining cased characters are lowercase.
See Also:
char.title
tofile(fid, sep="", format="%s")
Write array to a file as text or binary (default).
Data is always written in C order, independent of the order of a. The data produced by this method can
be recovered using the function fromfile().
Parameters
fid : file or str
An open file object, or a string containing a filename.
sep : str
Separator between array items for text output. If (empty), a binary file is written,
equivalent to file.write(a.tostring()).
format : str
Format string for text file output. Each entry in the array is formatted to text by first
converting it to the closest Python type, and then using format % item.

Notes
This is a convenience function for quick storage of array data. Information on endianness and precision
is lost, so this method is not a good choice for files intended to archive data or transport data between
machines with different endianness. Some of these problems can be overcome by outputting the data as
text files, at the expense of speed and file size.
tolist()
Return the array as a (possibly nested) list.

1.5. Standard array subclasses 143

NumPy Reference, Release 2.0.0.dev8464

Return a copy of the array data as a (nested) Python list. Data items are converted to the nearest compatible
Python type.
Parameters
none :
Returns
y : list
The possibly nested list of array elements.

Notes
The array may be recreated, a = np.array(a.tolist()).

Examples

>>> a = np.array([1, 2])

>>> a.tolist()
[1, 2]
>>> a = np.array([[1, 2], [3, 4]])
>>> list(a)
[array([1, 2]), array([3, 4])]
>>> a.tolist()
[[1, 2], [3, 4]]

tostring(order=C)
Construct a Python string containing the raw data bytes in the array.
Constructs a Python string showing a copy of the raw contents of data memory. The string can be produced
in either C or Fortran, or Any order (the default is C-order). Any order means C-order unless the
F_CONTIGUOUS flag in the array is set, in which case it means Fortran order.
Parameters
order : {C, F, None}, optional
Order of the data for multidimensional arrays: C, Fortran, or the same as for the original
array.
Returns
s : str
A Python string exhibiting a copy of as raw data.

Examples

>>> x = np.array([[0, 1], [2, 3]])

>>> x.tostring()
\x00\x00\x00\x00\x01\x00\x00\x00\x02\x00\x00\x00\x03\x00\x00\x00
>>> x.tostring(C) == x.tostring()
True
>>> x.tostring(F)
\x00\x00\x00\x00\x02\x00\x00\x00\x01\x00\x00\x00\x03\x00\x00\x00

translate(table, deletechars=None)
For each element in self, return a copy of the string where all characters occurring in the optional argument
deletechars are removed, and the remaining characters have been mapped through the given translation
table.
See Also:

144 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

char.translate
transpose(*axes)
Returns a view of the array with axes transposed.
For a 1-D array, this has no effect. (To change between column and row vectors, first cast the 1-D ar-
ray into a matrix object.) For a 2-D array, this is the usual matrix transpose. For an n-D array, if axes
are given, their order indicates how the axes are permuted (see Examples). If axes are not provided
and a.shape = (i[0], i[1], ... i[n-2], i[n-1]), then a.transpose().shape
= (i[n-1], i[n-2], ... i[1], i[0]).
Parameters
axes : None, tuple of ints, or n ints
None or no argument: reverses the order of the axes.
tuple of ints: i in the j-th place in the tuple means as i-th axis becomes a.transpose()s
j-th axis.
n ints: same as an n-tuple of the same ints (this form is intended simply as a convenience
alternative to the tuple form)
Returns
out : ndarray
View of a, with axes suitably permuted.
See Also:

ndarray.T
Array property returning the array transposed.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> a
array([[1, 2],
[3, 4]])
>>> a.transpose()
array([[1, 3],
[2, 4]])
>>> a.transpose((1, 0))
array([[1, 3],
[2, 4]])
>>> a.transpose(1, 0)
array([[1, 3],
[2, 4]])

upper()
Return an array with the elements of self converted to uppercase.
See Also:
char.upper
view(dtype=None, type=None)
New view of array with the same data.
Parameters
dtype : data-type
Data-type descriptor of the returned view, e.g. float32 or int16.

1.5. Standard array subclasses 145

NumPy Reference, Release 2.0.0.dev8464

type : python type

Type of the returned view, e.g. ndarray or matrix.

Notes
a.view() is used two different ways.
a.view(some_dtype) or a.view(dtype=some_dtype) constructs a view of the arrays memory with a different
dtype. This can cause a reinterpretation of the bytes of memory.
a.view(ndarray_subclass), or a.view(type=ndarray_subclass), just returns an instance of ndarray_subclass
that looks at the same array (same shape, dtype, etc.). This does not cause a reinterpretation of the memory.

Examples

>>> x = np.array([(1, 2)], dtype=[(a, np.int8), (b, np.int8)])

Viewing array data using a different type and dtype:

>>> y = x.view(dtype=np.int16, type=np.matrix)

>>> y
matrix([[513]], dtype=int16)
>>> print type(y)
<class numpy.matrixlib.defmatrix.matrix>

Creating a view on a structured array so it can be used in calculations

>>> x = np.array([(1, 2),(3,4)], dtype=[(a, np.int8), (b, np.int8)])

>>> xv = x.view(dtype=np.int8).reshape(-1,2)
>>> xv
array([[1, 2],
[3, 4]], dtype=int8)
>>> xv.mean(0)
array([ 2., 3.])

Making changes to the view changes the underlying array

>>> xv[0,1] = 20
>>> print x
[(1, 20) (3, 4)]

Using a view to convert an array to a record array:

>>> z = x.view(np.recarray)
>>> z.a
array([1], dtype=int8)

Views share data:

>>> x[0] = (9, 10)

>>> z[0]
(9, 10)

zfill(width)
Return the numeric string left-filled with zeros in a string of length width.
See Also:

146 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

char.zfill
array(obj, itemsize=None, copy=True, unicode=None, order=None)
Create a chararray.
Note: This class is provided for numarray backward-compatibility. New code (not concerned with numarray
compatibility) should use arrays of type string_ or unicode_ and use the free functions in numpy.char for
fast vectorized string operations instead.
Versus a regular Numpy array of type str or unicode, this class adds the following functionality:
1.values automatically have whitespace removed from the end when indexed
2.comparison operators automatically remove whitespace from the end when comparing values
3.vectorized string operations are provided as methods (e.g. str.endswith) and infix operators (e.g. +, *, %)

Parameters
obj : array of str or unicode-like
itemsize : int, optional
itemsize is the number of characters per scalar in the resulting array. If itemsize is
None, and obj is an object array or a Python list, the itemsize will be automatically
determined. If itemsize is provided and obj is of type str or unicode, then the obj string
will be chunked into itemsize pieces.
copy : bool, optional
If true (default), then the object is copied. Otherwise, a copy will only be made if
__array__ returns a copy, if obj is a nested sequence, or if a copy is needed to satisfy
any of the other requirements (itemsize, unicode, order, etc.).
unicode : bool, optional
When true, the resulting chararray can contain Unicode characters, when false only
8-bit characters. If unicode is None and obj is one of the following:
a chararray,
an ndarray of type str or unicode
a Python str or unicode object,
then the unicode setting of the output array will be automatically determined.
order : {C, F, A}, optional
Specify the order of the array. If order is C (default), then the array will be in C-
contiguous order (last-index varies the fastest). If order is F, then the returned array
will be in Fortran-contiguous order (first-index varies the fastest). If order is A, then
the returned array may be in any order (either C-, Fortran-contiguous, or even discon-
tiguous).

Another difference with the standard ndarray of str data-type is that the chararray inherits the feature introduced by
Numarray that white-space at the end of any element in the array will be ignored on item retrieval and comparison
operations.

1.5.5 Record arrays (numpy.rec)

See Also:
Creating record arrays (numpy.rec), Data type routines, Data type objects (dtype).

1.5. Standard array subclasses 147

NumPy Reference, Release 2.0.0.dev8464

Numpy provides the recarray class which allows accessing the fields of a record/structured array as attributes, and
a corresponding scalar data type object record.
recarray Construct an ndarray that allows field access using attributes.
record A data-type scalar that allows field access as attribute lookup.
class recarray()
Construct an ndarray that allows field access using attributes.
Arrays may have a data-types containing fields, analogous to columns in a spread sheet. An example is [(x,
int), (y, float)], where each entry in the array is a pair of (int, float). Normally, these attributes
are accessed using dictionary lookups such as arr[x] and arr[y]. Record arrays allow the fields to be
accessed as members of the array, using arr.x and arr.y.
Parameters
shape : tuple
Shape of output array.
dtype : data-type, optional
The desired data-type. By default, the data-type is determined from formats, names,
titles, aligned and byteorder.
formats : list of data-types, optional
A list containing the data-types for the different columns, e.g. [i4, f8,
i4]. formats does not support the new convention of using types directly, i.e.
(int, float, int). Note that formats must be a list, not a tuple. Given that
formats is somewhat limited, we recommend specifying dtype instead.
names : tuple of str, optional
The name of each column, e.g. (x, y, z).
buf : buffer, optional
By default, a new array is created of the given shape and data-type. If buf is specified
and is an object exposing the buffer interface, the array will use the memory from the
existing buffer. In this case, the offset and strides keywords are available.
Returns
rec : recarray
Empty array of the given shape and type.
See Also:

rec.fromrecords
Construct a record array from data.
record
fundamental data-type for recarray.
format_parser
determine a data-type from formats, names, titles.

Notes
This constructor can be compared to empty: it creates a new record array but does not fill it with data. To create
a record array from data, use one of the following methods:
1.Create a standard ndarray and convert it to a record array, using arr.view(np.recarray)

148 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

2.Use the buf keyword.

3.Use np.rec.fromrecords.

Examples
Create an array with two fields, x and y:

>>> x = np.array([(1.0, 2), (3.0, 4)], dtype=[(x, float), (y, int)])

>>> x
array([(1.0, 2), (3.0, 4)],
dtype=[(x, <f8), (y, <i4)])

>>> x[x]
array([ 1., 3.])

View the array as a record array:

>>> x = x.view(np.recarray)

>>> x.x
array([ 1., 3.])

>>> x.y
array([2, 4])

Create a new, empty record array:

>>> np.recarray((2,),
... dtype=[(x, int), (y, float), (z, int)])
rec.array([(-1073741821, 1.2249118382103472e-301, 24547520),
(3471280, 1.2134086255804012e-316, 0)],
dtype=[(x, <i4), (y, <f8), (z, <i4)])

Methods

all([axis, out]) Returns True if all elements evaluate to True.

any([axis, out]) Returns True if any of the elements of a evaluate to True.
argmax([axis, out]) Return indices of the maximum values along the given axis.
argmin([axis, out]) Return indices of the minimum values along the given axis of a.
argsort([axis, kind, order]) Returns the indices that would sort this array.
astype(t) Copy of the array, cast to a specified type.
byteswap(inplace) Swap the bytes of the array elements
choose(choices[, out, mode]) Use an index array to construct a new array from a set of choices.
clip(a_min, a_max[, out]) Return an array whose values are limited to [a_min, a_max].
compress(condition[, axis, out]) Return selected slices of this array along given axis.
conj() Complex-conjugate all elements.
conjugate() Return the complex conjugate, element-wise.
copy([order]) Return a copy of the array.
cumprod([axis, dtype, out]) Return the cumulative product of the elements along the given axis.
cumsum([axis, dtype, out]) Return the cumulative sum of the elements along the given axis.
diagonal([offset, axis1, axis2]) Return specified diagonals.
dot
dump(file) Dump a pickle of the array to the specified file.
Continued on next page

1.5. Standard array subclasses 149

NumPy Reference, Release 2.0.0.dev8464

Table 1.5 continued from previous page

dumps() Returns the pickle of the array as a string.
field(attr[, val])
fill(value) Fill the array with a scalar value.
flatten([order]) Return a copy of the array collapsed into one dimension.
getfield(dtype, offset) Returns a field of the given array as a certain type.
item(*args) Copy an element of an array to a standard Python scalar and return it.
itemset
max([axis, out]) Return the maximum along a given axis.
mean([axis, dtype, out]) Returns the average of the array elements along given axis.
min([axis, out]) Return the minimum along a given axis.
newbyteorder([new_order]) Return the array with the same data viewed with a different byte order.
nonzero() Return the indices of the elements that are non-zero.
prod([axis, dtype, out]) Return the product of the array elements over the given axis
ptp([axis, out]) Peak to peak (maximum - minimum) value along a given axis.
put(indices, values[, mode]) Set a.flat[n] = values[n] for all n in indices.
ravel() Return a flattened array.
repeat(repeats[, axis]) Repeat elements of an array.
reshape(shape[, order]) Returns an array containing the same data with a new shape.
resize(new_shape[, refcheck]) Change shape and size of array in-place.
round([decimals, out]) Return an array rounded a to the given number of decimals.
searchsorted(v[, side]) Find indices where elements of v should be inserted in a to maintain order.
setfield(val, dtype[, offset]) Put a value into a specified place in a field defined by a data-type.
setflags([write, align, uic]) Set array flags WRITEABLE, ALIGNED, and UPDATEIFCOPY, respectively.
sort([axis, kind, order]) Sort an array, in-place.
squeeze() Remove single-dimensional entries from the shape of a.
std([axis, dtype, out, ddof]) Returns the standard deviation of the array elements along given axis.
sum([axis, dtype, out]) Return the sum of the array elements over the given axis.
swapaxes(axis1, axis2) Return a view of the array with axis1 and axis2 interchanged.
take(indices[, axis, out, mode]) Return an array formed from the elements of a at the given indices.
tofile(fid[, sep, format]) Write array to a file as text or binary (default).
tolist() Return the array as a (possibly nested) list.
tostring([order]) Construct a Python string containing the raw data bytes in the array.
trace([offset, axis1, axis2, dtype, out]) Return the sum along diagonals of the array.
transpose(*axes) Returns a view of the array with axes transposed.
var([axis, dtype, out, ddof]) Returns the variance of the array elements, along given axis.
view([dtype, type])

all(axis=None, out=None)
Returns True if all elements evaluate to True.
Refer to numpy.all for full documentation.
See Also:

numpy.all
equivalent function

any(axis=None, out=None)
Returns True if any of the elements of a evaluate to True.
Refer to numpy.any for full documentation.
See Also:

150 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

numpy.any
equivalent function

argmax(axis=None, out=None)
Return indices of the maximum values along the given axis.
Refer to numpy.argmax for full documentation.
See Also:

numpy.argmax
equivalent function

argmin(axis=None, out=None)
Return indices of the minimum values along the given axis of a.
Refer to numpy.argmin for detailed documentation.
See Also:

numpy.argmin
equivalent function

argsort(axis=-1, kind=quicksort, order=None)

Returns the indices that would sort this array.
Refer to numpy.argsort for full documentation.
See Also:

numpy.argsort
equivalent function

astype(t)
Copy of the array, cast to a specified type.
Parameters
t : string or dtype
Typecode or data-type to which the array is cast.

Examples

>>> x = np.array([1, 2, 2.5])

>>> x
array([ 1. , 2. , 2.5])

>>> x.astype(int)
array([1, 2, 2])

byteswap(inplace)
Swap the bytes of the array elements
Toggle between low-endian and big-endian data representation by returning a byteswapped array, option-
ally swapped in-place.
Parameters
inplace: bool, optional :

1.5. Standard array subclasses 151

NumPy Reference, Release 2.0.0.dev8464

If True, swap bytes in-place, default is False.

Returns
out: ndarray :
The byteswapped array. If inplace is True, this is a view to self.

Examples

>>> A = np.array([1, 256, 8755], dtype=np.int16)

>>> map(hex, A)
[0x1, 0x100, 0x2233]
>>> A.byteswap(True)
array([ 256, 1, 13090], dtype=int16)
>>> map(hex, A)
[0x100, 0x1, 0x3322]

Arrays of strings are not swapped

>>> A = np.array([ceg, fac])

>>> A.byteswap()
array([ceg, fac],
dtype=|S3)

choose(choices, out=None, mode=raise)

Use an index array to construct a new array from a set of choices.
Refer to numpy.choose for full documentation.
See Also:

numpy.choose
equivalent function

clip(a_min, a_max, out=None)

Return an array whose values are limited to [a_min, a_max].
Refer to numpy.clip for full documentation.
See Also:

numpy.clip
equivalent function

compress(condition, axis=None, out=None)

Return selected slices of this array along given axis.
Refer to numpy.compress for full documentation.
See Also:

numpy.compress
equivalent function

conj()
Complex-conjugate all elements.
Refer to numpy.conjugate for full documentation.
See Also:

152 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

numpy.conjugate
equivalent function

conjugate()
Return the complex conjugate, element-wise.
Refer to numpy.conjugate for full documentation.
See Also:

numpy.conjugate
equivalent function

copy(order=C)
Return a copy of the array.
Parameters
order : {C, F, A}, optional
By default, the result is stored in C-contiguous (row-major) order in memory. If order is
F, the result has Fortran (column-major) order. If order is A (Any), then the result
has the same order as the input.

Examples

>>> x = np.array([[1,2,3],[4,5,6]], order=F)

>>> y = x.copy()

>>> x.fill(0)

>>> x
array([[0, 0, 0],
[0, 0, 0]])

>>> y
array([[1, 2, 3],
[4, 5, 6]])

>>> y.flags[C_CONTIGUOUS]
True

cumprod(axis=None, dtype=None, out=None)

Return the cumulative product of the elements along the given axis.
Refer to numpy.cumprod for full documentation.
See Also:

numpy.cumprod
equivalent function

cumsum(axis=None, dtype=None, out=None)

Return the cumulative sum of the elements along the given axis.
Refer to numpy.cumsum for full documentation.

1.5. Standard array subclasses 153

NumPy Reference, Release 2.0.0.dev8464

See Also:

numpy.cumsum
equivalent function

diagonal(offset=0, axis1=0, axis2=1)

Return specified diagonals.
Refer to numpy.diagonal for full documentation.
See Also:

numpy.diagonal
equivalent function

dot()

dump(file)
Dump a pickle of the array to the specified file. The array can be read back with pickle.load or numpy.load.
Parameters
file : str
A string naming the dump file.
dumps()
Returns the pickle of the array as a string. pickle.loads or numpy.loads will convert the string back to an
array.
Parameters
None :
field(attr, val=None)

fill(value)
Fill the array with a scalar value.
Parameters
value : scalar
All elements of a will be assigned this value.

Examples

>>> a = np.array([1, 2])

>>> a.fill(0)
>>> a
array([0, 0])
>>> a = np.empty(2)
>>> a.fill(1)
>>> a
array([ 1., 1.])

flatten(order=C)
Return a copy of the array collapsed into one dimension.
Parameters
order : {C, F}, optional

154 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Whether to flatten in C (row-major) or Fortran (column-major) order. The default is C.

Returns
y : ndarray
A copy of the input array, flattened to one dimension.
See Also:

ravel
Return a flattened array.
flat
A 1-D flat iterator over the array.

Examples

>>> a = np.array([[1,2], [3,4]])

>>> a.flatten()
array([1, 2, 3, 4])
>>> a.flatten(F)
array([1, 3, 2, 4])

getfield(dtype, offset)
Returns a field of the given array as a certain type.
A field is a view of the array data with each itemsize determined by the given type and the offset into the
current array, i.e. from offset * dtype.itemsize to (offset+1) * dtype.itemsize.
Parameters
dtype : str
String denoting the data type of the field.
offset : int
Number of dtype.itemsizes to skip before beginning the element view.

Examples

>>> x = np.diag([1.+1.j]*2)
>>> x
array([[ 1.+1.j, 0.+0.j],
[ 0.+0.j, 1.+1.j]])
>>> x.dtype
dtype(complex128)

>>> x.getfield(complex64, 0) # Note how this != x

array([[ 0.+1.875j, 0.+0.j ],
[ 0.+0.j , 0.+1.875j]], dtype=complex64)

>>> x.getfield(complex64,1) # Note how different this is than x

array([[ 0. +5.87173204e-39j, 0. +0.00000000e+00j],
[ 0. +0.00000000e+00j, 0. +5.87173204e-39j]], dtype=complex64)

>>> x.getfield(complex128, 0) # == x
array([[ 1.+1.j, 0.+0.j],
[ 0.+0.j, 1.+1.j]])

1.5. Standard array subclasses 155

NumPy Reference, Release 2.0.0.dev8464

If the argument dtype is the same as x.dtype, then offset != 0 raises a ValueError:

>>> x.getfield(complex128, 1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: Need 0 <= offset <= 0 for requested type but received offset = 1

>>> x.getfield(float64, 0)
array([[ 1., 0.],
[ 0., 1.]])

>>> x.getfield(float64, 1)
array([[ 1.77658241e-307, 0.00000000e+000],
[ 0.00000000e+000, 1.77658241e-307]])

item(*args)
Copy an element of an array to a standard Python scalar and return it.
Parameters
*args : Arguments (variable number and type)
none: in this case, the method only works for arrays with one element (a.size == 1), which
element is copied into a standard Python scalar object and returned.
int_type: this argument is interpreted as a flat index into the array, specifying which ele-
ment to copy and return.
tuple of int_types: functions as does a single int_type argument, except that the argument
is interpreted as an nd-index into the array.
Returns
z : Standard Python scalar object
A copy of the specified element of the array as a suitable Python scalar

Notes
When the data type of a is longdouble or clongdouble, item() returns a scalar array object because there is
no available Python scalar that would not lose information. Void arrays return a buffer object for item(),
unless fields are defined, in which case a tuple is returned.
item is very similar to a[args], except, instead of an array scalar, a standard Python scalar is returned. This
can be useful for speeding up access to elements of the array and doing arithmetic on elements of the array
using Pythons optimized math.

Examples

>>> x = np.random.randint(9, size=(3, 3))

>>> x
array([[3, 1, 7],
[2, 8, 3],
[8, 5, 3]])
>>> x.item(3)
2
>>> x.item(7)
5
>>> x.item((0, 1))
1

156 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> x.item((2, 2))

3

itemset()

max(axis=None, out=None)
Return the maximum along a given axis.
Refer to numpy.amax for full documentation.
See Also:

numpy.amax
equivalent function

mean(axis=None, dtype=None, out=None)

Returns the average of the array elements along given axis.
Refer to numpy.mean for full documentation.
See Also:

numpy.mean
equivalent function

min(axis=None, out=None)
Return the minimum along a given axis.
Refer to numpy.amin for full documentation.
See Also:

numpy.amin
equivalent function

newbyteorder(new_order=S)
Return the array with the same data viewed with a different byte order.
Equivalent to:

arr.view(arr.dtype.newbytorder(new_order))

Changes are also made in all fields and sub-arrays of the array data type.
Parameters
new_order : string, optional
Byte order to force; a value from the byte order specifications above. new_order codes
can be any of:

* S - swap dtype from current to opposite endian

* {<, L} - little endian
* {>, B} - big endian
* {=, N} - native order
* {|, I} - ignore (no change to byte order)

1.5. Standard array subclasses 157

NumPy Reference, Release 2.0.0.dev8464

The default value (S) results in swapping the current byte order. The code does a
case-insensitive check on the first letter of new_order for the alternatives above. For
example, any of B or b or biggish are valid to specify big-endian.
Returns
new_arr : array
New array object with the dtype reflecting given change to the byte order.
nonzero()
Return the indices of the elements that are non-zero.
Refer to numpy.nonzero for full documentation.
See Also:

numpy.nonzero
equivalent function

prod(axis=None, dtype=None, out=None)

Return the product of the array elements over the given axis
Refer to numpy.prod for full documentation.
See Also:

numpy.prod
equivalent function

ptp(axis=None, out=None)
Peak to peak (maximum - minimum) value along a given axis.
Refer to numpy.ptp for full documentation.
See Also:

numpy.ptp
equivalent function

put(indices, values, mode=raise)

Set a.flat[n] = values[n] for all n in indices.
Refer to numpy.put for full documentation.
See Also:

numpy.put
equivalent function

ravel([order])
Return a flattened array.
Refer to numpy.ravel for full documentation.
See Also:

numpy.ravel
equivalent function
ndarray.flat
a flat iterator on the array.

158 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

repeat(repeats, axis=None)
Repeat elements of an array.
Refer to numpy.repeat for full documentation.
See Also:

numpy.repeat
equivalent function

reshape(shape, order=C)
Returns an array containing the same data with a new shape.
Refer to numpy.reshape for full documentation.
See Also:

numpy.reshape
equivalent function

resize(new_shape, refcheck=True)
Change shape and size of array in-place.
Parameters
new_shape : tuple of ints, or n ints
Shape of resized array.
refcheck : bool, optional
If False, reference count will not be checked. Default is True.
Returns
None :
Raises
ValueError :
If a does not own its own data or references or views to it exist, and the data memory
must be changed.
SystemError :
If the order keyword argument is specified. This behaviour is a bug in NumPy.
See Also:

resize
Return a new array with the specified shape.

Notes
This reallocates space for the data area if necessary.
Only contiguous arrays (data elements consecutive in memory) can be resized.
The purpose of the reference count check is to make sure you do not use this array as a buffer for another
Python object and then reallocate the memory. However, reference counts can increase in other ways so if
you are sure that you have not shared the memory for this array with another Python object, then you may
safely set refcheck to False.

1.5. Standard array subclasses 159

NumPy Reference, Release 2.0.0.dev8464

Examples
Shrinking an array: array is flattened (in the order that the data are stored in memory), resized, and re-
shaped:

>>> a = np.array([[0, 1], [2, 3]], order=C)

>>> a.resize((2, 1))
>>> a
array([[0],
[1]])

>>> a = np.array([[0, 1], [2, 3]], order=F)

>>> a.resize((2, 1))
>>> a
array([[0],
[2]])

Enlarging an array: as above, but missing entries are filled with zeros:

>>> b = np.array([[0, 1], [2, 3]])

>>> b.resize(2, 3) # new_shape parameter doesnt have to be a tuple
>>> b
array([[0, 1, 2],
[3, 0, 0]])

Referencing an array prevents resizing...

>>> c = a
>>> a.resize((1, 1))
...
ValueError: cannot resize an array that has been referenced ...

Unless refcheck is False:

>>> a.resize((1, 1), refcheck=False)

>>> a
array([[0]])
>>> c
array([[0]])

round(decimals=0, out=None)
Return an array rounded a to the given number of decimals.
Refer to numpy.around for full documentation.
See Also:

numpy.around
equivalent function

searchsorted(v, side=left)
Find indices where elements of v should be inserted in a to maintain order.
For full documentation, see numpy.searchsorted
See Also:

160 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

numpy.searchsorted
equivalent function

setfield(val, dtype, offset=0)

Put a value into a specified place in a field defined by a data-type.
Place val into as field defined by dtype and beginning offset bytes into the field.
Parameters
val : object
Value to be placed in field.
dtype : dtype object
Data-type of the field in which to place val.
offset : int, optional
The number of bytes into the field at which to place val.
Returns
None :
See Also:
getfield

Examples

>>> x = np.eye(3)
>>> x.getfield(np.float64)
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])
>>> x.setfield(3, np.int32)
>>> x.getfield(np.int32)
array([[3, 3, 3],
[3, 3, 3],
[3, 3, 3]])
>>> x
array([[ 1.00000000e+000, 1.48219694e-323, 1.48219694e-323],
[ 1.48219694e-323, 1.00000000e+000, 1.48219694e-323],
[ 1.48219694e-323, 1.48219694e-323, 1.00000000e+000]])
>>> x.setfield(np.eye(3), np.int32)
>>> x
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])

setflags(write=None, align=None, uic=None)

Set array flags WRITEABLE, ALIGNED, and UPDATEIFCOPY, respectively.
These Boolean-valued flags affect how numpy interprets the memory area used by a (see Notes below).
The ALIGNED flag can only be set to True if the data is actually aligned according to the type. The
UPDATEIFCOPY flag can never be set to True. The flag WRITEABLE can only be set to True if the array
owns its own memory, or the ultimate owner of the memory exposes a writeable buffer interface, or is a
string. (The exception for string is made so that unpickling can be done without copying memory.)
Parameters
write : bool, optional

1.5. Standard array subclasses 161

NumPy Reference, Release 2.0.0.dev8464

Describes whether or not a can be written to.

align : bool, optional
Describes whether or not a is aligned properly for its type.
uic : bool, optional
Describes whether or not a is a copy of another base array.

Notes
Array flags provide information about how the memory area used for the array is to be interpreted. There
are 6 Boolean flags in use, only three of which can be changed by the user: UPDATEIFCOPY, WRITE-
ABLE, and ALIGNED.
WRITEABLE (W) the data area can be written to;
ALIGNED (A) the data and strides are aligned appropriately for the hardware (as determined by the com-
piler);
UPDATEIFCOPY (U) this array is a copy of some other array (referenced by .base). When this array is
deallocated, the base array will be updated with the contents of this array.
All flags can be accessed using their first (upper case) letter as well as the full name.

Examples

>>> y
array([[3, 1, 7],
[2, 0, 0],
[8, 5, 9]])
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : True
WRITEABLE : True
ALIGNED : True
UPDATEIFCOPY : False
>>> y.setflags(write=0, align=0)
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : True
WRITEABLE : False
ALIGNED : False
UPDATEIFCOPY : False
>>> y.setflags(uic=1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: cannot set UPDATEIFCOPY flag to True

sort(axis=-1, kind=quicksort, order=None)

Sort an array, in-place.
Parameters
axis : int, optional
Axis along which to sort. Default is -1, which means sort along the last axis.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm. Default is quicksort.

162 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

order : list, optional

When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
See Also:

numpy.sort
Return a sorted copy of an array.
argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in sorted array.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = np.array([[1,4], [3,1]])

>>> a.sort(axis=1)
>>> a
array([[1, 4],
[1, 3]])
>>> a.sort(axis=0)
>>> a
array([[1, 3],
[1, 4]])

Use the order keyword to specify a field to use when sorting a structured array:

>>> a = np.array([(a, 2), (c, 1)], dtype=[(x, S1), (y, int)])

>>> a.sort(order=y)
>>> a
array([(c, 1), (a, 2)],
dtype=[(x, |S1), (y, <i4)])

squeeze()
Remove single-dimensional entries from the shape of a.
Refer to numpy.squeeze for full documentation.
See Also:

numpy.squeeze
equivalent function

std(axis=None, dtype=None, out=None, ddof=0)

Returns the standard deviation of the array elements along given axis.
Refer to numpy.std for full documentation.
See Also:

1.5. Standard array subclasses 163

NumPy Reference, Release 2.0.0.dev8464

numpy.std
equivalent function

sum(axis=None, dtype=None, out=None)

Return the sum of the array elements over the given axis.
Refer to numpy.sum for full documentation.
See Also:

numpy.sum
equivalent function

swapaxes(axis1, axis2)
Return a view of the array with axis1 and axis2 interchanged.
Refer to numpy.swapaxes for full documentation.
See Also:

numpy.swapaxes
equivalent function

take(indices, axis=None, out=None, mode=raise)

Return an array formed from the elements of a at the given indices.
Refer to numpy.take for full documentation.
See Also:

numpy.take
equivalent function

tofile(fid, sep="", format="%s")

Write array to a file as text or binary (default).
Data is always written in C order, independent of the order of a. The data produced by this method can
be recovered using the function fromfile().
Parameters
fid : file or str
An open file object, or a string containing a filename.
sep : str
Separator between array items for text output. If (empty), a binary file is written,
equivalent to file.write(a.tostring()).
format : str
Format string for text file output. Each entry in the array is formatted to text by first
converting it to the closest Python type, and then using format % item.

Notes
This is a convenience function for quick storage of array data. Information on endianness and precision
is lost, so this method is not a good choice for files intended to archive data or transport data between
machines with different endianness. Some of these problems can be overcome by outputting the data as
text files, at the expense of speed and file size.

164 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

tolist()
Return the array as a (possibly nested) list.
Return a copy of the array data as a (nested) Python list. Data items are converted to the nearest compatible
Python type.
Parameters
none :
Returns
y : list
The possibly nested list of array elements.

Notes
The array may be recreated, a = np.array(a.tolist()).

Examples

>>> a = np.array([1, 2])

>>> a.tolist()
[1, 2]
>>> a = np.array([[1, 2], [3, 4]])
>>> list(a)
[array([1, 2]), array([3, 4])]
>>> a.tolist()
[[1, 2], [3, 4]]

tostring(order=C)
Construct a Python string containing the raw data bytes in the array.
Constructs a Python string showing a copy of the raw contents of data memory. The string can be produced
in either C or Fortran, or Any order (the default is C-order). Any order means C-order unless the
F_CONTIGUOUS flag in the array is set, in which case it means Fortran order.
Parameters
order : {C, F, None}, optional
Order of the data for multidimensional arrays: C, Fortran, or the same as for the original
array.
Returns
s : str
A Python string exhibiting a copy of as raw data.

Examples

>>> x = np.array([[0, 1], [2, 3]])

>>> x.tostring()
\x00\x00\x00\x00\x01\x00\x00\x00\x02\x00\x00\x00\x03\x00\x00\x00
>>> x.tostring(C) == x.tostring()
True
>>> x.tostring(F)
\x00\x00\x00\x00\x02\x00\x00\x00\x01\x00\x00\x00\x03\x00\x00\x00

trace(offset=0, axis1=0, axis2=1, dtype=None, out=None)

Return the sum along diagonals of the array.
Refer to numpy.trace for full documentation.

1.5. Standard array subclasses 165

NumPy Reference, Release 2.0.0.dev8464

See Also:

numpy.trace
equivalent function

transpose(*axes)
Returns a view of the array with axes transposed.
For a 1-D array, this has no effect. (To change between column and row vectors, first cast the 1-D ar-
ray into a matrix object.) For a 2-D array, this is the usual matrix transpose. For an n-D array, if axes
are given, their order indicates how the axes are permuted (see Examples). If axes are not provided
and a.shape = (i[0], i[1], ... i[n-2], i[n-1]), then a.transpose().shape
= (i[n-1], i[n-2], ... i[1], i[0]).
Parameters
axes : None, tuple of ints, or n ints
None or no argument: reverses the order of the axes.
tuple of ints: i in the j-th place in the tuple means as i-th axis becomes a.transpose()s
j-th axis.
n ints: same as an n-tuple of the same ints (this form is intended simply as a convenience
alternative to the tuple form)
Returns
out : ndarray
View of a, with axes suitably permuted.
See Also:

ndarray.T
Array property returning the array transposed.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> a
array([[1, 2],
[3, 4]])
>>> a.transpose()
array([[1, 3],
[2, 4]])
>>> a.transpose((1, 0))
array([[1, 3],
[2, 4]])
>>> a.transpose(1, 0)
array([[1, 3],
[2, 4]])

var(axis=None, dtype=None, out=None, ddof=0)

Returns the variance of the array elements, along given axis.
Refer to numpy.var for full documentation.
See Also:

numpy.var
equivalent function

166 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

view(dtype=None, type=None)

class record()
A data-type scalar that allows field access as attribute lookup.

Methods

all Not implemented (virtual attribute)

any Not implemented (virtual attribute)
argmax Not implemented (virtual attribute)
argmin Not implemented (virtual attribute)
argsort Not implemented (virtual attribute)
astype Not implemented (virtual attribute)
byteswap Not implemented (virtual attribute)
choose Not implemented (virtual attribute)
clip Not implemented (virtual attribute)
compress Not implemented (virtual attribute)
conj
conjugate Not implemented (virtual attribute)
copy Not implemented (virtual attribute)
cumprod Not implemented (virtual attribute)
cumsum Not implemented (virtual attribute)
diagonal Not implemented (virtual attribute)
dump Not implemented (virtual attribute)
dumps Not implemented (virtual attribute)
fill Not implemented (virtual attribute)
flatten Not implemented (virtual attribute)
getfield
item Not implemented (virtual attribute)
itemset Not implemented (virtual attribute)
max Not implemented (virtual attribute)
mean Not implemented (virtual attribute)
min Not implemented (virtual attribute)
newbyteorder([new_order]) Return a new dtype with a different byte order.
nonzero Not implemented (virtual attribute)
pprint() Pretty-print all fields.
prod Not implemented (virtual attribute)
ptp Not implemented (virtual attribute)
put Not implemented (virtual attribute)
ravel Not implemented (virtual attribute)
repeat Not implemented (virtual attribute)
reshape Not implemented (virtual attribute)
resize Not implemented (virtual attribute)
round Not implemented (virtual attribute)
searchsorted Not implemented (virtual attribute)
setfield
setflags Not implemented (virtual attribute)
sort Not implemented (virtual attribute)
squeeze Not implemented (virtual attribute)
std Not implemented (virtual attribute)
sum Not implemented (virtual attribute)
swapaxes Not implemented (virtual attribute)
take Not implemented (virtual attribute)
Continued on next page

1.5. Standard array subclasses 167

NumPy Reference, Release 2.0.0.dev8464

Table 1.6 continued from previous page

tofile Not implemented (virtual attribute)
tolist Not implemented (virtual attribute)
tostring Not implemented (virtual attribute)
trace Not implemented (virtual attribute)
transpose Not implemented (virtual attribute)
var Not implemented (virtual attribute)
view Not implemented (virtual attribute)

all()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
any()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
argmax()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
argmin()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
argsort()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
astype()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.

168 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

See Also:
The
byteswap()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
choose()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
clip()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
compress()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
conj()

conjugate()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
copy()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The

1.5. Standard array subclasses 169

NumPy Reference, Release 2.0.0.dev8464

cumprod()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
cumsum()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
diagonal()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
dump()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
dumps()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
fill()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
flatten()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:

170 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

The
getfield()

item()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
itemset()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
max()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
mean()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
min()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
newbyteorder(new_order=S)
Return a new dtype with a different byte order.
Changes are also made in all fields and sub-arrays of the data type.
The new_order code can be any from the following:
{<, L} - little endian
{>, B} - big endian
{=, N} - native order

1.5. Standard array subclasses 171

NumPy Reference, Release 2.0.0.dev8464

S - swap dtype from current to opposite endian

{|, I} - ignore (no change to byte order)

Parameters
new_order : str, optional
Byte order to force; a value from the byte order specifications above. The default value
(S) results in swapping the current byte order. The code does a case-insensitive check
on the first letter of new_order for the alternatives above. For example, any of B or b
or biggish are valid to specify big-endian.
Returns
new_dtype : dtype
New dtype object with the given change to the byte order.

nonzero()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
pprint()
Pretty-print all fields.
prod()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
ptp()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
put()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
ravel()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.

172 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

See Also:
The
repeat()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
reshape()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
resize()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
round()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
searchsorted()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
setfield()

setflags()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The

1.5. Standard array subclasses 173

NumPy Reference, Release 2.0.0.dev8464

sort()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
squeeze()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
std()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
sum()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
swapaxes()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
take()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
tofile()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:

174 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

The
tolist()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
tostring()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
trace()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
transpose()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
var()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The
view()
Not implemented (virtual attribute)
Class generic exists solely to derive numpy scalars from, and possesses, albeit unimplemented, all the
attributes of the ndarray class so as to provide a uniform API.
See Also:
The

1.5. Standard array subclasses 175

NumPy Reference, Release 2.0.0.dev8464

1.5.6 Masked arrays (numpy.ma)

See Also:
Masked arrays

1.5.7 Standard container class

For backward compatibility and as a standard container class, the UserArray from Numeric has been brought over to
NumPy and named numpy.lib.user_array.container The container class is a Python class whose self.array
attribute is an ndarray. Multiple inheritance is probably easier with numpy.lib.user_array.container than with the
ndarray itself and so it is included by default. It is not documented here beyond mentioning its existence because you
are encouraged to use the ndarray class directly if you can.
numpy.lib.user_array.container(data[, ...])
class container(data, dtype=None, copy=True)

1.5.8 Array Iterators

Iterators are a powerful concept for array processing. Essentially, iterators implement a generalized for-loop. If myiter
is an iterator object, then the Python code:

for val in myiter:

...
some code involving val
...

calls val = myiter.next() repeatedly until StopIteration is raised by the iterator. There are several ways
to iterate over an array that may be useful: default iteration, flat iteration, and N -dimensional enumeration.

Default iteration

The default iterator of an ndarray object is the default Python iterator of a sequence type. Thus, when the array object
itself is used as an iterator. The default behavior is equivalent to:

for i in arr.shape[0]:
val = arr[i]

This default iterator selects a sub-array of dimension N 1 from the array. This can be a useful construct for defining
recursive algorithms. To loop over the entire array requires N for-loops.

>>> a = arange(24).reshape(3,2,4)+10
>>> for val in a:
... print item:, val
item: [[10 11 12 13]
[14 15 16 17]]
item: [[18 19 20 21]
[22 23 24 25]]
item: [[26 27 28 29]
[30 31 32 33]]

176 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Flat iteration

ndarray.flat A 1-D iterator over the array.

flat
A 1-D iterator over the array.
This is a numpy.flatiter instance, which acts similarly to, but is not a subclass of, Pythons built-in iterator
object.
See Also:

flatten
Return a copy of the array collapsed into one dimension.

flatiter

Examples

>>> x = np.arange(1, 7).reshape(2, 3)

>>> x
array([[1, 2, 3],
[4, 5, 6]])
>>> x.flat[3]
4
>>> x.T
array([[1, 4],
[2, 5],
[3, 6]])
>>> x.T.flat[3]
5
>>> type(x.flat)
<type numpy.flatiter>

An assignment example:

>>> x.flat = 3; x
array([[3, 3, 3],
[3, 3, 3]])
>>> x.flat[[1,4]] = 1; x
array([[3, 1, 3],
[3, 1, 3]])

As mentioned previously, the flat attribute of ndarray objects returns an iterator that will cycle over the entire array in
C-style contiguous order.

>>> for i, val in enumerate(a.flat):

... if i%5 == 0: print i, val
0 10
5 15
10 20
15 25
20 30

Here, Ive used the built-in enumerate iterator to return the iterator index as well as the value.

1.5. Standard array subclasses 177

NumPy Reference, Release 2.0.0.dev8464

N-dimensional enumeration

ndenumerate(arr) Multidimensional index iterator.

class ndenumerate(arr)
Multidimensional index iterator.
Return an iterator yielding pairs of array coordinates and values.
Parameters
a : ndarray
Input array.
See Also:
ndindex, flatiter

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> for index, x in np.ndenumerate(a):
... print index, x
(0, 0) 1
(0, 1) 2
(1, 0) 3
(1, 1) 4

Methods
next() Standard iterator method, returns the index tuple and array value.
next()
Standard iterator method, returns the index tuple and array value.
Returns
coords : tuple of ints
The indices of the current iteration.
val : scalar
The array element of the current iteration.
Sometimes it may be useful to get the N-dimensional index while iterating. The ndenumerate iterator can achieve this.

>>> for i, val in ndenumerate(a):

... if sum(i)%5 == 0: print i, val
(0, 0, 0) 10
(1, 1, 3) 25
(2, 0, 3) 29
(2, 1, 2) 32

Iterator for broadcasting

broadcast Produce an object that mimics broadcasting.

class broadcast()
Produce an object that mimics broadcasting.

178 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Parameters
in1, in2, ... : array_like
Input parameters.
Returns
b : broadcast object
Broadcast the input parameters against one another, and return an object that encapsu-
lates the result. Amongst others, it has shape and nd properties, and may be used as
an iterator.

Examples
Manually adding two vectors, using broadcasting:

>>> x = np.array([[1], [2], [3]])

>>> y = np.array([4, 5, 6])
>>> b = np.broadcast(x, y)

>>> out = np.empty(b.shape)

>>> out.flat = [u+v for (u,v) in b]
>>> out
array([[ 5., 6., 7.],
[ 6., 7., 8.],
[ 7., 8., 9.]])

Compare against built-in broadcasting:

>>> x + y
array([[5, 6, 7],
[6, 7, 8],
[7, 8, 9]])

Methods
next x.next() -> the next value, or raise StopIteration
reset
next()
x.next() -> the next value, or raise StopIteration
reset()

The general concept of broadcasting is also available from Python using the broadcast iterator. This object takes
N objects as inputs and returns an iterator that returns tuples providing each of the input sequence elements in the
broadcasted result.

>>> for val in broadcast([[1,0],[2,3]],[0,1]):

... print val
(1, 0)
(0, 1)
(2, 0)
(3, 1)

1.5. Standard array subclasses 179

NumPy Reference, Release 2.0.0.dev8464

1.6 Masked arrays

Masked arrays are arrays that may have missing or invalid entries. The numpy.ma module provides a nearly work-
alike replacement for numpy that supports data arrays with masks.

1.6.1 The numpy.ma module

Rationale

Masked arrays are arrays that may have missing or invalid entries. The numpy.ma module provides a nearly work-
alike replacement for numpy that supports data arrays with masks.

What is a masked array?

In many circumstances, datasets can be incomplete or tainted by the presence of invalid data. For example, a sensor
may have failed to record a data, or recorded an invalid value. The numpy.ma module provides a convenient way to
address this issue, by introducing masked arrays.
A masked array is the combination of a standard numpy.ndarray and a mask. A mask is either nomask, indicating
that no value of the associated array is invalid, or an array of booleans that determines for each element of the associated
array whether the value is valid or not. When an element of the mask is False, the corresponding element of the
associated array is valid and is said to be unmasked. When an element of the mask is True, the corresponding element
of the associated array is said to be masked (invalid).
The package ensures that masked entries are not used in computations.
As an illustration, lets consider the following dataset:

>>> import numpy as np

>>> import numpy.ma as ma
>>> x = np.array([1, 2, 3, -1, 5])

We wish to mark the fourth entry as invalid. The easiest is to create a masked array:

>>> mx = ma.masked_array(x, mask=[0, 0, 0, 1, 0])

We can now compute the mean of the dataset, without taking the invalid data into account:

>>> mx.mean()
2.75

The numpy.ma module

The main feature of the numpy.ma module is the MaskedArray class, which is a subclass of numpy.ndarray.
The class, its attributes and methods are described in more details in the MaskedArray class section.
The numpy.ma module can be used as an addition to numpy:

>>> import numpy as np

>>> import numpy.ma as ma

To create an array with the second element invalid, we would do:

180 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> y = ma.array([1, 2, 3], mask = [0, 1, 0])

To create a masked array where all values close to 1.e20 are invalid, we would do:

>>> z = masked_values([1.0, 1.e20, 3.0, 4.0], 1.e20)

For a complete discussion of creation methods for masked arrays please see section Constructing masked arrays.

1.6.2 Using numpy.ma

Constructing masked arrays

There are several ways to construct a masked array.

A first possibility is to directly invoke the MaskedArray class.
A second possibility is to use the two masked array constructors, array and masked_array.
array(data[, dtype, copy, order, mask, ...]) An array class with possibly masked values.
masked_array An array class with possibly masked values.
array(data, dtype=None, copy=False, order=False, mask=False, fill_value=None, keep_mask=True,
hard_mask=False, shrink=True, subok=True, ndmin=0)
An array class with possibly masked values.
Masked values of True exclude the corresponding element from any computation.
Construction:

x = MaskedArray(data, mask=nomask, dtype=None, copy=True,

fill_value=None, keep_mask=True, hard_mask=False,
shrink=True)

Parameters
data : array_like
Input data.
mask : sequence, optional
Mask. Must be convertible to an array of booleans with the same shape as data. True
indicates a masked (i.e. invalid) data.
dtype : dtype, optional
Data type of the output. If dtype is None, the type of the data argument (data.dtype)
is used. If dtype is not None and different from data.dtype, a copy is performed.
copy : bool, optional
Whether to copy the input data (True), or to use a reference instead. Default is False.
subok : bool, optional
Whether to return a subclass of MaskedArray if possible (True) or a plain MaskedArray.
Default is True.
ndmin : int, optional
Minimum number of dimensions. Default is 0.

1.6. Masked arrays 181

NumPy Reference, Release 2.0.0.dev8464

fill_value : scalar, optional

Value used to fill in the masked values when necessary. If None, a default based on the
data-type is used.
keep_mask : bool, optional
Whether to combine mask with the mask of the input data, if any (True), or to use only
mask for the output (False). Default is True.
hard_mask : bool, optional
Whether to use a hard mask or not. With a hard mask, masked values cannot be un-
masked. Default is False.
shrink : bool, optional
Whether to force compression of an empty mask. Default is True.

masked_array
alias of MaskedArray
A third option is to take the view of an existing array. In that case, the mask of the view is set to nomask if the
array has no named fields, or an array of boolean with the same structure as the array otherwise.

>>> x = np.array([1, 2, 3])

>>> x.view(ma.MaskedArray)
masked_array(data = [1 2 3],
mask = False,
fill_value = 999999)
>>> x = np.array([(1, 1.), (2, 2.)], dtype=[(a,int), (b, float)])
>>> x.view(ma.MaskedArray)
masked_array(data = [(1, 1.0) (2, 2.0)],
mask = [(False, False) (False, False)],
fill_value = (999999, 1e+20),
dtype = [(a, <i4), (b, <f8)])

Yet another possibility is to use any of the following functions:

asarray(a[, dtype, order]) Convert the input to a masked array of the given data-type.
asanyarray(a[, dtype]) Convert the input to a masked array, conserving subclasses.
fix_invalid(a[, mask, copy, fill_value]) Return input with invalid data masked and replaced by a fill
value.
masked_equal(x, value[, copy]) Mask an array where equal to a given value.
masked_greater(x, value[, copy]) Mask an array where greater than a given value.
masked_greater_equal(x, value[, copy]) Mask an array where greater than or equal to a given value.
masked_inside(x, v1, v2[, copy]) Mask an array inside a given interval.
masked_invalid(a[, copy]) Mask an array where invalid values occur (NaNs or infs).
masked_less(x, value[, copy]) Mask an array where less than a given value.
masked_less_equal(x, value[, copy]) Mask an array where less than or equal to a given value.
masked_not_equal(x, value[, copy]) Mask an array where not equal to a given value.
masked_object(x, value[, copy, shrink]) Mask the array x where the data are exactly equal to value.
masked_outside(x, v1, v2[, copy]) Mask an array outside a given interval.
masked_values(x, value[, rtol, atol, copy, Mask using floating point equality.
...])
masked_where(condition, a[, copy]) Mask an array where a condition is met.
asarray(a, dtype=None, order=None)
Convert the input to a masked array of the given data-type.

182 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

No copy is performed if the input is already an ndarray. If a is a subclass of MaskedArray, a base class
MaskedArray is returned.
Parameters
a : array_like
Input data, in any form that can be converted to a masked array. This includes lists, lists
of tuples, tuples, tuples of tuples, tuples of lists, ndarrays and masked arrays.
dtype : dtype, optional
By default, the data-type is inferred from the input data.
order : {C, F}, optional
Whether to use row-major (C) or column-major (FORTRAN) memory representa-
tion. Default is C.
Returns
out : MaskedArray
Masked array interpretation of a.
See Also:

asanyarray
Similar to asarray, but conserves subclasses.

Examples

>>> x = np.arange(10.).reshape(2, 5)
>>> x
array([[ 0., 1., 2., 3., 4.],
[ 5., 6., 7., 8., 9.]])
>>> np.ma.asarray(x)
masked_array(data =
[[ 0. 1. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]],
mask =
False,
fill_value = 1e+20)
>>> type(np.ma.asarray(x))
<class numpy.ma.core.MaskedArray>

asanyarray(a, dtype=None)
Convert the input to a masked array, conserving subclasses.
If a is a subclass of MaskedArray, its class is conserved. No copy is performed if the input is already an
ndarray.
Parameters
a : array_like
Input data, in any form that can be converted to an array.
dtype : dtype, optional
By default, the data-type is inferred from the input data.
order : {C, F}, optional
Whether to use row-major (C) or column-major (FORTRAN) memory representa-
tion. Default is C.

1.6. Masked arrays 183

NumPy Reference, Release 2.0.0.dev8464

Returns
out : MaskedArray
MaskedArray interpretation of a.
See Also:

asarray
Similar to asanyarray, but does not conserve subclass.

Examples

>>> x = np.arange(10.).reshape(2, 5)
>>> x
array([[ 0., 1., 2., 3., 4.],
[ 5., 6., 7., 8., 9.]])
>>> np.ma.asanyarray(x)
masked_array(data =
[[ 0. 1. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]],
mask =
False,
fill_value = 1e+20)
>>> type(np.ma.asanyarray(x))
<class numpy.ma.core.MaskedArray>

fix_invalid(a, mask=False, copy=True, fill_value=None)

Return input with invalid data masked and replaced by a fill value.
Invalid data means values of nan, inf, etc.
Parameters
a : array_like
Input array, a (subclass of) ndarray.
copy : bool, optional
Whether to use a copy of a (True) or to fix a in place (False). Default is True.
fill_value : scalar, optional
Value used for fixing invalid data. Default is None, in which case the a.fill_value
is used.
Returns
b : MaskedArray
The input array with invalid entries fixed.

Notes
A copy is performed by default.

Examples

>>> x = np.ma.array([1., -1, np.nan, np.inf], mask=[1] + [0]*3)

>>> x
masked_array(data = [-- -1.0 nan inf],
mask = [ True False False False],
fill_value = 1e+20)

184 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> np.ma.fix_invalid(x)
masked_array(data = [-- -1.0 -- --],
mask = [ True False True True],
fill_value = 1e+20)

>>> fixed = np.ma.fix_invalid(x)

>>> fixed.data
array([ 1.00000000e+00, -1.00000000e+00, 1.00000000e+20,
1.00000000e+20])
>>> x.data
array([ 1., -1., NaN, Inf])

masked_equal(x, value, copy=True)

Mask an array where equal to a given value.
This function is a shortcut to masked_where, with condition = (x == value). For floating point arrays,
consider using masked_values(x, value).
See Also:

masked_where
Mask where a condition is met.
masked_values
Mask using floating point equality.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_equal(a, 2)
masked_array(data = [0 1 -- 3],
mask = [False False True False],
fill_value=999999)

masked_greater(x, value, copy=True)

Mask an array where greater than a given value.
This function is a shortcut to masked_where, with condition = (x > value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_greater(a, 2)
masked_array(data = [0 1 2 --],
mask = [False False False True],
fill_value=999999)

1.6. Masked arrays 185

NumPy Reference, Release 2.0.0.dev8464

masked_greater_equal(x, value, copy=True)

Mask an array where greater than or equal to a given value.
This function is a shortcut to masked_where, with condition = (x >= value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_greater_equal(a, 2)
masked_array(data = [0 1 -- --],
mask = [False False True True],
fill_value=999999)

masked_inside(x, v1, v2, copy=True)

Mask an array inside a given interval.
Shortcut to masked_where, where condition is True for x inside the interval [v1,v2] (v1 <= x <= v2).
The boundaries v1 and v2 can be given in either order.
See Also:

masked_where
Mask where a condition is met.

Notes
The array x is prefilled with its filling value.

Examples

>>> import numpy.ma as ma

>>> x = [0.31, 1.2, 0.01, 0.2, -0.4, -1.1]
>>> ma.masked_inside(x, -0.3, 0.3)
masked_array(data = [0.31 1.2 -- -- -0.4 -1.1],
mask = [False False True True False False],
fill_value=1e+20)

The order of v1 and v2 doesnt matter.

>>> ma.masked_inside(x, 0.3, -0.3)

masked_array(data = [0.31 1.2 -- -- -0.4 -1.1],
mask = [False False True True False False],
fill_value=1e+20)

masked_invalid(a, copy=True)
Mask an array where invalid values occur (NaNs or infs).
This function is a shortcut to masked_where, with condition = ~(np.isfinite(a)). Any pre-existing mask
is conserved. Only applies to arrays with a dtype where NaNs or infs make sense (i.e. floating point types),
but accepts any array_like object.

186 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(5, dtype=np.float)
>>> a[2] = np.NaN
>>> a[3] = np.PINF
>>> a
array([ 0., 1., NaN, Inf, 4.])
>>> ma.masked_invalid(a)
masked_array(data = [0.0 1.0 -- -- 4.0],
mask = [False False True True False],
fill_value=1e+20)

masked_less(x, value, copy=True)

Mask an array where less than a given value.
This function is a shortcut to masked_where, with condition = (x < value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_less(a, 2)
masked_array(data = [-- -- 2 3],
mask = [ True True False False],
fill_value=999999)

masked_less_equal(x, value, copy=True)

Mask an array where less than or equal to a given value.
This function is a shortcut to masked_where, with condition = (x <= value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_less_equal(a, 2)
masked_array(data = [-- -- -- 3],

1.6. Masked arrays 187

NumPy Reference, Release 2.0.0.dev8464

mask = [ True True True False],

fill_value=999999)

masked_not_equal(x, value, copy=True)

Mask an array where not equal to a given value.
This function is a shortcut to masked_where, with condition = (x != value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_not_equal(a, 2)
masked_array(data = [-- -- 2 --],
mask = [ True True False True],
fill_value=999999)

masked_object(x, value, copy=True, shrink=True)

Mask the array x where the data are exactly equal to value.
This function is similar to masked_values, but only suitable for object arrays: for floating point, use
masked_values instead.
Parameters
x : array_like
Array to mask
value : object
Comparison value
copy : {True, False}, optional
Whether to return a copy of x.
shrink : {True, False}, optional
Whether to collapse a mask full of False to nomask
Returns
result : MaskedArray
The result of masking x where equal to value.
See Also:

masked_where
Mask where a condition is met.
masked_equal
Mask where equal to a given value (integers).
masked_values
Mask using floating point equality.

188 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> import numpy.ma as ma

>>> food = np.array([green_eggs, ham], dtype=object)
>>> # dont eat spoiled food
>>> eat = ma.masked_object(food, green_eggs)
>>> print eat
[-- ham]
>>> # plain ol ham is boring
>>> fresh_food = np.array([cheese, ham, pineapple], dtype=object)
>>> eat = ma.masked_object(fresh_food, green_eggs)
>>> print eat
[cheese ham pineapple]

Note that mask is set to nomask if possible.

>>> eat
masked_array(data = [cheese ham pineapple],
mask = False,
fill_value=?)

masked_outside(x, v1, v2, copy=True)

Mask an array outside a given interval.
Shortcut to masked_where, where condition is True for x outside the interval [v1,v2] (x < v1)|(x > v2).
The boundaries v1 and v2 can be given in either order.
See Also:

masked_where
Mask where a condition is met.

Notes
The array x is prefilled with its filling value.

Examples

>>> import numpy.ma as ma

>>> x = [0.31, 1.2, 0.01, 0.2, -0.4, -1.1]
>>> ma.masked_outside(x, -0.3, 0.3)
masked_array(data = [-- -- 0.01 0.2 -- --],
mask = [ True True False False True True],
fill_value=1e+20)

The order of v1 and v2 doesnt matter.

>>> ma.masked_outside(x, 0.3, -0.3)

masked_array(data = [-- -- 0.01 0.2 -- --],
mask = [ True True False False True True],
fill_value=1e+20)

masked_values(x, value, rtol=1.0000000000000001e-05, atol=1e-08, copy=True, shrink=True)

Mask using floating point equality.
Return a MaskedArray, masked where the data in array x are approximately equal to value, i.e. where the
following condition is True
(abs(x - value) <= atol+rtol*abs(value))

1.6. Masked arrays 189

NumPy Reference, Release 2.0.0.dev8464

The fill_value is set to value and the mask is set to nomask if possible. For integers, consider using
masked_equal.
Parameters
x : array_like
Array to mask.
value : float
Masking value.
rtol : float, optional
Tolerance parameter.
atol : float, optional
Tolerance parameter (1e-8).
copy : bool, optional
Whether to return a copy of x.
shrink : bool, optional
Whether to collapse a mask full of False to nomask.
Returns
result : MaskedArray
The result of masking x where approximately equal to value.
See Also:

masked_where
Mask where a condition is met.
masked_equal
Mask where equal to a given value (integers).

Examples

>>> import numpy.ma as ma

>>> x = np.array([1, 1.1, 2, 1.1, 3])
>>> ma.masked_values(x, 1.1)
masked_array(data = [1.0 -- 2.0 -- 3.0],
mask = [False True False True False],
fill_value=1.1)

Note that mask is set to nomask if possible.

>>> ma.masked_values(x, 1.5)

masked_array(data = [ 1. 1.1 2. 1.1 3. ],
mask = False,
fill_value=1.5)

For integers, the fill value will be different in general to the result of masked_equal.

>>> x = np.arange(5)
>>> x
array([0, 1, 2, 3, 4])
>>> ma.masked_values(x, 2)

190 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

masked_array(data = [0 1 -- 3 4],
mask = [False False True False False],
fill_value=2)
>>> ma.masked_equal(x, 2)
masked_array(data = [0 1 -- 3 4],
mask = [False False True False False],
fill_value=999999)

masked_where(condition, a, copy=True)
Mask an array where a condition is met.
Return a as an array masked where condition is True. Any masked values of a or condition are also masked
in the output.
Parameters
condition : array_like
Masking condition. When condition tests floating point values for equality, consider
using masked_values instead.
a : array_like
Array to mask.
copy : bool
If True (default) make a copy of a in the result. If False modify a in place and return a
view.
Returns
result : MaskedArray
The result of masking a where condition is True.
See Also:

masked_values
Mask using floating point equality.
masked_equal
Mask where equal to a given value.
masked_not_equal
Mask where not equal to a given value.
masked_less_equal
Mask where less than or equal to a given value.
masked_greater_equal
Mask where greater than or equal to a given value.
masked_less
Mask where less than a given value.
masked_greater
Mask where greater than a given value.
masked_inside
Mask inside a given interval.
masked_outside
Mask outside a given interval.

1.6. Masked arrays 191

NumPy Reference, Release 2.0.0.dev8464

masked_invalid
Mask invalid values (NaNs or infs).

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_where(a <= 2, a)
masked_array(data = [-- -- -- 3],
mask = [ True True True False],
fill_value=999999)

Mask array b conditional on a.

>>> b = [a, b, c, d]
>>> ma.masked_where(a == 2, b)
masked_array(data = [a b -- d],
mask = [False False True False],
fill_value=N/A)

Effect of the copy argument.

>>> c = ma.masked_where(a <= 2, a)

>>> c
masked_array(data = [-- -- -- 3],
mask = [ True True True False],
fill_value=999999)
>>> c[0] = 99
>>> c
masked_array(data = [99 -- -- 3],
mask = [False True True False],
fill_value=999999)
>>> a
array([0, 1, 2, 3])
>>> c = ma.masked_where(a <= 2, a, copy=False)
>>> c[0] = 99
>>> c
masked_array(data = [99 -- -- 3],
mask = [False True True False],
fill_value=999999)
>>> a
array([99, 1, 2, 3])

When condition or a contain masked values.

>>> a = np.arange(4)
>>> a = ma.masked_where(a == 2, a)
>>> a
masked_array(data = [0 1 -- 3],
mask = [False False True False],
fill_value=999999)
>>> b = np.arange(4)
>>> b = ma.masked_where(b == 0, b)
>>> b
masked_array(data = [-- 1 2 3],

192 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

mask = [ True False False False],

fill_value=999999)
>>> ma.masked_where(a == 3, b)
masked_array(data = [-- 1 -- --],
mask = [ True False True True],
fill_value=999999)

Accessing the data

The underlying data of a masked array can be accessed in several ways:

through the data attribute. The output is a view of the array as a numpy.ndarray or one of its subclasses,
depending on the type of the underlying data at the masked array creation.
through the __array__ method. The output is then a numpy.ndarray.
by directly taking a view of the masked array as a numpy.ndarray or one of its subclass (which is actually
what using the data attribute does).
by using the getdata function.
None of these methods is completely satisfactory if some entries have been marked as invalid. As a general rule,
where a representation of the array is required without any masked entries, it is recommended to fill the array with the
filled method.

Accessing the mask

The mask of a masked array is accessible through its mask attribute. We must keep in mind that a True entry in the
mask indicates an invalid data.
Another possibility is to use the getmask and getmaskarray functions. getmask(x) outputs the mask of x if
x is a masked array, and the special value nomask otherwise. getmaskarray(x) outputs the mask of x if x is a
masked array. If x has no invalid entry or is not a masked array, the function outputs a boolean array of False with
as many elements as x.

Accessing only the valid entries

To retrieve only the valid entries, we can use the inverse of the mask as an index. The inverse of the mask can be
calculated with the numpy.logical_not function or simply with the ~ operator:

>>> x = ma.array([[1, 2], [3, 4]], mask=[[0, 1], [1, 0]])

>>> x[~x.mask]
masked_array(data = [1 4],
mask = [False False],
fill_value = 999999)

Another way to retrieve the valid data is to use the compressed method, which returns a one-dimensional ndarray
(or one of its subclasses, depending on the value of the baseclass attribute):

>>> x.compressed()
array([1, 4])

Note that the output of compressed is always 1D.

1.6. Masked arrays 193

NumPy Reference, Release 2.0.0.dev8464

Modifying the mask

Masking an entry
The recommended way to mark one or several specific entries of a masked array as invalid is to assign the special
value masked to them:

>>> x = ma.array([1, 2, 3])

>>> x[0] = ma.masked
>>> x
masked_array(data = [-- 2 3],
mask = [ True False False],
fill_value = 999999)
>>> y = ma.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
>>> y[(0, 1, 2), (1, 2, 0)] = ma.masked
>>> y
masked_array(data =
[[1 -- 3]
[4 5 --]
[-- 8 9]],
mask =
[[False True False]
[False False True]
[ True False False]],
fill_value = 999999)
>>> z = ma.array([1, 2, 3, 4])
>>> z[:-2] = ma.masked
>>> z
masked_array(data = [-- -- 3 4],
mask = [ True True False False],
fill_value = 999999)

A second possibility is to modify the mask directly, but this usage is discouraged.
Note: When creating a new masked array with a simple, non-structured datatype, the mask is initially set to the
special value nomask, that corresponds roughly to the boolean False. Trying to set an element of nomask will fail
with a TypeError exception, as a boolean does not support item assignment.
All the entries of an array can be masked at once by assigning True to the mask:

>>> x = ma.array([1, 2, 3], mask=[0, 0, 1])

>>> x.mask = True
>>> x
masked_array(data = [-- -- --],
mask = [ True True True],
fill_value = 999999)

Finally, specific entries can be masked and/or unmasked by assigning to the mask a sequence of booleans:

>>> x = ma.array([1, 2, 3])

>>> x.mask = [0, 1, 0]
>>> x
masked_array(data = [1 -- 3],
mask = [False True False],
fill_value = 999999)

194 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Unmasking an entry
To unmask one or several specific entries, we can just assign one or several new valid values to them:

>>> x = ma.array([1, 2, 3], mask=[0, 0, 1])

>>> x
masked_array(data = [1 2 --],
mask = [False False True],
fill_value = 999999)
>>> x[-1] = 5
>>> x
masked_array(data = [1 2 5],
mask = [False False False],
fill_value = 999999)

Note: Unmasking an entry by direct assignment will silently fail if the masked array has a hard mask, as shown by
the hardmask attribute. This feature was introduced to prevent overwriting the mask. To force the unmasking of an
entry where the array has a hard mask, the mask must first to be softened using the soften_mask method before the
allocation. It can be re-hardened with harden_mask:

>>> x = ma.array([1, 2, 3], mask=[0, 0, 1], hard_mask=True)

>>> x
masked_array(data = [1 2 --],
mask = [False False True],
fill_value = 999999)
>>> x[-1] = 5
>>> x
masked_array(data = [1 2 --],
mask = [False False True],
fill_value = 999999)
>>> x.soften_mask()
>>> x[-1] = 5
>>> x
masked_array(data = [1 2 --],
mask = [False False True],
fill_value = 999999)
>>> x.harden_mask()

To unmask all masked entries of a masked array (provided the mask isnt a hard mask), the simplest solution is to
assign the constant nomask to the mask:

>>> x = ma.array([1, 2, 3], mask=[0, 0, 1])

>>> x
masked_array(data = [1 2 --],
mask = [False False True],
fill_value = 999999)
>>> x.mask = ma.nomask
>>> x
masked_array(data = [1 2 3],
mask = [False False False],
fill_value = 999999)

Indexing and slicing

As a MaskedArray is a subclass of numpy.ndarray, it inherits its mechanisms for indexing and slicing.

1.6. Masked arrays 195

NumPy Reference, Release 2.0.0.dev8464

When accessing a single entry of a masked array with no named fields, the output is either a scalar (if the corresponding
entry of the mask is False) or the special value masked (if the corresponding entry of the mask is True):

>>> x = ma.array([1, 2, 3], mask=[0, 0, 1])

>>> x[0]
1
>>> x[-1]
masked_array(data = --,
mask = True,
fill_value = 1e+20)
>>> x[-1] is ma.masked
True

If the masked array has named fields, accessing a single entry returns a numpy.void object if none of the fields are
masked, or a 0d masked array with the same dtype as the initial array if at least one of the fields is masked.

>>> y = ma.masked_array([(1,2), (3, 4)],

... mask=[(0, 0), (0, 1)],
... dtype=[(a, int), (b, int)])
>>> y[0]
(1, 2)
>>> y[-1]
masked_array(data = (3, --),
mask = (False, True),
fill_value = (999999, 999999),
dtype = [(a, <i4), (b, <i4)])

When accessing a slice, the output is a masked array whose data attribute is a view of the original data, and whose
mask is either nomask (if there was no invalid entries in the original array) or a copy of the corresponding slice of the
original mask. The copy is required to avoid propagation of any modification of the mask to the original.

>>> x = ma.array([1, 2, 3, 4, 5], mask=[0, 1, 0, 0, 1])

>>> mx = x[:3]
>>> mx
masked_array(data = [1 -- 3],
mask = [False True False],
fill_value = 999999)
>>> mx[1] = -1
>>> mx
masked_array(data = [1 -1 3],
mask = [False True False],
fill_value = 999999)
>>> x.mask
array([False, True, False, False, True], dtype=bool)
>>> x.data
array([ 1, -1, 3, 4, 5])

Accessing a field of a masked array with structured datatype returns a MaskedArray.

Operations on masked arrays

Arithmetic and comparison operations are supported by masked arrays. As much as possible, invalid entries of a
masked array are not processed, meaning that the corresponding data entries should be the same before and after the
operation.

196 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Warning: We need to stress that this behavior may not be systematic, that masked data may be affected by the
operation in some cases and therefore users should not rely on this data remaining unchanged.

The numpy.ma module comes with a specific implementation of most ufuncs. Unary and binary functions that have a
validity domain (such as log or divide) return the masked constant whenever the input is masked or falls outside
the validity domain:

>>> ma.log([-1, 0, 1, 2])

masked_array(data = [-- -- 0.0 0.69314718056],
mask = [ True True False False],
fill_value = 1e+20)

Masked arrays also support standard numpy ufuncs. The output is then a masked array. The result of a unary ufunc is
masked wherever the input is masked. The result of a binary ufunc is masked wherever any of the input is masked. If
the ufunc also returns the optional context output (a 3-element tuple containing the name of the ufunc, its arguments
and its domain), the context is processed and entries of the output masked array are masked wherever the corresponding
input fall outside the validity domain:

>>> x = ma.array([-1, 1, 0, 2, 3], mask=[0, 0, 0, 0, 1])

>>> np.log(x)
masked_array(data = [-- -- 0.0 0.69314718056 --],
mask = [ True True False False True],
fill_value = 1e+20)

1.6.3 Examples

Data with a given value representing missing data

Lets consider a list of elements, x, where values of -9999. represent missing data. We wish to compute the average
value of the data and the vector of anomalies (deviations from the average):

>>> import numpy.ma as ma

>>> x = [0.,1.,-9999.,3.,4.]
>>> mx = ma.masked_values (x, -9999.)
>>> print mx.mean()
2.0
>>> print mx - mx.mean()
[-2.0 -1.0 -- 1.0 2.0]
>>> print mx.anom()
[-2.0 -1.0 -- 1.0 2.0]

Filling in the missing data

Suppose now that we wish to print that same data, but with the missing values replaced by the average value.

>>> print mx.filled(mx.mean())

[ 0. 1. 2. 3. 4.]

1.6. Masked arrays 197

NumPy Reference, Release 2.0.0.dev8464

Numerical operations

Numerical operations can be easily performed without worrying about missing values, dividing by zero, square roots
of negative numbers, etc.:

>>> import numpy as np, numpy.ma as ma

>>> x = ma.array([1., -1., 3., 4., 5., 6.], mask=[0,0,0,0,1,0])
>>> y = ma.array([1., 2., 0., 4., 5., 6.], mask=[0,0,0,0,0,1])
>>> print np.sqrt(x/y)
[1.0 -- -- 1.0 -- --]

Four values of the output are invalid: the first one comes from taking the square root of a negative number, the second
from the division by zero, and the last two where the inputs were masked.

Ignoring extreme values

Lets consider an array d of random floats between 0 and 1. We wish to compute the average of the values of d while
ignoring any data outside the range [0.1, 0.9]:

>>> print ma.masked_outside(d, 0.1, 0.9).mean()

1.6.4 Constants of the numpy.ma module

In addition to the MaskedArray class, the numpy.ma module defines several constants.
masked
The masked constant is a special case of MaskedArray, with a float datatype and a null shape. It is used to
test whether a specific entry of a masked array is masked, or to mask one or several entries of a masked array:

>>> x = ma.array([1, 2, 3], mask=[0, 1, 0])

>>> x[1] is ma.masked
True
>>> x[-1] = ma.masked
>>> x
masked_array(data = [1 -- --],
mask = [False True True],
fill_value = 999999)

nomask
Value indicating that a masked array has no invalid entry. nomask is used internally to speed up computations
when the mask is not needed.
masked_print_options
String used in lieu of missing data when a masked array is printed. By default, this string is --.

1.6.5 The MaskedArray class

class MaskedArray()

A subclass of ndarray designed to manipulate numerical arrays with missing data.

An instance of MaskedArray can be thought as the combination of several elements:
The data, as a regular numpy.ndarray of any shape or datatype (the data).

198 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

A boolean mask with the same shape as the data, where a True value indicates that the corresponding element
of the data is invalid. The special value nomask is also acceptable for arrays without named fields, and indicates
that no data is invalid.
A fill_value, a value that may be used to replace the invalid entries in order to return a standard
numpy.ndarray.

Attributes and properties of masked arrays

See Also:
Array Attributes
data
Returns the underlying data, as a view of the masked array. If the underlying data is a subclass of
numpy.ndarray, it is returned as such.

>>> x = ma.array(np.matrix([[1, 2], [3, 4]]), mask=[[0, 1], [1, 0]])

>>> x.data
matrix([[1, 2],
[3, 4]])

The type of the data can be accessed through the baseclass attribute.
mask
Returns the underlying mask, as an array with the same shape and structure as the data, but where all fields are
atomically booleans. A value of True indicates an invalid entry.
recordmask
Returns the mask of the array if it has no named fields. For structured arrays, returns a ndarray of booleans
where entries are True if all the fields are masked, False otherwise:

>>> x = ma.array([(1, 1), (2, 2), (3, 3), (4, 4), (5, 5)],
... mask=[(0, 0), (1, 0), (1, 1), (0, 1), (0, 0)],
... dtype=[(a, int), (b, int)])
>>> x.recordmask
array([False, False, True, False, False], dtype=bool)

fill_value
Returns the value used to fill the invalid entries of a masked array. The value is either a scalar (if the masked
array has no named fields), or a 0-D ndarray with the same dtype as the masked array if it has named fields.
The default filling value depends on the datatype of the array:
datatype default
bool True
int 999999
float 1.e20
complex 1.e20+0j
object ?
string N/A
baseclass
Returns the class of the underlying data.

>>> x = ma.array(np.matrix([[1, 2], [3, 4]]), mask=[[0, 0], [1, 0]])

>>> x.baseclass
<class numpy.matrixlib.defmatrix.matrix>

1.6. Masked arrays 199

NumPy Reference, Release 2.0.0.dev8464

sharedmask
Returns whether the mask of the array is shared between several masked arrays. If this is the case, any modifi-
cation to the mask of one array will be propagated to the others.
hardmask
Returns whether the mask is hard (True) or soft (False). When the mask is hard, masked entries cannot be
unmasked.
As MaskedArray is a subclass of ndarray, a masked array also inherits all the attributes and properties of a
ndarray instance.
MaskedArray.base Base object if memory is from some other object.
MaskedArray.ctypes An object to simplify the interaction of the array with the ctypes
module.
MaskedArray.dtype Data-type of the arrays elements.
MaskedArray.flags Information about the memory layout of the array.
MaskedArray.itemsize Length of one array element in bytes.
MaskedArray.nbytes Total bytes consumed by the elements of the array.
MaskedArray.ndim Number of array dimensions.
MaskedArray.shape Tuple of array dimensions.
MaskedArray.size Number of elements in the array.
MaskedArray.strides Tuple of bytes to step in each dimension when traversing an array.
MaskedArray.imag Imaginary part.
MaskedArray.real Real part
MaskedArray.flat Flat version of the array.
MaskedArray.__array_priority__
base
Base object if memory is from some other object.

Examples
The base of an array that owns its memory is None:

>>> x = np.array([1,2,3,4])
>>> x.base is None
True

Slicing creates a view, whose memory is shared with x:

>>> y = x[2:]
>>> y.base is x
True

ctypes
An object to simplify the interaction of the array with the ctypes module.
This attribute creates an object that makes it easier to use arrays when calling shared libraries with the ctypes
module. The returned object has, among others, data, shape, and strides attributes (see Notes below) which
themselves return ctypes objects that can be used as arguments to a shared library.
Parameters
None :
Returns
c : Python object
Possessing attributes data, shape, strides, etc.

200 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

See Also:
numpy.ctypeslib

Notes
Below are the public attributes of this object which were documented in Guide to NumPy (we have omitted
undocumented public attributes, as well as documented private attributes):
data: A pointer to the memory area of the array as a Python integer. This memory area may contain data
that is not aligned, or not in correct byte-order. The memory area may not even be writeable. The array
flags and data-type of this array should be respected when passing this attribute to arbitrary C-code to avoid
trouble that can include Python crashing. User Beware! The value of this attribute is exactly the same as
self._array_interface_[data][0].
shape (c_intp*self.ndim): A ctypes array of length self.ndim where the basetype is the C-integer corre-
sponding to dtype(p) on this platform. This base-type could be c_int, c_long, or c_longlong depending
on the platform. The c_intp type is defined accordingly in numpy.ctypeslib. The ctypes array contains the
shape of the underlying array.
strides (c_intp*self.ndim): A ctypes array of length self.ndim where the basetype is the same as for the
shape attribute. This ctypes array contains the strides information from the underlying array. This strides
information is important for showing how many bytes must be jumped to get to the next element in the
array.
data_as(obj): Return the data pointer cast to a particular c-types object. For example, calling
self._as_parameter_ is equivalent to self.data_as(ctypes.c_void_p). Perhaps you want to use the data as
a pointer to a ctypes array of floating-point data: self.data_as(ctypes.POINTER(ctypes.c_double)).
shape_as(obj): Return the shape tuple as an array of some other c-types type. For example:
self.shape_as(ctypes.c_short).
strides_as(obj): Return the strides tuple as an array of some other c-types type. For example:
self.strides_as(ctypes.c_longlong).
Be careful using the ctypes attribute - especially on temporary arrays or arrays constructed on the fly. For
example, calling (a+b).ctypes.data_as(ctypes.c_void_p) returns a pointer to memory that is
invalid because the array created as (a+b) is deallocated before the next Python statement. You can avoid this
problem using either c=a+b or ct=(a+b).ctypes. In the latter case, ct will hold a reference to the array
until ct is deleted or re-assigned.
If the ctypes module is not available, then the ctypes attribute of array objects still returns something useful, but
ctypes objects are not returned and errors may be raised instead. In particular, the object will still have the as
parameter attribute which will return an integer equal to the data attribute.

Examples

>>> import ctypes

>>> x
array([[0, 1],
[2, 3]])
>>> x.ctypes.data
30439712
>>> x.ctypes.data_as(ctypes.POINTER(ctypes.c_long))
<ctypes.LP_c_long object at 0x01F01300>
>>> x.ctypes.data_as(ctypes.POINTER(ctypes.c_long)).contents
c_long(0)
>>> x.ctypes.data_as(ctypes.POINTER(ctypes.c_longlong)).contents
c_longlong(4294967296L)
>>> x.ctypes.shape

1.6. Masked arrays 201

NumPy Reference, Release 2.0.0.dev8464

<numpy.core._internal.c_long_Array_2 object at 0x01FFD580>

>>> x.ctypes.shape_as(ctypes.c_long)
<numpy.core._internal.c_long_Array_2 object at 0x01FCE620>
>>> x.ctypes.strides
<numpy.core._internal.c_long_Array_2 object at 0x01FCE620>
>>> x.ctypes.strides_as(ctypes.c_longlong)
<numpy.core._internal.c_longlong_Array_2 object at 0x01F01300>

dtype
Data-type of the arrays elements.
Parameters
None :
Returns
d : numpy dtype object
See Also:
numpy.dtype

Examples

>>> x
array([[0, 1],
[2, 3]])
>>> x.dtype
dtype(int32)
>>> type(x.dtype)
<type numpy.dtype>

flags
Information about the memory layout of the array.

Notes
The flags object can be accessed dictionary-like (as in a.flags[WRITEABLE]), or by using lowercased
attribute names (as in a.flags.writeable). Short flag names are only supported in dictionary access.
Only the UPDATEIFCOPY, WRITEABLE, and ALIGNED flags can be changed by the user, via direct assign-
ment to the attribute or dictionary entry, or by calling ndarray.setflags.
The array flags cannot be set arbitrarily:
UPDATEIFCOPY can only be set False.
ALIGNED can only be set True if the data is truly aligned.
WRITEABLE can only be set True if the array owns its own memory or the ultimate owner of the memory
exposes a writeable buffer interface or is a string.

Attributes
itemsize
Length of one array element in bytes.

Examples

>>> x = np.array([1,2,3], dtype=np.float64)

>>> x.itemsize
8

202 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> x = np.array([1,2,3], dtype=np.complex128)

>>> x.itemsize
16

nbytes
Total bytes consumed by the elements of the array.

Notes
Does not include memory consumed by non-element attributes of the array object.

Examples

>>> x = np.zeros((3,5,2), dtype=np.complex128)

>>> x.nbytes
480
>>> np.prod(x.shape) * x.itemsize
480

ndim
Number of array dimensions.

Examples

>>> x = np.array([1, 2, 3])

>>> x.ndim
1
>>> y = np.zeros((2, 3, 4))
>>> y.ndim
3

shape
Tuple of array dimensions.

Notes
May be used to reshape the array, as long as this would not require a change in the total number of elements

Examples

>>> x = np.array([1, 2, 3, 4])

>>> x.shape
(4,)
>>> y = np.zeros((2, 3, 4))
>>> y.shape
(2, 3, 4)
>>> y.shape = (3, 8)
>>> y
array([[ 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 0., 0.]])
>>> y.shape = (3, 6)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: total size of new array must be unchanged

size
Number of elements in the array.

1.6. Masked arrays 203

NumPy Reference, Release 2.0.0.dev8464

Equivalent to np.prod(a.shape), i.e., the product of the arrays dimensions.

Examples

>>> x = np.zeros((3, 5, 2), dtype=np.complex128)

>>> x.size
30
>>> np.prod(x.shape)
30

strides
Tuple of bytes to step in each dimension when traversing an array.
The byte offset of element (i[0], i[1], ..., i[n]) in an array a is:

offset = sum(np.array(i) * a.strides)

A more detailed explanation of strides can be found in the ndarray.rst file in the NumPy reference guide.
See Also:
numpy.lib.stride_tricks.as_strided

Notes
Imagine an array of 32-bit integers (each 4 bytes):

x = np.array([[0, 1, 2, 3, 4],
[5, 6, 7, 8, 9]], dtype=np.int32)

This array is stored in memory as 40 bytes, one after the other (known as a contiguous block of memory). The
strides of an array tell us how many bytes we have to skip in memory to move to the next position along a certain
axis. For example, we have to skip 4 bytes (1 value) to move to the next column, but 20 bytes (5 values) to get
to the same position in the next row. As such, the strides for the array x will be (20, 4).

Examples

>>> y = np.reshape(np.arange(2*3*4), (2,3,4))

>>> y
array([[[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]],
[[12, 13, 14, 15],
[16, 17, 18, 19],
[20, 21, 22, 23]]])
>>> y.strides
(48, 16, 4)
>>> y[1,1,1]
17
>>> offset=sum(y.strides * np.array((1,1,1)))
>>> offset/y.itemsize
17

>>> x = np.reshape(np.arange(5*6*7*8), (5,6,7,8)).transpose(2,3,1,0)

>>> x.strides
(32, 4, 224, 1344)
>>> i = np.array([3,5,2,2])
>>> offset = sum(i * x.strides)

204 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> x[3,5,2,2]
813
>>> offset / x.itemsize
813

imag
Imaginary part.
real
Real part
flat
Flat version of the array.

1.6.6 MaskedArray methods

See Also:
Array methods

Conversion

MaskedArray.__float__() Convert to float.

MaskedArray.__hex__(x)
MaskedArray.__int__() Convert to int.
MaskedArray.__long__(x)
MaskedArray.__oct__(x)
MaskedArray.view([dtype, type]) New view of array with the same data.
MaskedArray.astype(newtype) Returns a copy of the MaskedArray cast to given newtype.
MaskedArray.byteswap(inplace) Swap the bytes of the array elements
MaskedArray.compressed() Return all the non-masked data as a 1-D array.
MaskedArray.filled([fill_value]) Return a copy of self, with masked values filled with a given
value.
MaskedArray.tofile(fid[, sep, Save a masked array to a file in binary format.
format])
MaskedArray.toflex() Transforms a masked array into a flexible-type array.
MaskedArray.tolist([fill_value]) Return the data portion of the masked array as a hierarchical
Python list.
MaskedArray.torecords() Transforms a masked array into a flexible-type array.
MaskedArray.tostring([fill_value, Return the array data as a string containing the raw bytes in the
order]) array.
__float__()
Convert to float.
__hex__() <==> hex(x)

__int__()
Convert to int.
__long__() <==> long(x)

__oct__() <==> oct(x)

1.6. Masked arrays 205

NumPy Reference, Release 2.0.0.dev8464

view(dtype=None, type=None)
New view of array with the same data.
Parameters
dtype : data-type
Data-type descriptor of the returned view, e.g. float32 or int16.
type : python type
Type of the returned view, e.g. ndarray or matrix.

Notes
a.view() is used two different ways.
a.view(some_dtype) or a.view(dtype=some_dtype) constructs a view of the arrays memory with a different
dtype. This can cause a reinterpretation of the bytes of memory.
a.view(ndarray_subclass), or a.view(type=ndarray_subclass), just returns an instance of ndarray_subclass that
looks at the same array (same shape, dtype, etc.). This does not cause a reinterpretation of the memory.

Examples

>>> x = np.array([(1, 2)], dtype=[(a, np.int8), (b, np.int8)])

Viewing array data using a different type and dtype:

>>> y = x.view(dtype=np.int16, type=np.matrix)

>>> y
matrix([[513]], dtype=int16)
>>> print type(y)
<class numpy.matrixlib.defmatrix.matrix>

Creating a view on a structured array so it can be used in calculations

>>> x = np.array([(1, 2),(3,4)], dtype=[(a, np.int8), (b, np.int8)])

>>> xv = x.view(dtype=np.int8).reshape(-1,2)
>>> xv
array([[1, 2],
[3, 4]], dtype=int8)
>>> xv.mean(0)
array([ 2., 3.])

Making changes to the view changes the underlying array

>>> xv[0,1] = 20
>>> print x
[(1, 20) (3, 4)]

Using a view to convert an array to a record array:

>>> z = x.view(np.recarray)
>>> z.a
array([1], dtype=int8)

Views share data:

206 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> x[0] = (9, 10)

>>> z[0]
(9, 10)

astype(newtype)
Returns a copy of the MaskedArray cast to given newtype.
Returns
output : MaskedArray
A copy of self cast to input newtype. The returned record shape matches self.shape.

Examples

>>> x = np.ma.array([[1,2,3.1],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1.0 -- 3.1]
[-- 5.0 --]
[7.0 -- 9.0]]
>>> print x.astype(int32)
[[1 -- 3]
[-- 5 --]
[7 -- 9]]

byteswap(inplace)
Swap the bytes of the array elements
Toggle between low-endian and big-endian data representation by returning a byteswapped array, optionally
swapped in-place.
Parameters
inplace: bool, optional :
If True, swap bytes in-place, default is False.
Returns
out: ndarray :
The byteswapped array. If inplace is True, this is a view to self.

Examples

>>> A = np.array([1, 256, 8755], dtype=np.int16)

>>> map(hex, A)
[0x1, 0x100, 0x2233]
>>> A.byteswap(True)
array([ 256, 1, 13090], dtype=int16)
>>> map(hex, A)
[0x100, 0x1, 0x3322]

Arrays of strings are not swapped

>>> A = np.array([ceg, fac])

>>> A.byteswap()
array([ceg, fac],
dtype=|S3)

compressed()
Return all the non-masked data as a 1-D array.

1.6. Masked arrays 207

NumPy Reference, Release 2.0.0.dev8464

Returns
data : ndarray
A new ndarray holding the non-masked data is returned.

Notes
The result is not a MaskedArray!

Examples

>>> x = np.ma.array(np.arange(5), mask=[0]*2 + [1]*3)

>>> x.compressed()
array([0, 1])
>>> type(x.compressed())
<type numpy.ndarray>

filled(fill_value=None)
Return a copy of self, with masked values filled with a given value.
Parameters
fill_value : scalar, optional
The value to use for invalid entries (None by default). If None, the fill_value attribute
of the array is used instead.

Notes
The result is not a MaskedArray!

Examples

>>> x = np.ma.array([1,2,3,4,5], mask=[0,0,1,0,1], fill_value=-999)

>>> x.filled()
array([1, 2, -999, 4, -999])
>>> type(x.filled())
<type numpy.ndarray>

Subclassing is preserved. This means that if the data part of the masked array is a matrix, filled returns a matrix:

>>> x = np.ma.array(np.matrix([[1, 2], [3, 4]]), mask=[[0, 1], [1, 0]])

>>> x.filled()
matrix([[ 1, 999999],
[999999, 4]])

tofile(fid, sep=, format=%s)

Save a masked array to a file in binary format.

Warning: This function is not implemented yet.

Raises
NotImplementedError :
When tofile is called.
toflex()
Transforms a masked array into a flexible-type array.
The flexible type array that is returned will have two fields:

208 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

the _data field stores the _data part of the array.

the _mask field stores the _mask part of the array.

Parameters
None :
Returns
record : ndarray
A new flexible-type ndarray with two fields: the first element containing a value, the
second element containing the corresponding mask boolean. The returned record shape
matches self.shape.

Notes
A side-effect of transforming a masked array into a flexible ndarray is that meta information (fill_value,
...) will be lost.

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.toflex()
[[(1, False) (2, True) (3, False)]
[(4, True) (5, False) (6, True)]
[(7, False) (8, True) (9, False)]]

tolist(fill_value=None)
Return the data portion of the masked array as a hierarchical Python list.
Data items are converted to the nearest compatible Python type. Masked values are converted to fill_value. If
fill_value is None, the corresponding entries in the output list will be None.
Parameters
fill_value : scalar, optional
The value to use for invalid entries. Default is None.
Returns
result : list
The Python list representation of the masked array.

Examples

>>> x = np.ma.array([[1,2,3], [4,5,6], [7,8,9]], mask=[0] + [1,0]*4)

>>> x.tolist()
[[1, None, 3], [None, 5, None], [7, None, 9]]
>>> x.tolist(-999)
[[1, -999, 3], [-999, 5, -999], [7, -999, 9]]

torecords()
Transforms a masked array into a flexible-type array.
The flexible type array that is returned will have two fields:
the _data field stores the _data part of the array.

1.6. Masked arrays 209

NumPy Reference, Release 2.0.0.dev8464

the _mask field stores the _mask part of the array.

Parameters
None :
Returns
record : ndarray
A new flexible-type ndarray with two fields: the first element containing a value, the
second element containing the corresponding mask boolean. The returned record shape
matches self.shape.

Notes
A side-effect of transforming a masked array into a flexible ndarray is that meta information (fill_value,
...) will be lost.

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.toflex()
[[(1, False) (2, True) (3, False)]
[(4, True) (5, False) (6, True)]
[(7, False) (8, True) (9, False)]]

tostring(fill_value=None, order=C)
Return the array data as a string containing the raw bytes in the array.
The array is filled with a fill value before the string conversion.
Parameters
fill_value : scalar, optional
Value used to fill in the masked values. Deafult is None, in which case MaskedAr-
ray.fill_value is used.
order : {C,F,A}, optional
Order of the data item in the copy. Default is C.
C C order (row major).
F Fortran order (column major).
A Any, current order of array.
None Same as A.
See Also:
ndarray.tostring, tolist, tofile

Notes
As for ndarray.tostring, information about the shape, dtype, etc., but also about fill_value, will be lost.

210 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.ma.array(np.array([[1, 2], [3, 4]]), mask=[[0, 1], [1, 0]])

>>> x.tostring()
\x01\x00\x00\x00?B\x0f\x00?B\x0f\x00\x04\x00\x00\x00

Shape manipulation

For reshape, resize, and transpose, the single tuple argument may be replaced with n integers which will be interpreted
as an n-tuple.
MaskedArray.flatten([order]) Return a copy of the array collapsed into one dimension.
MaskedArray.ravel() Returns a 1D version of self, as a view.
MaskedArray.reshape(*s, **kwargs) Give a new shape to the array without changing its data.
MaskedArray.resize(newshape[, refcheck,
order])
MaskedArray.squeeze() Remove single-dimensional entries from the shape of a.
MaskedArray.swapaxes(axis1, axis2) Return a view of the array with axis1 and axis2
interchanged.
MaskedArray.transpose(*axes) Returns a view of the array with axes transposed.
MaskedArray.T
flatten(order=C)
Return a copy of the array collapsed into one dimension.
Parameters
order : {C, F}, optional
Whether to flatten in C (row-major) or Fortran (column-major) order. The default is
C.
Returns
y : ndarray
A copy of the input array, flattened to one dimension.
See Also:

ravel
Return a flattened array.
flat
A 1-D flat iterator over the array.

Examples

>>> a = np.array([[1,2], [3,4]])

>>> a.flatten()
array([1, 2, 3, 4])
>>> a.flatten(F)
array([1, 3, 2, 4])

ravel()
Returns a 1D version of self, as a view.
Returns
MaskedArray :

1.6. Masked arrays 211

NumPy Reference, Release 2.0.0.dev8464

Output view is of shape (self.size,) (or

(np.ma.product(self.shape),)).

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.ravel()
[1 -- 3 -- 5 -- 7 -- 9]

reshape(*s, **kwargs)
Give a new shape to the array without changing its data.
Returns a masked array containing the same data, but with a new shape. The result is a view on the original
array; if this is not possible, a ValueError is raised.
Parameters
shape : int or tuple of ints
The new shape should be compatible with the original shape. If an integer is supplied,
then the result will be a 1-D array of that length.
order : {C, F}, optional
Determines whether the array data should be viewed as in C (row-major) or FORTRAN
(column-major) order.
Returns
reshaped_array : array
A new view on the array.
See Also:

reshape
Equivalent function in the masked array module.
numpy.ndarray.reshape
Equivalent method on ndarray object.
numpy.reshape
Equivalent function in the NumPy module.

Notes
The reshaping operation cannot guarantee that a copy will not be made, to modify the shape in place, use
a.shape = s

Examples

>>> x = np.ma.array([[1,2],[3,4]], mask=[1,0,0,1])

>>> print x
[[-- 2]
[3 --]]
>>> x = x.reshape((4,1))
>>> print x
[[--]
[2]

212 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

[3]
[--]]

resize(newshape, refcheck=True, order=False)

Warning: This method does nothing, except raise a ValueError exception. A masked array does not own
its data and therefore cannot safely be resized in place. Use the numpy.ma.resize function instead.

This method is difficult to implement safely and may be deprecated in future releases of NumPy.
squeeze()
Remove single-dimensional entries from the shape of a.
Refer to numpy.squeeze for full documentation.
See Also:

numpy.squeeze
equivalent function

swapaxes(axis1, axis2)
Return a view of the array with axis1 and axis2 interchanged.
Refer to numpy.swapaxes for full documentation.
See Also:

numpy.swapaxes
equivalent function

transpose(*axes)
Returns a view of the array with axes transposed.
For a 1-D array, this has no effect. (To change between column and row vectors, first cast the 1-D array into a ma-
trix object.) For a 2-D array, this is the usual matrix transpose. For an n-D array, if axes are given, their order in-
dicates how the axes are permuted (see Examples). If axes are not provided and a.shape = (i[0], i[1],
... i[n-2], i[n-1]), then a.transpose().shape = (i[n-1], i[n-2], ... i[1],
i[0]).
Parameters
axes : None, tuple of ints, or n ints
None or no argument: reverses the order of the axes.
tuple of ints: i in the j-th place in the tuple means as i-th axis becomes a.transpose()s j-th
axis.
n ints: same as an n-tuple of the same ints (this form is intended simply as a convenience
alternative to the tuple form)
Returns
out : ndarray
View of a, with axes suitably permuted.
See Also:

ndarray.T
Array property returning the array transposed.

1.6. Masked arrays 213

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> a
array([[1, 2],
[3, 4]])
>>> a.transpose()
array([[1, 3],
[2, 4]])
>>> a.transpose((1, 0))
array([[1, 3],
[2, 4]])
>>> a.transpose(1, 0)
array([[1, 3],
[2, 4]])

Item selection and manipulation

For array methods that take an axis keyword, it defaults to None. If axis is None, then the array is treated as a 1-D
array. Any other value for axis represents the dimension along which the operation should proceed.
MaskedArray.argmax([axis, fill_value, Returns array of indices of the maximum values along the given
out]) axis.
MaskedArray.argmin([axis, fill_value, Return array of indices to the minimum values along the given
out]) axis.
MaskedArray.argsort([axis, kind, Return an ndarray of indices that sort the array along the
order, ...]) specified axis.
MaskedArray.choose(choices[, out, Use an index array to construct a new array from a set of choices.
mode])
MaskedArray.compress(condition[, Return a where condition is True.
axis, out])
MaskedArray.diagonal([offset, axis1, Return specified diagonals.
axis2])
MaskedArray.fill(value) Fill the array with a scalar value.
MaskedArray.item(*args) Copy an element of an array to a standard Python scalar and
return it.
MaskedArray.nonzero() Return the indices of unmasked elements that are not zero.
MaskedArray.put(indices, values[, Set storage-indexed locations to corresponding values.
mode])
MaskedArray.repeat(repeats[, axis]) Repeat elements of an array.
MaskedArray.searchsorted(v[, side]) Find indices where elements of v should be inserted in a to
maintain order.
MaskedArray.sort([axis, kind, order, Sort the array, in-place
...])
MaskedArray.take(indices[, axis, out,
mode])
argmax(axis=None, fill_value=None, out=None)
Returns array of indices of the maximum values along the given axis. Masked values are treated as if they had
the value fill_value.
Parameters
axis : {None, integer}

214 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

If None, the index is into the flattened array, otherwise along the specified axis
fill_value : {var}, optional
Value used to fill in the masked values. If None, the output of maxi-
mum_fill_value(self._data) is used instead.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
Returns
index_array : {integer_array}

Examples

>>> a = np.arange(6).reshape(2,3)
>>> a.argmax()
5
>>> a.argmax(0)
array([1, 1, 1])
>>> a.argmax(1)
array([2, 2])

argmin(axis=None, fill_value=None, out=None)

Return array of indices to the minimum values along the given axis.
Parameters
axis : {None, integer}
If None, the index is into the flattened array, otherwise along the specified axis
fill_value : {var}, optional
Value used to fill in the masked values. If None, the output of mini-
mum_fill_value(self._data) is used instead.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
Returns
{ndarray, scalar} :
If multi-dimension input, returns a new ndarray of indices to the minimum values along
the given axis. Otherwise, returns a scalar of index to the minimum values along the
given axis.

Examples

>>> x = np.ma.array(arange(4), mask=[1,1,0,0])

>>> x.shape = (2,2)
>>> print x
[[-- --]
[2 3]]
>>> print x.argmin(axis=0, fill_value=-1)
[0 0]
>>> print x.argmin(axis=0, fill_value=9)
[1 1]

1.6. Masked arrays 215

NumPy Reference, Release 2.0.0.dev8464

argsort(axis=None, kind=quicksort, order=None, fill_value=None)

Return an ndarray of indices that sort the array along the specified axis. Masked values are filled beforehand to
fill_value.
Parameters
axis : int, optional
Axis along which to sort. The default is -1 (last axis). If None, the flattened array is
used.
fill_value : var, optional
Value used to fill the array before sorting. The default is the fill_value attribute of the
input array.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm.
order : list, optional
When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
Returns
index_array : ndarray, int
Array of indices that sort a along the specified axis. In other words,
a[index_array] yields a sorted a.
See Also:

sort
Describes sorting algorithms used.
lexsort
Indirect stable sort with multiple keys.
ndarray.sort
Inplace sort.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = np.ma.array([3,2,1], mask=[False, False, True])

>>> a
masked_array(data = [3 2 --],
mask = [False False True],
fill_value = 999999)
>>> a.argsort()
array([1, 0, 2])

choose(choices, out=None, mode=raise)

Use an index array to construct a new array from a set of choices.
Refer to numpy.choose for full documentation.
See Also:

216 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

numpy.choose
equivalent function

compress(condition, axis=None, out=None)

Return a where condition is True.
If condition is a MaskedArray, missing values are considered as False.
Parameters
condition : var
Boolean 1-d array selecting which entries to return. If len(condition) is less than the
size of a along the axis, then output is truncated to length of condition array.
axis : {None, int}, optional
Axis along which the operation must be performed.
out : {None, ndarray}, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary.
Returns
result : MaskedArray
A MaskedArray object.

Notes
Please note the difference with compressed ! The output of compress has a mask, the output of
compressed does not.

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> x.compress([1, 0, 1])
masked_array(data = [1 3],
mask = [False False],
fill_value=999999)

>>> x.compress([1, 0, 1], axis=1)

masked_array(data =
[[1 3]
[-- --]
[7 9]],
mask =
[[False False]
[ True True]
[False False]],
fill_value=999999)

diagonal(offset=0, axis1=0, axis2=1)

Return specified diagonals.
Refer to numpy.diagonal for full documentation.

1.6. Masked arrays 217

NumPy Reference, Release 2.0.0.dev8464

See Also:

numpy.diagonal
equivalent function

fill(value)
Fill the array with a scalar value.
Parameters
value : scalar
All elements of a will be assigned this value.

Examples

>>> a = np.array([1, 2])

>>> a.fill(0)
>>> a
array([0, 0])
>>> a = np.empty(2)
>>> a.fill(1)
>>> a
array([ 1., 1.])

item(*args)
Copy an element of an array to a standard Python scalar and return it.
Parameters
*args : Arguments (variable number and type)
none: in this case, the method only works for arrays with one element (a.size == 1), which
element is copied into a standard Python scalar object and returned.
int_type: this argument is interpreted as a flat index into the array, specifying which element
to copy and return.
tuple of int_types: functions as does a single int_type argument, except that the argument is
interpreted as an nd-index into the array.
Returns
z : Standard Python scalar object
A copy of the specified element of the array as a suitable Python scalar

Notes
When the data type of a is longdouble or clongdouble, item() returns a scalar array object because there is no
available Python scalar that would not lose information. Void arrays return a buffer object for item(), unless
fields are defined, in which case a tuple is returned.
item is very similar to a[args], except, instead of an array scalar, a standard Python scalar is returned. This can
be useful for speeding up access to elements of the array and doing arithmetic on elements of the array using
Pythons optimized math.

Examples

>>> x = np.random.randint(9, size=(3, 3))

>>> x
array([[3, 1, 7],
[2, 8, 3],

218 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

[8, 5, 3]])
>>> x.item(3)
2
>>> x.item(7)
5
>>> x.item((0, 1))
1
>>> x.item((2, 2))
3

nonzero()
Return the indices of unmasked elements that are not zero.
Returns a tuple of arrays, one for each dimension, containing the indices of the non-zero elements in that
dimension. The corresponding non-zero values can be obtained with:

a[a.nonzero()]

To group the indices by element, rather than dimension, use instead:

np.transpose(a.nonzero())

The result of this is always a 2d array, with a row for each non-zero element.
Parameters
None :
Returns
tuple_of_arrays : tuple
Indices of elements that are non-zero.
See Also:

numpy.nonzero
Function operating on ndarrays.
flatnonzero
Return indices that are non-zero in the flattened version of the input array.
ndarray.nonzero
Equivalent ndarray method.

Examples

>>> import numpy.ma as ma

>>> x = ma.array(np.eye(3))
>>> x
masked_array(data =
[[ 1. 0. 0.]
[ 0. 1. 0.]
[ 0. 0. 1.]],
mask =
False,
fill_value=1e+20)
>>> x.nonzero()
(array([0, 1, 2]), array([0, 1, 2]))

Masked elements are ignored.

1.6. Masked arrays 219

NumPy Reference, Release 2.0.0.dev8464

>>> x[1, 1] = ma.masked

>>> x
masked_array(data =
[[1.0 0.0 0.0]
[0.0 -- 0.0]
[0.0 0.0 1.0]],
mask =
[[False False False]
[False True False]
[False False False]],
fill_value=1e+20)
>>> x.nonzero()
(array([0, 2]), array([0, 2]))

Indices can also be grouped by element.

>>> np.transpose(x.nonzero())
array([[0, 0],
[2, 2]])

A common use for nonzero is to find the indices of an array, where a condition is True. Given an array a, the
condition a > 3 is a boolean array and since False is interpreted as 0, ma.nonzero(a > 3) yields the indices of the
a where the condition is true.

>>> a = ma.array([[1,2,3],[4,5,6],[7,8,9]])
>>> a > 3
masked_array(data =
[[False False False]
[ True True True]
[ True True True]],
mask =
False,
fill_value=999999)
>>> ma.nonzero(a > 3)
(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

The nonzero method of the condition array can also be called.

>>> (a > 3).nonzero()

(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

put(indices, values, mode=raise)

Set storage-indexed locations to corresponding values.
Sets self._data.flat[n] = values[n] for each n in indices. If values is shorter than indices then it will repeat. If
values has some masked values, the initial mask is updated in consequence, else the corresponding values are
unmasked.
Parameters
indices : 1-D array_like
Target indices, interpreted as integers.
values : array_like
Values to place in self._data copy at target indices.
mode : {raise, wrap, clip}, optional

220 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Specifies how out-of-bounds indices will behave. raise : raise an error. wrap : wrap
around. clip : clip to the range.

Notes
values can be a scalar or length 1 array.

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> x.put([0,4,8],[10,20,30])
>>> print x
[[10 -- 3]
[-- 20 --]
[7 -- 30]]

>>> x.put(4,999)
>>> print x
[[10 -- 3]
[-- 999 --]
[7 -- 30]]

repeat(repeats, axis=None)
Repeat elements of an array.
Refer to numpy.repeat for full documentation.
See Also:

numpy.repeat
equivalent function

searchsorted(v, side=left)
Find indices where elements of v should be inserted in a to maintain order.
For full documentation, see numpy.searchsorted
See Also:

numpy.searchsorted
equivalent function

sort(axis=-1, kind=quicksort, order=None, endwith=True, fill_value=None)

Sort the array, in-place
Parameters
a : array_like
Array to be sorted.
axis : int, optional
Axis along which to sort. If None, the array is flattened before sorting. The default is
-1, which sorts along the last axis.
kind : {quicksort, mergesort, heapsort}, optional

1.6. Masked arrays 221

NumPy Reference, Release 2.0.0.dev8464

Sorting algorithm. Default is quicksort.

order : list, optional
When a is a structured array, this argument specifies which fields to compare first, sec-
ond, and so on. This list does not need to include all of the fields.
endwith : {True, False}, optional
Whether missing values (if any) should be forced in the upper indices (at the end of the
array) (True) or lower indices (at the beginning).
fill_value : {var}, optional
Value used internally for the masked values. If fill_value is not None, it supersedes
endwith.
Returns
sorted_array : ndarray
Array of the same type and shape as a.
See Also:

ndarray.sort
Method to sort an array in-place.
argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in a sorted array.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # Default
>>> a.sort()
>>> print a
[1 3 5 -- --]

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # Put missing values in the front
>>> a.sort(endwith=False)
>>> print a
[-- -- 1 3 5]

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # fill_value takes over endwith
>>> a.sort(endwith=False, fill_value=3)
>>> print a
[1 -- -- 3 5]

222 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

take(indices, axis=None, out=None, mode=raise)

Pickling and copy

MaskedArray.copy([order]) Return a copy of the array.

MaskedArray.dump(file) Dump a pickle of the array to the specified file.
MaskedArray.dumps() Returns the pickle of the array as a string.
copy(order=C)
Return a copy of the array.
Parameters
order : {C, F, A}, optional
By default, the result is stored in C-contiguous (row-major) order in memory. If order is
F, the result has Fortran (column-major) order. If order is A (Any), then the result
has the same order as the input.

Examples

>>> x = np.array([[1,2,3],[4,5,6]], order=F)

>>> y = x.copy()

>>> x.fill(0)

>>> x
array([[0, 0, 0],
[0, 0, 0]])

>>> y
array([[1, 2, 3],
[4, 5, 6]])

>>> y.flags[C_CONTIGUOUS]
True

dump(file)
Dump a pickle of the array to the specified file. The array can be read back with pickle.load or numpy.load.
Parameters
file : str
A string naming the dump file.
dumps()
Returns the pickle of the array as a string. pickle.loads or numpy.loads will convert the string back to an array.
Parameters
None :

1.6. Masked arrays 223

NumPy Reference, Release 2.0.0.dev8464

Calculations

MaskedArray.all([axis, out]) Check if all of the elements of a are true.

MaskedArray.anom([axis, dtype]) Compute the anomalies (deviations from the arithmetic mean) along
the given axis.
MaskedArray.any([axis, out]) Check if any of the elements of a are true.
MaskedArray.clip(a_min, a_max[, Return an array whose values are limited to [a_min, a_max].
out])
MaskedArray.conj() Complex-conjugate all elements.
MaskedArray.conjugate() Return the complex conjugate, element-wise.
MaskedArray.cumprod([axis, dtype, Return the cumulative product of the elements along the given axis.
out])
MaskedArray.cumsum([axis, dtype, Return the cumulative sum of the elements along the given axis.
out])
MaskedArray.max([axis, out, Return the maximum along a given axis.
fill_value])
MaskedArray.mean([axis, dtype, Returns the average of the array elements.
out])
MaskedArray.min([axis, out, Return the minimum along a given axis.
fill_value])
MaskedArray.prod([axis, dtype, Return the product of the array elements over the given axis.
out])
MaskedArray.product([axis, dtype, Return the product of the array elements over the given axis.
out])
MaskedArray.ptp([axis, out, Return (maximum - minimum) along the the given dimension (i.e.
fill_value])
MaskedArray.round([decimals, out]) Return an array rounded a to the given number of decimals.
MaskedArray.std([axis, dtype, out, Compute the standard deviation along the specified axis.
ddof])
MaskedArray.sum([axis, dtype, out]) Return the sum of the array elements over the given axis.
MaskedArray.trace([offset, axis1, Return the sum along diagonals of the array.
axis2, ...])
MaskedArray.var([axis, dtype, out, Compute the variance along the specified axis.
ddof])
all(axis=None, out=None)
Check if all of the elements of a are true.
Performs a logical_and over the given axis and returns the result. Masked values are considered as True
during computation. For convenience, the output array is masked where ALL the values along the current axis
are masked: if the output would have been a scalar and that all the values are masked, then the output is masked.
Parameters
axis : {None, integer}
Axis to perform the operation over. If None, perform over flattened array.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
See Also:

all
equivalent function

224 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.ma.array([1,2,3]).all()
True
>>> a = np.ma.array([1,2,3], mask=True)
>>> (a.all() is np.ma.masked)
True

anom(axis=None, dtype=None)
Compute the anomalies (deviations from the arithmetic mean) along the given axis.
Returns an array of anomalies, with the same shape as the input and where the arithmetic mean is computed
along the given axis.
Parameters
axis : int, optional
Axis over which the anomalies are taken. The default is to use the mean of the flattened
array as reference.
dtype : dtype, optional
Type to use in computing the variance. For arrays of integer type
the default is float32; for arrays of float types it is the same as the array type.
See Also:

mean
Compute the mean of the array.

Examples

>>> a = np.ma.array([1,2,3])
>>> a.anom()
masked_array(data = [-1. 0. 1.],
mask = False,
fill_value = 1e+20)

any(axis=None, out=None)
Check if any of the elements of a are true.
Performs a logical_or over the given axis and returns the result. Masked values are considered as False during
computation.
Parameters
axis : {None, integer}
Axis to perform the operation over. If None, perform over flattened array and return a
scalar.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
See Also:

any
equivalent function

1.6. Masked arrays 225

NumPy Reference, Release 2.0.0.dev8464

clip(a_min, a_max, out=None)

Return an array whose values are limited to [a_min, a_max].
Refer to numpy.clip for full documentation.
See Also:

numpy.clip
equivalent function

conj()
Complex-conjugate all elements.
Refer to numpy.conjugate for full documentation.
See Also:

numpy.conjugate
equivalent function

conjugate()
Return the complex conjugate, element-wise.
Refer to numpy.conjugate for full documentation.
See Also:

numpy.conjugate
equivalent function

cumprod(axis=None, dtype=None, out=None)

Return the cumulative product of the elements along the given axis. The cumulative product is taken over the
flattened array by default, otherwise over the specified axis.
Masked values are set to 1 internally during the computation. However, their position is saved, and the result
will be masked at the same locations.
Parameters
axis : {None, -1, int}, optional
Axis along which the product is computed. The default (axis = None) is to compute
over the flattened array.
dtype : {None, dtype}, optional
Determines the type of the returned array and of the accumulator where the elements
are multiplied. If dtype has the value None and the type of a is an integer type
of precision less than the default platform integer, then the default platform integer
precision is used. Otherwise, the dtype is the same as that of a.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
cumprod : ndarray
A new array holding the result is returned unless out is specified, in which case a refer-
ence to out is returned.

226 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Notes
The mask is lost if out is not a valid MaskedArray !
Arithmetic is modular when using integer types, and no error is raised on overflow.
cumsum(axis=None, dtype=None, out=None)
Return the cumulative sum of the elements along the given axis. The cumulative sum is calculated over the
flattened array by default, otherwise over the specified axis.
Masked values are set to 0 internally during the computation. However, their position is saved, and the result
will be masked at the same locations.
Parameters
axis : {None, -1, int}, optional
Axis along which the sum is computed. The default (axis = None) is to compute over
the flattened array. axis may be negative, in which case it counts from the last to the
first axis.
dtype : {None, dtype}, optional
Type of the returned array and of the accumulator in which the elements are summed. If
dtype is not specified, it defaults to the dtype of a, unless a has an integer dtype with a
precision less than that of the default platform integer. In that case, the default platform
integer is used.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
cumsum : ndarray.
A new array holding the result is returned unless out is specified, in which case a
reference to out is returned.

Notes
The mask is lost if out is not a valid MaskedArray !
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> marr = np.ma.array(np.arange(10), mask=[0,0,0,1,1,1,0,0,0,0])

>>> print marr.cumsum()
[0 1 3 -- -- -- 9 16 24 33]

max(axis=None, out=None, fill_value=None)

Return the maximum along a given axis.
Parameters
axis : {None, int}, optional
Axis along which to operate. By default, axis is None and the flattened input is used.
out : array_like, optional
Alternative output array in which to place the result. Must be of the same shape and
buffer length as the expected output.
fill_value : {var}, optional

1.6. Masked arrays 227

NumPy Reference, Release 2.0.0.dev8464

Value used to fill in the masked values. If None, use the output of maxi-
mum_fill_value().
Returns
amax : array_like
New array holding the result. If out was specified, out is returned.
See Also:

maximum_fill_value
Returns the maximum filling value for a given datatype.

mean(axis=None, dtype=None, out=None)

Returns the average of the array elements.
Masked entries are ignored. The average is taken over the flattened array by default, otherwise over the specified
axis. Refer to numpy.mean for the full documentation.
Parameters
a : array_like
Array containing numbers whose mean is desired. If a is not an array, a conversion is
attempted.
axis : int, optional
Axis along which the means are computed. The default is to compute the mean of the
flattened array.
dtype : dtype, optional
Type to use in computing the mean. For integer inputs, the default is float64; for floating
point, inputs it is the same as the input dtype.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary.
Returns
mean : ndarray, see dtype parameter above
If out=None, returns a new array containing the mean values, otherwise a reference to
the output array is returned.
See Also:

numpy.ma.mean
Equivalent function.
numpy.mean
Equivalent function on non-masked arrays.
numpy.ma.average
Weighted average.

Examples

>>> a = np.ma.array([1,2,3], mask=[False, False, True])

>>> a
masked_array(data = [1 2 --],

228 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

mask = [False False True],

fill_value = 999999)
>>> a.mean()
1.5

min(axis=None, out=None, fill_value=None)

Return the minimum along a given axis.
Parameters
axis : {None, int}, optional
Axis along which to operate. By default, axis is None and the flattened input is used.
out : array_like, optional
Alternative output array in which to place the result. Must be of the same shape and
buffer length as the expected output.
fill_value : {var}, optional
Value used to fill in the masked values. If None, use the output of minimum_fill_value.
Returns
amin : array_like
New array holding the result. If out was specified, out is returned.
See Also:

minimum_fill_value
Returns the minimum filling value for a given datatype.

prod(axis=None, dtype=None, out=None)

Return the product of the array elements over the given axis. Masked elements are set to 1 internally for
computation.
Parameters
axis : {None, int}, optional
Axis over which the product is taken. If None is used, then the product is over all the
array elements.
dtype : {None, dtype}, optional
Determines the type of the returned array and of the accumulator where the elements are
multiplied. If dtype has the value None and the type of a is an integer type of precision
less than the default platform integer, then the default platform integer precision is used.
Otherwise, the dtype is the same as that of a.
out : {None, array}, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary.
Returns
product_along_axis : {array, scalar}, see dtype parameter above.
Returns an array whose shape is the same as a with the specified axis removed. Returns
a 0d array when a is 1d or axis=None. Returns a reference to the specified output array
if specified.
See Also:

1.6. Masked arrays 229

NumPy Reference, Release 2.0.0.dev8464

prod
equivalent function

Notes
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> np.prod([1.,2.])
2.0
>>> np.prod([1.,2.], dtype=np.int32)
2
>>> np.prod([[1.,2.],[3.,4.]])
24.0
>>> np.prod([[1.,2.],[3.,4.]], axis=1)
array([ 2., 12.])

product(axis=None, dtype=None, out=None)

Return the product of the array elements over the given axis. Masked elements are set to 1 internally for
computation.
Parameters
axis : {None, int}, optional
Axis over which the product is taken. If None is used, then the product is over all the
array elements.
dtype : {None, dtype}, optional
Determines the type of the returned array and of the accumulator where the elements are
multiplied. If dtype has the value None and the type of a is an integer type of precision
less than the default platform integer, then the default platform integer precision is used.
Otherwise, the dtype is the same as that of a.
out : {None, array}, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary.
Returns
product_along_axis : {array, scalar}, see dtype parameter above.
Returns an array whose shape is the same as a with the specified axis removed. Returns
a 0d array when a is 1d or axis=None. Returns a reference to the specified output array
if specified.
See Also:

prod
equivalent function

Notes
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> np.prod([1.,2.])
2.0

230 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> np.prod([1.,2.], dtype=np.int32)

2
>>> np.prod([[1.,2.],[3.,4.]])
24.0
>>> np.prod([[1.,2.],[3.,4.]], axis=1)
array([ 2., 12.])

ptp(axis=None, out=None, fill_value=None)

Return (maximum - minimum) along the the given dimension (i.e. peak-to-peak value).
Parameters
axis : {None, int}, optional
Axis along which to find the peaks. If None (default) the flattened array is used.
out : {None, array_like}, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
fill_value : {var}, optional
Value used to fill in the masked values.
Returns
ptp : ndarray.
A new array holding the result, unless out was specified, in which case a reference to
out is returned.
round(decimals=0, out=None)
Return an array rounded a to the given number of decimals.
Refer to numpy.around for full documentation.
See Also:

numpy.around
equivalent function

std(axis=None, dtype=None, out=None, ddof=0)

Compute the standard deviation along the specified axis.
Returns the standard deviation, a measure of the spread of a distribution, of the array elements. The standard
deviation is computed for the flattened array by default, otherwise over the specified axis.
Parameters
a : array_like
Calculate the standard deviation of these values.
axis : int, optional
Axis along which the standard deviation is computed. The default is to compute the
standard deviation of the flattened array.
dtype : dtype, optional
Type to use in computing the standard deviation. For arrays of integer type the default
is float64, for arrays of float types it is the same as the array type.
out : ndarray, optional

1.6. Masked arrays 231

NumPy Reference, Release 2.0.0.dev8464

Alternative output array in which to place the result. It must have the same shape as the
expected output but the type (of the calculated values) will be cast if necessary.
ddof : int, optional
Means Delta Degrees of Freedom. The divisor used in calculations is N - ddof,
where N represents the number of elements. By default ddof is zero.
Returns
standard_deviation : ndarray, see dtype parameter above.
If out is None, return a new array containing the standard deviation, otherwise return a
reference to the output array.
See Also:
var, mean
numpy.doc.ufuncs
Section Output arguments

Notes
The standard deviation is the square root of the average of the squared deviations from the mean, i.e., std =
sqrt(mean(abs(x - x.mean())**2)).
The average squared deviation is normally calculated as x.sum() / N, where N = len(x). If, however,
ddof is specified, the divisor N - ddof is used instead. In standard statistical practice, ddof=1 provides an
unbiased estimator of the variance of the infinite population. ddof=0 provides a maximum likelihood estimate
of the variance for normally distributed variables. The standard deviation computed in this function is the
square root of the estimated variance, so even with ddof=1, it will not be an unbiased estimate of the standard
deviation per se.
Note that, for complex numbers, std takes the absolute value before squaring, so that the result is always real
and nonnegative.
For floating-point input, the std is computed using the same precision the input has. Depending on the input
data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a higher-
accuracy accumulator using the dtype keyword can alleviate this issue.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> np.std(a)
1.1180339887498949
>>> np.std(a, axis=0)
array([ 1., 1.])
>>> np.std(a, axis=1)
array([ 0.5, 0.5])

In single precision, std() can be inaccurate:

>>> a = np.zeros((2,512*512), dtype=np.float32)

>>> a[0,:] = 1.0
>>> a[1,:] = 0.1
>>> np.std(a)
0.45172946707416706

Computing the standard deviation in float64 is more accurate:

232 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> np.std(a, dtype=np.float64)

0.44999999925552653

sum(axis=None, dtype=None, out=None)

Return the sum of the array elements over the given axis. Masked elements are set to 0 internally.
Parameters
axis : {None, -1, int}, optional
Axis along which the sum is computed. The default (axis = None) is to compute over
the flattened array.
dtype : {None, dtype}, optional
Determines the type of the returned array and of the accumulator where the elements
are summed. If dtype has the value None and the type of a is an integer type of precision
less than the default platform integer, then the default platform integer precision is used.
Otherwise, the dtype is the same as that of a.
out : {None, ndarray}, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
sum_along_axis : MaskedArray or scalar
An array with the same shape as self, with the specified axis removed. If self is a 0-d
array, or if axis is None, a scalar is returned. If an output array is specified, a reference
to out is returned.

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.sum()
25
>>> print x.sum(axis=1)
[4 5 16]
>>> print x.sum(axis=0)
[8 5 12]
>>> print type(x.sum(axis=0, dtype=np.int64)[0])
<type numpy.int64>

trace(offset=0, axis1=0, axis2=1, dtype=None, out=None)

Return the sum along diagonals of the array.
Refer to numpy.trace for full documentation.
See Also:

numpy.trace
equivalent function

var(axis=None, dtype=None, out=None, ddof=0)

Compute the variance along the specified axis.

1.6. Masked arrays 233

NumPy Reference, Release 2.0.0.dev8464

Returns the variance of the array elements, a measure of the spread of a distribution. The variance is computed
for the flattened array by default, otherwise over the specified axis.
Parameters
a : array_like
Array containing numbers whose variance is desired. If a is not an array, a conversion
is attempted.
axis : int, optional
Axis along which the variance is computed. The default is to compute the variance of
the flattened array.
dtype : dtype, optional
Type to use in computing the variance. For arrays of integer type the default is float32;
for arrays of float types it is the same as the array type.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type is cast if necessary.
ddof : int, optional
Delta Degrees of Freedom: the divisor used in calculation is N - ddof, where N
represents the number of elements. By default ddof is zero.
Returns
variance : ndarray, see dtype parameter above
If out=None, returns a new array containing the variance; otherwise a reference to the
output array is returned.
See Also:

std
Standard deviation
mean
Average
numpy.doc.ufuncs
Section Output arguments

Notes
The variance is the average of the squared deviations from the mean, i.e., var = mean(abs(x -
x.mean())**2).
The mean is normally calculated as x.sum() / N, where N = len(x). If, however, ddof is specified, the
divisor N - ddof is used instead. In standard statistical practice, ddof=1 provides an unbiased estimator of
the variance of the infinite population. ddof=0 provides a maximum likelihood estimate of the variance for
normally distributed variables.
Note that for complex numbers, the absolute value is taken before squaring, so that the result is always real and
nonnegative.
For floating-point input, the variance is computed using the same precision the input has. Depending on the
input data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a
higher-accuracy accumulator using the dtype keyword can alleviate this issue.

234 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.array([[1,2],[3,4]])
>>> np.var(a)
1.25
>>> np.var(a,0)
array([ 1., 1.])
>>> np.var(a,1)
array([ 0.25, 0.25])

In single precision, var() can be inaccurate:

>>> a = np.zeros((2,512*512), dtype=np.float32)

>>> a[0,:] = 1.0
>>> a[1,:] = 0.1
>>> np.var(a)
0.20405951142311096

Computing the standard deviation in float64 is more accurate:

>>> np.var(a, dtype=np.float64)

0.20249999932997387
>>> ((1-0.55)**2 + (0.1-0.55)**2)/2
0.20250000000000001

Arithmetic and comparison operations

Comparison operators:
MaskedArray.__lt__ x.__lt__(y) <==> x<y
MaskedArray.__le__ x.__le__(y) <==> x<=y
MaskedArray.__gt__ x.__gt__(y) <==> x>y
MaskedArray.__ge__ x.__ge__(y) <==> x>=y
MaskedArray.__eq__(other) Check whether other equals self elementwise
MaskedArray.__ne__(other) Check whether other doesnt equal self elementwise
__lt__()
x.__lt__(y) <==> x<y
__le__()
x.__le__(y) <==> x<=y
__gt__()
x.__gt__(y) <==> x>y
__ge__()
x.__ge__(y) <==> x>=y
__eq__(other)
Check whether other equals self elementwise
__ne__(other)
Check whether other doesnt equal self elementwise

Truth value of an array (bool):

MaskedArray.__nonzero__ x.__nonzero__() <==> x != 0

1.6. Masked arrays 235

NumPy Reference, Release 2.0.0.dev8464

__nonzero__()
x.__nonzero__() <==> x != 0

Arithmetic:
MaskedArray.__abs__(x)
MaskedArray.__add__(other) Add other to self, and return a new masked array.
MaskedArray.__radd__(other) Add other to self, and return a new masked array.
MaskedArray.__sub__(other) Subtract other to self, and return a new masked array.
MaskedArray.__rsub__(other) Subtract other to self, and return a new masked array.
MaskedArray.__mul__(other) Multiply other by self, and return a new masked array.
MaskedArray.__rmul__(other) Multiply other by self, and return a new masked array.
MaskedArray.__div__(other) Divide other into self, and return a new masked array.
MaskedArray.__rdiv__ x.__rdiv__(y) <==> y/x
MaskedArray.__truediv__(other) Divide other into self, and return a new masked array.
MaskedArray.__rtruediv__(other) Divide other into self, and return a new masked array.
MaskedArray.__floordiv__(other) Divide other into self, and return a new masked array.
MaskedArray.__rfloordiv__(other) Divide other into self, and return a new masked array.
MaskedArray.__mod__ x.__mod__(y) <==> x%y
MaskedArray.__rmod__ x.__rmod__(y) <==> y%x
MaskedArray.__divmod__(x, y)
MaskedArray.__rdivmod__(y, x)
MaskedArray.__pow__(other) Raise self to the power other, masking the potential NaNs/Infs
MaskedArray.__rpow__(other) Raise self to the power other, masking the potential NaNs/Infs
MaskedArray.__lshift__ x.__lshift__(y) <==> x<<y
MaskedArray.__rlshift__ x.__rlshift__(y) <==> y<<x
MaskedArray.__rshift__ x.__rshift__(y) <==> x>>y
MaskedArray.__rrshift__ x.__rrshift__(y) <==> y>>x
MaskedArray.__and__ x.__and__(y) <==> x&y
MaskedArray.__rand__ x.__rand__(y) <==> y&x
MaskedArray.__or__ x.__or__(y) <==> x|y
MaskedArray.__ror__ x.__ror__(y) <==> y|x
MaskedArray.__xor__ x.__xor__(y) <==> x^y
MaskedArray.__rxor__ x.__rxor__(y) <==> y^x
__abs__() <==> abs(x)

__add__(other)
Add other to self, and return a new masked array.
__radd__(other)
Add other to self, and return a new masked array.
__sub__(other)
Subtract other to self, and return a new masked array.
__rsub__(other)
Subtract other to self, and return a new masked array.
__mul__(other)
Multiply other by self, and return a new masked array.
__rmul__(other)
Multiply other by self, and return a new masked array.
__div__(other)
Divide other into self, and return a new masked array.

236 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

__rdiv__()
x.__rdiv__(y) <==> y/x
__truediv__(other)
Divide other into self, and return a new masked array.
__rtruediv__(other)
Divide other into self, and return a new masked array.
__floordiv__(other)
Divide other into self, and return a new masked array.
__rfloordiv__(other)
Divide other into self, and return a new masked array.
__mod__()
x.__mod__(y) <==> x%y
__rmod__()
x.__rmod__(y) <==> y%x
__divmod__(y) <==> divmod(x, y)

__rdivmod__(y) <==> divmod(y, x)

__pow__(other)
Raise self to the power other, masking the potential NaNs/Infs
__rpow__(other)
Raise self to the power other, masking the potential NaNs/Infs
__lshift__()
x.__lshift__(y) <==> x<<y
__rlshift__()
x.__rlshift__(y) <==> y<<x
__rshift__()
x.__rshift__(y) <==> x>>y
__rrshift__()
x.__rrshift__(y) <==> y>>x
__and__()
x.__and__(y) <==> x&y
__rand__()
x.__rand__(y) <==> y&x
__or__()
x.__or__(y) <==> x|y
__ror__()
x.__ror__(y) <==> y|x
__xor__()
x.__xor__(y) <==> x^y
__rxor__()
x.__rxor__(y) <==> y^x

1.6. Masked arrays 237

NumPy Reference, Release 2.0.0.dev8464

Arithmetic, in-place:
MaskedArray.__iadd__(other) Add other to self in-place.
MaskedArray.__isub__(other) Subtract other from self in-place.
MaskedArray.__imul__(other) Multiply self by other in-place.
MaskedArray.__idiv__(other) Divide self by other in-place.
MaskedArray.__itruediv__(other) True divide self by other in-place.
MaskedArray.__ifloordiv__(other) Floor divide self by other in-place.
MaskedArray.__imod__ x.__imod__(y) <==> x%y
MaskedArray.__ipow__(other) Raise self to the power other, in place.
MaskedArray.__ilshift__ x.__ilshift__(y) <==> x<<y
MaskedArray.__irshift__ x.__irshift__(y) <==> x>>y
MaskedArray.__iand__ x.__iand__(y) <==> x&y
MaskedArray.__ior__ x.__ior__(y) <==> x|y
MaskedArray.__ixor__ x.__ixor__(y) <==> x^y
__iadd__(other)
Add other to self in-place.
__isub__(other)
Subtract other from self in-place.
__imul__(other)
Multiply self by other in-place.
__idiv__(other)
Divide self by other in-place.
__itruediv__(other)
True divide self by other in-place.
__ifloordiv__(other)
Floor divide self by other in-place.
__imod__()
x.__imod__(y) <==> x%y
__ipow__(other)
Raise self to the power other, in place.
__ilshift__()
x.__ilshift__(y) <==> x<<y
__irshift__()
x.__irshift__(y) <==> x>>y
__iand__()
x.__iand__(y) <==> x&y
__ior__()
x.__ior__(y) <==> x|y
__ixor__()
x.__ixor__(y) <==> x^y

238 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Representation

MaskedArray.__repr__() Literal string representation.

MaskedArray.__str__() String representation.
MaskedArray.ids() Return the addresses of the data and mask areas.
MaskedArray.iscontiguous() Return a boolean indicating whether the data is contiguous.
__repr__()
Literal string representation.
__str__()
String representation.
ids()
Return the addresses of the data and mask areas.
Parameters
None :

Examples

>>> x = np.ma.array([1, 2, 3], mask=[0, 1, 1])

>>> x.ids()
(166670640, 166659832)

If the array has no mask, the address of nomask is returned. This address is typically not close to the data in
memory:

>>> x = np.ma.array([1, 2, 3])

>>> x.ids()
(166691080, 3083169284L)

iscontiguous()
Return a boolean indicating whether the data is contiguous.
Parameters
None :

Examples

>>> x = np.ma.array([1, 2, 3])

>>> x.iscontiguous()
True

iscontiguous returns one of the flags of the masked array:

>>> x.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : True
OWNDATA : False
WRITEABLE : True
ALIGNED : True
UPDATEIFCOPY : False

Special methods

For standard library functions:

1.6. Masked arrays 239

NumPy Reference, Release 2.0.0.dev8464

MaskedArray.__copy__() Return a copy of the array.

MaskedArray.__deepcopy__([memo])
MaskedArray.__getstate__() Return the internal state of the masked array, for pickling
MaskedArray.__reduce__() Return a 3-tuple for pickling a MaskedArray.
MaskedArray.__setstate__(state) Restore the internal state of the masked array, for pickling purposes.
__copy__([order])
Return a copy of the array.
Parameters
order : {C, F, A}, optional
If order is C (False) then the result is contiguous (default). If order is Fortran (True)
then the result has fortran order. If order is Any (None) then the result has fortran
order only if the array already is in fortran order.
__deepcopy__(memo=None)

__getstate__()
Return the internal state of the masked array, for pickling purposes.
__reduce__()
Return a 3-tuple for pickling a MaskedArray.
__setstate__(state)
Restore the internal state of the masked array, for pickling purposes. state is typically the output of the
__getstate__ output, and is a 5-tuple:
class name
a tuple giving the shape of the data
a typecode for the data
a binary string for the data
a binary string for the mask.
Basic customization:
MaskedArray.__new__
MaskedArray.__array__ a.__array__(|dtype) -> reference if type unchanged, copy
otherwise.
MaskedArray.__array_wrap__(obj[, Special hook for ufuncs.
context])
__array__()
a.__array__(|dtype) -> reference if type unchanged, copy otherwise.
Returns either a new reference to self if dtype is not given or a new array of provided data type if dtype is
different from the current dtype of the array.
__array_wrap__(obj, context=None)
Special hook for ufuncs. Wraps the numpy array and sets the mask according to context.
Container customization: (see Indexing)

240 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

MaskedArray.__len__(x)
MaskedArray.__getitem__(indx) x.__getitem__(y) <==> x[y]
MaskedArray.__setitem__(indx, value) x.__setitem__(i, y) <==> x[i]=y
MaskedArray.__delitem__ x.__delitem__(y) <==> del x[y]
MaskedArray.__getslice__(i, j) x.__getslice__(i, j) <==> x[i:j]
MaskedArray.__setslice__(i, j, value) x.__setslice__(i, j, value) <==> x[i:j]=value
MaskedArray.__contains__ x.__contains__(y) <==> y in x
__len__() <==> len(x)

__getitem__(indx)
x.__getitem__(y) <==> x[y]
Return the item described by i, as a masked array.
__setitem__(indx, value)
x.__setitem__(i, y) <==> x[i]=y
Set item described by index. If value is masked, masks those locations.
__delitem__()
x.__delitem__(y) <==> del x[y]
__getslice__(i, j)
x.__getslice__(i, j) <==> x[i:j]
Return the slice described by (i, j). The use of negative indices is not supported.
__setslice__(i, j, value)
x.__setslice__(i, j, value) <==> x[i:j]=value
Set the slice (i,j) of a to value. If value is masked, mask those locations.
__contains__()
x.__contains__(y) <==> y in x

Specific methods

Handling the mask

The following methods can be used to access information about the mask or to manipulate the mask.
MaskedArray.__setmask__(mask[, copy]) Set the mask.
MaskedArray.harden_mask() Force the mask to hard.
MaskedArray.soften_mask() Force the mask to soft.
MaskedArray.unshare_mask() Copy the mask and set the sharedmask flag to False.
MaskedArray.shrink_mask() Reduce a mask to nomask when possible.
__setmask__(mask, copy=False)
Set the mask.
harden_mask()
Force the mask to hard.
Whether the mask of a masked array is hard or soft is determined by its hardmask property. harden_mask sets
hardmask to True.
See Also:
hardmask

1.6. Masked arrays 241

NumPy Reference, Release 2.0.0.dev8464

soften_mask()
Force the mask to soft.
Whether the mask of a masked array is hard or soft is determined by its hardmask property. soften_mask sets
hardmask to False.
See Also:
hardmask
unshare_mask()
Copy the mask and set the sharedmask flag to False.
Whether the mask is shared between masked arrays can be seen from the sharedmask property. unshare_mask
ensures the mask is not shared. A copy of the mask is only made if it was shared.
See Also:
sharedmask
shrink_mask()
Reduce a mask to nomask when possible.
Parameters
None :
Returns
None :

Examples

>>> x = np.ma.array([[1,2 ], [3, 4]], mask=[0]*4)

>>> x.mask
array([[False, False],
[False, False]], dtype=bool)
>>> x.shrink_mask()
>>> x.mask
False

Handling the fill_value

MaskedArray.get_fill_value() Return the filling value of the masked array.
MaskedArray.set_fill_value([value]) Set the filling value of the masked array.
get_fill_value()
Return the filling value of the masked array.
Returns
fill_value : scalar
The filling value.

Examples

>>> for dt in [np.int32, np.int64, np.float64, np.complex128]:

... np.ma.array([0, 1], dtype=dt).get_fill_value()
...
999999
999999
1e+20
(1e+20+0j)

242 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> x = np.ma.array([0, 1.], fill_value=-np.inf)

>>> x.get_fill_value()
-inf

set_fill_value(value=None)
Set the filling value of the masked array.
Parameters
value : scalar, optional
The new filling value. Default is None, in which case a default based on the data type
is used.
See Also:

ma.set_fill_value
Equivalent function.

Examples

>>> x = np.ma.array([0, 1.], fill_value=-np.inf)

>>> x.fill_value
-inf
>>> x.set_fill_value(np.pi)
>>> x.fill_value
3.1415926535897931

Reset to default:

>>> x.set_fill_value()
>>> x.fill_value
1e+20

Counting the missing elements

MaskedArray.count([axis]) Count the non-masked elements of the array along the given axis.
count(axis=None)
Count the non-masked elements of the array along the given axis.
Parameters
axis : int, optional
Axis along which to count the non-masked elements. If axis is None, all non-masked
elements are counted.
Returns
result : int or ndarray
If axis is None, an integer count is returned. When axis is not None, an array with shape
determined by the lengths of the remaining axes, is returned.
See Also:

count_masked
Count masked elements in array or along a given axis.

1.6. Masked arrays 243

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> import numpy.ma as ma

>>> a = ma.arange(6).reshape((2, 3))
>>> a[1, :] = ma.masked
>>> a
masked_array(data =
[[0 1 2]
[-- -- --]],
mask =
[[False False False]
[ True True True]],
fill_value = 999999)
>>> a.count()
3

When the axis keyword is specified an array of appropriate size is returned.

>>> a.count(axis=0)
array([1, 1, 1])
>>> a.count(axis=1)
array([3, 0])

1.6.7 Masked array operations

Constants

ma.MaskType Numpys Boolean type. Character code: ?. Alias: bool8

MaskType
alias of bool_

Creation

From existing data

ma.masked_array An array class with possibly masked values.
ma.array(data[, dtype, copy, order, mask, ...]) An array class with possibly masked values.
ma.copy copy
ma.frombuffer(buffer[, dtype, count, offset]) Interpret a buffer as a 1-dimensional array.
ma.fromfunction(function, shape, **kwargs) Construct an array by executing a function over each coordinate.
ma.MaskedArray.copy([order]) Return a copy of the array.
masked_array
alias of MaskedArray
array(data, dtype=None, copy=False, order=False, mask=False, fill_value=None, keep_mask=True,
hard_mask=False, shrink=True, subok=True, ndmin=0)
An array class with possibly masked values.
Masked values of True exclude the corresponding element from any computation.
Construction:

244 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

x = MaskedArray(data, mask=nomask, dtype=None, copy=True,

fill_value=None, keep_mask=True, hard_mask=False,
shrink=True)

Parameters
data : array_like
Input data.
mask : sequence, optional
Mask. Must be convertible to an array of booleans with the same shape as data. True
indicates a masked (i.e. invalid) data.
dtype : dtype, optional
Data type of the output. If dtype is None, the type of the data argument (data.dtype)
is used. If dtype is not None and different from data.dtype, a copy is performed.
copy : bool, optional
Whether to copy the input data (True), or to use a reference instead. Default is False.
subok : bool, optional
Whether to return a subclass of MaskedArray if possible (True) or a plain MaskedArray.
Default is True.
ndmin : int, optional
Minimum number of dimensions. Default is 0.
fill_value : scalar, optional
Value used to fill in the masked values when necessary. If None, a default based on the
data-type is used.
keep_mask : bool, optional
Whether to combine mask with the mask of the input data, if any (True), or to use only
mask for the output (False). Default is True.
hard_mask : bool, optional
Whether to use a hard mask or not. With a hard mask, masked values cannot be un-
masked. Default is False.
shrink : bool, optional
Whether to force compression of an empty mask. Default is True.

copy
copy a.copy(order=C)
Return a copy of the array.

Parameters
order : {C, F, A}, optional
By default, the result is stored in C-contiguous (row-major) order in memory. If order is
F, the result has Fortran (column-major) order. If order is A (Any), then the result
has the same order as the input.

1.6. Masked arrays 245

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.array([[1,2,3],[4,5,6]], order=F)

>>> y = x.copy()

>>> x.fill(0)

>>> x
array([[0, 0, 0],
[0, 0, 0]])

>>> y
array([[1, 2, 3],
[4, 5, 6]])

>>> y.flags[C_CONTIGUOUS]
True

frombuffer
Interpret a buffer as a 1-dimensional array.
Parameters
buffer :
An object that exposes the buffer interface.
dtype : data-type, optional
Data type of the returned array.
count : int, optional
Number of items to read. -1 means all data in the buffer.
offset : int, optional
Start reading the buffer from this offset.

Notes
If the buffer has data that is not in machine byte-order, this should be specified as part of the data-type, e.g.:

>>> dt = np.dtype(int)
>>> dt = dt.newbyteorder(>)
>>> np.frombuffer(buf, dtype=dt)

The data of the resulting array will not be byteswapped, but will be interpreted correctly.

Examples

>>> s = hello world

>>> np.frombuffer(s, dtype=S1, count=5, offset=6)
array([w, o, r, l, d],
dtype=|S1)

246 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

fromfunction
Construct an array by executing a function over each coordinate.
The resulting array therefore has a value fn(x, y, z) at coordinate (x, y, z).
Parameters
function : callable
The function is called with N parameters, each of which represents the coordinates of
the array varying along a specific axis. For example, if shape were (2, 2), then the
parameters would be two arrays, [[0, 0], [1, 1]] and [[0, 1], [0, 1]].
function must be capable of operating on arrays, and should return a scalar value.
shape : (N,) tuple of ints
Shape of the output array, which also determines the shape of the coordinate arrays
passed to function.
dtype : data-type, optional
Data-type of the coordinate arrays passed to function. By default, dtype is float.
Returns
out : any
The result of the call to function is passed back directly. Therefore the type and shape
of out is completely determined by function.
See Also:
indices, meshgrid

Notes
Keywords other than shape and dtype are passed to function.

Examples

>>> np.fromfunction(lambda i, j: i == j, (3, 3), dtype=int)

array([[ True, False, False],
[False, True, False],
[False, False, True]], dtype=bool)

>>> np.fromfunction(lambda i, j: i + j, (3, 3), dtype=int)

array([[0, 1, 2],
[1, 2, 3],
[2, 3, 4]])

copy(order=C)
Return a copy of the array.
Parameters
order : {C, F, A}, optional
By default, the result is stored in C-contiguous (row-major) order in memory. If order is
F, the result has Fortran (column-major) order. If order is A (Any), then the result
has the same order as the input.

1.6. Masked arrays 247

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.array([[1,2,3],[4,5,6]], order=F)

>>> y = x.copy()

>>> x.fill(0)

>>> x
array([[0, 0, 0],
[0, 0, 0]])

>>> y
array([[1, 2, 3],
[4, 5, 6]])

>>> y.flags[C_CONTIGUOUS]
True

Ones and zeros

ma.empty(shape[, dtype, order]) Return a new array of given shape and type, without initializing entries.
ma.empty_like(a) Return a new array with the same shape and type as a given array.
ma.masked_all(shape[, dtype]) Empty masked array with all elements masked.
ma.masked_all_like(arr) Empty masked array with the properties of an existing array.
ma.ones(shape[, dtype, order]) Return a new array of given shape and type, filled with ones.
ma.zeros(shape[, dtype, order]) Return a new array of given shape and type, filled with zeros.
empty
Return a new array of given shape and type, without initializing entries.
Parameters
shape : int or tuple of int
Shape of the empty array
dtype : data-type, optional
Desired output data-type.
order : {C, F}, optional
Whether to store multi-dimensional data in C (row-major) or Fortran (column-major)
order in memory.
See Also:
empty_like, zeros, ones

Notes
empty, unlike zeros, does not set the array values to zero, and may therefore be marginally faster. On the other
hand, it requires the user to manually set all the values in the array, and should be used with caution.

Examples

248 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> np.empty([2, 2])

array([[ -9.74499359e+001, 6.69583040e-309],
[ 2.13182611e-314, 3.06959433e-309]]) #random

>>> np.empty([2, 2], dtype=int)

array([[-1073741821, -1067949133],
[ 496041986, 19249760]]) #random

empty_like
Return a new array with the same shape and type as a given array.
Parameters
a : array_like
The shape and data-type of a define the parameters of the returned array.
Returns
out : ndarray
Array of random data with the same shape and type as a.
See Also:

ones_like
Return an array of ones with shape and type of input.
zeros_like
Return an array of zeros with shape and type of input.
empty
Return a new uninitialized array.
ones
Return a new array setting values to one.
zeros
Return a new array setting values to zero.

Notes
This function does not initialize the returned array; to do that use zeros_like or ones_like instead. It may be
marginally faster than the functions that do set the array values.

Examples

>>> a = ([1,2,3], [4,5,6]) # a is array-like

>>> np.empty_like(a)
array([[-1073741821, -1073741821, 3], #random
[ 0, 0, -1073741821]])
>>> a = np.array([[1., 2., 3.],[4.,5.,6.]])
>>> np.empty_like(a)
array([[ -2.00000715e+000, 1.48219694e-323, -2.00000572e+000], #random
[ 4.38791518e-305, -2.00000715e+000, 4.17269252e-309]])

masked_all(shape, dtype=<type float>)

Empty masked array with all elements masked.
Return an empty masked array of the given shape and dtype, where all the data are masked.

1.6. Masked arrays 249

NumPy Reference, Release 2.0.0.dev8464

Parameters
shape : tuple
Shape of the required MaskedArray.
dtype : dtype, optional
Data type of the output.
Returns
a : MaskedArray
A masked array with all data masked.
See Also:

masked_all_like
Empty masked array modelled on an existing array.

Examples

>>> import numpy.ma as ma

>>> ma.masked_all((3, 3))
masked_array(data =
[[-- -- --]
[-- -- --]
[-- -- --]],
mask =
[[ True True True]
[ True True True]
[ True True True]],
fill_value=1e+20)

The dtype parameter defines the underlying data type.

>>> a = ma.masked_all((3, 3))

>>> a.dtype
dtype(float64)
>>> a = ma.masked_all((3, 3), dtype=np.int32)
>>> a.dtype
dtype(int32)

masked_all_like(arr)
Empty masked array with the properties of an existing array.
Return an empty masked array of the same shape and dtype as the array arr, where all the data are masked.
Parameters
arr : ndarray
An array describing the shape and dtype of the required MaskedArray.
Returns
a : MaskedArray
A masked array with all data masked.
Raises
AttributeError :
If arr doesnt have a shape attribute (i.e. not an ndarray)

250 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

See Also:

masked_all
Empty masked array with all elements masked.

Examples

>>> import numpy.ma as ma

>>> arr = np.zeros((2, 3), dtype=np.float32)
>>> arr
array([[ 0., 0., 0.],
[ 0., 0., 0.]], dtype=float32)
>>> ma.masked_all_like(arr)
masked_array(data =
[[-- -- --]
[-- -- --]],
mask =
[[ True True True]
[ True True True]],
fill_value=1e+20)

The dtype of the masked array matches the dtype of arr.

>>> arr.dtype
dtype(float32)
>>> ma.masked_all_like(arr).dtype
dtype(float32)

ones
Return a new array of given shape and type, filled with ones.
Please refer to the documentation for zeros.
See Also:
zeros

Examples

>>> np.ones(5)
array([ 1., 1., 1., 1., 1.])

>>> np.ones((5,), dtype=np.int)

array([1, 1, 1, 1, 1])

>>> np.ones((2, 1))

array([[ 1.],
[ 1.]])

>>> s = (2,2)
>>> np.ones(s)
array([[ 1., 1.],
[ 1., 1.]])

zeros
Return a new array of given shape and type, filled with zeros.

1.6. Masked arrays 251

NumPy Reference, Release 2.0.0.dev8464

Parameters
shape : int or sequence of ints
Shape of the new array, e.g., (2, 3) or 2.
dtype : data-type, optional
The desired data-type for the array, e.g., numpy.int8. Default is numpy.float64.
order : {C, F}, optional
Whether to store multidimensional data in C- or Fortran-contiguous (row- or column-
wise) order in memory.
Returns
out : ndarray
Array of zeros with the given shape, dtype, and order.
See Also:

zeros_like
Return an array of zeros with shape and type of input.
ones_like
Return an array of ones with shape and type of input.
empty_like
Return an empty array with shape and type of input.
ones
Return a new array setting values to one.
empty
Return a new uninitialized array.

Examples

>>> np.zeros(5)
array([ 0., 0., 0., 0., 0.])

>>> np.zeros((5,), dtype=numpy.int)

array([0, 0, 0, 0, 0])

>>> np.zeros((2, 1))

array([[ 0.],
[ 0.]])

>>> s = (2,2)
>>> np.zeros(s)
array([[ 0., 0.],
[ 0., 0.]])

>>> np.zeros((2,), dtype=[(x, i4), (y, i4)]) # custom dtype

array([(0, 0), (0, 0)],
dtype=[(x, <i4), (y, <i4)])

252 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Inspecting the array

ma.all(self[, axis, out]) Check if all of the elements of a are true.

ma.any(self[, axis, out]) Check if any of the elements of a are true.
ma.count(a[, axis]) Count the non-masked elements of the array along the given axis.
ma.count_masked(arr[, axis]) Count the number of masked elements along the given axis.
ma.getmask(a) Return the mask of a masked array, or nomask.
ma.getmaskarray(arr) Return the mask of a masked array, or full boolean array of False.
ma.getdata(a[, subok]) Return the data of a masked array as an ndarray.
ma.nonzero(self) Return the indices of unmasked elements that are not zero.
ma.shape(obj) Return the shape of an array.
ma.size(obj[, axis]) Return the number of elements along a given axis.
ma.MaskedArray.data
ma.MaskedArray.mask Mask
ma.MaskedArray.recordmask
ma.MaskedArray.all([axis, out]) Check if all of the elements of a are true.
ma.MaskedArray.any([axis, out]) Check if any of the elements of a are true.
ma.MaskedArray.count([axis]) Count the non-masked elements of the array along the given axis.
ma.MaskedArray.nonzero() Return the indices of unmasked elements that are not zero.
ma.shape(obj) Return the shape of an array.
ma.size(obj[, axis]) Return the number of elements along a given axis.
all
Check if all of the elements of a are true.
Performs a logical_and over the given axis and returns the result. Masked values are considered as True
during computation. For convenience, the output array is masked where ALL the values along the current axis
are masked: if the output would have been a scalar and that all the values are masked, then the output is masked.
Parameters
axis : {None, integer}
Axis to perform the operation over. If None, perform over flattened array.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
See Also:

all
equivalent function

Examples

>>> np.ma.array([1,2,3]).all()
True
>>> a = np.ma.array([1,2,3], mask=True)
>>> (a.all() is np.ma.masked)
True

any
Check if any of the elements of a are true.
Performs a logical_or over the given axis and returns the result. Masked values are considered as False during
computation.

1.6. Masked arrays 253

NumPy Reference, Release 2.0.0.dev8464

Parameters
axis : {None, integer}
Axis to perform the operation over. If None, perform over flattened array and return a
scalar.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
See Also:

any
equivalent function

count(a, axis=None)
Count the non-masked elements of the array along the given axis.
Parameters
axis : int, optional
Axis along which to count the non-masked elements. If axis is None, all non-masked
elements are counted.
Returns
result : int or ndarray
If axis is None, an integer count is returned. When axis is not None, an array with shape
determined by the lengths of the remaining axes, is returned.
See Also:

count_masked
Count masked elements in array or along a given axis.

Examples

>>> import numpy.ma as ma

>>> a = ma.arange(6).reshape((2, 3))
>>> a[1, :] = ma.masked
>>> a
masked_array(data =
[[0 1 2]
[-- -- --]],
mask =
[[False False False]
[ True True True]],
fill_value = 999999)
>>> a.count()
3

When the axis keyword is specified an array of appropriate size is returned.

>>> a.count(axis=0)
array([1, 1, 1])
>>> a.count(axis=1)
array([3, 0])

254 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

count_masked(arr, axis=None)
Count the number of masked elements along the given axis.
Parameters
arr : array_like
An array with (possibly) masked elements.
axis : int, optional
Axis along which to count. If None (default), a flattened version of the array is used.
Returns
count : int, ndarray
The total number of masked elements (axis=None) or the number of masked elements
along each slice of the given axis.
See Also:

MaskedArray.count
Count non-masked elements.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(9).reshape((3,3))
>>> a = ma.array(a)
>>> a[1, 0] = ma.masked
>>> a[1, 2] = ma.masked
>>> a[2, 1] = ma.masked
>>> a
masked_array(data =
[[0 1 2]
[-- 4 --]
[6 -- 8]],
mask =
[[False False False]
[ True False True]
[False True False]],
fill_value=999999)
>>> ma.count_masked(a)
3

When the axis keyword is used an array is returned.

>>> ma.count_masked(a, axis=0)

array([1, 1, 1])
>>> ma.count_masked(a, axis=1)
array([0, 2, 1])

getmask(a)
Return the mask of a masked array, or nomask.
Return the mask of a as an ndarray if a is a MaskedArray and the mask is not nomask, else return nomask. To
guarantee a full array of booleans of the same shape as a, use getmaskarray.
Parameters
a : array_like
Input MaskedArray for which the mask is required.

1.6. Masked arrays 255

NumPy Reference, Release 2.0.0.dev8464

See Also:

getdata
Return the data of a masked array as an ndarray.
getmaskarray
Return the mask of a masked array, or full array of False.

Examples

>>> import numpy.ma as ma

>>> a = ma.masked_equal([[1,2],[3,4]], 2)
>>> a
masked_array(data =
[[1 --]
[3 4]],
mask =
[[False True]
[False False]],
fill_value=999999)
>>> ma.getmask(a)
array([[False, True],
[False, False]], dtype=bool)

Equivalently use the MaskedArray mask attribute.

>>> a.mask
array([[False, True],
[False, False]], dtype=bool)

Result when mask == nomask

>>> b = ma.masked_array([[1,2],[3,4]])
>>> b
masked_array(data =
[[1 2]
[3 4]],
mask =
False,
fill_value=999999)
>>> ma.nomask
False
>>> ma.getmask(b) == ma.nomask
True
>>> b.mask == ma.nomask
True

getmaskarray(arr)
Return the mask of a masked array, or full boolean array of False.
Return the mask of arr as an ndarray if arr is a MaskedArray and the mask is not nomask, else return a full
boolean array of False of the same shape as arr.
Parameters
arr : array_like
Input MaskedArray for which the mask is required.
See Also:

256 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

getmask
Return the mask of a masked array, or nomask.
getdata
Return the data of a masked array as an ndarray.

Examples

>>> import numpy.ma as ma

>>> a = ma.masked_equal([[1,2],[3,4]], 2)
>>> a
masked_array(data =
[[1 --]
[3 4]],
mask =
[[False True]
[False False]],
fill_value=999999)
>>> ma.getmaskarray(a)
array([[False, True],
[False, False]], dtype=bool)

Result when mask == nomask

>>> b = ma.masked_array([[1,2],[3,4]])
>>> b
masked_array(data =
[[1 2]
[3 4]],
mask =
False,
fill_value=999999)
>>> >ma.getmaskarray(b)
array([[False, False],
[False, False]], dtype=bool)

getdata(a, subok=True)
Return the data of a masked array as an ndarray.
Return the data of a (if any) as an ndarray if a is a MaskedArray, else return a as a ndarray or subclass
(depending on subok) if not.
Parameters
a : array_like
Input MaskedArray, alternatively a ndarray or a subclass thereof.
subok : bool
Whether to force the output to be a pure ndarray (False) or to return a subclass of ndarray
if appropriate (True, default).
See Also:

getmask
Return the mask of a masked array, or nomask.
getmaskarray
Return the mask of a masked array, or full array of False.

1.6. Masked arrays 257

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> import numpy.ma as ma

>>> a = ma.masked_equal([[1,2],[3,4]], 2)
>>> a
masked_array(data =
[[1 --]
[3 4]],
mask =
[[False True]
[False False]],
fill_value=999999)
>>> ma.getdata(a)
array([[1, 2],
[3, 4]])

Equivalently use the MaskedArray data attribute.

>>> a.data
array([[1, 2],
[3, 4]])

nonzero
Return the indices of unmasked elements that are not zero.
Returns a tuple of arrays, one for each dimension, containing the indices of the non-zero elements in that
dimension. The corresponding non-zero values can be obtained with:

a[a.nonzero()]

To group the indices by element, rather than dimension, use instead:

np.transpose(a.nonzero())

The result of this is always a 2d array, with a row for each non-zero element.
Parameters
None :
Returns
tuple_of_arrays : tuple
Indices of elements that are non-zero.
See Also:

numpy.nonzero
Function operating on ndarrays.
flatnonzero
Return indices that are non-zero in the flattened version of the input array.
ndarray.nonzero
Equivalent ndarray method.

Examples

258 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> import numpy.ma as ma

>>> x = ma.array(np.eye(3))
>>> x
masked_array(data =
[[ 1. 0. 0.]
[ 0. 1. 0.]
[ 0. 0. 1.]],
mask =
False,
fill_value=1e+20)
>>> x.nonzero()
(array([0, 1, 2]), array([0, 1, 2]))

Masked elements are ignored.

>>> x[1, 1] = ma.masked

>>> x
masked_array(data =
[[1.0 0.0 0.0]
[0.0 -- 0.0]
[0.0 0.0 1.0]],
mask =
[[False False False]
[False True False]
[False False False]],
fill_value=1e+20)
>>> x.nonzero()
(array([0, 2]), array([0, 2]))

Indices can also be grouped by element.

>>> np.transpose(x.nonzero())
array([[0, 0],
[2, 2]])

A common use for nonzero is to find the indices of an array, where a condition is True. Given an array a, the
condition a > 3 is a boolean array and since False is interpreted as 0, ma.nonzero(a > 3) yields the indices of the
a where the condition is true.

>>> a = ma.array([[1,2,3],[4,5,6],[7,8,9]])
>>> a > 3
masked_array(data =
[[False False False]
[ True True True]
[ True True True]],
mask =
False,
fill_value=999999)
>>> ma.nonzero(a > 3)
(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

The nonzero method of the condition array can also be called.

>>> (a > 3).nonzero()

(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

1.6. Masked arrays 259

NumPy Reference, Release 2.0.0.dev8464

shape(obj)
Return the shape of an array.
Parameters
a : array_like
Input array.
Returns
shape : tuple of ints
The elements of the shape tuple give the lengths of the corresponding array dimensions.
See Also:
alen
ndarray.shape
Equivalent array method.

Examples

>>> np.shape(np.eye(3))
(3, 3)
>>> np.shape([[1, 2]])
(1, 2)
>>> np.shape([0])
(1,)
>>> np.shape(0)
()

>>> a = np.array([(1, 2), (3, 4)], dtype=[(x, i4), (y, i4)])

>>> np.shape(a)
(2,)
>>> a.shape
(2,)

size(obj, axis=None)
Return the number of elements along a given axis.
Parameters
a : array_like
Input data.
axis : int, optional
Axis along which the elements are counted. By default, give the total number of ele-
ments.
Returns
element_count : int
Number of elements along the specified axis.
See Also:

shape
dimensions of array
ndarray.shape
dimensions of array

260 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

ndarray.size
number of elements in array

Examples

>>> a = np.array([[1,2,3],[4,5,6]])
>>> np.size(a)
6
>>> np.size(a,1)
3
>>> np.size(a,0)
2

data

mask
Mask
recordmask

all(axis=None, out=None)
Check if all of the elements of a are true.
Performs a logical_and over the given axis and returns the result. Masked values are considered as True
during computation. For convenience, the output array is masked where ALL the values along the current axis
are masked: if the output would have been a scalar and that all the values are masked, then the output is masked.
Parameters
axis : {None, integer}
Axis to perform the operation over. If None, perform over flattened array.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
See Also:

all
equivalent function

Examples

>>> np.ma.array([1,2,3]).all()
True
>>> a = np.ma.array([1,2,3], mask=True)
>>> (a.all() is np.ma.masked)
True

any(axis=None, out=None)
Check if any of the elements of a are true.
Performs a logical_or over the given axis and returns the result. Masked values are considered as False during
computation.
Parameters
axis : {None, integer}

1.6. Masked arrays 261

NumPy Reference, Release 2.0.0.dev8464

Axis to perform the operation over. If None, perform over flattened array and return a
scalar.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
See Also:

any
equivalent function

count(axis=None)
Count the non-masked elements of the array along the given axis.
Parameters
axis : int, optional
Axis along which to count the non-masked elements. If axis is None, all non-masked
elements are counted.
Returns
result : int or ndarray
If axis is None, an integer count is returned. When axis is not None, an array with shape
determined by the lengths of the remaining axes, is returned.
See Also:

count_masked
Count masked elements in array or along a given axis.

Examples

>>> import numpy.ma as ma

>>> a = ma.arange(6).reshape((2, 3))
>>> a[1, :] = ma.masked
>>> a
masked_array(data =
[[0 1 2]
[-- -- --]],
mask =
[[False False False]
[ True True True]],
fill_value = 999999)
>>> a.count()
3

When the axis keyword is specified an array of appropriate size is returned.

>>> a.count(axis=0)
array([1, 1, 1])
>>> a.count(axis=1)
array([3, 0])

nonzero()
Return the indices of unmasked elements that are not zero.

262 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Returns a tuple of arrays, one for each dimension, containing the indices of the non-zero elements in that
dimension. The corresponding non-zero values can be obtained with:

a[a.nonzero()]

To group the indices by element, rather than dimension, use instead:

np.transpose(a.nonzero())

The result of this is always a 2d array, with a row for each non-zero element.
Parameters
None :
Returns
tuple_of_arrays : tuple
Indices of elements that are non-zero.
See Also:

numpy.nonzero
Function operating on ndarrays.
flatnonzero
Return indices that are non-zero in the flattened version of the input array.
ndarray.nonzero
Equivalent ndarray method.

Examples

>>> import numpy.ma as ma

>>> x = ma.array(np.eye(3))
>>> x
masked_array(data =
[[ 1. 0. 0.]
[ 0. 1. 0.]
[ 0. 0. 1.]],
mask =
False,
fill_value=1e+20)
>>> x.nonzero()
(array([0, 1, 2]), array([0, 1, 2]))

Masked elements are ignored.

>>> x[1, 1] = ma.masked

>>> x
masked_array(data =
[[1.0 0.0 0.0]
[0.0 -- 0.0]
[0.0 0.0 1.0]],
mask =
[[False False False]
[False True False]
[False False False]],
fill_value=1e+20)

1.6. Masked arrays 263

NumPy Reference, Release 2.0.0.dev8464

>>> x.nonzero()
(array([0, 2]), array([0, 2]))

Indices can also be grouped by element.

>>> np.transpose(x.nonzero())
array([[0, 0],
[2, 2]])

A common use for nonzero is to find the indices of an array, where a condition is True. Given an array a, the
condition a > 3 is a boolean array and since False is interpreted as 0, ma.nonzero(a > 3) yields the indices of the
a where the condition is true.

>>> a = ma.array([[1,2,3],[4,5,6],[7,8,9]])
>>> a > 3
masked_array(data =
[[False False False]
[ True True True]
[ True True True]],
mask =
False,
fill_value=999999)
>>> ma.nonzero(a > 3)
(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

The nonzero method of the condition array can also be called.

>>> (a > 3).nonzero()

(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

shape(obj)
Return the shape of an array.
Parameters
a : array_like
Input array.
Returns
shape : tuple of ints
The elements of the shape tuple give the lengths of the corresponding array dimensions.
See Also:
alen
ndarray.shape
Equivalent array method.

Examples

>>> np.shape(np.eye(3))
(3, 3)
>>> np.shape([[1, 2]])
(1, 2)
>>> np.shape([0])
(1,)

264 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> np.shape(0)
()

>>> a = np.array([(1, 2), (3, 4)], dtype=[(x, i4), (y, i4)])

>>> np.shape(a)
(2,)
>>> a.shape
(2,)

size(obj, axis=None)
Return the number of elements along a given axis.
Parameters
a : array_like
Input data.
axis : int, optional
Axis along which the elements are counted. By default, give the total number of ele-
ments.
Returns
element_count : int
Number of elements along the specified axis.
See Also:

shape
dimensions of array
ndarray.shape
dimensions of array
ndarray.size
number of elements in array

Examples

>>> a = np.array([[1,2,3],[4,5,6]])
>>> np.size(a)
6
>>> np.size(a,1)
3
>>> np.size(a,0)
2

1.6. Masked arrays 265

NumPy Reference, Release 2.0.0.dev8464

Manipulating a MaskedArray

Changing the shape

ma.ravel(self) Returns a 1D version of self, as a view.
ma.reshape(a, new_shape[, order]) Returns an array containing the same data with a new
shape.
ma.resize(x, new_shape) Return a new masked array with the specified size and
shape.
ma.MaskedArray.flatten([order]) Return a copy of the array collapsed into one dimension.
ma.MaskedArray.ravel() Returns a 1D version of self, as a view.
ma.MaskedArray.reshape(*s, **kwargs) Give a new shape to the array without changing its data.
ma.MaskedArray.resize(newshape[, refcheck,
...])
ravel
Returns a 1D version of self, as a view.
Returns
MaskedArray :
Output view is of shape (self.size,) (or
(np.ma.product(self.shape),)).

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.ravel()
[1 -- 3 -- 5 -- 7 -- 9]

reshape(a, new_shape, order=C)

Returns an array containing the same data with a new shape.
Refer to MaskedArray.reshape for full documentation.
See Also:

MaskedArray.reshape
equivalent function

resize(x, new_shape)
Return a new masked array with the specified size and shape.
This is the masked equivalent of the numpy.resize function. The new array is filled with repeated copies of
x (in the order that the data are stored in memory). If x is masked, the new array will be masked, and the new
mask will be a repetition of the old one.
See Also:

numpy.resize
Equivalent function in the top level NumPy module.

266 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> import numpy.ma as ma

>>> a = ma.array([[1, 2] ,[3, 4]])
>>> a[0, 1] = ma.masked
>>> a
masked_array(data =
[[1 --]
[3 4]],
mask =
[[False True]
[False False]],
fill_value = 999999)
>>> np.resize(a, (3, 3))
array([[1, 2, 3],
[4, 1, 2],
[3, 4, 1]])
>>> ma.resize(a, (3, 3))
masked_array(data =
[[1 -- 3]
[4 1 --]
[3 4 1]],
mask =
[[False True False]
[False False True]
[False False False]],
fill_value = 999999)

A MaskedArray is always returned, regardless of the input type.

>>> a = np.array([[1, 2] ,[3, 4]])

>>> ma.resize(a, (3, 3))
masked_array(data =
[[1 2 3]
[4 1 2]
[3 4 1]],
mask =
False,
fill_value = 999999)

flatten(order=C)
Return a copy of the array collapsed into one dimension.
Parameters
order : {C, F}, optional
Whether to flatten in C (row-major) or Fortran (column-major) order. The default is
C.
Returns
y : ndarray
A copy of the input array, flattened to one dimension.
See Also:

ravel
Return a flattened array.

1.6. Masked arrays 267

NumPy Reference, Release 2.0.0.dev8464

flat
A 1-D flat iterator over the array.

Examples

>>> a = np.array([[1,2], [3,4]])

>>> a.flatten()
array([1, 2, 3, 4])
>>> a.flatten(F)
array([1, 3, 2, 4])

ravel()
Returns a 1D version of self, as a view.
Returns
MaskedArray :
Output view is of shape (self.size,) (or
(np.ma.product(self.shape),)).

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.ravel()
[1 -- 3 -- 5 -- 7 -- 9]

reshape(*s, **kwargs)
Give a new shape to the array without changing its data.
Returns a masked array containing the same data, but with a new shape. The result is a view on the original
array; if this is not possible, a ValueError is raised.
Parameters
shape : int or tuple of ints
The new shape should be compatible with the original shape. If an integer is supplied,
then the result will be a 1-D array of that length.
order : {C, F}, optional
Determines whether the array data should be viewed as in C (row-major) or FORTRAN
(column-major) order.
Returns
reshaped_array : array
A new view on the array.
See Also:

reshape
Equivalent function in the masked array module.
numpy.ndarray.reshape
Equivalent method on ndarray object.

268 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

numpy.reshape
Equivalent function in the NumPy module.

Notes
The reshaping operation cannot guarantee that a copy will not be made, to modify the shape in place, use
a.shape = s

Examples

>>> x = np.ma.array([[1,2],[3,4]], mask=[1,0,0,1])

>>> print x
[[-- 2]
[3 --]]
>>> x = x.reshape((4,1))
>>> print x
[[--]
[2]
[3]
[--]]

resize(newshape, refcheck=True, order=False)

Warning: This method does nothing, except raise a ValueError exception. A masked array does not own
its data and therefore cannot safely be resized in place. Use the numpy.ma.resize function instead.

This method is difficult to implement safely and may be deprecated in future releases of NumPy.

Modifying axes
ma.swapaxes swapaxes
ma.transpose(a[, axes]) Permute the dimensions of an array.
ma.MaskedArray.swapaxes(axis1, axis2) Return a view of the array with axis1 and axis2 interchanged.
ma.MaskedArray.transpose(*axes) Returns a view of the array with axes transposed.
swapaxes
swapaxes a.swapaxes(axis1, axis2)
Return a view of the array with axis1 and axis2 interchanged.
Refer to numpy.swapaxes for full documentation.
See Also:

numpy.swapaxes
equivalent function

transpose(a, axes=None)
Permute the dimensions of an array.
This function is exactly equivalent to numpy.transpose.
See Also:

numpy.transpose
Equivalent function in top-level NumPy module.

1.6. Masked arrays 269

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> import numpy.ma as ma

>>> x = ma.arange(4).reshape((2,2))
>>> x[1, 1] = ma.masked
>>>> x
masked_array(data =
[[0 1]
[2 --]],
mask =
[[False False]
[False True]],
fill_value = 999999)
>>> ma.transpose(x)
masked_array(data =
[[0 2]
[1 --]],
mask =
[[False False]
[False True]],
fill_value = 999999)

swapaxes(axis1, axis2)
Return a view of the array with axis1 and axis2 interchanged.
Refer to numpy.swapaxes for full documentation.
See Also:

numpy.swapaxes
equivalent function

transpose(*axes)
Returns a view of the array with axes transposed.
For a 1-D array, this has no effect. (To change between column and row vectors, first cast the 1-D array into a ma-
trix object.) For a 2-D array, this is the usual matrix transpose. For an n-D array, if axes are given, their order in-
dicates how the axes are permuted (see Examples). If axes are not provided and a.shape = (i[0], i[1],
... i[n-2], i[n-1]), then a.transpose().shape = (i[n-1], i[n-2], ... i[1],
i[0]).
Parameters
axes : None, tuple of ints, or n ints
None or no argument: reverses the order of the axes.
tuple of ints: i in the j-th place in the tuple means as i-th axis becomes a.transpose()s j-th
axis.
n ints: same as an n-tuple of the same ints (this form is intended simply as a convenience
alternative to the tuple form)
Returns
out : ndarray
View of a, with axes suitably permuted.
See Also:

270 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

ndarray.T
Array property returning the array transposed.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> a
array([[1, 2],
[3, 4]])
>>> a.transpose()
array([[1, 3],
[2, 4]])
>>> a.transpose((1, 0))
array([[1, 3],
[2, 4]])
>>> a.transpose(1, 0)
array([[1, 3],
[2, 4]])

Changing the number of dimensions

ma.atleast_1d(*arys) Convert inputs to arrays with at least one dimension.
ma.atleast_2d(*arys) View inputs as arrays with at least two dimensions.
ma.atleast_3d(*arys) View inputs as arrays with at least three dimensions.
ma.expand_dims(x, axis) Expand the shape of an array.
ma.squeeze(a) Remove single-dimensional entries from the shape of an array.
ma.MaskedArray.squeeze() Remove single-dimensional entries from the shape of a.
ma.column_stack(tup) Stack 1-D arrays as columns into a 2-D array.
ma.concatenate(arrays[, axis]) Concatenate a sequence of arrays along the given axis.
ma.dstack(tup) Stack arrays in sequence depth wise (along third axis).
ma.hstack(tup) Stack arrays in sequence horizontally (column wise).
ma.hsplit(ary, indices_or_sections) Split an array into multiple sub-arrays horizontally (column-wise).
ma.mr_ Translate slice objects to concatenation along the first axis.
ma.row_stack(tup) Stack arrays in sequence vertically (row wise).
ma.vstack(tup) Stack arrays in sequence vertically (row wise).
atleast_1d(*arys)
Convert inputs to arrays with at least one dimension.
Scalar inputs are converted to 1-dimensional arrays, whilst higher-dimensional inputs are preserved.
Parameters
array1, array2, ... : array_like
One or more input arrays.
Returns
ret : ndarray
An array, or sequence of arrays, each with a.ndim >= 1. Copies are made only if
necessary.
See Also:
atleast_2d, atleast_3d

1.6. Masked arrays 271

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.atleast_1d(1.0)
array([ 1.])

>>> x = np.arange(9.0).reshape(3,3)
>>> np.atleast_1d(x)
array([[ 0., 1., 2.],
[ 3., 4., 5.],
[ 6., 7., 8.]])
>>> np.atleast_1d(x) is x
True

>>> np.atleast_1d(1, [3, 4])

[array([1]), array([3, 4])]

atleast_2d(*arys)
View inputs as arrays with at least two dimensions.
Parameters
array1, array2, ... : array_like
One or more array-like sequences. Non-array inputs are converted to arrays. Arrays
that already have two or more dimensions are preserved.
Returns
res, res2, ... : ndarray
An array, or tuple of arrays, each with a.ndim >= 2. Copies are avoided where
possible, and views with two or more dimensions are returned.
See Also:
atleast_1d, atleast_3d

Examples

>>> np.atleast_2d(3.0)
array([[ 3.]])

>>> x = np.arange(3.0)
>>> np.atleast_2d(x)
array([[ 0., 1., 2.]])
>>> np.atleast_2d(x).base is x
True

>>> np.atleast_2d(1, [1, 2], [[1, 2]])

[array([[1]]), array([[1, 2]]), array([[1, 2]])]

atleast_3d(*arys)
View inputs as arrays with at least three dimensions.
Parameters
array1, array2, ... : array_like
One or more array-like sequences. Non-array inputs are converted to arrays. Arrays
that already have three or more dimensions are preserved.

272 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Returns
res1, res2, ... : ndarray
An array, or tuple of arrays, each with a.ndim >= 3. Copies are avoided where
possible, and views with three or more dimensions are returned. For example, a 1-D
array of shape N becomes a view of shape (1, N, 1). A 2-D array of shape (M, N)
becomes a view of shape (M, N, 1).
See Also:
atleast_1d, atleast_2d

Examples

>>> np.atleast_3d(3.0)
array([[[ 3.]]])

>>> x = np.arange(3.0)
>>> np.atleast_3d(x).shape
(1, 3, 1)

>>> x = np.arange(12.0).reshape(4,3)
>>> np.atleast_3d(x).shape
(4, 3, 1)
>>> np.atleast_3d(x).base is x
True

>>> for arr in np.atleast_3d([1, 2], [[1, 2]], [[[1, 2]]]):

... print arr, arr.shape
...
[[[1]
[2]]] (1, 2, 1)
[[[1]
[2]]] (1, 2, 1)
[[[1 2]]] (1, 1, 2)

expand_dims(x, axis)
Expand the shape of an array.
Expands the shape of the array by including a new axis before the one specified by the axis parameter. This
function behaves the same as numpy.expand_dims but preserves masked elements.
See Also:

numpy.expand_dims
Equivalent function in top-level NumPy module.

Examples

>>> import numpy.ma as ma

>>> x = ma.array([1, 2, 4])
>>> x[1] = ma.masked
>>> x
masked_array(data = [1 -- 4],
mask = [False True False],
fill_value = 999999)
>>> np.expand_dims(x, axis=0)

1.6. Masked arrays 273

NumPy Reference, Release 2.0.0.dev8464

array([[1, 2, 4]])
>>> ma.expand_dims(x, axis=0)
masked_array(data =
[[1 -- 4]],
mask =
[[False True False]],
fill_value = 999999)

The same result can be achieved using slicing syntax with np.newaxis.

>>> x[np.newaxis, :]
masked_array(data =
[[1 -- 4]],
mask =
[[False True False]],
fill_value = 999999)

squeeze(a)
Remove single-dimensional entries from the shape of an array.
Parameters
a : array_like
Input data.
Returns
squeezed : ndarray
The input array, but with with all dimensions of length 1 removed. Whenever possible,
a view on a is returned.

Examples

>>> x = np.array([[[0], [1], [2]]])

>>> x.shape
(1, 3, 1)
>>> np.squeeze(x).shape
(3,)

squeeze()
Remove single-dimensional entries from the shape of a.
Refer to numpy.squeeze for full documentation.
See Also:

numpy.squeeze
equivalent function

column_stack

Stack 1-D arrays as columns into a 2-D array.

Take a sequence of 1-D arrays and stack them as columns to make a single 2-D array. 2-D arrays are
stacked as-is, just like with hstack. 1-D arrays are turned into 2-D columns first.

Parameters
tup : sequence of 1-D or 2-D arrays.

274 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Arrays to stack. All of them must have the same first dimension.
Returns
stacked : 2-D array
The array formed by stacking the given arrays.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.column_stack((a,b))
array([[1, 2],
[2, 3],
[3, 4]])

concatenate(arrays, axis=0)
Concatenate a sequence of arrays along the given axis.
Parameters
arrays : sequence of array_like
The arrays must have the same shape, except in the dimension corresponding to axis
(the first, by default).
axis : int, optional
The axis along which the arrays will be joined. Default is 0.
Returns
result : MaskedArray
The concatenated array with any masked entries preserved.
See Also:

numpy.concatenate
Equivalent function in the top-level NumPy module.

Examples

>>> import numpy.ma as ma

>>> a = ma.arange(3)
>>> a[1] = ma.masked
>>> b = ma.arange(2, 5)
>>> a
masked_array(data = [0 -- 2],
mask = [False True False],
fill_value = 999999)
>>> b
masked_array(data = [2 3 4],
mask = False,
fill_value = 999999)
>>> ma.concatenate([a, b])
masked_array(data = [0 -- 2 2 3 4],

1.6. Masked arrays 275

NumPy Reference, Release 2.0.0.dev8464

mask = [False True False False False False],

fill_value = 999999)

dstack

Stack arrays in sequence depth wise (along third axis).

Takes a sequence of arrays and stack them along the third axis to make a single array. Rebuilds
arrays divided by dsplit. This is a simple way to stack 2D arrays (images) into a single 3D array for
processing.

Parameters
tup : sequence of arrays
Arrays to stack. All of them must have the same shape along all but the third axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

See Also:

vstack
Stack along first axis.
hstack
Stack along second axis.
concatenate
Join arrays.
dsplit
Split array along third axis.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.dstack((a,b))
array([[[1, 2],
[2, 3],
[3, 4]]])

>>> a = np.array([[1],[2],[3]])
>>> b = np.array([[2],[3],[4]])
>>> np.dstack((a,b))
array([[[1, 2]],
[[2, 3]],
[[3, 4]]])

hstack

276 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Stack arrays in sequence horizontally (column wise).

Take a sequence of arrays and stack them horizontally to make a single array. Rebuild arrays divided
by hsplit.

Parameters
tup : sequence of ndarrays
All arrays must have the same shape along all but the second axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

See Also:

vstack
Stack arrays in sequence vertically (row wise).
dstack
Stack arrays in sequence depth wise (along third axis).
concatenate
Join a sequence of arrays together.
hsplit
Split array along second axis.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.hstack((a,b))
array([1, 2, 3, 2, 3, 4])
>>> a = np.array([[1],[2],[3]])
>>> b = np.array([[2],[3],[4]])
>>> np.hstack((a,b))
array([[1, 2],
[2, 3],
[3, 4]])

hsplit

Split an array into multiple sub-arrays horizontally (column-wise).

Please refer to the split documentation. hsplit is equivalent to split with axis=1, the array is always
split along the second axis regardless of the array dimension.
See Also:

split
Split an array into multiple sub-arrays of equal size.

1.6. Masked arrays 277

NumPy Reference, Release 2.0.0.dev8464

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> x = np.arange(16.0).reshape(4, 4)
>>> x
array([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.],
[ 12., 13., 14., 15.]])
>>> np.hsplit(x, 2)
[array([[ 0., 1.],
[ 4., 5.],
[ 8., 9.],
[ 12., 13.]]),
array([[ 2., 3.],
[ 6., 7.],
[ 10., 11.],
[ 14., 15.]])]
>>> np.hsplit(x, np.array([3, 6]))
[array([[ 0., 1., 2.],
[ 4., 5., 6.],
[ 8., 9., 10.],
[ 12., 13., 14.]]),
array([[ 3.],
[ 7.],
[ 11.],
[ 15.]]),
array([], dtype=float64)]

With a higher dimensional array the split is still along the second axis.

>>> x = np.arange(8.0).reshape(2, 2, 2)
>>> x
array([[[ 0., 1.],
[ 2., 3.]],
[[ 4., 5.],
[ 6., 7.]]])
>>> np.hsplit(x, 2)
[array([[[ 0., 1.]],
[[ 4., 5.]]]),
array([[[ 2., 3.]],
[[ 6., 7.]]])]

mr_
Translate slice objects to concatenation along the first axis.
This is the masked array version of lib.index_tricks.RClass.
See Also:
lib.index_tricks.RClass

Examples

>>> np.ma.mr_[np.ma.array([1,2,3]), 0, 0, np.ma.array([4,5,6])]

array([1, 2, 3, 0, 0, 4, 5, 6])

278 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

row_stack

Stack arrays in sequence vertically (row wise).

Take a sequence of arrays and stack them vertically to make a single array. Rebuild arrays divided
by vsplit.

Parameters
tup : sequence of ndarrays
Tuple containing arrays to be stacked. The arrays must have the same shape along all
but the first axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

See Also:

hstack
Stack arrays in sequence horizontally (column wise).
dstack
Stack arrays in sequence depth wise (along third dimension).
concatenate
Join a sequence of arrays together.
vsplit
Split array into a list of multiple sub-arrays vertically.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array([1, 2, 3])

>>> b = np.array([2, 3, 4])
>>> np.vstack((a,b))
array([[1, 2, 3],
[2, 3, 4]])

>>> a = np.array([[1], [2], [3]])

>>> b = np.array([[2], [3], [4]])
>>> np.vstack((a,b))
array([[1],
[2],
[3],
[2],
[3],
[4]])

vstack

1.6. Masked arrays 279

NumPy Reference, Release 2.0.0.dev8464

Stack arrays in sequence vertically (row wise).

Take a sequence of arrays and stack them vertically to make a single array. Rebuild arrays divided
by vsplit.

Parameters
tup : sequence of ndarrays
Tuple containing arrays to be stacked. The arrays must have the same shape along all
but the first axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

See Also:

hstack
Stack arrays in sequence horizontally (column wise).
dstack
Stack arrays in sequence depth wise (along third dimension).
concatenate
Join a sequence of arrays together.
vsplit
Split array into a list of multiple sub-arrays vertically.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array([1, 2, 3])

>>> b = np.array([2, 3, 4])
>>> np.vstack((a,b))
array([[1, 2, 3],
[2, 3, 4]])

>>> a = np.array([[1], [2], [3]])

>>> b = np.array([[2], [3], [4]])
>>> np.vstack((a,b))
array([[1],
[2],
[3],
[2],
[3],
[4]])

Joining arrays
ma.column_stack(tup) Stack 1-D arrays as columns into a 2-D array.
ma.concatenate(arrays[, axis]) Concatenate a sequence of arrays along the given axis.
ma.dstack(tup) Stack arrays in sequence depth wise (along third axis).
ma.hstack(tup) Stack arrays in sequence horizontally (column wise).
ma.vstack(tup) Stack arrays in sequence vertically (row wise).

280 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

column_stack

Stack 1-D arrays as columns into a 2-D array.

Take a sequence of 1-D arrays and stack them as columns to make a single 2-D array. 2-D arrays are
stacked as-is, just like with hstack. 1-D arrays are turned into 2-D columns first.

Parameters
tup : sequence of 1-D or 2-D arrays.
Arrays to stack. All of them must have the same first dimension.
Returns
stacked : 2-D array
The array formed by stacking the given arrays.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.column_stack((a,b))
array([[1, 2],
[2, 3],
[3, 4]])

concatenate(arrays, axis=0)
Concatenate a sequence of arrays along the given axis.
Parameters
arrays : sequence of array_like
The arrays must have the same shape, except in the dimension corresponding to axis
(the first, by default).
axis : int, optional
The axis along which the arrays will be joined. Default is 0.
Returns
result : MaskedArray
The concatenated array with any masked entries preserved.
See Also:

numpy.concatenate
Equivalent function in the top-level NumPy module.

Examples

>>> import numpy.ma as ma

>>> a = ma.arange(3)
>>> a[1] = ma.masked
>>> b = ma.arange(2, 5)
>>> a

1.6. Masked arrays 281

NumPy Reference, Release 2.0.0.dev8464

masked_array(data = [0 -- 2],
mask = [False True False],
fill_value = 999999)
>>> b
masked_array(data = [2 3 4],
mask = False,
fill_value = 999999)
>>> ma.concatenate([a, b])
masked_array(data = [0 -- 2 2 3 4],
mask = [False True False False False False],
fill_value = 999999)

dstack

Stack arrays in sequence depth wise (along third axis).

Takes a sequence of arrays and stack them along the third axis to make a single array. Rebuilds
arrays divided by dsplit. This is a simple way to stack 2D arrays (images) into a single 3D array for
processing.

Parameters
tup : sequence of arrays
Arrays to stack. All of them must have the same shape along all but the third axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

See Also:

vstack
Stack along first axis.
hstack
Stack along second axis.
concatenate
Join arrays.
dsplit
Split array along third axis.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.dstack((a,b))
array([[[1, 2],
[2, 3],
[3, 4]]])

282 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> a = np.array([[1],[2],[3]])
>>> b = np.array([[2],[3],[4]])
>>> np.dstack((a,b))
array([[[1, 2]],
[[2, 3]],
[[3, 4]]])

hstack

Stack arrays in sequence horizontally (column wise).

Take a sequence of arrays and stack them horizontally to make a single array. Rebuild arrays divided
by hsplit.

Parameters
tup : sequence of ndarrays
All arrays must have the same shape along all but the second axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

See Also:

vstack
Stack arrays in sequence vertically (row wise).
dstack
Stack arrays in sequence depth wise (along third axis).
concatenate
Join a sequence of arrays together.
hsplit
Split array along second axis.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.hstack((a,b))
array([1, 2, 3, 2, 3, 4])
>>> a = np.array([[1],[2],[3]])
>>> b = np.array([[2],[3],[4]])
>>> np.hstack((a,b))
array([[1, 2],
[2, 3],
[3, 4]])

vstack

1.6. Masked arrays 283

NumPy Reference, Release 2.0.0.dev8464

Stack arrays in sequence vertically (row wise).

Take a sequence of arrays and stack them vertically to make a single array. Rebuild arrays divided
by vsplit.

Parameters
tup : sequence of ndarrays
Tuple containing arrays to be stacked. The arrays must have the same shape along all
but the first axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

See Also:

hstack
Stack arrays in sequence horizontally (column wise).
dstack
Stack arrays in sequence depth wise (along third dimension).
concatenate
Join a sequence of arrays together.
vsplit
Split array into a list of multiple sub-arrays vertically.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array([1, 2, 3])

>>> b = np.array([2, 3, 4])
>>> np.vstack((a,b))
array([[1, 2, 3],
[2, 3, 4]])

>>> a = np.array([[1], [2], [3]])

>>> b = np.array([[2], [3], [4]])
>>> np.vstack((a,b))
array([[1],
[2],
[3],
[2],
[3],
[4]])

284 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Operations on masks

Creating a mask
ma.make_mask(m[, copy, shrink, flag, dtype]) Create a boolean mask from an array.
ma.make_mask_none(newshape[, dtype]) Return a boolean mask of the given shape, filled with False.
ma.mask_or(m1, m2[, copy, shrink]) Combine two masks with the logical_or operator.
ma.make_mask_descr(ndtype) Construct a dtype description list from a given dtype.
make_mask(m, copy=False, shrink=True, flag=None, dtype=<type numpy.bool_>)
Create a boolean mask from an array.
Return m as a boolean mask, creating a copy if necessary or requested. The function can accept any sequence
that is convertible to integers, or nomask. Does not require that contents must be 0s and 1s, values of 0 are
interepreted as False, everything else as True.
Parameters
m : array_like
Potential mask.
copy : bool, optional
Whether to return a copy of m (True) or m itself (False).
shrink : bool, optional
Whether to shrink m to nomask if all its values are False.
flag : bool, optional
Deprecated equivalent of shrink.
dtype : dtype, optional
Data-type of the output mask. By default, the output mask has a dtype of MaskType
(bool). If the dtype is flexible, each field has a boolean dtype.
Returns
result : ndarray
A boolean mask derived from m.

Examples

>>> import numpy.ma as ma

>>> m = [True, False, True, True]
>>> ma.make_mask(m)
array([ True, False, True, True], dtype=bool)
>>> m = [1, 0, 1, 1]
>>> ma.make_mask(m)
array([ True, False, True, True], dtype=bool)
>>> m = [1, 0, 2, -3]
>>> ma.make_mask(m)
array([ True, False, True, True], dtype=bool)

Effect of the shrink parameter.

>>> m = np.zeros(4)
>>> m
array([ 0., 0., 0., 0.])
>>> ma.make_mask(m)

1.6. Masked arrays 285

NumPy Reference, Release 2.0.0.dev8464

False
>>> ma.make_mask(m, shrink=False)
array([False, False, False, False], dtype=bool)

Using a flexible dtype.

>>> m = [1, 0, 1, 1]
>>> n = [0, 1, 0, 0]
>>> arr = []
>>> for man, mouse in zip(m, n):
... arr.append((man, mouse))
>>> arr
[(1, 0), (0, 1), (1, 0), (1, 0)]
>>> dtype = np.dtype({names:[man, mouse],
formats:[np.int, np.int]})
>>> arr = np.array(arr, dtype=dtype)
>>> arr
array([(1, 0), (0, 1), (1, 0), (1, 0)],
dtype=[(man, <i4), (mouse, <i4)])
>>> ma.make_mask(arr, dtype=dtype)
array([(True, False), (False, True), (True, False), (True, False)],
dtype=[(man, |b1), (mouse, |b1)])

make_mask_none(newshape, dtype=None)
Return a boolean mask of the given shape, filled with False.
This function returns a boolean ndarray with all entries False, that can be used in common mask manipulations.
If a complex dtype is specified, the type of each field is converted to a boolean type.
Parameters
newshape : tuple
A tuple indicating the shape of the mask.
dtype: {None, dtype}, optional :
If None, use a MaskType instance. Otherwise, use a new datatype with the same fields
as dtype, converted to boolean types.
Returns
result : ndarray
An ndarray of appropriate shape and dtype, filled with False.
See Also:

make_mask
Create a boolean mask from an array.
make_mask_descr
Construct a dtype description list from a given dtype.

Examples

>>> import numpy.ma as ma

>>> ma.make_mask_none((3,))
array([False, False, False], dtype=bool)

Defining a more complex dtype.

286 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> dtype = np.dtype({names:[foo, bar],

formats:[np.float32, np.int]})
>>> dtype
dtype([(foo, <f4), (bar, <i4)])
>>> ma.make_mask_none((3,), dtype=dtype)
array([(False, False), (False, False), (False, False)],
dtype=[(foo, |b1), (bar, |b1)])

mask_or(m1, m2, copy=False, shrink=True)

Combine two masks with the logical_or operator.
The result may be a view on m1 or m2 if the other is nomask (i.e. False).
Parameters
m1, m2 : array_like
Input masks.
copy : bool, optional
If copy is False and one of the inputs is nomask, return a view of the other input mask.
Defaults to False.
shrink : bool, optional
Whether to shrink the output to nomask if all its values are False. Defaults to True.
Returns
mask : output mask
The result masks values that are masked in either m1 or m2.
Raises
ValueError :
If m1 and m2 have different flexible dtypes.

Examples

>>> m1 = np.ma.make_mask([0, 1, 1, 0])

>>> m2 = np.ma.make_mask([1, 0, 0, 0])
>>> np.ma.mask_or(m1, m2)
array([ True, True, True, False], dtype=bool)

make_mask_descr(ndtype)
Construct a dtype description list from a given dtype.
Returns a new dtype object, with the type of all fields in ndtype to a boolean type. Field names are not altered.
Parameters
ndtype : dtype
The dtype to convert.
Returns
result : dtype
A dtype that looks like ndtype, the type of all fields is boolean.

1.6. Masked arrays 287

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> import numpy.ma as ma

>>> dtype = np.dtype({names:[foo, bar],
formats:[np.float32, np.int]})
>>> dtype
dtype([(foo, <f4), (bar, <i4)])
>>> ma.make_mask_descr(dtype)
dtype([(foo, |b1), (bar, |b1)])
>>> ma.make_mask_descr(np.float32)
<type numpy.bool_>

Accessing a mask
ma.getmask(a) Return the mask of a masked array, or nomask.
ma.getmaskarray(arr) Return the mask of a masked array, or full boolean array of False.
ma.masked_array.mask Mask
getmask(a)
Return the mask of a masked array, or nomask.
Return the mask of a as an ndarray if a is a MaskedArray and the mask is not nomask, else return nomask. To
guarantee a full array of booleans of the same shape as a, use getmaskarray.
Parameters
a : array_like
Input MaskedArray for which the mask is required.
See Also:

getdata
Return the data of a masked array as an ndarray.
getmaskarray
Return the mask of a masked array, or full array of False.

Examples

>>> import numpy.ma as ma

>>> a = ma.masked_equal([[1,2],[3,4]], 2)
>>> a
masked_array(data =
[[1 --]
[3 4]],
mask =
[[False True]
[False False]],
fill_value=999999)
>>> ma.getmask(a)
array([[False, True],
[False, False]], dtype=bool)

Equivalently use the MaskedArray mask attribute.

>>> a.mask
array([[False, True],
[False, False]], dtype=bool)

288 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Result when mask == nomask

>>> b = ma.masked_array([[1,2],[3,4]])
>>> b
masked_array(data =
[[1 2]
[3 4]],
mask =
False,
fill_value=999999)
>>> ma.nomask
False
>>> ma.getmask(b) == ma.nomask
True
>>> b.mask == ma.nomask
True

getmaskarray(arr)
Return the mask of a masked array, or full boolean array of False.
Return the mask of arr as an ndarray if arr is a MaskedArray and the mask is not nomask, else return a full
boolean array of False of the same shape as arr.
Parameters
arr : array_like
Input MaskedArray for which the mask is required.
See Also:

getmask
Return the mask of a masked array, or nomask.
getdata
Return the data of a masked array as an ndarray.

Examples

>>> import numpy.ma as ma

>>> a = ma.masked_equal([[1,2],[3,4]], 2)
>>> a
masked_array(data =
[[1 --]
[3 4]],
mask =
[[False True]
[False False]],
fill_value=999999)
>>> ma.getmaskarray(a)
array([[False, True],
[False, False]], dtype=bool)

Result when mask == nomask

>>> b = ma.masked_array([[1,2],[3,4]])
>>> b
masked_array(data =
[[1 2]
[3 4]],

1.6. Masked arrays 289

NumPy Reference, Release 2.0.0.dev8464

mask =
False,
fill_value=999999)
>>> >ma.getmaskarray(b)
array([[False, False],
[False, False]], dtype=bool)

mask
Mask

Finding masked data

ma.flatnotmasked_contiguous(a) Find contiguous unmasked data in a masked array along the given axis.
ma.flatnotmasked_edges(a) Find the indices of the first and last unmasked values.
ma.notmasked_contiguous(a[, axis]) Find contiguous unmasked data in a masked array along the given axis.
ma.notmasked_edges(a[, axis]) Find the indices of the first and last unmasked values along an axis.
flatnotmasked_contiguous(a)
Find contiguous unmasked data in a masked array along the given axis.
Parameters
a : narray
The input array.
Returns
slice_list : list
A sorted sequence of slices (start index, end index).
See Also:
flatnotmasked_edges, notmasked_contiguous, notmasked_edges

Notes
Only accepts 2-D arrays at most.

Examples

>>> a = np.arange(10)
>>> mask = (a < 3) | (a > 8) | (a == 5)
>>> ma = np.ma.array(a, mask=mask)
>>> np.array(ma[~ma.mask])
array([3, 4, 6, 7, 8])

>>> np.ma.extras.flatnotmasked_contiguous(ma)
[slice(3, 4, None), slice(6, 8, None)]
>>> ma = np.ma.array(a, mask=np.ones_like(a))
>>> print np.ma.extras.flatnotmasked_edges(ma)
None

flatnotmasked_edges(a)
Find the indices of the first and last unmasked values.
Expects a 1-D MaskedArray, returns None if all values are masked.
Parameters
arr : array_like
Input 1-D MaskedArray

290 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Returns
edges : ndarray or None
The indices of first and last non-masked value in the array. Returns None if all values
are masked.
See Also:
flatnotmasked_contiguous, notmasked_contiguous, notmasked_edges

Notes
Only accepts 1-D arrays.

Examples

>>> a = np.arange(10)
>>> mask = (a < 3) | (a > 8) | (a == 5)

>>> ma = np.ma.array(a, mask=m)

>>> np.array(ma[~ma.mask])
array([3, 4, 6, 7, 8])

>>> flatnotmasked_edges(ma)
array([3, 8])

>>> ma = np.ma.array(a, mask=np.ones_like(a))

>>> print flatnotmasked_edges(ma)
None

notmasked_contiguous(a, axis=None)
Find contiguous unmasked data in a masked array along the given axis.
Parameters
a : array_like
The input array.
axis : int, optional
Axis along which to perform the operation. If None (default), applies to a flattened
version of the array.
Returns
endpoints : list
A list of slices (start and end indexes) of unmasked indexes in the array.
See Also:
flatnotmasked_edges, flatnotmasked_contiguous, notmasked_edges

Notes
Only accepts 2-D arrays at most.

Examples

1.6. Masked arrays 291

NumPy Reference, Release 2.0.0.dev8464

>>> a = np.arange(9).reshape((3, 3))

>>> mask = np.zeros_like(a)
>>> mask[1:, 1:] = 1

>>> ma = np.ma.array(a, mask=mask)

>>> np.array(ma[~ma.mask])
array([0, 1, 2, 3, 6])

>>> np.ma.extras.notmasked_contiguous(ma)
[slice(0, 3, None), slice(6, 6, None)]

notmasked_edges(a, axis=None)
Find the indices of the first and last unmasked values along an axis.
If all values are masked, return None. Otherwise, return a list of two tuples, corresponding to the indices of the
first and last unmasked values respectively.
Parameters
a : array_like
The input array.
axis : int, optional
Axis along which to perform the operation. If None (default), applies to a flattened
version of the array.
Returns
edges : ndarray or list
An array of start and end indexes if there are any masked data in the array. If there are
no masked data in the array, edges is a list of the first and last index.
See Also:
flatnotmasked_contiguous, flatnotmasked_edges, notmasked_contiguous

Examples

>>> a = np.arange(9).reshape((3, 3))

>>> m = np.zeros_like(a)
>>> m[1:, 1:] = 1

>>> ma = np.ma.array(a, mask=m)

>>> np.array(ma[~ma.mask])
array([0, 1, 2, 3, 6])

>>> np.ma.extras.notmasked_edges(ma)
array([0, 6])

292 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Modifying a mask
ma.mask_cols(a[, axis]) Mask columns of a 2D array that contain masked values.
ma.mask_or(m1, m2[, copy, shrink]) Combine two masks with the logical_or operator.
ma.mask_rowcols(a[, axis]) Mask rows and/or columns of a 2D array that contain masked values.
ma.mask_rows(a[, axis]) Mask rows of a 2D array that contain masked values.
ma.harden_mask(self) Force the mask to hard.
ma.soften_mask(self) Force the mask to soft.
ma.MaskedArray.harden_mask() Force the mask to hard.
ma.MaskedArray.soften_mask() Force the mask to soft.
ma.MaskedArray.shrink_mask() Reduce a mask to nomask when possible.
ma.MaskedArray.unshare_mask() Copy the mask and set the sharedmask flag to False.
mask_cols(a, axis=None)
Mask columns of a 2D array that contain masked values.
This function is a shortcut to mask_rowcols with axis equal to 1.
See Also:

mask_rowcols
Mask rows and/or columns of a 2D array.
masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.zeros((3, 3), dtype=np.int)
>>> a[1, 1] = 1
>>> a
array([[0, 0, 0],
[0, 1, 0],
[0, 0, 0]])
>>> a = ma.masked_equal(a, 1)
>>> a
masked_array(data =
[[0 0 0]
[0 -- 0]
[0 0 0]],
mask =
[[False False False]
[False True False]
[False False False]],
fill_value=999999)
>>> ma.mask_cols(a)
masked_array(data =
[[0 -- 0]
[0 -- 0]
[0 -- 0]],
mask =
[[False True False]
[False True False]
[False True False]],
fill_value=999999)

mask_or(m1, m2, copy=False, shrink=True)

Combine two masks with the logical_or operator.

1.6. Masked arrays 293

NumPy Reference, Release 2.0.0.dev8464

The result may be a view on m1 or m2 if the other is nomask (i.e. False).

Parameters
m1, m2 : array_like
Input masks.
copy : bool, optional
If copy is False and one of the inputs is nomask, return a view of the other input mask.
Defaults to False.
shrink : bool, optional
Whether to shrink the output to nomask if all its values are False. Defaults to True.
Returns
mask : output mask
The result masks values that are masked in either m1 or m2.
Raises
ValueError :
If m1 and m2 have different flexible dtypes.

Examples

>>> m1 = np.ma.make_mask([0, 1, 1, 0])

>>> m2 = np.ma.make_mask([1, 0, 0, 0])
>>> np.ma.mask_or(m1, m2)
array([ True, True, True, False], dtype=bool)

mask_rowcols(a, axis=None)
Mask rows and/or columns of a 2D array that contain masked values.
Mask whole rows and/or columns of a 2D array that contain masked values. The masking behavior is selected
using the axis parameter.
If axis is None, rows and columns are masked.
If axis is 0, only rows are masked.
If axis is 1 or -1, only columns are masked.

Parameters
a : array_like, MaskedArray
The array to mask. If not a MaskedArray instance (or if no array elements are masked).
The result is a MaskedArray with mask set to nomask (False). Must be a 2D array.
axis : int, optional
Axis along which to perform the operation. If None, applies to a flattened version of the
array.
Returns
a : MaskedArray
A modified version of the input array, masked depending on the value of the axis pa-
rameter.
Raises
NotImplementedError :

294 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

If input array a is not 2D.

See Also:

mask_rows
Mask rows of a 2D array that contain masked values.
mask_cols
Mask cols of a 2D array that contain masked values.
masked_where
Mask where a condition is met.

Notes
The input arrays mask is modified by this function.

Examples

>>> import numpy.ma as ma

>>> a = np.zeros((3, 3), dtype=np.int)
>>> a[1, 1] = 1
>>> a
array([[0, 0, 0],
[0, 1, 0],
[0, 0, 0]])
>>> a = ma.masked_equal(a, 1)
>>> a
masked_array(data =
[[0 0 0]
[0 -- 0]
[0 0 0]],
mask =
[[False False False]
[False True False]
[False False False]],
fill_value=999999)
>>> ma.mask_rowcols(a)
masked_array(data =
[[0 -- 0]
[-- -- --]
[0 -- 0]],
mask =
[[False True False]
[ True True True]
[False True False]],
fill_value=999999)

mask_rows(a, axis=None)
Mask rows of a 2D array that contain masked values.
This function is a shortcut to mask_rowcols with axis equal to 0.
See Also:

mask_rowcols
Mask rows and/or columns of a 2D array.

1.6. Masked arrays 295

NumPy Reference, Release 2.0.0.dev8464

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.zeros((3, 3), dtype=np.int)
>>> a[1, 1] = 1
>>> a
array([[0, 0, 0],
[0, 1, 0],
[0, 0, 0]])
>>> a = ma.masked_equal(a, 1)
>>> a
masked_array(data =
[[0 0 0]
[0 -- 0]
[0 0 0]],
mask =
[[False False False]
[False True False]
[False False False]],
fill_value=999999)
>>> ma.mask_rows(a)
masked_array(data =
[[0 0 0]
[-- -- --]
[0 0 0]],
mask =
[[False False False]
[ True True True]
[False False False]],
fill_value=999999)

harden_mask
Force the mask to hard.
Whether the mask of a masked array is hard or soft is determined by its hardmask property. harden_mask sets
hardmask to True.
See Also:
hardmask
soften_mask
Force the mask to soft.
Whether the mask of a masked array is hard or soft is determined by its hardmask property. soften_mask sets
hardmask to False.
See Also:
hardmask
harden_mask()
Force the mask to hard.
Whether the mask of a masked array is hard or soft is determined by its hardmask property. harden_mask sets
hardmask to True.
See Also:

296 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

hardmask
soften_mask()
Force the mask to soft.
Whether the mask of a masked array is hard or soft is determined by its hardmask property. soften_mask sets
hardmask to False.
See Also:
hardmask
shrink_mask()
Reduce a mask to nomask when possible.
Parameters
None :
Returns
None :

Examples

>>> x = np.ma.array([[1,2 ], [3, 4]], mask=[0]*4)

>>> x.mask
array([[False, False],
[False, False]], dtype=bool)
>>> x.shrink_mask()
>>> x.mask
False

unshare_mask()
Copy the mask and set the sharedmask flag to False.
Whether the mask is shared between masked arrays can be seen from the sharedmask property. unshare_mask
ensures the mask is not shared. A copy of the mask is only made if it was shared.
See Also:
sharedmask

1.6. Masked arrays 297

NumPy Reference, Release 2.0.0.dev8464

Conversion operations

> to a masked array

ma.asarray(a[, dtype, order]) Convert the input to a masked array of the given data-type.
ma.asanyarray(a[, dtype]) Convert the input to a masked array, conserving subclasses.
ma.fix_invalid(a[, mask, copy, fill_value]) Return input with invalid data masked and replaced by a fill
value.
ma.masked_equal(x, value[, copy]) Mask an array where equal to a given value.
ma.masked_greater(x, value[, copy]) Mask an array where greater than a given value.
ma.masked_greater_equal(x, value[, Mask an array where greater than or equal to a given value.
copy])
ma.masked_inside(x, v1, v2[, copy]) Mask an array inside a given interval.
ma.masked_invalid(a[, copy]) Mask an array where invalid values occur (NaNs or infs).
ma.masked_less(x, value[, copy]) Mask an array where less than a given value.
ma.masked_less_equal(x, value[, copy]) Mask an array where less than or equal to a given value.
ma.masked_not_equal(x, value[, copy]) Mask an array where not equal to a given value.
ma.masked_object(x, value[, copy, shrink]) Mask the array x where the data are exactly equal to value.
ma.masked_outside(x, v1, v2[, copy]) Mask an array outside a given interval.
ma.masked_values(x, value[, rtol, atol, ...]) Mask using floating point equality.
ma.masked_where(condition, a[, copy]) Mask an array where a condition is met.
asarray(a, dtype=None, order=None)
Convert the input to a masked array of the given data-type.
No copy is performed if the input is already an ndarray. If a is a subclass of MaskedArray, a base class
MaskedArray is returned.
Parameters
a : array_like
Input data, in any form that can be converted to a masked array. This includes lists, lists
of tuples, tuples, tuples of tuples, tuples of lists, ndarrays and masked arrays.
dtype : dtype, optional
By default, the data-type is inferred from the input data.
order : {C, F}, optional
Whether to use row-major (C) or column-major (FORTRAN) memory representa-
tion. Default is C.
Returns
out : MaskedArray
Masked array interpretation of a.
See Also:

asanyarray
Similar to asarray, but conserves subclasses.

Examples

>>> x = np.arange(10.).reshape(2, 5)
>>> x
array([[ 0., 1., 2., 3., 4.],
[ 5., 6., 7., 8., 9.]])
>>> np.ma.asarray(x)

298 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

masked_array(data =
[[ 0. 1. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]],
mask =
False,
fill_value = 1e+20)
>>> type(np.ma.asarray(x))
<class numpy.ma.core.MaskedArray>

asanyarray(a, dtype=None)
Convert the input to a masked array, conserving subclasses.
If a is a subclass of MaskedArray, its class is conserved. No copy is performed if the input is already an ndarray.
Parameters
a : array_like
Input data, in any form that can be converted to an array.
dtype : dtype, optional
By default, the data-type is inferred from the input data.
order : {C, F}, optional
Whether to use row-major (C) or column-major (FORTRAN) memory representa-
tion. Default is C.
Returns
out : MaskedArray
MaskedArray interpretation of a.
See Also:

asarray
Similar to asanyarray, but does not conserve subclass.

Examples

>>> x = np.arange(10.).reshape(2, 5)
>>> x
array([[ 0., 1., 2., 3., 4.],
[ 5., 6., 7., 8., 9.]])
>>> np.ma.asanyarray(x)
masked_array(data =
[[ 0. 1. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]],
mask =
False,
fill_value = 1e+20)
>>> type(np.ma.asanyarray(x))
<class numpy.ma.core.MaskedArray>

fix_invalid(a, mask=False, copy=True, fill_value=None)

Return input with invalid data masked and replaced by a fill value.
Invalid data means values of nan, inf, etc.
Parameters
a : array_like

1.6. Masked arrays 299

NumPy Reference, Release 2.0.0.dev8464

Input array, a (subclass of) ndarray.

copy : bool, optional
Whether to use a copy of a (True) or to fix a in place (False). Default is True.
fill_value : scalar, optional
Value used for fixing invalid data. Default is None, in which case the a.fill_value
is used.
Returns
b : MaskedArray
The input array with invalid entries fixed.

Notes
A copy is performed by default.

Examples

>>> x = np.ma.array([1., -1, np.nan, np.inf], mask=[1] + [0]*3)

>>> x
masked_array(data = [-- -1.0 nan inf],
mask = [ True False False False],
fill_value = 1e+20)
>>> np.ma.fix_invalid(x)
masked_array(data = [-- -1.0 -- --],
mask = [ True False True True],
fill_value = 1e+20)

>>> fixed = np.ma.fix_invalid(x)

>>> fixed.data
array([ 1.00000000e+00, -1.00000000e+00, 1.00000000e+20,
1.00000000e+20])
>>> x.data
array([ 1., -1., NaN, Inf])

masked_equal(x, value, copy=True)

Mask an array where equal to a given value.
This function is a shortcut to masked_where, with condition = (x == value). For floating point arrays, consider
using masked_values(x, value).
See Also:

masked_where
Mask where a condition is met.
masked_values
Mask using floating point equality.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_equal(a, 2)

300 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

masked_array(data = [0 1 -- 3],
mask = [False False True False],
fill_value=999999)

masked_greater(x, value, copy=True)

Mask an array where greater than a given value.
This function is a shortcut to masked_where, with condition = (x > value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_greater(a, 2)
masked_array(data = [0 1 2 --],
mask = [False False False True],
fill_value=999999)

masked_greater_equal(x, value, copy=True)

Mask an array where greater than or equal to a given value.
This function is a shortcut to masked_where, with condition = (x >= value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_greater_equal(a, 2)
masked_array(data = [0 1 -- --],
mask = [False False True True],
fill_value=999999)

masked_inside(x, v1, v2, copy=True)

Mask an array inside a given interval.
Shortcut to masked_where, where condition is True for x inside the interval [v1,v2] (v1 <= x <= v2). The
boundaries v1 and v2 can be given in either order.
See Also:

masked_where
Mask where a condition is met.

1.6. Masked arrays 301

NumPy Reference, Release 2.0.0.dev8464

Notes
The array x is prefilled with its filling value.

Examples

>>> import numpy.ma as ma

>>> x = [0.31, 1.2, 0.01, 0.2, -0.4, -1.1]
>>> ma.masked_inside(x, -0.3, 0.3)
masked_array(data = [0.31 1.2 -- -- -0.4 -1.1],
mask = [False False True True False False],
fill_value=1e+20)

The order of v1 and v2 doesnt matter.

>>> ma.masked_inside(x, 0.3, -0.3)

masked_array(data = [0.31 1.2 -- -- -0.4 -1.1],
mask = [False False True True False False],
fill_value=1e+20)

masked_invalid(a, copy=True)
Mask an array where invalid values occur (NaNs or infs).
This function is a shortcut to masked_where, with condition = ~(np.isfinite(a)). Any pre-existing mask is
conserved. Only applies to arrays with a dtype where NaNs or infs make sense (i.e. floating point types), but
accepts any array_like object.
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(5, dtype=np.float)
>>> a[2] = np.NaN
>>> a[3] = np.PINF
>>> a
array([ 0., 1., NaN, Inf, 4.])
>>> ma.masked_invalid(a)
masked_array(data = [0.0 1.0 -- -- 4.0],
mask = [False False True True False],
fill_value=1e+20)

masked_less(x, value, copy=True)

Mask an array where less than a given value.
This function is a shortcut to masked_where, with condition = (x < value).
See Also:

masked_where
Mask where a condition is met.

302 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_less(a, 2)
masked_array(data = [-- -- 2 3],
mask = [ True True False False],
fill_value=999999)

masked_less_equal(x, value, copy=True)

Mask an array where less than or equal to a given value.
This function is a shortcut to masked_where, with condition = (x <= value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_less_equal(a, 2)
masked_array(data = [-- -- -- 3],
mask = [ True True True False],
fill_value=999999)

masked_not_equal(x, value, copy=True)

Mask an array where not equal to a given value.
This function is a shortcut to masked_where, with condition = (x != value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_not_equal(a, 2)
masked_array(data = [-- -- 2 --],
mask = [ True True False True],
fill_value=999999)

masked_object(x, value, copy=True, shrink=True)

Mask the array x where the data are exactly equal to value.
This function is similar to masked_values, but only suitable for object arrays: for floating point, use
masked_values instead.

1.6. Masked arrays 303

NumPy Reference, Release 2.0.0.dev8464

Parameters
x : array_like
Array to mask
value : object
Comparison value
copy : {True, False}, optional
Whether to return a copy of x.
shrink : {True, False}, optional
Whether to collapse a mask full of False to nomask
Returns
result : MaskedArray
The result of masking x where equal to value.
See Also:

masked_where
Mask where a condition is met.
masked_equal
Mask where equal to a given value (integers).
masked_values
Mask using floating point equality.

Examples

>>> import numpy.ma as ma

>>> food = np.array([green_eggs, ham], dtype=object)
>>> # dont eat spoiled food
>>> eat = ma.masked_object(food, green_eggs)
>>> print eat
[-- ham]
>>> # plain ol ham is boring
>>> fresh_food = np.array([cheese, ham, pineapple], dtype=object)
>>> eat = ma.masked_object(fresh_food, green_eggs)
>>> print eat
[cheese ham pineapple]

Note that mask is set to nomask if possible.

>>> eat
masked_array(data = [cheese ham pineapple],
mask = False,
fill_value=?)

masked_outside(x, v1, v2, copy=True)

Mask an array outside a given interval.
Shortcut to masked_where, where condition is True for x outside the interval [v1,v2] (x < v1)|(x > v2). The
boundaries v1 and v2 can be given in either order.
See Also:

304 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

masked_where
Mask where a condition is met.

Notes
The array x is prefilled with its filling value.

Examples

>>> import numpy.ma as ma

>>> x = [0.31, 1.2, 0.01, 0.2, -0.4, -1.1]
>>> ma.masked_outside(x, -0.3, 0.3)
masked_array(data = [-- -- 0.01 0.2 -- --],
mask = [ True True False False True True],
fill_value=1e+20)

The order of v1 and v2 doesnt matter.

>>> ma.masked_outside(x, 0.3, -0.3)

masked_array(data = [-- -- 0.01 0.2 -- --],
mask = [ True True False False True True],
fill_value=1e+20)

masked_values(x, value, rtol=1.0000000000000001e-05, atol=1e-08, copy=True, shrink=True)

Mask using floating point equality.
Return a MaskedArray, masked where the data in array x are approximately equal to value, i.e. where the
following condition is True
(abs(x - value) <= atol+rtol*abs(value))
The fill_value is set to value and the mask is set to nomask if possible. For integers, consider using
masked_equal.
Parameters
x : array_like
Array to mask.
value : float
Masking value.
rtol : float, optional
Tolerance parameter.
atol : float, optional
Tolerance parameter (1e-8).
copy : bool, optional
Whether to return a copy of x.
shrink : bool, optional
Whether to collapse a mask full of False to nomask.
Returns
result : MaskedArray
The result of masking x where approximately equal to value.

1.6. Masked arrays 305

NumPy Reference, Release 2.0.0.dev8464

See Also:

masked_where
Mask where a condition is met.
masked_equal
Mask where equal to a given value (integers).

Examples

>>> import numpy.ma as ma

>>> x = np.array([1, 1.1, 2, 1.1, 3])
>>> ma.masked_values(x, 1.1)
masked_array(data = [1.0 -- 2.0 -- 3.0],
mask = [False True False True False],
fill_value=1.1)

Note that mask is set to nomask if possible.

>>> ma.masked_values(x, 1.5)

masked_array(data = [ 1. 1.1 2. 1.1 3. ],
mask = False,
fill_value=1.5)

For integers, the fill value will be different in general to the result of masked_equal.

>>> x = np.arange(5)
>>> x
array([0, 1, 2, 3, 4])
>>> ma.masked_values(x, 2)
masked_array(data = [0 1 -- 3 4],
mask = [False False True False False],
fill_value=2)
>>> ma.masked_equal(x, 2)
masked_array(data = [0 1 -- 3 4],
mask = [False False True False False],
fill_value=999999)

masked_where(condition, a, copy=True)
Mask an array where a condition is met.
Return a as an array masked where condition is True. Any masked values of a or condition are also masked in
the output.
Parameters
condition : array_like
Masking condition. When condition tests floating point values for equality, consider
using masked_values instead.
a : array_like
Array to mask.
copy : bool
If True (default) make a copy of a in the result. If False modify a in place and return a
view.

306 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Returns
result : MaskedArray
The result of masking a where condition is True.
See Also:

masked_values
Mask using floating point equality.
masked_equal
Mask where equal to a given value.
masked_not_equal
Mask where not equal to a given value.
masked_less_equal
Mask where less than or equal to a given value.
masked_greater_equal
Mask where greater than or equal to a given value.
masked_less
Mask where less than a given value.
masked_greater
Mask where greater than a given value.
masked_inside
Mask inside a given interval.
masked_outside
Mask outside a given interval.
masked_invalid
Mask invalid values (NaNs or infs).

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_where(a <= 2, a)
masked_array(data = [-- -- -- 3],
mask = [ True True True False],
fill_value=999999)

Mask array b conditional on a.

>>> b = [a, b, c, d]
>>> ma.masked_where(a == 2, b)
masked_array(data = [a b -- d],
mask = [False False True False],
fill_value=N/A)

Effect of the copy argument.

1.6. Masked arrays 307

NumPy Reference, Release 2.0.0.dev8464

>>> c = ma.masked_where(a <= 2, a)

>>> c
masked_array(data = [-- -- -- 3],
mask = [ True True True False],
fill_value=999999)
>>> c[0] = 99
>>> c
masked_array(data = [99 -- -- 3],
mask = [False True True False],
fill_value=999999)
>>> a
array([0, 1, 2, 3])
>>> c = ma.masked_where(a <= 2, a, copy=False)
>>> c[0] = 99
>>> c
masked_array(data = [99 -- -- 3],
mask = [False True True False],
fill_value=999999)
>>> a
array([99, 1, 2, 3])

When condition or a contain masked values.

>>> a = np.arange(4)
>>> a = ma.masked_where(a == 2, a)
>>> a
masked_array(data = [0 1 -- 3],
mask = [False False True False],
fill_value=999999)
>>> b = np.arange(4)
>>> b = ma.masked_where(b == 0, b)
>>> b
masked_array(data = [-- 1 2 3],
mask = [ True False False False],
fill_value=999999)
>>> ma.masked_where(a == 3, b)
masked_array(data = [-- 1 -- --],
mask = [ True False True True],
fill_value=999999)

> to a ndarray
ma.compress_cols(a) Suppress whole columns of a 2-D array that contain masked values.
ma.compress_rowcols(x[, axis]) Suppress the rows and/or columns of a 2-D array that contain
ma.compress_rows(a) Suppress whole rows of a 2-D array that contain masked values.
ma.compressed(x) Return all the non-masked data as a 1-D array.
ma.filled(a[, fill_value]) Return input as an array with masked data replaced by a fill value.
ma.MaskedArray.compressed() Return all the non-masked data as a 1-D array.
ma.MaskedArray.filled([fill_value]) Return a copy of self, with masked values filled with a given value.
compress_cols(a)
Suppress whole columns of a 2-D array that contain masked values.
This is equivalent to np.ma.extras.compress_rowcols(a, 1), see extras.compress_rowcols for de-
tails.
See Also:

308 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

extras.compress_rowcols
compress_rowcols(x, axis=None)
Suppress the rows and/or columns of a 2-D array that contain masked values.
The suppression behavior is selected with the axis parameter.
If axis is None, both rows and columns are suppressed.
If axis is 0, only rows are suppressed.
If axis is 1 or -1, only columns are suppressed.

Parameters
axis : int, optional
Axis along which to perform the operation. Default is None.
Returns
compressed_array : ndarray
The compressed array.

Examples

>>> x = np.ma.array(np.arange(9).reshape(3, 3), mask=[[1, 0, 0],

... [1, 0, 0],
... [0, 0, 0]])
>>> x
masked_array(data =
[[-- 1 2]
[-- 4 5]
[6 7 8]],
mask =
[[ True False False]
[ True False False]
[False False False]],
fill_value = 999999)

>>> np.ma.extras.compress_rowcols(x)
array([[7, 8]])
>>> np.ma.extras.compress_rowcols(x, 0)
array([[6, 7, 8]])
>>> np.ma.extras.compress_rowcols(x, 1)
array([[1, 2],
[4, 5],
[7, 8]])

compress_rows(a)
Suppress whole rows of a 2-D array that contain masked values.
This is equivalent to np.ma.extras.compress_rowcols(a, 0), see extras.compress_rowcols for de-
tails.
See Also:
extras.compress_rowcols
compressed(x)
Return all the non-masked data as a 1-D array.

1.6. Masked arrays 309

NumPy Reference, Release 2.0.0.dev8464

This function is equivalent to calling the compressed method of a MaskedArray, see MaskedArray.compressed
for details.
See Also:

MaskedArray.compressed
Equivalent method.

filled(a, fill_value=None)
Return input as an array with masked data replaced by a fill value.
If a is not a MaskedArray, a itself is returned. If a is a MaskedArray and fill_value is None, fill_value is set to
a.fill_value.
Parameters
a : MaskedArray or array_like
An input object.
fill_value : scalar, optional
Filling value. Default is None.
Returns
a : ndarray
The filled array.
See Also:
compressed

Examples

>>> x = np.ma.array(np.arange(9).reshape(3, 3), mask=[[1, 0, 0],

... [1, 0, 0],
... [0, 0, 0]])
>>> x.filled()
array([[999999, 1, 2],
[999999, 4, 5],
[ 6, 7, 8]])

compressed()
Return all the non-masked data as a 1-D array.
Returns
data : ndarray
A new ndarray holding the non-masked data is returned.

Notes
The result is not a MaskedArray!

Examples

>>> x = np.ma.array(np.arange(5), mask=[0]*2 + [1]*3)

>>> x.compressed()
array([0, 1])
>>> type(x.compressed())
<type numpy.ndarray>

310 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

filled(fill_value=None)
Return a copy of self, with masked values filled with a given value.
Parameters
fill_value : scalar, optional
The value to use for invalid entries (None by default). If None, the fill_value attribute
of the array is used instead.

Notes
The result is not a MaskedArray!

Examples

>>> x = np.ma.array([1,2,3,4,5], mask=[0,0,1,0,1], fill_value=-999)

>>> x.filled()
array([1, 2, -999, 4, -999])
>>> type(x.filled())
<type numpy.ndarray>

Subclassing is preserved. This means that if the data part of the masked array is a matrix, filled returns a matrix:

>>> x = np.ma.array(np.matrix([[1, 2], [3, 4]]), mask=[[0, 1], [1, 0]])

>>> x.filled()
matrix([[ 1, 999999],
[999999, 4]])

> to another object

ma.MaskedArray.tofile(fid[, sep, Save a masked array to a file in binary format.
format])
ma.MaskedArray.tolist([fill_value]) Return the data portion of the masked array as a hierarchical
Python list.
ma.MaskedArray.torecords() Transforms a masked array into a flexible-type array.
ma.MaskedArray.tostring([fill_value, Return the array data as a string containing the raw bytes in the
order]) array.
tofile(fid, sep=, format=%s)
Save a masked array to a file in binary format.

Warning: This function is not implemented yet.

Raises
NotImplementedError :
When tofile is called.
tolist(fill_value=None)
Return the data portion of the masked array as a hierarchical Python list.
Data items are converted to the nearest compatible Python type. Masked values are converted to fill_value. If
fill_value is None, the corresponding entries in the output list will be None.
Parameters
fill_value : scalar, optional
The value to use for invalid entries. Default is None.

1.6. Masked arrays 311

NumPy Reference, Release 2.0.0.dev8464

Returns
result : list
The Python list representation of the masked array.

Examples

>>> x = np.ma.array([[1,2,3], [4,5,6], [7,8,9]], mask=[0] + [1,0]*4)

>>> x.tolist()
[[1, None, 3], [None, 5, None], [7, None, 9]]
>>> x.tolist(-999)
[[1, -999, 3], [-999, 5, -999], [7, -999, 9]]

torecords()
Transforms a masked array into a flexible-type array.
The flexible type array that is returned will have two fields:
the _data field stores the _data part of the array.
the _mask field stores the _mask part of the array.

Parameters
None :
Returns
record : ndarray
A new flexible-type ndarray with two fields: the first element containing a value, the
second element containing the corresponding mask boolean. The returned record shape
matches self.shape.

Notes
A side-effect of transforming a masked array into a flexible ndarray is that meta information (fill_value,
...) will be lost.

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.toflex()
[[(1, False) (2, True) (3, False)]
[(4, True) (5, False) (6, True)]
[(7, False) (8, True) (9, False)]]

tostring(fill_value=None, order=C)
Return the array data as a string containing the raw bytes in the array.
The array is filled with a fill value before the string conversion.
Parameters
fill_value : scalar, optional
Value used to fill in the masked values. Deafult is None, in which case MaskedAr-
ray.fill_value is used.
order : {C,F,A}, optional

312 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Order of the data item in the copy. Default is C.

C C order (row major).
F Fortran order (column major).
A Any, current order of array.
None Same as A.
See Also:
ndarray.tostring, tolist, tofile

Notes
As for ndarray.tostring, information about the shape, dtype, etc., but also about fill_value, will be lost.

Examples

>>> x = np.ma.array(np.array([[1, 2], [3, 4]]), mask=[[0, 1], [1, 0]])

>>> x.tostring()
\x01\x00\x00\x00?B\x0f\x00?B\x0f\x00\x04\x00\x00\x00

Pickling and unpickling

ma.dump(a, F) Pickle a masked array to a file.
ma.dumps(a) Return a string corresponding to the pickling of a masked array.
ma.load(F) Wrapper around cPickle.load which accepts either a file-like object
ma.loads(strg) Load a pickle from the current string.
dump(a, F)
Pickle a masked array to a file.
This is a wrapper around cPickle.dump.
Parameters
a : MaskedArray
The array to be pickled.
F : str or file-like object
The file to pickle a to. If a string, the full path to the file.
dumps(a)
Return a string corresponding to the pickling of a masked array.
This is a wrapper around cPickle.dumps.
Parameters
a : MaskedArray
The array for which the string representation of the pickle is returned.
load(F)
Wrapper around cPickle.load which accepts either a file-like object or a filename.
Parameters
F : str or file
The file or file name to load.
See Also:

1.6. Masked arrays 313

NumPy Reference, Release 2.0.0.dev8464

dump
Pickle an array

Notes
This is different from numpy.load, which does not use cPickle but loads the NumPy binary .npy format.
loads(strg)
Load a pickle from the current string.
The result of cPickle.loads(strg) is returned.
Parameters
strg : str
The string to load.
See Also:

dumps
Return a string corresponding to the pickling of a masked array.

Filling a masked array

ma.common_fill_value(a, b) Return the common filling value of two masked arrays, if any.
ma.default_fill_value(obj) Return the default fill value for the argument object.
ma.maximum_fill_value(obj) Return the minimum value that can be represented by the dtype of
an object.
ma.maximum_fill_value(obj) Return the minimum value that can be represented by the dtype of
an object.
ma.set_fill_value(a, fill_value) Set the filling value of a, if a is a masked array.
ma.MaskedArray.get_fill_value() Return the filling value of the masked array.
Set the filling value of the masked array.
ma.MaskedArray.set_fill_value([value])
ma.MaskedArray.fill_value Filling value.
common_fill_value(a, b)
Return the common filling value of two masked arrays, if any.
If a.fill_value == b.fill_value, return the fill value, otherwise return None.
Parameters
a, b : MaskedArray
The masked arrays for which to compare fill values.
Returns
fill_value : scalar or None
The common fill value, or None.

Examples

>>> x = np.ma.array([0, 1.], fill_value=3)

>>> y = np.ma.array([0, 1.], fill_value=3)
>>> np.ma.common_fill_value(x, y)
3.0

default_fill_value(obj)
Return the default fill value for the argument object.
The default filling value depends on the datatype of the input array or the type of the input scalar:

314 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

datatype default
bool True
int 999999
float 1.e20
complex 1.e20+0j
object ?
string N/A

Parameters
obj : ndarray, dtype or scalar
The array data-type or scalar for which the default fill value is returned.
Returns
fill_value : scalar
The default fill value.

Examples

>>> np.ma.default_fill_value(1)
999999
>>> np.ma.default_fill_value(np.array([1.1, 2., np.pi]))
1e+20
>>> np.ma.default_fill_value(np.dtype(complex))
(1e+20+0j)

maximum_fill_value(obj)
Return the minimum value that can be represented by the dtype of an object.
This function is useful for calculating a fill value suitable for taking the maximum of an array with a given dtype.
Parameters
obj : {ndarray, dtype}
An object that can be queried for its numeric type.
Returns
val : scalar
The minimum representable value.
Raises
TypeError :
If obj isnt a suitable numeric type.
See Also:

minimum_fill_value
The inverse function.
set_fill_value
Set the filling value of a masked array.
MaskedArray.fill_value
Return current fill value.

1.6. Masked arrays 315

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> import numpy.ma as ma

>>> a = np.int8()
>>> ma.maximum_fill_value(a)
-128
>>> a = np.int32()
>>> ma.maximum_fill_value(a)
-2147483648

An array of numeric data can also be passed.

>>> a = np.array([1, 2, 3], dtype=np.int8)

>>> ma.maximum_fill_value(a)
-128
>>> a = np.array([1, 2, 3], dtype=np.float32)
>>> ma.maximum_fill_value(a)
-inf

maximum_fill_value(obj)
Return the minimum value that can be represented by the dtype of an object.
This function is useful for calculating a fill value suitable for taking the maximum of an array with a given dtype.
Parameters
obj : {ndarray, dtype}
An object that can be queried for its numeric type.
Returns
val : scalar
The minimum representable value.
Raises
TypeError :
If obj isnt a suitable numeric type.
See Also:

minimum_fill_value
The inverse function.
set_fill_value
Set the filling value of a masked array.
MaskedArray.fill_value
Return current fill value.

Examples

>>> import numpy.ma as ma

>>> a = np.int8()
>>> ma.maximum_fill_value(a)
-128
>>> a = np.int32()
>>> ma.maximum_fill_value(a)
-2147483648

An array of numeric data can also be passed.

316 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> a = np.array([1, 2, 3], dtype=np.int8)

>>> ma.maximum_fill_value(a)
-128
>>> a = np.array([1, 2, 3], dtype=np.float32)
>>> ma.maximum_fill_value(a)
-inf

set_fill_value(a, fill_value)
Set the filling value of a, if a is a masked array.
This function changes the fill value of the masked array a in place. If a is not a masked array, the function
returns silently, without doing anything.
Parameters
a : array_like
Input array.
fill_value : dtype
Filling value. A consistency test is performed to make sure the value is compatible with
the dtype of a.
Returns
None :
Nothing returned by this function.
See Also:

maximum_fill_value
Return the default fill value for a dtype.
MaskedArray.fill_value
Return current fill value.
MaskedArray.set_fill_value
Equivalent method.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(5)
>>> a
array([0, 1, 2, 3, 4])
>>> a = ma.masked_where(a < 3, a)
>>> a
masked_array(data = [-- -- -- 3 4],
mask = [ True True True False False],
fill_value=999999)
>>> ma.set_fill_value(a, -999)
>>> a
masked_array(data = [-- -- -- 3 4],
mask = [ True True True False False],
fill_value=-999)

Nothing happens if a is not a masked array.

1.6. Masked arrays 317

NumPy Reference, Release 2.0.0.dev8464

>>> a = range(5)
>>> a
[0, 1, 2, 3, 4]
>>> ma.set_fill_value(a, 100)
>>> a
[0, 1, 2, 3, 4]
>>> a = np.arange(5)
>>> a
array([0, 1, 2, 3, 4])
>>> ma.set_fill_value(a, 100)
>>> a
array([0, 1, 2, 3, 4])

get_fill_value()
Return the filling value of the masked array.
Returns
fill_value : scalar
The filling value.

Examples

>>> for dt in [np.int32, np.int64, np.float64, np.complex128]:

... np.ma.array([0, 1], dtype=dt).get_fill_value()
...
999999
999999
1e+20
(1e+20+0j)

>>> x = np.ma.array([0, 1.], fill_value=-np.inf)

>>> x.get_fill_value()
-inf

set_fill_value(value=None)
Set the filling value of the masked array.
Parameters
value : scalar, optional
The new filling value. Default is None, in which case a default based on the data type
is used.
See Also:

ma.set_fill_value
Equivalent function.

Examples

>>> x = np.ma.array([0, 1.], fill_value=-np.inf)

>>> x.fill_value
-inf
>>> x.set_fill_value(np.pi)
>>> x.fill_value
3.1415926535897931

318 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Reset to default:

>>> x.set_fill_value()
>>> x.fill_value
1e+20

fill_value
Filling value.

Masked arrays arithmetics

Arithmetics
ma.anom(self[, axis, dtype]) Compute the anomalies (deviations from the arithmetic mean) along
the given axis.
ma.anomalies(self[, axis, dtype]) Compute the anomalies (deviations from the arithmetic mean) along
the given axis.
ma.average(a[, axis, weights, Return the weighted average of array over the given axis.
returned])
ma.conjugate() Return the complex conjugate, element-wise.
ma.corrcoef(x[, y, rowvar, bias, Return correlation coefficients of the input array.
allow_masked])
ma.cov(x[, y, rowvar, bias, Estimate the covariance matrix.
allow_masked])
ma.cumsum(self[, axis, dtype, out]) Return the cumulative sum of the elements along the given axis.
ma.cumprod(self[, axis, dtype, out]) Return the cumulative product of the elements along the given axis.
ma.mean(self[, axis, dtype, out]) Returns the average of the array elements.
ma.median(a[, axis, out, Compute the median along the specified axis.
overwrite_input])
ma.power(a, b[, third]) Returns element-wise base array raised to power from second array.
ma.prod(self[, axis, dtype, out]) Return the product of the array elements over the given axis.
ma.std(self[, axis, dtype, out, ddof]) Compute the standard deviation along the specified axis.
ma.sum(self[, axis, dtype, out]) Return the sum of the array elements over the given axis.
ma.var(self[, axis, dtype, out, ddof]) Compute the variance along the specified axis.
ma.MaskedArray.anom([axis, dtype]) Compute the anomalies (deviations from the arithmetic mean) along
the given axis.
ma.MaskedArray.cumprod([axis, Return the cumulative product of the elements along the given axis.
dtype, out])
ma.MaskedArray.cumsum([axis, Return the cumulative sum of the elements along the given axis.
dtype, out])
ma.MaskedArray.mean([axis, dtype, Returns the average of the array elements.
out])
ma.MaskedArray.prod([axis, dtype, Return the product of the array elements over the given axis.
out])
ma.MaskedArray.std([axis, dtype, Compute the standard deviation along the specified axis.
out, ddof])
ma.MaskedArray.sum([axis, dtype, Return the sum of the array elements over the given axis.
out])
ma.MaskedArray.var([axis, dtype, Compute the variance along the specified axis.
out, ddof])
anom
Compute the anomalies (deviations from the arithmetic mean) along the given axis.

1.6. Masked arrays 319

NumPy Reference, Release 2.0.0.dev8464

Returns an array of anomalies, with the same shape as the input and where the arithmetic mean is computed
along the given axis.
Parameters
axis : int, optional
Axis over which the anomalies are taken. The default is to use the mean of the flattened
array as reference.
dtype : dtype, optional
Type to use in computing the variance. For arrays of integer type
the default is float32; for arrays of float types it is the same as the array type.
See Also:

mean
Compute the mean of the array.

Examples

>>> a = np.ma.array([1,2,3])
>>> a.anom()
masked_array(data = [-1. 0. 1.],
mask = False,
fill_value = 1e+20)

anomalies
Compute the anomalies (deviations from the arithmetic mean) along the given axis.
Returns an array of anomalies, with the same shape as the input and where the arithmetic mean is computed
along the given axis.
Parameters
axis : int, optional
Axis over which the anomalies are taken. The default is to use the mean of the flattened
array as reference.
dtype : dtype, optional
Type to use in computing the variance. For arrays of integer type
the default is float32; for arrays of float types it is the same as the array type.
See Also:

mean
Compute the mean of the array.

Examples

>>> a = np.ma.array([1,2,3])
>>> a.anom()
masked_array(data = [-1. 0. 1.],
mask = False,
fill_value = 1e+20)

average(a, axis=None, weights=None, returned=False)

Return the weighted average of array over the given axis.

320 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Parameters
a : array_like
Data to be averaged. Masked entries are not taken into account in the computation.
axis : int, optional
Axis along which the variance is computed. The default is to compute the variance of
the flattened array.
weights : array_like, optional
The importance that each element has in the computation of the average. The weights
array can either be 1-D (in which case its length must be the size of a along the given
axis) or of the same shape as a. If weights=None, then all data in a are assumed to
have a weight equal to one.
returned : bool, optional
Flag indicating whether a tuple (result, sum of weights) should be returned
as output (True), or just the result (False). Default is False.
Returns
average, [sum_of_weights] : (tuple of) scalar or MaskedArray
The average along the specified axis. When returned is True, return a tuple with the
average as the first element and the sum of the weights as the second element. The
return type is np.float64 if a is of integer type, otherwise it is of the same type as a. If
returned, sum_of_weights is of the same type as average.

Examples

>>> a = np.ma.array([1., 2., 3., 4.], mask=[False, False, True, True])

>>> np.ma.average(a, weights=[3, 1, 0, 0])
1.25

>>> x = np.ma.arange(6.).reshape(3, 2)
>>> print x
[[ 0. 1.]
[ 2. 3.]
[ 4. 5.]]
>>> avg, sumweights = np.ma.average(x, axis=0, weights=[1, 2, 3],
... returned=True)
>>> print avg
[2.66666666667 3.66666666667]

conjugate
Return the complex conjugate, element-wise.
The complex conjugate of a complex number is obtained by changing the sign of its imaginary part.
Parameters
x : array_like
Input value.
Returns
y : ndarray
The complex conjugate of x, with same dtype as y.

1.6. Masked arrays 321

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.conjugate(1+2j)
(1-2j)

>>> x = np.eye(2) + 1j * np.eye(2)

>>> np.conjugate(x)
array([[ 1.-1.j, 0.-0.j],
[ 0.-0.j, 1.-1.j]])

corrcoef(x, y=None, rowvar=True, bias=False, allow_masked=True)

Return correlation coefficients of the input array.
Except for the handling of missing data this function does the same as numpy.corrcoef. For more details
and examples, see numpy.corrcoef.
Parameters
x : array_like
A 1-D or 2-D array containing multiple variables and observations. Each row of x
represents a variable, and each column a single observation of all those variables. Also
see rowvar below.
y : array_like, optional
An additional set of variables and observations. y has the same shape as x.
rowvar : bool, optional
If rowvar is True (default), then each row represents a variable, with observations in the
columns. Otherwise, the relationship is transposed: each column represents a variable,
while the rows contain observations.
bias : bool, optional
Default normalization (False) is by (N-1), where N is the number of observations given
(unbiased estimate). If bias is 1, then normalization is by N.
allow_masked : bool, optional
If True, masked values are propagated pair-wise: if a value is masked in x, the corre-
sponding value is masked in y. If False, raises an exception.
See Also:

numpy.corrcoef
Equivalent function in top-level NumPy module.
cov
Estimate the covariance matrix.

cov(x, y=None, rowvar=True, bias=False, allow_masked=True)

Estimate the covariance matrix.
Except for the handling of missing data this function does the same as numpy.cov. For more details and
examples, see numpy.cov.
By default, masked values are recognized as such. If x and y have the same shape, a common mask is allocated:
if x[i,j] is masked, then y[i,j] will also be masked. Setting allow_masked to False will raise an exception
if values are missing in either of the input arrays.

322 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Parameters
x : array_like
A 1-D or 2-D array containing multiple variables and observations. Each row of x
represents a variable, and each column a single observation of all those variables. Also
see rowvar below.
y : array_like, optional
An additional set of variables and observations. y has the same form as x.
rowvar : bool, optional
If rowvar is True (default), then each row represents a variable, with observations in the
columns. Otherwise, the relationship is transposed: each column represents a variable,
while the rows contain observations.
bias : bool, optional
Default normalization (False) is by (N-1), where N is the number of observations given
(unbiased estimate). If bias is True, then normalization is by N.
allow_masked : bool, optional
If True, masked values are propagated pair-wise: if a value is masked in x, the cor-
responding value is masked in y. If False, raises a ValueError exception when some
values are missing.
Raises
ValueError: :
Raised if some values are missing and allow_masked is False.
See Also:
numpy.cov
cumsum
Return the cumulative sum of the elements along the given axis. The cumulative sum is calculated over the
flattened array by default, otherwise over the specified axis.
Masked values are set to 0 internally during the computation. However, their position is saved, and the result
will be masked at the same locations.
Parameters
axis : {None, -1, int}, optional
Axis along which the sum is computed. The default (axis = None) is to compute over
the flattened array. axis may be negative, in which case it counts from the last to the
first axis.
dtype : {None, dtype}, optional
Type of the returned array and of the accumulator in which the elements are summed. If
dtype is not specified, it defaults to the dtype of a, unless a has an integer dtype with a
precision less than that of the default platform integer. In that case, the default platform
integer is used.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
cumsum : ndarray.

1.6. Masked arrays 323

NumPy Reference, Release 2.0.0.dev8464

A new array holding the result is returned unless out is specified, in which case a
reference to out is returned.

Notes
The mask is lost if out is not a valid MaskedArray !
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> marr = np.ma.array(np.arange(10), mask=[0,0,0,1,1,1,0,0,0,0])

>>> print marr.cumsum()
[0 1 3 -- -- -- 9 16 24 33]

cumprod
Return the cumulative product of the elements along the given axis. The cumulative product is taken over the
flattened array by default, otherwise over the specified axis.
Masked values are set to 1 internally during the computation. However, their position is saved, and the result
will be masked at the same locations.
Parameters
axis : {None, -1, int}, optional
Axis along which the product is computed. The default (axis = None) is to compute
over the flattened array.
dtype : {None, dtype}, optional
Determines the type of the returned array and of the accumulator where the elements
are multiplied. If dtype has the value None and the type of a is an integer type
of precision less than the default platform integer, then the default platform integer
precision is used. Otherwise, the dtype is the same as that of a.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
cumprod : ndarray
A new array holding the result is returned unless out is specified, in which case a refer-
ence to out is returned.

Notes
The mask is lost if out is not a valid MaskedArray !
Arithmetic is modular when using integer types, and no error is raised on overflow.
mean
Returns the average of the array elements.
Masked entries are ignored. The average is taken over the flattened array by default, otherwise over the specified
axis. Refer to numpy.mean for the full documentation.
Parameters
a : array_like
Array containing numbers whose mean is desired. If a is not an array, a conversion is
attempted.

324 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

axis : int, optional

Axis along which the means are computed. The default is to compute the mean of the
flattened array.
dtype : dtype, optional
Type to use in computing the mean. For integer inputs, the default is float64; for floating
point, inputs it is the same as the input dtype.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary.
Returns
mean : ndarray, see dtype parameter above
If out=None, returns a new array containing the mean values, otherwise a reference to
the output array is returned.
See Also:

numpy.ma.mean
Equivalent function.
numpy.mean
Equivalent function on non-masked arrays.
numpy.ma.average
Weighted average.

Examples

>>> a = np.ma.array([1,2,3], mask=[False, False, True])

>>> a
masked_array(data = [1 2 --],
mask = [False False True],
fill_value = 999999)
>>> a.mean()
1.5

median(a, axis=None, out=None, overwrite_input=False)

Compute the median along the specified axis.
Returns the median of the array elements.
Parameters
a : array_like
Input array or object that can be converted to an array.
axis : int, optional
Axis along which the medians are computed. The default (None) is to compute the
median along a flattened version of the array.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
overwrite_input : bool, optional

1.6. Masked arrays 325

NumPy Reference, Release 2.0.0.dev8464

If True, then allow use of memory of input array (a) for calculations. The input array
will be modified by the call to median. This will save memory when you do not need to
preserve the contents of the input array. Treat the input as undefined, but it will probably
be fully or partially sorted. Default is False. Note that, if overwrite_input is True, and
the input is not already an ndarray, an error will be raised.
Returns
median : ndarray
A new array holding the result is returned unless out is specified, in which case a refer-
ence to out is returned. Return data-type is float64 for integers and floats smaller than
float64, or the input data-type, otherwise.
See Also:
mean

Notes
Given a vector V with N non masked values, the median of V is the middle value of a sorted copy of V (Vs) - i.e.
Vs[(N-1)/2], when N is odd, or {Vs[N/2 - 1] + Vs[N/2]}/2 when N is even.

Examples

>>> x = np.ma.array(np.arange(8), mask=[0]*4 + [1]*4)

>>> np.ma.extras.median(x)
1.5

>>> x = np.ma.array(np.arange(10).reshape(2, 5), mask=[0]*6 + [1]*4)

>>> np.ma.extras.median(x)
2.5
>>> np.ma.extras.median(x, axis=-1, overwrite_input=True)
masked_array(data = [ 2. 5.],
mask = False,
fill_value = 1e+20)

power(a, b, third=None)
Returns element-wise base array raised to power from second array.
This is the masked array version of numpy.power. For details see numpy.power.
See Also:
numpy.power

Notes
The out argument to numpy.power is not supported, third has to be None.
prod
Return the product of the array elements over the given axis. Masked elements are set to 1 internally for
computation.
Parameters
axis : {None, int}, optional
Axis over which the product is taken. If None is used, then the product is over all the
array elements.
dtype : {None, dtype}, optional

326 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Determines the type of the returned array and of the accumulator where the elements are
multiplied. If dtype has the value None and the type of a is an integer type of precision
less than the default platform integer, then the default platform integer precision is used.
Otherwise, the dtype is the same as that of a.
out : {None, array}, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary.
Returns
product_along_axis : {array, scalar}, see dtype parameter above.
Returns an array whose shape is the same as a with the specified axis removed. Returns
a 0d array when a is 1d or axis=None. Returns a reference to the specified output array
if specified.
See Also:

prod
equivalent function

Notes
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> np.prod([1.,2.])
2.0
>>> np.prod([1.,2.], dtype=np.int32)
2
>>> np.prod([[1.,2.],[3.,4.]])
24.0
>>> np.prod([[1.,2.],[3.,4.]], axis=1)
array([ 2., 12.])

std
Compute the standard deviation along the specified axis.
Returns the standard deviation, a measure of the spread of a distribution, of the array elements. The standard
deviation is computed for the flattened array by default, otherwise over the specified axis.
Parameters
a : array_like
Calculate the standard deviation of these values.
axis : int, optional
Axis along which the standard deviation is computed. The default is to compute the
standard deviation of the flattened array.
dtype : dtype, optional
Type to use in computing the standard deviation. For arrays of integer type the default
is float64, for arrays of float types it is the same as the array type.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type (of the calculated values) will be cast if necessary.

1.6. Masked arrays 327

NumPy Reference, Release 2.0.0.dev8464

ddof : int, optional

Means Delta Degrees of Freedom. The divisor used in calculations is N - ddof,
where N represents the number of elements. By default ddof is zero.
Returns
standard_deviation : ndarray, see dtype parameter above.
If out is None, return a new array containing the standard deviation, otherwise return a
reference to the output array.
See Also:
var, mean
numpy.doc.ufuncs
Section Output arguments

Notes
The standard deviation is the square root of the average of the squared deviations from the mean, i.e., std =
sqrt(mean(abs(x - x.mean())**2)).
The average squared deviation is normally calculated as x.sum() / N, where N = len(x). If, however,
ddof is specified, the divisor N - ddof is used instead. In standard statistical practice, ddof=1 provides an
unbiased estimator of the variance of the infinite population. ddof=0 provides a maximum likelihood estimate
of the variance for normally distributed variables. The standard deviation computed in this function is the
square root of the estimated variance, so even with ddof=1, it will not be an unbiased estimate of the standard
deviation per se.
Note that, for complex numbers, std takes the absolute value before squaring, so that the result is always real
and nonnegative.
For floating-point input, the std is computed using the same precision the input has. Depending on the input
data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a higher-
accuracy accumulator using the dtype keyword can alleviate this issue.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> np.std(a)
1.1180339887498949
>>> np.std(a, axis=0)
array([ 1., 1.])
>>> np.std(a, axis=1)
array([ 0.5, 0.5])

In single precision, std() can be inaccurate:

>>> a = np.zeros((2,512*512), dtype=np.float32)

>>> a[0,:] = 1.0
>>> a[1,:] = 0.1
>>> np.std(a)
0.45172946707416706

Computing the standard deviation in float64 is more accurate:

>>> np.std(a, dtype=np.float64)

0.44999999925552653

328 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

sum
Return the sum of the array elements over the given axis. Masked elements are set to 0 internally.
Parameters
axis : {None, -1, int}, optional
Axis along which the sum is computed. The default (axis = None) is to compute over
the flattened array.
dtype : {None, dtype}, optional
Determines the type of the returned array and of the accumulator where the elements
are summed. If dtype has the value None and the type of a is an integer type of precision
less than the default platform integer, then the default platform integer precision is used.
Otherwise, the dtype is the same as that of a.
out : {None, ndarray}, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
sum_along_axis : MaskedArray or scalar
An array with the same shape as self, with the specified axis removed. If self is a 0-d
array, or if axis is None, a scalar is returned. If an output array is specified, a reference
to out is returned.

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.sum()
25
>>> print x.sum(axis=1)
[4 5 16]
>>> print x.sum(axis=0)
[8 5 12]
>>> print type(x.sum(axis=0, dtype=np.int64)[0])
<type numpy.int64>

var
Compute the variance along the specified axis.
Returns the variance of the array elements, a measure of the spread of a distribution. The variance is computed
for the flattened array by default, otherwise over the specified axis.
Parameters
a : array_like
Array containing numbers whose variance is desired. If a is not an array, a conversion
is attempted.
axis : int, optional
Axis along which the variance is computed. The default is to compute the variance of
the flattened array.
dtype : dtype, optional

1.6. Masked arrays 329

NumPy Reference, Release 2.0.0.dev8464

Type to use in computing the variance. For arrays of integer type the default is float32;
for arrays of float types it is the same as the array type.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type is cast if necessary.
ddof : int, optional
Delta Degrees of Freedom: the divisor used in calculation is N - ddof, where N
represents the number of elements. By default ddof is zero.
Returns
variance : ndarray, see dtype parameter above
If out=None, returns a new array containing the variance; otherwise a reference to the
output array is returned.
See Also:

std
Standard deviation
mean
Average
numpy.doc.ufuncs
Section Output arguments

Notes
The variance is the average of the squared deviations from the mean, i.e., var = mean(abs(x -
x.mean())**2).
The mean is normally calculated as x.sum() / N, where N = len(x). If, however, ddof is specified, the
divisor N - ddof is used instead. In standard statistical practice, ddof=1 provides an unbiased estimator of
the variance of the infinite population. ddof=0 provides a maximum likelihood estimate of the variance for
normally distributed variables.
Note that for complex numbers, the absolute value is taken before squaring, so that the result is always real and
nonnegative.
For floating-point input, the variance is computed using the same precision the input has. Depending on the
input data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a
higher-accuracy accumulator using the dtype keyword can alleviate this issue.

Examples

>>> a = np.array([[1,2],[3,4]])
>>> np.var(a)
1.25
>>> np.var(a,0)
array([ 1., 1.])
>>> np.var(a,1)
array([ 0.25, 0.25])

In single precision, var() can be inaccurate:

330 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> a = np.zeros((2,512*512), dtype=np.float32)

>>> a[0,:] = 1.0
>>> a[1,:] = 0.1
>>> np.var(a)
0.20405951142311096

Computing the standard deviation in float64 is more accurate:

>>> np.var(a, dtype=np.float64)

0.20249999932997387
>>> ((1-0.55)**2 + (0.1-0.55)**2)/2
0.20250000000000001

anom(axis=None, dtype=None)
Compute the anomalies (deviations from the arithmetic mean) along the given axis.
Returns an array of anomalies, with the same shape as the input and where the arithmetic mean is computed
along the given axis.
Parameters
axis : int, optional
Axis over which the anomalies are taken. The default is to use the mean of the flattened
array as reference.
dtype : dtype, optional
Type to use in computing the variance. For arrays of integer type
the default is float32; for arrays of float types it is the same as the array type.
See Also:

mean
Compute the mean of the array.

Examples

>>> a = np.ma.array([1,2,3])
>>> a.anom()
masked_array(data = [-1. 0. 1.],
mask = False,
fill_value = 1e+20)

cumprod(axis=None, dtype=None, out=None)

Return the cumulative product of the elements along the given axis. The cumulative product is taken over the
flattened array by default, otherwise over the specified axis.
Masked values are set to 1 internally during the computation. However, their position is saved, and the result
will be masked at the same locations.
Parameters
axis : {None, -1, int}, optional
Axis along which the product is computed. The default (axis = None) is to compute
over the flattened array.
dtype : {None, dtype}, optional
Determines the type of the returned array and of the accumulator where the elements
are multiplied. If dtype has the value None and the type of a is an integer type

1.6. Masked arrays 331

NumPy Reference, Release 2.0.0.dev8464

of precision less than the default platform integer, then the default platform integer
precision is used. Otherwise, the dtype is the same as that of a.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
cumprod : ndarray
A new array holding the result is returned unless out is specified, in which case a refer-
ence to out is returned.

Notes
The mask is lost if out is not a valid MaskedArray !
Arithmetic is modular when using integer types, and no error is raised on overflow.
cumsum(axis=None, dtype=None, out=None)
Return the cumulative sum of the elements along the given axis. The cumulative sum is calculated over the
flattened array by default, otherwise over the specified axis.
Masked values are set to 0 internally during the computation. However, their position is saved, and the result
will be masked at the same locations.
Parameters
axis : {None, -1, int}, optional
Axis along which the sum is computed. The default (axis = None) is to compute over
the flattened array. axis may be negative, in which case it counts from the last to the
first axis.
dtype : {None, dtype}, optional
Type of the returned array and of the accumulator in which the elements are summed. If
dtype is not specified, it defaults to the dtype of a, unless a has an integer dtype with a
precision less than that of the default platform integer. In that case, the default platform
integer is used.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
cumsum : ndarray.
A new array holding the result is returned unless out is specified, in which case a
reference to out is returned.

Notes
The mask is lost if out is not a valid MaskedArray !
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> marr = np.ma.array(np.arange(10), mask=[0,0,0,1,1,1,0,0,0,0])

>>> print marr.cumsum()
[0 1 3 -- -- -- 9 16 24 33]

332 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

mean(axis=None, dtype=None, out=None)

Returns the average of the array elements.
Masked entries are ignored. The average is taken over the flattened array by default, otherwise over the specified
axis. Refer to numpy.mean for the full documentation.
Parameters
a : array_like
Array containing numbers whose mean is desired. If a is not an array, a conversion is
attempted.
axis : int, optional
Axis along which the means are computed. The default is to compute the mean of the
flattened array.
dtype : dtype, optional
Type to use in computing the mean. For integer inputs, the default is float64; for floating
point, inputs it is the same as the input dtype.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary.
Returns
mean : ndarray, see dtype parameter above
If out=None, returns a new array containing the mean values, otherwise a reference to
the output array is returned.
See Also:

numpy.ma.mean
Equivalent function.
numpy.mean
Equivalent function on non-masked arrays.
numpy.ma.average
Weighted average.

Examples

>>> a = np.ma.array([1,2,3], mask=[False, False, True])

>>> a
masked_array(data = [1 2 --],
mask = [False False True],
fill_value = 999999)
>>> a.mean()
1.5

prod(axis=None, dtype=None, out=None)

Return the product of the array elements over the given axis. Masked elements are set to 1 internally for
computation.
Parameters
axis : {None, int}, optional
Axis over which the product is taken. If None is used, then the product is over all the
array elements.

1.6. Masked arrays 333

NumPy Reference, Release 2.0.0.dev8464

dtype : {None, dtype}, optional

Determines the type of the returned array and of the accumulator where the elements are
multiplied. If dtype has the value None and the type of a is an integer type of precision
less than the default platform integer, then the default platform integer precision is used.
Otherwise, the dtype is the same as that of a.
out : {None, array}, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary.
Returns
product_along_axis : {array, scalar}, see dtype parameter above.
Returns an array whose shape is the same as a with the specified axis removed. Returns
a 0d array when a is 1d or axis=None. Returns a reference to the specified output array
if specified.
See Also:

prod
equivalent function

Notes
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> np.prod([1.,2.])
2.0
>>> np.prod([1.,2.], dtype=np.int32)
2
>>> np.prod([[1.,2.],[3.,4.]])
24.0
>>> np.prod([[1.,2.],[3.,4.]], axis=1)
array([ 2., 12.])

std(axis=None, dtype=None, out=None, ddof=0)

Compute the standard deviation along the specified axis.
Returns the standard deviation, a measure of the spread of a distribution, of the array elements. The standard
deviation is computed for the flattened array by default, otherwise over the specified axis.
Parameters
a : array_like
Calculate the standard deviation of these values.
axis : int, optional
Axis along which the standard deviation is computed. The default is to compute the
standard deviation of the flattened array.
dtype : dtype, optional
Type to use in computing the standard deviation. For arrays of integer type the default
is float64, for arrays of float types it is the same as the array type.
out : ndarray, optional

334 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Alternative output array in which to place the result. It must have the same shape as the
expected output but the type (of the calculated values) will be cast if necessary.
ddof : int, optional
Means Delta Degrees of Freedom. The divisor used in calculations is N - ddof,
where N represents the number of elements. By default ddof is zero.
Returns
standard_deviation : ndarray, see dtype parameter above.
If out is None, return a new array containing the standard deviation, otherwise return a
reference to the output array.
See Also:
var, mean
numpy.doc.ufuncs
Section Output arguments

Notes
The standard deviation is the square root of the average of the squared deviations from the mean, i.e., std =
sqrt(mean(abs(x - x.mean())**2)).
The average squared deviation is normally calculated as x.sum() / N, where N = len(x). If, however,
ddof is specified, the divisor N - ddof is used instead. In standard statistical practice, ddof=1 provides an
unbiased estimator of the variance of the infinite population. ddof=0 provides a maximum likelihood estimate
of the variance for normally distributed variables. The standard deviation computed in this function is the
square root of the estimated variance, so even with ddof=1, it will not be an unbiased estimate of the standard
deviation per se.
Note that, for complex numbers, std takes the absolute value before squaring, so that the result is always real
and nonnegative.
For floating-point input, the std is computed using the same precision the input has. Depending on the input
data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a higher-
accuracy accumulator using the dtype keyword can alleviate this issue.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> np.std(a)
1.1180339887498949
>>> np.std(a, axis=0)
array([ 1., 1.])
>>> np.std(a, axis=1)
array([ 0.5, 0.5])

In single precision, std() can be inaccurate:

>>> a = np.zeros((2,512*512), dtype=np.float32)

>>> a[0,:] = 1.0
>>> a[1,:] = 0.1
>>> np.std(a)
0.45172946707416706

Computing the standard deviation in float64 is more accurate:

1.6. Masked arrays 335

NumPy Reference, Release 2.0.0.dev8464

>>> np.std(a, dtype=np.float64)

0.44999999925552653

sum(axis=None, dtype=None, out=None)

Return the sum of the array elements over the given axis. Masked elements are set to 0 internally.
Parameters
axis : {None, -1, int}, optional
Axis along which the sum is computed. The default (axis = None) is to compute over
the flattened array.
dtype : {None, dtype}, optional
Determines the type of the returned array and of the accumulator where the elements
are summed. If dtype has the value None and the type of a is an integer type of precision
less than the default platform integer, then the default platform integer precision is used.
Otherwise, the dtype is the same as that of a.
out : {None, ndarray}, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
sum_along_axis : MaskedArray or scalar
An array with the same shape as self, with the specified axis removed. If self is a 0-d
array, or if axis is None, a scalar is returned. If an output array is specified, a reference
to out is returned.

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.sum()
25
>>> print x.sum(axis=1)
[4 5 16]
>>> print x.sum(axis=0)
[8 5 12]
>>> print type(x.sum(axis=0, dtype=np.int64)[0])
<type numpy.int64>

var(axis=None, dtype=None, out=None, ddof=0)

Compute the variance along the specified axis.
Returns the variance of the array elements, a measure of the spread of a distribution. The variance is computed
for the flattened array by default, otherwise over the specified axis.
Parameters
a : array_like
Array containing numbers whose variance is desired. If a is not an array, a conversion
is attempted.
axis : int, optional

336 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Axis along which the variance is computed. The default is to compute the variance of
the flattened array.
dtype : dtype, optional
Type to use in computing the variance. For arrays of integer type the default is float32;
for arrays of float types it is the same as the array type.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type is cast if necessary.
ddof : int, optional
Delta Degrees of Freedom: the divisor used in calculation is N - ddof, where N
represents the number of elements. By default ddof is zero.
Returns
variance : ndarray, see dtype parameter above
If out=None, returns a new array containing the variance; otherwise a reference to the
output array is returned.
See Also:

std
Standard deviation
mean
Average
numpy.doc.ufuncs
Section Output arguments

Notes
The variance is the average of the squared deviations from the mean, i.e., var = mean(abs(x -
x.mean())**2).
The mean is normally calculated as x.sum() / N, where N = len(x). If, however, ddof is specified, the
divisor N - ddof is used instead. In standard statistical practice, ddof=1 provides an unbiased estimator of
the variance of the infinite population. ddof=0 provides a maximum likelihood estimate of the variance for
normally distributed variables.
Note that for complex numbers, the absolute value is taken before squaring, so that the result is always real and
nonnegative.
For floating-point input, the variance is computed using the same precision the input has. Depending on the
input data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a
higher-accuracy accumulator using the dtype keyword can alleviate this issue.

Examples

>>> a = np.array([[1,2],[3,4]])
>>> np.var(a)
1.25
>>> np.var(a,0)
array([ 1., 1.])
>>> np.var(a,1)
array([ 0.25, 0.25])

1.6. Masked arrays 337

NumPy Reference, Release 2.0.0.dev8464

In single precision, var() can be inaccurate:

>>> a = np.zeros((2,512*512), dtype=np.float32)

>>> a[0,:] = 1.0
>>> a[1,:] = 0.1
>>> np.var(a)
0.20405951142311096

Computing the standard deviation in float64 is more accurate:

>>> np.var(a, dtype=np.float64)

0.20249999932997387
>>> ((1-0.55)**2 + (0.1-0.55)**2)/2
0.20250000000000001

Minimum/maximum
ma.argmax(a[, axis, fill_value]) Function version of the eponymous method.
ma.argmin(a[, axis, fill_value]) Returns array of indices of the maximum values along the
given axis.
ma.max(obj[, axis, out, fill_value]) Return the maximum along a given axis.
ma.min(obj[, axis, out, fill_value]) Return the minimum along a given axis.
ma.ptp(obj[, axis, out, fill_value]) Return (maximum - minimum) along the the given dimension
(i.e.
ma.MaskedArray.argmax([axis, fill_value, Returns array of indices of the maximum values along the
out]) given axis.
ma.MaskedArray.argmin([axis, fill_value, Return array of indices to the minimum values along the given
out]) axis.
ma.MaskedArray.max([axis, out, Return the maximum along a given axis.
fill_value])
ma.MaskedArray.min([axis, out, Return the minimum along a given axis.
fill_value])
ma.MaskedArray.ptp([axis, out, Return (maximum - minimum) along the the given dimension
fill_value]) (i.e.
argmax(a, axis=None, fill_value=None)
Function version of the eponymous method.
argmin(a, axis=None, fill_value=None)
Returns array of indices of the maximum values along the given axis. Masked values are treated as if they had
the value fill_value.
Parameters
axis : {None, integer}
If None, the index is into the flattened array, otherwise along the specified axis
fill_value : {var}, optional
Value used to fill in the masked values. If None, the output of maxi-
mum_fill_value(self._data) is used instead.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
Returns
index_array : {integer_array}

338 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.arange(6).reshape(2,3)
>>> a.argmax()
5
>>> a.argmax(0)
array([1, 1, 1])
>>> a.argmax(1)
array([2, 2])

max(obj, axis=None, out=None, fill_value=None)

Return the maximum along a given axis.
Parameters
axis : {None, int}, optional
Axis along which to operate. By default, axis is None and the flattened input is used.
out : array_like, optional
Alternative output array in which to place the result. Must be of the same shape and
buffer length as the expected output.
fill_value : {var}, optional
Value used to fill in the masked values. If None, use the output of maxi-
mum_fill_value().
Returns
amax : array_like
New array holding the result. If out was specified, out is returned.
See Also:

maximum_fill_value
Returns the maximum filling value for a given datatype.

min(obj, axis=None, out=None, fill_value=None)

Return the minimum along a given axis.
Parameters
axis : {None, int}, optional
Axis along which to operate. By default, axis is None and the flattened input is used.
out : array_like, optional
Alternative output array in which to place the result. Must be of the same shape and
buffer length as the expected output.
fill_value : {var}, optional
Value used to fill in the masked values. If None, use the output of minimum_fill_value.
Returns
amin : array_like
New array holding the result. If out was specified, out is returned.
See Also:

1.6. Masked arrays 339

NumPy Reference, Release 2.0.0.dev8464

minimum_fill_value
Returns the minimum filling value for a given datatype.

ptp(obj, axis=None, out=None, fill_value=None)

Return (maximum - minimum) along the the given dimension (i.e. peak-to-peak value).
Parameters
axis : {None, int}, optional
Axis along which to find the peaks. If None (default) the flattened array is used.
out : {None, array_like}, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
fill_value : {var}, optional
Value used to fill in the masked values.
Returns
ptp : ndarray.
A new array holding the result, unless out was specified, in which case a reference to
out is returned.
argmax(axis=None, fill_value=None, out=None)
Returns array of indices of the maximum values along the given axis. Masked values are treated as if they had
the value fill_value.
Parameters
axis : {None, integer}
If None, the index is into the flattened array, otherwise along the specified axis
fill_value : {var}, optional
Value used to fill in the masked values. If None, the output of maxi-
mum_fill_value(self._data) is used instead.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
Returns
index_array : {integer_array}

Examples

>>> a = np.arange(6).reshape(2,3)
>>> a.argmax()
5
>>> a.argmax(0)
array([1, 1, 1])
>>> a.argmax(1)
array([2, 2])

argmin(axis=None, fill_value=None, out=None)

Return array of indices to the minimum values along the given axis.
Parameters
axis : {None, integer}

340 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

If None, the index is into the flattened array, otherwise along the specified axis
fill_value : {var}, optional
Value used to fill in the masked values. If None, the output of mini-
mum_fill_value(self._data) is used instead.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
Returns
{ndarray, scalar} :
If multi-dimension input, returns a new ndarray of indices to the minimum values along
the given axis. Otherwise, returns a scalar of index to the minimum values along the
given axis.

Examples

>>> x = np.ma.array(arange(4), mask=[1,1,0,0])

>>> x.shape = (2,2)
>>> print x
[[-- --]
[2 3]]
>>> print x.argmin(axis=0, fill_value=-1)
[0 0]
>>> print x.argmin(axis=0, fill_value=9)
[1 1]

max(axis=None, out=None, fill_value=None)

Return the maximum along a given axis.
Parameters
axis : {None, int}, optional
Axis along which to operate. By default, axis is None and the flattened input is used.
out : array_like, optional
Alternative output array in which to place the result. Must be of the same shape and
buffer length as the expected output.
fill_value : {var}, optional
Value used to fill in the masked values. If None, use the output of maxi-
mum_fill_value().
Returns
amax : array_like
New array holding the result. If out was specified, out is returned.
See Also:

maximum_fill_value
Returns the maximum filling value for a given datatype.

min(axis=None, out=None, fill_value=None)

Return the minimum along a given axis.

1.6. Masked arrays 341

NumPy Reference, Release 2.0.0.dev8464

Parameters
axis : {None, int}, optional
Axis along which to operate. By default, axis is None and the flattened input is used.
out : array_like, optional
Alternative output array in which to place the result. Must be of the same shape and
buffer length as the expected output.
fill_value : {var}, optional
Value used to fill in the masked values. If None, use the output of minimum_fill_value.
Returns
amin : array_like
New array holding the result. If out was specified, out is returned.
See Also:

minimum_fill_value
Returns the minimum filling value for a given datatype.

ptp(axis=None, out=None, fill_value=None)

Return (maximum - minimum) along the the given dimension (i.e. peak-to-peak value).
Parameters
axis : {None, int}, optional
Axis along which to find the peaks. If None (default) the flattened array is used.
out : {None, array_like}, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
fill_value : {var}, optional
Value used to fill in the masked values.
Returns
ptp : ndarray.
A new array holding the result, unless out was specified, in which case a reference to
out is returned.

Sorting
ma.argsort(a[, axis, kind, order, fill_value]) Return an ndarray of indices that sort the array along the
specified axis.
ma.sort(a[, axis, kind, order, endwith, ...]) Sort the array, in-place
ma.MaskedArray.argsort([axis, kind, Return an ndarray of indices that sort the array along the
order, ...]) specified axis.
ma.MaskedArray.sort([axis, kind, order, Sort the array, in-place
...])
argsort(a, axis=None, kind=quicksort, order=None, fill_value=None)
Return an ndarray of indices that sort the array along the specified axis. Masked values are filled beforehand to
fill_value.
Parameters
axis : int, optional

342 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Axis along which to sort. The default is -1 (last axis). If None, the flattened array is
used.
fill_value : var, optional
Value used to fill the array before sorting. The default is the fill_value attribute of the
input array.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm.
order : list, optional
When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
Returns
index_array : ndarray, int
Array of indices that sort a along the specified axis. In other words,
a[index_array] yields a sorted a.
See Also:

sort
Describes sorting algorithms used.
lexsort
Indirect stable sort with multiple keys.
ndarray.sort
Inplace sort.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = np.ma.array([3,2,1], mask=[False, False, True])

>>> a
masked_array(data = [3 2 --],
mask = [False False True],
fill_value = 999999)
>>> a.argsort()
array([1, 0, 2])

sort(a, axis=-1, kind=quicksort, order=None, endwith=True, fill_value=None)

Sort the array, in-place
Parameters
a : array_like
Array to be sorted.
axis : int, optional
Axis along which to sort. If None, the array is flattened before sorting. The default is
-1, which sorts along the last axis.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm. Default is quicksort.

1.6. Masked arrays 343

NumPy Reference, Release 2.0.0.dev8464

order : list, optional

When a is a structured array, this argument specifies which fields to compare first, sec-
ond, and so on. This list does not need to include all of the fields.
endwith : {True, False}, optional
Whether missing values (if any) should be forced in the upper indices (at the end of the
array) (True) or lower indices (at the beginning).
fill_value : {var}, optional
Value used internally for the masked values. If fill_value is not None, it supersedes
endwith.
Returns
sorted_array : ndarray
Array of the same type and shape as a.
See Also:

ndarray.sort
Method to sort an array in-place.
argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in a sorted array.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # Default
>>> a.sort()
>>> print a
[1 3 5 -- --]

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # Put missing values in the front
>>> a.sort(endwith=False)
>>> print a
[-- -- 1 3 5]

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # fill_value takes over endwith
>>> a.sort(endwith=False, fill_value=3)
>>> print a
[1 -- -- 3 5]

344 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

argsort(axis=None, kind=quicksort, order=None, fill_value=None)

Return an ndarray of indices that sort the array along the specified axis. Masked values are filled beforehand to
fill_value.
Parameters
axis : int, optional
Axis along which to sort. The default is -1 (last axis). If None, the flattened array is
used.
fill_value : var, optional
Value used to fill the array before sorting. The default is the fill_value attribute of the
input array.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm.
order : list, optional
When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
Returns
index_array : ndarray, int
Array of indices that sort a along the specified axis. In other words,
a[index_array] yields a sorted a.
See Also:

sort
Describes sorting algorithms used.
lexsort
Indirect stable sort with multiple keys.
ndarray.sort
Inplace sort.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = np.ma.array([3,2,1], mask=[False, False, True])

>>> a
masked_array(data = [3 2 --],
mask = [False False True],
fill_value = 999999)
>>> a.argsort()
array([1, 0, 2])

sort(axis=-1, kind=quicksort, order=None, endwith=True, fill_value=None)

Sort the array, in-place
Parameters
a : array_like
Array to be sorted.

1.6. Masked arrays 345

NumPy Reference, Release 2.0.0.dev8464

axis : int, optional

Axis along which to sort. If None, the array is flattened before sorting. The default is
-1, which sorts along the last axis.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm. Default is quicksort.
order : list, optional
When a is a structured array, this argument specifies which fields to compare first, sec-
ond, and so on. This list does not need to include all of the fields.
endwith : {True, False}, optional
Whether missing values (if any) should be forced in the upper indices (at the end of the
array) (True) or lower indices (at the beginning).
fill_value : {var}, optional
Value used internally for the masked values. If fill_value is not None, it supersedes
endwith.
Returns
sorted_array : ndarray
Array of the same type and shape as a.
See Also:

ndarray.sort
Method to sort an array in-place.
argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in a sorted array.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # Default
>>> a.sort()
>>> print a
[1 3 5 -- --]

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # Put missing values in the front
>>> a.sort(endwith=False)
>>> print a
[-- -- 1 3 5]

346 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # fill_value takes over endwith
>>> a.sort(endwith=False, fill_value=3)
>>> print a
[1 -- -- 3 5]

Algebra
ma.diag(v[, k]) Extract a diagonal or construct a diagonal array.
ma.dot(a, b[, strict]) Return the dot product of two arrays.
ma.identity(n[, dtype]) Return the identity array.
ma.inner(a, b) Inner product of two arrays.
ma.innerproduct(a, b) Inner product of two arrays.
ma.outer(a, b) Compute the outer product of two vectors.
ma.outerproduct(a, b) Compute the outer product of two vectors.
ma.trace(self[, offset, axis1, axis2, ...]) Return the sum along diagonals of the array.
ma.transpose(a[, axes]) Permute the dimensions of an array.
ma.MaskedArray.trace([offset, axis1, axis2, ...]) Return the sum along diagonals of the array.
ma.MaskedArray.transpose(*axes) Returns a view of the array with axes transposed.
diag(v, k=0)
Extract a diagonal or construct a diagonal array.
This function is the equivalent of numpy.diag that takes masked values into account, see numpy.diag for
details.
See Also:

numpy.diag
Equivalent function for ndarrays.

dot(a, b, strict=False)
Return the dot product of two arrays.
Note: Works only with 2-D arrays at the moment.
This function is the equivalent of numpy.dot that takes masked values into account, see numpy.dot for
details.
Parameters
a, b : ndarray
Inputs arrays.
strict : bool, optional
Whether masked data are propagated (True) or set to 0 (False) for the computation.
Default is False. Propagating the mask means that if a masked value appears in a row
or column, the whole row or column is considered masked.
See Also:

numpy.dot
Equivalent function for ndarrays.

1.6. Masked arrays 347

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = ma.array([[1, 2, 3], [4, 5, 6]], mask=[[1, 0, 0], [0, 0, 0]])

>>> b = ma.array([[1, 2], [3, 4], [5, 6]], mask=[[1, 0], [0, 0], [0, 0]])
>>> np.ma.dot(a, b)
masked_array(data =
[[21 26]
[45 64]],
mask =
[[False False]
[False False]],
fill_value = 999999)
>>> np.ma.dot(a, b, strict=True)
masked_array(data =
[[-- --]
[-- 64]],
mask =
[[ True True]
[ True False]],
fill_value = 999999)

identity
Return the identity array.
The identity array is a square array with ones on the main diagonal.
Parameters
n : int
Number of rows (and columns) in n x n output.
dtype : data-type, optional
Data-type of the output. Defaults to float.
Returns
out : ndarray
n x n array with its main diagonal set to one, and all other elements 0.

Examples

>>> np.identity(3)
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])

inner(a, b)
Inner product of two arrays.
Ordinary inner product of vectors for 1-D arrays (without complex conjugation), in higher dimensions a sum
product over the last axes.
Parameters
a, b : array_like
If a and b are nonscalar, their last dimensions of must match.
Returns
out : ndarray
out.shape = a.shape[:-1] + b.shape[:-1]

348 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Raises
ValueError :
If the last dimension of a and b has different size.
See Also:

tensordot
Sum products over arbitrary axes.
dot
Generalised matrix product, using second last dimension of b.

Notes
Masked values are replaced by 0.

Examples
Ordinary inner product for vectors:

>>> a = np.array([1,2,3])
>>> b = np.array([0,1,0])
>>> np.inner(a, b)
2

A multidimensional example:

>>> a = np.arange(24).reshape((2,3,4))
>>> b = np.arange(4)
>>> np.inner(a, b)
array([[ 14, 38, 62],
[ 86, 110, 134]])

An example where b is a scalar:

>>> np.inner(np.eye(2), 7)
array([[ 7., 0.],
[ 0., 7.]])

innerproduct(a, b)
Inner product of two arrays.
Ordinary inner product of vectors for 1-D arrays (without complex conjugation), in higher dimensions a sum
product over the last axes.
Parameters
a, b : array_like
If a and b are nonscalar, their last dimensions of must match.
Returns
out : ndarray
out.shape = a.shape[:-1] + b.shape[:-1]
Raises
ValueError :
If the last dimension of a and b has different size.
See Also:

1.6. Masked arrays 349

NumPy Reference, Release 2.0.0.dev8464

tensordot
Sum products over arbitrary axes.
dot
Generalised matrix product, using second last dimension of b.

Notes
Masked values are replaced by 0.

Examples
Ordinary inner product for vectors:

>>> a = np.array([1,2,3])
>>> b = np.array([0,1,0])
>>> np.inner(a, b)
2

A multidimensional example:

>>> a = np.arange(24).reshape((2,3,4))
>>> b = np.arange(4)
>>> np.inner(a, b)
array([[ 14, 38, 62],
[ 86, 110, 134]])

An example where b is a scalar:

>>> np.inner(np.eye(2), 7)
array([[ 7., 0.],
[ 0., 7.]])

outer(a, b)
Compute the outer product of two vectors.
Given two vectors, a = [a0, a1, ..., aM] and b = [b0, b1, ..., bN], the outer product [R57]
is:

[[a0*b0 a0*b1 ... a0*bN ]

[a1*b0 .
[ ... .
[aM*b0 aM*bN ]]

Parameters
a, b : array_like, shape (M,), (N,)
First and second input vectors. Inputs are flattened if they are not already 1-dimensional.
Returns
out : ndarray, shape (M, N)
out[i, j] = a[i] * b[j]

Notes
Masked values are replaced by 0.

350 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

References
[R57]

Examples
Make a (very coarse) grid for computing a Mandelbrot set:

>>> rl = np.outer(np.ones((5,)), np.linspace(-2, 2, 5))

>>> rl
array([[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.]])
>>> im = np.outer(1j*np.linspace(2, -2, 5), np.ones((5,)))
>>> im
array([[ 0.+2.j, 0.+2.j, 0.+2.j, 0.+2.j, 0.+2.j],
[ 0.+1.j, 0.+1.j, 0.+1.j, 0.+1.j, 0.+1.j],
[ 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[ 0.-1.j, 0.-1.j, 0.-1.j, 0.-1.j, 0.-1.j],
[ 0.-2.j, 0.-2.j, 0.-2.j, 0.-2.j, 0.-2.j]])
>>> grid = rl + im
>>> grid
array([[-2.+2.j, -1.+2.j, 0.+2.j, 1.+2.j, 2.+2.j],
[-2.+1.j, -1.+1.j, 0.+1.j, 1.+1.j, 2.+1.j],
[-2.+0.j, -1.+0.j, 0.+0.j, 1.+0.j, 2.+0.j],
[-2.-1.j, -1.-1.j, 0.-1.j, 1.-1.j, 2.-1.j],
[-2.-2.j, -1.-2.j, 0.-2.j, 1.-2.j, 2.-2.j]])

An example using a vector of letters:

>>> x = np.array([a, b, c], dtype=object)

>>> np.outer(x, [1, 2, 3])
array([[a, aa, aaa],
[b, bb, bbb],
[c, cc, ccc]], dtype=object)

outerproduct(a, b)
Compute the outer product of two vectors.
Given two vectors, a = [a0, a1, ..., aM] and b = [b0, b1, ..., bN], the outer product [R58]
is:

[[a0*b0 a0*b1 ... a0*bN ]

[a1*b0 .
[ ... .
[aM*b0 aM*bN ]]

Parameters
a, b : array_like, shape (M,), (N,)
First and second input vectors. Inputs are flattened if they are not already 1-dimensional.
Returns
out : ndarray, shape (M, N)
out[i, j] = a[i] * b[j]

1.6. Masked arrays 351

NumPy Reference, Release 2.0.0.dev8464

Notes
Masked values are replaced by 0.

References
[R58]

Examples
Make a (very coarse) grid for computing a Mandelbrot set:

>>> rl = np.outer(np.ones((5,)), np.linspace(-2, 2, 5))

>>> rl
array([[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.]])
>>> im = np.outer(1j*np.linspace(2, -2, 5), np.ones((5,)))
>>> im
array([[ 0.+2.j, 0.+2.j, 0.+2.j, 0.+2.j, 0.+2.j],
[ 0.+1.j, 0.+1.j, 0.+1.j, 0.+1.j, 0.+1.j],
[ 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[ 0.-1.j, 0.-1.j, 0.-1.j, 0.-1.j, 0.-1.j],
[ 0.-2.j, 0.-2.j, 0.-2.j, 0.-2.j, 0.-2.j]])
>>> grid = rl + im
>>> grid
array([[-2.+2.j, -1.+2.j, 0.+2.j, 1.+2.j, 2.+2.j],
[-2.+1.j, -1.+1.j, 0.+1.j, 1.+1.j, 2.+1.j],
[-2.+0.j, -1.+0.j, 0.+0.j, 1.+0.j, 2.+0.j],
[-2.-1.j, -1.-1.j, 0.-1.j, 1.-1.j, 2.-1.j],
[-2.-2.j, -1.-2.j, 0.-2.j, 1.-2.j, 2.-2.j]])

An example using a vector of letters:

>>> x = np.array([a, b, c], dtype=object)

>>> np.outer(x, [1, 2, 3])
array([[a, aa, aaa],
[b, bb, bbb],
[c, cc, ccc]], dtype=object)

trace

Return the sum along diagonals of the array.

Refer to numpy.trace for full documentation.
See Also:

numpy.trace
equivalent function

transpose(a, axes=None)
Permute the dimensions of an array.
This function is exactly equivalent to numpy.transpose.
See Also:

352 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

numpy.transpose
Equivalent function in top-level NumPy module.

Examples

>>> import numpy.ma as ma

>>> x = ma.arange(4).reshape((2,2))
>>> x[1, 1] = ma.masked
>>>> x
masked_array(data =
[[0 1]
[2 --]],
mask =
[[False False]
[False True]],
fill_value = 999999)
>>> ma.transpose(x)
masked_array(data =
[[0 2]
[1 --]],
mask =
[[False False]
[False True]],
fill_value = 999999)

trace(offset=0, axis1=0, axis2=1, dtype=None, out=None)

Return the sum along diagonals of the array.
Refer to numpy.trace for full documentation.
See Also:

numpy.trace
equivalent function

transpose(*axes)
Returns a view of the array with axes transposed.
For a 1-D array, this has no effect. (To change between column and row vectors, first cast the 1-D array into a ma-
trix object.) For a 2-D array, this is the usual matrix transpose. For an n-D array, if axes are given, their order in-
dicates how the axes are permuted (see Examples). If axes are not provided and a.shape = (i[0], i[1],
... i[n-2], i[n-1]), then a.transpose().shape = (i[n-1], i[n-2], ... i[1],
i[0]).
Parameters
axes : None, tuple of ints, or n ints
None or no argument: reverses the order of the axes.
tuple of ints: i in the j-th place in the tuple means as i-th axis becomes a.transpose()s j-th
axis.
n ints: same as an n-tuple of the same ints (this form is intended simply as a convenience
alternative to the tuple form)
Returns
out : ndarray
View of a, with axes suitably permuted.

1.6. Masked arrays 353

NumPy Reference, Release 2.0.0.dev8464

See Also:

ndarray.T
Array property returning the array transposed.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> a
array([[1, 2],
[3, 4]])
>>> a.transpose()
array([[1, 3],
[2, 4]])
>>> a.transpose((1, 0))
array([[1, 3],
[2, 4]])
>>> a.transpose(1, 0)
array([[1, 3],
[2, 4]])

Polynomial fit
ma.vander(x[, n]) Generate a Van der Monde matrix.
ma.polyfit(x, y, deg[, rcond, full]) Least squares polynomial fit.
vander(x, n=None)
Generate a Van der Monde matrix.
The columns of the output matrix are decreasing powers of the input vector. Specifically, the i-th output column
is the input vector to the power of N - i - 1. Such a matrix with a geometric progression in each row is
named Van Der Monde, or Vandermonde matrix, from Alexandre-Theophile Vandermonde.
Parameters
x : array_like
1-D input array.
N : int, optional
Order of (number of columns in) the output. If N is not specified, a square array is
returned (N = len(x)).
Returns
out : ndarray
Van der Monde matrix of order N. The first column is x^(N-1), the second x^(N-2)
and so forth.

Notes
Masked values in the input array result in rows of zeros.

References
[R61]

354 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.array([1, 2, 3, 5])

>>> N = 3
>>> np.vander(x, N)
array([[ 1, 1, 1],
[ 4, 2, 1],
[ 9, 3, 1],
[25, 5, 1]])

>>> np.column_stack([x**(N-1-i) for i in range(N)])

array([[ 1, 1, 1],
[ 4, 2, 1],
[ 9, 3, 1],
[25, 5, 1]])

>>> x = np.array([1, 2, 3, 5])

>>> np.vander(x)
array([[ 1, 1, 1, 1],
[ 8, 4, 2, 1],
[ 27, 9, 3, 1],
[125, 25, 5, 1]])

The determinant of a square Vandermonde matrix is the product of the differences between the values of the
input vector:

>>> np.linalg.det(np.vander(x))
48.000000000000043
>>> (5-3)*(5-2)*(5-1)*(3-2)*(3-1)*(2-1)
48

polyfit(x, y, deg, rcond=None, full=False)

Least squares polynomial fit.
Fit a polynomial p(x) = p[0] * x**deg + ... + p[deg] of degree deg to points (x, y). Returns a
vector of coefficients p that minimises the squared error.
Parameters
x : array_like, shape (M,)
x-coordinates of the M sample points (x[i], y[i]).
y : array_like, shape (M,) or (M, K)
y-coordinates of the sample points. Several data sets of sample points sharing the same
x-coordinates can be fitted at once by passing in a 2D-array that contains one dataset
per column.
deg : int
Degree of the fitting polynomial
rcond : float, optional
Relative condition number of the fit. Singular values smaller than this relative to the
largest singular value will be ignored. The default value is len(x)*eps, where eps is the
relative precision of the float type, about 2e-16 in most cases.
full : bool, optional

1.6. Masked arrays 355

NumPy Reference, Release 2.0.0.dev8464

Switch determining nature of return value. When it is False (the default) just the coef-
ficients are returned, when True diagnostic information from the singular value decom-
position is also returned.
Returns
p : ndarray, shape (M,) or (M, K)
Polynomial coefficients, highest power first. If y was 2-D, the coefficients for k-th data
set are in p[:,k].
residuals, rank, singular_values, rcond : present only if full = True
Residuals of the least-squares fit, the effective rank of the scaled Vandermonde coeffi-
cient matrix, its singular values, and the specified value of rcond. For more details, see
linalg.lstsq.
See Also:

polyval
Computes polynomial values.
linalg.lstsq
Computes a least-squares fit.
scipy.interpolate.UnivariateSpline
Computes spline fits.

Notes
Any masked values in x is propagated in y, and vice-versa.

References
[R59], [R60]

Examples

>>> x = np.array([0.0, 1.0, 2.0, 3.0, 4.0, 5.0])

>>> y = np.array([0.0, 0.8, 0.9, 0.1, -0.8, -1.0])
>>> z = np.polyfit(x, y, 3)
>>> z
array([ 0.08703704, -0.81349206, 1.69312169, -0.03968254])

It is convenient to use poly1d objects for dealing with polynomials:

>>> p = np.poly1d(z)
>>> p(0.5)
0.6143849206349179
>>> p(3.5)
-0.34732142857143039
>>> p(10)
22.579365079365115

High-order polynomials may oscillate wildly:

>>> p30 = np.poly1d(np.polyfit(x, y, 30))

/... RankWarning: Polyfit may be poorly conditioned...
>>> p30(4)
-0.80000000000000204
>>> p30(5)

356 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

-0.99999999999999445
>>> p30(4.5)
-0.10547061179440398

Illustration:

>>> import matplotlib.pyplot as plt

>>> xp = np.linspace(-2, 6, 100)
>>> plt.plot(x, y, ., xp, p(xp), -, xp, p30(xp), --)
[<matplotlib.lines.Line2D object at 0x...>, <matplotlib.lines.Line2D object at 0x...>, <matplotl
>>> plt.ylim(-2,2)
(-2, 2)
>>> plt.show()

2.0
1.5
1.0
0.5
0.0
0.5
1.0
1.5
2.0
2 1 0 1 2 3 4 5 6

Clipping and rounding

ma.around Round an array to the given number of decimals.
ma.clip(a, a_min, a_max[, out]) Clip (limit) the values in an array.
ma.round(a[, decimals, out]) Return a copy of a, rounded to decimals places.
ma.MaskedArray.clip(a_min, a_max[, out]) Return an array whose values are limited to [a_min, a_max].
ma.MaskedArray.round([decimals, out]) Return an array rounded a to the given number of decimals.
around
Round an array to the given number of decimals.
Refer to around for full documentation.
See Also:

around
equivalent function

clip(a, a_min, a_max, out=None)

Clip (limit) the values in an array.
Given an interval, values outside the interval are clipped to the interval edges. For example, if an interval of
[0, 1] is specified, values smaller than 0 become 0, and values larger than 1 become 1.

1.6. Masked arrays 357

NumPy Reference, Release 2.0.0.dev8464

Parameters
a : array_like
Array containing elements to clip.
a_min : scalar or array_like
Minimum value.
a_max : scalar or array_like
Maximum value. If a_min or a_max are array_like, then they will be broadcasted to the
shape of a.
out : ndarray, optional
The results will be placed in this array. It may be the input array for in-place clipping.
out must be of the right shape to hold the output. Its type is preserved.
Returns
clipped_array : ndarray
An array with the elements of a, but where values < a_min are replaced with a_min, and
those > a_max with a_max.
See Also:

numpy.doc.ufuncs
Section Output arguments

Examples

>>> a = np.arange(10)
>>> np.clip(a, 1, 8)
array([1, 1, 2, 3, 4, 5, 6, 7, 8, 8])
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> np.clip(a, 3, 6, out=a)
array([3, 3, 3, 3, 4, 5, 6, 6, 6, 6])
>>> a = np.arange(10)
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> np.clip(a, [3,4,1,1,1,4,4,4,4,4], 8)
array([3, 4, 2, 3, 4, 5, 6, 7, 8, 8])

round(a, decimals=0, out=None)

Return a copy of a, rounded to decimals places.
When decimals is negative, it specifies the number of positions to the left of the decimal point. The real and
imaginary parts of complex numbers are rounded separately. Nothing is done if the array is not of float type and
decimals is greater than or equal to 0.
Parameters
decimals : int
Number of decimals to round to. May be negative.
out : array_like
Existing array to use for output. If not given, returns a default copy of a.

358 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

Notes
If out is given and does not have a mask attribute, the mask of a is lost!
clip(a_min, a_max, out=None)
Return an array whose values are limited to [a_min, a_max].
Refer to numpy.clip for full documentation.
See Also:

numpy.clip
equivalent function

round(decimals=0, out=None)
Return an array rounded a to the given number of decimals.
Refer to numpy.around for full documentation.
See Also:

numpy.around
equivalent function

Miscellanea
ma.allequal(a, b[, fill_value]) Return True if all entries of a and b are equal, using
ma.allclose(a, b[, masked_equal, rtol, Returns True if two arrays are element-wise equal within a
...]) tolerance.
ma.apply_along_axis(func1d, axis, Apply a function to 1-D slices along the given axis.
arr, ...)
ma.arange([dtype]) Return evenly spaced values within a given interval.
ma.choose(indices, choices[, out, mode]) Use an index array to construct a new array from a set of choices.
ma.ediff1d(arr[, to_end, to_begin]) Compute the differences between consecutive elements of an array.
ma.indices(dimensions[, dtype]) Return an array representing the indices of a grid.
ma.where(condition[, x, y]) Return a masked array with elements from x or y, depending on
condition.
allequal(a, b, fill_value=True)
Return True if all entries of a and b are equal, using fill_value as a truth value where either or both are masked.
Parameters
a, b : array_like
Input arrays to compare.
fill_value : bool, optional
Whether masked values in a or b are considered equal (True) or not (False).
Returns
y : bool
Returns True if the two arrays are equal within the given tolerance, False otherwise. If
either array contains NaN, then False is returned.
See Also:
all, any, numpy.ma.allclose

1.6. Masked arrays 359

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = ma.array([1e10, 1e-7, 42.0], mask=[0, 0, 1])

>>> a
masked_array(data = [10000000000.0 1e-07 --],
mask = [False False True],
fill_value=1e+20)

>>> b = array([1e10, 1e-7, -42.0])

>>> b
array([ 1.00000000e+10, 1.00000000e-07, -4.20000000e+01])
>>> ma.allequal(a, b, fill_value=False)
False
>>> ma.allequal(a, b)
True

allclose(a, b, masked_equal=True, rtol=1.0000000000000001e-05, atol=1e-08, fill_value=None)

Returns True if two arrays are element-wise equal within a tolerance.
This function is equivalent to allclose except that masked values are treated as equal (default) or unequal,
depending on the masked_equal argument.
Parameters
a, b : array_like
Input arrays to compare.
masked_equal : bool, optional
Whether masked values in a and b are considered equal (True) or not (False). They are
considered equal by default.
rtol : float, optional
Relative tolerance. The relative difference is equal to rtol * b. Default is 1e-5.
atol : float, optional
Absolute tolerance. The absolute difference is equal to atol. Default is 1e-8.
fill_value : bool, optional
Deprecated - Whether masked values in a or b are considered equal (True) or not
(False).
Returns
y : bool
Returns True if the two arrays are equal within the given tolerance, False otherwise. If
either array contains NaN, then False is returned.
See Also:
all, any
numpy.allclose
the non-masked allclose.

Notes
If the following equation is element-wise True, then allclose returns True:

360 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

absolute(a - b) <= (atol + rtol * absolute(b))

Return True if all elements of a and b are equal subject to given tolerances.

Examples

>>> a = ma.array([1e10, 1e-7, 42.0], mask=[0, 0, 1])

>>> a
masked_array(data = [10000000000.0 1e-07 --],
mask = [False False True],
fill_value = 1e+20)
>>> b = ma.array([1e10, 1e-8, -42.0], mask=[0, 0, 1])
>>> ma.allclose(a, b)
False

>>> a = ma.array([1e10, 1e-8, 42.0], mask=[0, 0, 1])

>>> b = ma.array([1.00001e10, 1e-9, -42.0], mask=[0, 0, 1])
>>> ma.allclose(a, b)
True
>>> ma.allclose(a, b, masked_equal=False)
False

Masked values are not compared directly.

>>> a = ma.array([1e10, 1e-8, 42.0], mask=[0, 0, 1])

>>> b = ma.array([1.00001e10, 1e-9, 42.0], mask=[0, 0, 1])
>>> ma.allclose(a, b)
True
>>> ma.allclose(a, b, masked_equal=False)
False

apply_along_axis(func1d, axis, arr, *args, **kwargs)

Apply a function to 1-D slices along the given axis.
Execute func1d(a, *args) where func1d operates on 1-D arrays and a is a 1-D slice of arr along axis.
Parameters
func1d : function
This function should accept 1-D arrays. It is applied to 1-D slices of arr along the
specified axis.
axis : integer
Axis along which arr is sliced.
arr : ndarray
Input array.
args : any
Additional arguments to func1d.
Returns
outarr : ndarray
The output array. The shape of outarr is identical to the shape of arr, except along the
axis dimension, where the length of outarr is equal to the size of the return value of
func1d. If func1d returns a scalar outarr will have one fewer dimensions than arr.

1.6. Masked arrays 361

NumPy Reference, Release 2.0.0.dev8464

See Also:

apply_over_axes
Apply a function repeatedly over multiple axes.

Examples

>>> def my_func(a):

... """Average first and last element of a 1-D array"""
... return (a[0] + a[-1]) * 0.5
>>> b = np.array([[1,2,3], [4,5,6], [7,8,9]])
>>> np.apply_along_axis(my_func, 0, b)
array([ 4., 5., 6.])
>>> np.apply_along_axis(my_func, 1, b)
array([ 2., 5., 8.])

For a function that doesnt return a scalar, the number of dimensions in outarr is the same as arr.

>>> def new_func(a):

... """Divide elements of a by 2."""
... return a * 0.5
>>> b = np.array([[1,2,3], [4,5,6], [7,8,9]])
>>> np.apply_along_axis(new_func, 0, b)
array([[ 0.5, 1. , 1.5],
[ 2. , 2.5, 3. ],
[ 3.5, 4. , 4.5]])

arange
Return evenly spaced values within a given interval.
Values are generated within the half-open interval [start, stop) (in other words, the interval including
start but excluding stop). For integer arguments the function is equivalent to the Python built-in range function,
but returns a ndarray rather than a list.
Parameters
start : number, optional
Start of interval. The interval includes this value. The default start value is 0.
stop : number
End of interval. The interval does not include this value.
step : number, optional
Spacing between values. For any output out, this is the distance between two adjacent
values, out[i+1] - out[i]. The default step size is 1. If step is specified, start
must also be given.
dtype : dtype
The type of the output array. If dtype is not given, infer the data type from the other
input arguments.
Returns
out : ndarray
Array of evenly spaced values.

362 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

For floating point arguments, the length of the result is ceil((stop -

start)/step). Because of floating point overflow, this rule may result in the last
element of out being greater than stop.
See Also:

linspace
Evenly spaced numbers with careful handling of endpoints.
ogrid
Arrays of evenly spaced numbers in N-dimensions
mgrid
Grid-shaped arrays of evenly spaced numbers in N-dimensions

Examples

>>> np.arange(3)
array([0, 1, 2])
>>> np.arange(3.0)
array([ 0., 1., 2.])
>>> np.arange(3,7)
array([3, 4, 5, 6])
>>> np.arange(3,7,2)
array([3, 5])

choose(indices, choices, out=None, mode=raise)

Use an index array to construct a new array from a set of choices.
Given an array of integers and a set of n choice arrays, this method will create a new array that merges each of
the choice arrays. Where a value in a is i, the new array will have the value that choices[i] contains in the same
place.
Parameters
a : ndarray of ints
This array must contain integers in [0, n-1], where n is the number of choices.
choices : sequence of arrays
Choice arrays. The index array and all of the choices should be broadcastable to the
same shape.
out : array, optional
If provided, the result will be inserted into this array. It should be of the appropriate
shape and dtype.
mode : {raise, wrap, clip}, optional
Specifies how out-of-bounds indices will behave.
raise : raise an error
wrap : wrap around
clip : clip to the range
Returns
merged_array : array
See Also:

1.6. Masked arrays 363

NumPy Reference, Release 2.0.0.dev8464

choose
equivalent function

Examples

>>> choice = np.array([[1,1,1], [2,2,2], [3,3,3]])

>>> a = np.array([2, 1, 0])
>>> np.ma.choose(a, choice)
masked_array(data = [3 2 1],
mask = False,
fill_value=999999)

ediff1d(arr, to_end=None, to_begin=None)

Compute the differences between consecutive elements of an array.
This function is the equivalent of numpy.ediff1d that takes masked values into account, see
numpy.ediff1d for details.
See Also:

numpy.ediff1d
Equivalent function for ndarrays.

indices(dimensions, dtype=<type int>)

Return an array representing the indices of a grid.
Compute an array where the subarrays contain index values 0,1,... varying only along the corresponding axis.
Parameters
dimensions : sequence of ints
The shape of the grid.
dtype : dtype, optional
Data type of the result.
Returns
grid : ndarray
The array of grid indices, grid.shape = (len(dimensions),) +
tuple(dimensions).
See Also:
mgrid, meshgrid

Notes
The output shape is obtained by prepending the number of dimensions in front of the tuple of dimensions, i.e. if
dimensions is a tuple (r0, ..., rN-1) of length N, the output shape is (N,r0,...,rN-1).
The subarrays grid[k] contains the N-D array of indices along the k-th axis. Explicitly:

grid[k,i0,i1,...,iN-1] = ik

Examples

>>> grid = np.indices((2, 3))

>>> grid.shape
(2, 2, 3)

364 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

>>> grid[0] # row indices

array([[0, 0, 0],
[1, 1, 1]])
>>> grid[1] # column indices
array([[0, 1, 2],
[0, 1, 2]])

The indices can be used as an index into an array.

>>> x = np.arange(20).reshape(5, 4)
>>> row, col = np.indices((2, 3))
>>> x[row, col]
array([[0, 1, 2],
[4, 5, 6]])

Note that it would be more straightforward in the above example to extract the required elements directly with
x[:2, :3].
where(condition, x=None, y=None)
Return a masked array with elements from x or y, depending on condition.
Returns a masked array, shaped like condition, where the elements are from x when condition is True, and from
y otherwise. If neither x nor y are given, the function returns a tuple of indices where condition is True (the
result of condition.nonzero()).
Parameters
condition : array_like, bool
The condition to meet. For each True element, yield the corresponding element from x,
otherwise from y.
x, y : array_like, optional
Values from which to choose. x and y need to have the same shape as condition, or be
broadcast-able to that shape.
Returns
out : MaskedArray or tuple of ndarrays
The resulting masked array if x and y were given, otherwise the result of
condition.nonzero().
See Also:

numpy.where
Equivalent function in the top-level NumPy module.

Examples

>>> x = np.ma.array(np.arange(9.).reshape(3, 3), mask=[[0, 1, 0],

... [1, 0, 1],
... [0, 1, 0]])
>>> print x
[[0.0 -- 2.0]
[-- 4.0 --]
[6.0 -- 8.0]]
>>> np.ma.where(x > 5) # return the indices where x > 5
(array([2, 2]), array([0, 2]))

1.6. Masked arrays 365

NumPy Reference, Release 2.0.0.dev8464

>>> print np.ma.where(x > 5, x, -3.1416)

[[-3.1416 -- -3.1416]
[-- -3.1416 --]
[6.0 -- 8.0]]

1.7 The Array Interface

Warning: This page describes the old, deprecated array interface. Everything still works as described as of numpy
1.2 and on into the foreseeable future, but new development should target PEP 3118 The Revised Buffer
Protocol. PEP 3118 was incorporated into Python 2.6 and 3.0, and is additionally supported by Cythons
numpy buffer support. (See the Cython numpy tutorial.) Cython provides a way to write code that supports the
buffer protocol with Python versions older than 2.6 because it has a backward-compatible implementation utilizing
the legacy array interface described here.

version
3
The array interface (sometimes called array protocol) was created in 2005 as a means for array-like Python objects to
re-use each others data buffers intelligently whenever possible. The homogeneous N-dimensional array interface is
a default mechanism for objects to share N-dimensional array memory and information. The interface consists of a
Python-side and a C-side using two attributes. Objects wishing to be considered an N-dimensional array in application
code should support at least one of these attributes. Objects wishing to support an N-dimensional array in application
code should look for at least one of these attributes and use the information provided appropriately.
This interface describes homogeneous arrays in the sense that each item of the array has the same type. This type
can be very simple or it can be a quite arbitrary and complicated C-like structure.
There are two ways to use the interface: A Python side and a C-side. Both are separate attributes.

1.7.1 Python side

This approach to the interface consists of the object having an __array_interface__ attribute.
__array_interface__
A dictionary of items (3 required and 5 optional). The optional keys in the dictionary have implied defaults if
they are not provided.
The keys are:
shape (required)
Tuple whose elements are the array size in each dimension. Each entry is an integer (a Python int or
long). Note that these integers could be larger than the platform int or long could hold (a Python
int is a C long). It is up to the code using this attribute to handle this appropriately; either by raising
an error when overflow is possible, or by using Py_LONG_LONG as the C type for the shapes.
typestr (required)
A string providing the basic type of the homogenous array The basic string format consists of 3 parts:
a character describing the byteorder of the data (<: little-endian, >: big-endian, |: not-relevant), a
character code giving the basic type of the array, and an integer providing the number of bytes the
type uses.
The basic type character codes are:

366 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

t Bit field (following integer gives the number of bits in the bit field).
b Boolean (integer type where all values are only True or False)
i Integer
u Unsigned integer
f Floating point
c Complex floating point
O Object (i.e. the memory contains a pointer to PyObject)
S String (fixed-length sequence of char)
U Unicode (fixed-length sequence of Py_UNICODE)
V Other (void * each item is a fixed-size chunk of memory)
descr (optional)
A list of tuples providing a more detailed description of the memory layout for each item in the
homogeneous array. Each tuple in the list has two or three elements. Normally, this attribute would
be used when typestr is V[0-9]+, but this is not a requirement. The only requirement is that the
number of bytes represented in the typestr key is the same as the total number of bytes represented
here. The idea is to support descriptions of C-like structs (records) that make up array elements. The
elements of each tuple in the list are
1.A string providing a name associated with this portion of the record. This could also be a tuple
of (full name, basic_name) where basic name would be a valid Python variable
name representing the full name of the field.
2.Either a basic-type description string as in typestr or another list (for nested records)
3.An optional shape tuple providing how many times this part of the record should be repeated. No
repeats are assumed if this is not given. Very complicated structures can be described using this
generic interface. Notice, however, that each element of the array is still of the same data-type.
Some examples of using this interface are given below.
Default: [(, typestr)]
data (optional)
A 2-tuple whose first argument is an integer (a long integer if necessary) that points to the data-area
storing the array contents. This pointer must point to the first element of data (in other words any
offset is always ignored in this case). The second entry in the tuple is a read-only flag (true means
the data area is read-only).
This attribute can also be an object exposing the buffer interface which will be used to share
the data. If this key is not present (or returns None), then memory sharing will be done through the
buffer interface of the object itself. In this case, the offset key can be used to indicate the start of the
buffer. A reference to the object exposing the array interface must be stored by the new object if the
memory area is to be secured.
Default: None
strides (optional)
Either None to indicate a C-style contiguous array or a Tuple of strides which provides the number of
bytes needed to jump to the next array element in the corresponding dimension. Each entry must be
an integer (a Python int or long). As with shape, the values may be larger than can be represented
by a C int or long; the calling code should handle this appropiately, either by raising an error, or
by using Py_LONG_LONG in C. The default is None which implies a C-style contiguous memory
buffer. In this model, the last dimension of the array varies the fastest. For example, the default
strides tuple for an object whose array entries are 8 bytes long and whose shape is (10,20,30) would
be (4800, 240, 8)
Default: None (C-style contiguous)

1.7. The Array Interface 367

NumPy Reference, Release 2.0.0.dev8464

mask (optional)
None or an object exposing the array interface. All elements of the mask array should be interpreted
only as true or not true indicating which elements of this array are valid. The shape of this object
should be broadcastable to the shape of the original array.
Default: None (All array values are valid)
offset (optional)
An integer offset into the array data region. This can only be used when data is None or returns a
buffer object.
Default: 0.
version (required)
An integer showing the version of the interface (i.e. 3 for this version). Be careful not to use this to
invalidate objects exposing future versions of the interface.

1.7.2 C-struct access

This approach to the array interface allows for faster access to an array using only one attribute lookup and a well-
defined C-structure.
__array_struct__
A PyCObject whose voidptr member contains a pointer to a filled PyArrayInterface structure.
Memory for the structure is dynamically created and the PyCObject is also created with an appropriate
destructor so the retriever of this attribute simply has to apply Py_DECREF to the object returned by this
attribute when it is finished. Also, either the data needs to be copied out, or a reference to the object exposing
this attribute must be held to ensure the data is not freed. Objects exposing the __array_struct__ interface
must also not reallocate their memory if other objects are referencing them.
The PyArrayInterface structure is defined in numpy/ndarrayobject.h as:

typedef struct {
int two; /* contains the integer 2 -- simple sanity check */
int nd; /* number of dimensions */
char typekind; /* kind in array --- character code of typestr */
int itemsize; /* size of each element */
int flags; /* flags indicating how the data should be interpreted */
/* must set ARR_HAS_DESCR bit to validate descr */
Py_intptr_t *shape; /* A length-nd array of shape information */
Py_intptr_t *strides; /* A length-nd array of stride information */
void *data; /* A pointer to the first element of the array */
PyObject *descr; /* NULL or data-description (same as descr key
of __array_interface__) -- must set ARR_HAS_DESCR
flag or this will be ignored. */
} PyArrayInterface;

The flags member may consist of 5 bits showing how the data should be interpreted and one bit showing how
the Interface should be interpreted. The data-bits are CONTIGUOUS (0x1), FORTRAN (0x2), ALIGNED (0x100),
NOTSWAPPED (0x200), and WRITEABLE (0x400). A final flag ARR_HAS_DESCR (0x800) indicates whether or not
this structure has the arrdescr field. The field should not be accessed unless this flag is present.
New since June 16, 2006:
In the past most implementations used the desc member of the PyCObject itself (do not confuse this with the
descr member of the PyArrayInterface structure above they are two separate things) to hold the pointer to

368 Chapter 1. Array objects

 NumPy Reference, Release 2.0.0.dev8464

the object exposing the interface. This is now an explicit part of the interface. Be sure to own a reference to the object
when the PyCObject is created using PyCObject_FromVoidPtrAndDesc.

1.7.3 Type description examples

For clarity it is useful to provide some examples of the type description and corresponding __array_interface__
descr entries. Thanks to Scott Gilbert for these examples:
In every case, the descr key is optional, but of course provides more information which may be important for various
applications:

* Float data
typestr == >f4
descr == [(,>f4)]

* Complex double
typestr == >c8
descr == [(real,>f4), (imag,>f4)]

* RGB Pixel data

typestr == |V3
descr == [(r,|u1), (g,|u1), (b,|u1)]

* Mixed endian (weird but could happen).

typestr == |V8 (or >u8)
descr == [(big,>i4), (little,<i4)]

* Nested structure
struct {
int ival;
struct {
unsigned short sval;
unsigned char bval;
unsigned char cval;
} sub;
}
typestr == |V8 (or <u8 if you want)
descr == [(ival,<i4), (sub, [(sval,<u2), (bval,|u1), (cval,|u1) ]) ]

* Nested array
struct {
int ival;
double data[16*4];
}
typestr == |V516
descr == [(ival,>i4), (data,>f8,(16,4))]

* Padded structure
struct {
int ival;
double dval;
}
typestr == |V16
descr == [(ival,>i4),(,|V4),(dval,>f8)]

It should be clear that any record type could be described using this interface.

1.7. The Array Interface 369

NumPy Reference, Release 2.0.0.dev8464

1.7.4 Differences with Array interface (Version 2)

The version 2 interface was very similar. The differences were largely asthetic. In particular:
1. The PyArrayInterface structure had no descr member at the end (and therefore no flag ARR_HAS_DESCR)
2. The desc member of the PyCObject returned from __array_struct__ was not specified. Usually, it was the object
exposing the array (so that a reference to it could be kept and destroyed when the C-object was destroyed). Now
it must be a tuple whose first element is a string with PyArrayInterface Version # and whose second element
is the object exposing the array.
3. The tuple returned from __array_interface__[data] used to be a hex-string (now it is an integer or a long
integer).
4. There was no __array_interface__ attribute instead all of the keys (except for version) in the __array_interface__
dictionary were their own attribute: Thus to obtain the Python-side information you had to access separately the
attributes:
__array_data__
__array_shape__
__array_strides__
__array_typestr__
__array_descr__
__array_offset__
__array_mask__

370 Chapter 1. Array objects

 CHAPTER

TWO

UNIVERSAL FUNCTIONS (UFUNC)

A universal function (or ufunc for short) is a function that operates on ndarrays in an element-by-element fashion,
supporting array broadcasting, type casting, and several other standard features. That is, a ufunc is a vectorized
wrapper for a function that takes a fixed number of scalar inputs and produces a fixed number of scalar outputs.
In Numpy, universal functions are instances of the numpy.ufunc class. Many of the built-in functions are imple-
mented in compiled C code, but ufunc instances can also be produced using the frompyfunc factory function.

2.1 Broadcasting

Each universal function takes array inputs and produces array outputs by performing the core function element-wise
on the inputs. Standard broadcasting rules are applied so that inputs not sharing exactly the same shapes can still be
usefully operated on. Broadcasting can be understood by four rules:
1. All input arrays with ndim smaller than the input array of largest ndim, have 1s prepended to their shapes.
2. The size in each dimension of the output shape is the maximum of all the input sizes in that dimension.
3. An input can be used in the calculation if its size in a particular dimension either matches the output size in that
dimension, or has value exactly 1.
4. If an input has a dimension size of 1 in its shape, the first data entry in that dimension will be used for all
calculations along that dimension. In other words, the stepping machinery of the ufunc will simply not step
along that dimension (the stride will be 0 for that dimension).
Broadcasting is used throughout NumPy to decide how to handle disparately shaped arrays; for example, all arith-
metic operations (+, -, *, ...) between ndarrays broadcast the arrays before operation. A set of arrays is called
broadcastable to the same shape if the above rules produce a valid result, i.e., one of the following is true:
1. The arrays all have exactly the same shape.
2. The arrays all have the same number of dimensions and the length of each dimensions is either a common length
or 1.
3. The arrays that have too few dimensions can have their shapes prepended with a dimension of length 1 to satisfy
property 2.
Example
If a.shape is (5,1), b.shape is (1,6), c.shape is (6,) and d.shape is () so that d is a scalar, then a, b, c, and d
are all broadcastable to dimension (5,6); and
a acts like a (5,6) array where a[:,0] is broadcast to the other columns,
b acts like a (5,6) array where b[0,:] is broadcast to the other rows,

371
NumPy Reference, Release 2.0.0.dev8464

c acts like a (1,6) array and therefore like a (5,6) array where c[:] is broadcast to every row, and finally,
d acts like a (5,6) array where the single value is repeated.

2.2 Output type determination

The output of the ufunc (and its methods) is not necessarily an ndarray, if all input arguments are not ndarrays.
All output arrays will be passed to the __array_prepare__ and __array_wrap__ methods of the in-
put (besides ndarrays, and scalars) that defines it and has the highest __array_priority__ of any other
input to the universal function. The default __array_priority__ of the ndarray is 0.0, and the default
__array_priority__ of a subtype is 1.0. Matrices have __array_priority__ equal to 10.0.
All ufuncs can also take output arguments. If necessary, output will be cast to the data-type(s) of the provided output
array(s). If a class with an __array__ method is used for the output, results will be written to the object returned
by __array__. Then, if the class also has an __array_prepare__ method, it is called so metadata may be
determined based on the context of the ufunc (the context consisting of the ufunc itself, the arguments passed to
the ufunc, and the ufunc domain.) The array object returned by __array_prepare__ is passed to the ufunc for
computation. Finally, if the class also has an __array_wrap__ method, the returned ndarray result will be
passed to that method just before passing control back to the caller.

2.3 Use of internal buffers

Internally, buffers are used for misaligned data, swapped data, and data that has to be converted from one data type to
another. The size of internal buffers is settable on a per-thread basis. There can be up to 2(ninputs + noutputs ) buffers
of the specified size created to handle the data from all the inputs and outputs of a ufunc. The default size of a buffer is
10,000 elements. Whenever buffer-based calculation would be needed, but all input arrays are smaller than the buffer
size, those misbehaved or incorrectly-typed arrays will be copied before the calculation proceeds. Adjusting the size of
the buffer may therefore alter the speed at which ufunc calculations of various sorts are completed. A simple interface
for setting this variable is accessible using the function
setbufsize(size) Set the size of the buffer used in ufuncs.
setbufsize(size)
Set the size of the buffer used in ufuncs.
Parameters
size : int
Size of buffer.

2.4 Error handling

Universal functions can trip special floating-point status registers in your hardware (such as divide-by-zero). If avail-
able on your platform, these registers will be regularly checked during calculation. Error handling is controlled on a
per-thread basis, and can be configured using the functions
seterr([all, divide, over, under, invalid]) Set how floating-point errors are handled.
seterrcall(func) Set the floating-point error callback function or log object.
seterr(all=None, divide=None, over=None, under=None, invalid=None)
Set how floating-point errors are handled.

372 Chapter 2. Universal functions (ufunc)

 NumPy Reference, Release 2.0.0.dev8464

Note that operations on integer scalar types (such as int16) are handled like floating point, and are affected by
these settings.
Parameters
all : {ignore, warn, raise, call, print, log}, optional
Set treatment for all types of floating-point errors at once:
ignore: Take no action when the exception occurs.
warn: Print a RuntimeWarning (via the Python warnings module).
raise: Raise a FloatingPointError.
call: Call a function specified using the seterrcall function.
print: Print a warning directly to stdout.
log: Record error in a Log object specified by seterrcall.
The default is not to change the current behavior.
divide : {ignore, warn, raise, call, print, log}, optional
Treatment for division by zero.
over : {ignore, warn, raise, call, print, log}, optional
Treatment for floating-point overflow.
under : {ignore, warn, raise, call, print, log}, optional
Treatment for floating-point underflow.
invalid : {ignore, warn, raise, call, print, log}, optional
Treatment for invalid floating-point operation.
Returns
old_settings : dict
Dictionary containing the old settings.
See Also:

seterrcall
Set a callback function for the call mode.

geterr, geterrcall

Notes
The floating-point exceptions are defined in the IEEE 754 standard [1]:
Division by zero: infinite result obtained from finite numbers.
Overflow: result too large to be expressed.
Underflow: result so close to zero that some precision was lost.
Invalid operation: result is not an expressible number, typically indicates that a NaN was produced.

Examples

2.4. Error handling 373

NumPy Reference, Release 2.0.0.dev8464

>>> old_settings = np.seterr(all=ignore) #seterr to known value

>>> np.seterr(over=raise)
{over: ignore, divide: ignore, invalid: ignore,
under: ignore}
>>> np.seterr(all=ignore) # reset to default
{over: raise, divide: ignore, invalid: ignore, under: ignore}

>>> np.int16(32000) * np.int16(3)

30464
>>> old_settings = np.seterr(all=warn, over=raise)
>>> np.int16(32000) * np.int16(3)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
FloatingPointError: overflow encountered in short_scalars

>>> old_settings = np.seterr(all=print)

>>> np.geterr()
{over: print, divide: print, invalid: print, under: print}
>>> np.int16(32000) * np.int16(3)
Warning: overflow encountered in short_scalars
30464

seterrcall(func)
Set the floating-point error callback function or log object.
There are two ways to capture floating-point error messages. The first is to set the error-handler to call, using
seterr. Then, set the function to call using this function.
The second is to set the error-handler to log, using seterr. Floating-point errors then trigger a call to the write
method of the provided object.
Parameters
func : callable f(err, flag) or object with write method
Function to call upon floating-point errors (call-mode) or object whose write method
is used to log such message (log-mode).
The call function takes two arguments. The first is the type of error (one of divide,
over, under, or invalid), and the second is the status flag. The flag is a byte, whose
least-significant bits indicate the status:

[0 0 0 0 invalid over under invalid]

In other words, flags = divide + 2*over + 4*under + 8*invalid.

If an object is provided, its write method should take one argument, a string.
Returns
h : callable, log instance or None
The old error handler.
See Also:
seterr, geterr, geterrcall

374 Chapter 2. Universal functions (ufunc)

 NumPy Reference, Release 2.0.0.dev8464

Examples
Callback upon error:

>>> def err_handler(type, flag):

... print "Floating point error (%s), with flag %s" % (type, flag)
...

>>> saved_handler = np.seterrcall(err_handler)

>>> save_err = np.seterr(all=call)

>>> np.array([1, 2, 3]) / 0.0

Floating point error (divide by zero), with flag 1
array([ Inf, Inf, Inf])

>>> np.seterrcall(saved_handler)
<function err_handler at 0x...>
>>> np.seterr(**save_err)
{over: call, divide: call, invalid: call, under: call}

Log error message:

>>> class Log(object):

... def write(self, msg):
... print "LOG: %s" % msg
...

>>> log = Log()

>>> saved_handler = np.seterrcall(log)
>>> save_err = np.seterr(all=log)

>>> np.array([1, 2, 3]) / 0.0

LOG: Warning: divide by zero encountered in divide
<BLANKLINE>
array([ Inf, Inf, Inf])

>>> np.seterrcall(saved_handler)
<__main__.Log object at 0x...>
>>> np.seterr(**save_err)
{over: log, divide: log, invalid: log, under: log}

2.5 Casting Rules

At the core of every ufunc is a one-dimensional strided loop that implements the actual function for a specific type
combination. When a ufunc is created, it is given a static list of inner loops and a corresponding list of type signatures
over which the ufunc operates. The ufunc machinery uses this list to determine which inner loop to use for a particular
case. You can inspect the .types attribute for a particular ufunc to see which type combinations have a defined inner
loop and which output type they produce (character codes are used in said output for brevity).
Casting must be done on one or more of the inputs whenever the ufunc does not have a core loop implementation for
the input types provided. If an implementation for the input types cannot be found, then the algorithm searches for an
implementation with a type signature to which all of the inputs can be cast safely. The first one it finds in its internal

2.5. Casting Rules 375

NumPy Reference, Release 2.0.0.dev8464

list of loops is selected and performed, after all necessary type casting. Recall that internal copies during ufuncs (even
for casting) are limited to the size of an internal buffer (which is user settable).
Note: Universal functions in NumPy are flexible enough to have mixed type signatures. Thus, for example, a universal
function could be defined that works with floating-point and integer values. See ldexp for an example.
By the above description, the casting rules are essentially implemented by the question of when a data type can be
cast safely to another data type. The answer to this question can be determined in Python with a function call:
can_cast(fromtype, totype). The Figure below shows the results of this call for the 21 internally supported
types on the authors 32-bit system. You can generate this table for your system with the code given in the Figure.
Figure
Code segment showing the can cast safely table for a 32-bit system.

>>> def print_table(ntypes):

... print X,
... for char in ntypes: print char,
... print
... for row in ntypes:
... print row,
... for col in ntypes:
... print int(np.can_cast(row, col)),
... print
>>> print_table(np.typecodes[All])
X ? b h i l q p B H I L Q P f d g F D G S U V O
? 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
b 0 1 1 1 1 1 1 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1
h 0 0 1 1 1 1 1 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1
i 0 0 0 1 1 1 1 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 1
l 0 0 0 1 1 1 1 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 1
q 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 1
p 0 0 0 1 1 1 1 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 1
B 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
H 0 0 0 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
I 0 0 0 0 0 1 0 0 0 1 1 1 1 0 1 1 0 1 1 1 1 1 1
L 0 0 0 0 0 1 0 0 0 1 1 1 1 0 1 1 0 1 1 1 1 1 1
Q 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 1 0 1 1 1 1 1 1
P 0 0 0 0 0 1 0 0 0 1 1 1 1 0 1 1 0 1 1 1 1 1 1
f 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1
d 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 1 1 1 1 1 1
g 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 1 1 1 1
F 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1
D 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1
G 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1
S 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1
U 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1
V 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1
O 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1

You should note that, while included in the table for completeness, the S, U, and V types cannot be operated on
by ufuncs. Also, note that on a 64-bit system the integer types may have different sizes, resulting in a slightly altered
table.
Mixed scalar-array operations use a different set of casting rules that ensure that a scalar cannot upcast an array
unless the scalar is of a fundamentally different kind of data (i.e., under a different hierarchy in the data-type hierarchy)
than the array. This rule enables you to use scalar constants in your code (which, as Python types, are interpreted
accordingly in ufuncs) without worrying about whether the precision of the scalar constant will cause upcasting on
your large (small precision) array.

376 Chapter 2. Universal functions (ufunc)

 NumPy Reference, Release 2.0.0.dev8464

2.6 ufunc

2.6.1 Optional keyword arguments

All ufuncs take optional keyword arguments. These represent rather advanced usage and will not typically be used by
most Numpy users. sig
Either a data-type, a tuple of data-types, or a special signature string indicating the input and output
types of a ufunc. This argument allows you to provide a specific signature for the 1-d loop to use in
the underlying calculation. If the loop specified does not exist for the ufunc, then a TypeError is raised.
Normally, a suitable loop is found automatically by comparing the input types with what is available and
searching for a loop with data-types to which all inputs can be cast safely. This keyword argument lets
you bypass that search and choose a particular loop. A list of available signatures is provided by the types
attribute of the ufunc object.
extobj
a list of length 1, 2, or 3 specifying the ufunc buffer-size, the error mode integer, and the error call-
back function. Normally, these values are looked up in a thread-specific dictionary. Passing them here
circumvents that look up and uses the low-level specification provided for the error mode. This may be
useful, for example, as an optimization for calculations requiring many ufunc calls on small arrays in a
loop.

2.6.2 Attributes

There are some informational attributes that universal functions possess. None of the attributes can be set.
__doc__ A docstring for each ufunc. The first part of the docstring is dynamically generated from the number of
outputs, the name, and the number of inputs. The second part of the docstring is provided at creation
time and stored with the ufunc.
__name__The name of the ufunc.
ufunc.nin The number of inputs.
ufunc.nout The number of outputs.
ufunc.nargs The number of arguments.
ufunc.ntypes The number of types.
ufunc.types Returns a list with types grouped input->output.
ufunc.identity The identity value.
nin
The number of inputs.
Data attribute containing the number of arguments the ufunc treats as input.

Examples

>>> np.add.nin
2
>>> np.multiply.nin
2
>>> np.power.nin
2
>>> np.exp.nin
1

nout
The number of outputs.

2.6. ufunc 377

NumPy Reference, Release 2.0.0.dev8464

Data attribute containing the number of arguments the ufunc treats as output.

Notes
Since all ufuncs can take output arguments, this will always be (at least) 1.

Examples

>>> np.add.nout
1
>>> np.multiply.nout
1
>>> np.power.nout
1
>>> np.exp.nout
1

nargs
The number of arguments.
Data attribute containing the number of arguments the ufunc takes, including optional ones.

Notes
Typically this value will be one more than what you might expect because all ufuncs take the optional out
argument.

Examples

>>> np.add.nargs
3
>>> np.multiply.nargs
3
>>> np.power.nargs
3
>>> np.exp.nargs
2

ntypes
The number of types.
The number of numerical NumPy types - of which there are 18 total - on which the ufunc can operate.
See Also:
numpy.ufunc.types

Examples

>>> np.add.ntypes
18
>>> np.multiply.ntypes
18
>>> np.power.ntypes
17
>>> np.exp.ntypes
7
>>> np.remainder.ntypes
14

378 Chapter 2. Universal functions (ufunc)

 NumPy Reference, Release 2.0.0.dev8464

types
Returns a list with types grouped input->output.
Data attribute listing the data-type Domain-Range groupings the ufunc can deliver. The data-types are given
using the character codes.
See Also:
numpy.ufunc.ntypes

Examples

>>> np.add.types
[??->?, bb->b, BB->B, hh->h, HH->H, ii->i, II->I, ll->l,
LL->L, qq->q, QQ->Q, ff->f, dd->d, gg->g, FF->F, DD->D,
GG->G, OO->O]

>>> np.multiply.types
[??->?, bb->b, BB->B, hh->h, HH->H, ii->i, II->I, ll->l,
LL->L, qq->q, QQ->Q, ff->f, dd->d, gg->g, FF->F, DD->D,
GG->G, OO->O]

>>> np.power.types
[bb->b, BB->B, hh->h, HH->H, ii->i, II->I, ll->l, LL->L,
qq->q, QQ->Q, ff->f, dd->d, gg->g, FF->F, DD->D, GG->G,
OO->O]

>>> np.exp.types
[f->f, d->d, g->g, F->F, D->D, G->G, O->O]

>>> np.remainder.types
[bb->b, BB->B, hh->h, HH->H, ii->i, II->I, ll->l, LL->L,
qq->q, QQ->Q, ff->f, dd->d, gg->g, OO->O]

identity
The identity value.
Data attribute containing the identity element for the ufunc, if it has one. If it does not, the attribute value is
None.

Examples

>>> np.add.identity
0
>>> np.multiply.identity
1
>>> np.power.identity
1
>>> print np.exp.identity
None

2.6.3 Methods

All ufuncs have four methods. However, these methods only make sense on ufuncs that take two input arguments
and return one output argument. Attempting to call these methods on other ufuncs will cause a ValueError. The

2.6. ufunc 379

NumPy Reference, Release 2.0.0.dev8464

reduce-like methods all take an axis keyword and a dtype keyword, and the arrays must all have dimension >= 1.
The axis keyword specifies the axis of the array over which the reduction will take place and may be negative, but
must be an integer. The dtype keyword allows you to manage a very common problem that arises when naively using
{op}.reduce. Sometimes you may have an array of a certain data type and wish to add up all of its elements, but
the result does not fit into the data type of the array. This commonly happens if you have an array of single-byte
integers. The dtype keyword allows you to alter the data type over which the reduction takes place (and therefore
the type of the output). Thus, you can ensure that the output is a data type with precision large enough to handle
your output. The responsibility of altering the reduce type is mostly up to you. There is one exception: if no
dtype is given for a reduction on the add or multiply operations, then if the input type is an integer (or Boolean)
data-type and smaller than the size of the int_ data type, it will be internally upcast to the int_ (or uint) data-type.
ufunc.reduce(a[, axis, dtype, out]) Reduces as dimension by one, by applying ufunc along one axis.
ufunc.accumulate(array[, axis, dtype, out]) Accumulate the result of applying the operator to all elements.
ufunc.reduceat(a, indices[, axis, dtype, out]) Performs a (local) reduce with specified slices over a single axis.
ufunc.outer(A, B) Apply the ufunc op to all pairs (a, b) with a in A and b in B.
reduce(a, axis=0, dtype=None, out=None)
Reduces as dimension by one, by applying ufunc along one axis.
Let a.shape = (N0 , ..., Ni , ..., NM 1 ). Then uf unc.reduce(a, axis = i)[k0 , .., ki1 , ki+1 , .., kM 1 ] = the
result of iterating j over range(Ni ), cumulatively applying ufunc to each a[k0 , .., ki1 , j, ki+1 , .., kM 1 ]. For a
one-dimensional array, reduce produces results equivalent to:

r = op.identity # op = ufunc
for i in xrange(len(A)):
r = op(r, A[i])
return r

For example, add.reduce() is equivalent to sum().

Parameters
a : array_like
The array to act on.
axis : int, optional
The axis along which to apply the reduction.
dtype : data-type code, optional
The type used to represent the intermediate results. Defaults to the data-type of the
output array if this is provided, or the data-type of the input array if no output array is
provided.
out : ndarray, optional
A location into which the result is stored. If not provided, a freshly-allocated array is
returned.
Returns
r : ndarray
The reduced array. If out was supplied, r is a reference to it.

Examples

>>> np.multiply.reduce([2,3,5])
30

A multi-dimensional array example:

380 Chapter 2. Universal functions (ufunc)

 NumPy Reference, Release 2.0.0.dev8464

>>> X = np.arange(8).reshape((2,2,2))
>>> X
array([[[0, 1],
[2, 3]],
[[4, 5],
[6, 7]]])
>>> np.add.reduce(X, 0)
array([[ 4, 6],
[ 8, 10]])
>>> np.add.reduce(X) # confirm: default axis value is 0
array([[ 4, 6],
[ 8, 10]])
>>> np.add.reduce(X, 1)
array([[ 2, 4],
[10, 12]])
>>> np.add.reduce(X, 2)
array([[ 1, 5],
[ 9, 13]])

accumulate(array, axis=0, dtype=None, out=None)

Accumulate the result of applying the operator to all elements.
For a one-dimensional array, accumulate produces results equivalent to:

r = np.empty(len(A))
t = op.identity # op = the ufunc being applied to As elements
for i in xrange(len(A)):
t = op(t, A[i])
r[i] = t
return r

For example, add.accumulate() is equivalent to np.cumsum().

For a multi-dimensional array, accumulate is applied along only one axis (axis zero by default; see Examples
below) so repeated use is necessary if one wants to accumulate over multiple axes.
Parameters
array : array_like
The array to act on.
axis : int, optional
The axis along which to apply the accumulation; default is zero.
dtype : data-type code, optional
The data-type used to represent the intermediate results. Defaults to the data-type of the
output array if such is provided, or the the data-type of the input array if no output array
is provided.
out : ndarray, optional
A location into which the result is stored. If not provided a freshly-allocated array is
returned.
Returns
r : ndarray
The accumulated values. If out was supplied, r is a reference to out.

2.6. ufunc 381

NumPy Reference, Release 2.0.0.dev8464

Examples
1-D array examples:

>>> np.add.accumulate([2, 3, 5])

array([ 2, 5, 10])
>>> np.multiply.accumulate([2, 3, 5])
array([ 2, 6, 30])

2-D array examples:

>>> I = np.eye(2)
>>> I
array([[ 1., 0.],
[ 0., 1.]])

Accumulate along axis 0 (rows), down columns:

>>> np.add.accumulate(I, 0)
array([[ 1., 0.],
[ 1., 1.]])
>>> np.add.accumulate(I) # no axis specified = axis zero
array([[ 1., 0.],
[ 1., 1.]])

Accumulate along axis 1 (columns), through rows:

>>> np.add.accumulate(I, 1)
array([[ 1., 1.],
[ 0., 1.]])

reduceat(a, indices, axis=0, dtype=None, out=None)

Performs a (local) reduce with specified slices over a single axis.
For i in range(len(indices)), reduceat computes ufunc.reduce(a[indices[i]:indices[i+1]]),
which becomes the i-th generalized row parallel to axis in the final result (i.e., in a 2-D array, for example, if
axis = 0, it becomes the i-th row, but if axis = 1, it becomes the i-th column). There are two exceptions to this:
when i = len(indices) - 1 (so for the last index), indices[i+1] = a.shape[axis].
if indices[i] >= indices[i + 1], the i-th generalized row is simply a[indices[i]].
The shape of the output depends on the size of indices, and may be larger than a (this happens if
len(indices) > a.shape[axis]).
Parameters
a : array_like
The array to act on.
indices : array_like
Paired indices, comma separated (not colon), specifying slices to reduce.
axis : int, optional
The axis along which to apply the reduceat.
dtype : data-type code, optional

382 Chapter 2. Universal functions (ufunc)

 NumPy Reference, Release 2.0.0.dev8464

The type used to represent the intermediate results. Defaults to the data type of the
output array if this is provided, or the data type of the input array if no output array is
provided.
out : ndarray, optional
A location into which the result is stored. If not provided a freshly-allocated array is
returned.
Returns
r : ndarray
The reduced values. If out was supplied, r is a reference to out.

Notes
A descriptive example:
If a is 1-D, the function ufunc.accumulate(a) is the same as ufunc.reduceat(a, indices)[::2]
where indices is range(len(array) - 1) with a zero placed in every other element: indices =
zeros(2 * len(a) - 1), indices[1::2] = range(1, len(a)).
Dont be fooled by this attributes name: reduceat(a) is not necessarily smaller than a.

Examples
To take the running sum of four successive values:

>>> np.add.reduceat(np.arange(8),[0,4, 1,5, 2,6, 3,7])[::2]

array([ 6, 10, 14, 18])

A 2-D example:

>>> x = np.linspace(0, 15, 16).reshape(4,4)

>>> x
array([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.],
[ 12., 13., 14., 15.]])

# reduce such that the result has the following five rows:
# [row1 + row2 + row3]
# [row4]
# [row2]
# [row3]
# [row1 + row2 + row3 + row4]

>>> np.add.reduceat(x, [0, 3, 1, 2, 0])

array([[ 12., 15., 18., 21.],
[ 12., 13., 14., 15.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.],
[ 24., 28., 32., 36.]])

# reduce such that result has the following two columns:

# [col1 * col2 * col3, col4]

2.6. ufunc 383

NumPy Reference, Release 2.0.0.dev8464

>>> np.multiply.reduceat(x, [0, 3], 1)

array([[ 0., 3.],
[ 120., 7.],
[ 720., 11.],
[ 2184., 15.]])

outer(A, B)
Apply the ufunc op to all pairs (a, b) with a in A and b in B.
Let M = A.ndim, N = B.ndim. Then the result, C, of op.outer(A, B) is an array of dimension M + N
such that:

C[i0 , ..., iM 1 , j0 , ..., jN 1 ] = op(A[i0 , ..., iM 1 ], B[j0 , ..., jN 1 ])

For A and B one-dimensional, this is equivalent to:

r = empty(len(A),len(B))
for i in xrange(len(A)):
for j in xrange(len(B)):
r[i,j] = op(A[i], B[j]) # op = ufunc in question

Parameters
A : array_like
First array
B : array_like
Second array
Returns
r : ndarray
Output array

See Also:
numpy.outer

Examples

>>> np.multiply.outer([1, 2, 3], [4, 5, 6])

array([[ 4, 5, 6],
[ 8, 10, 12],
[12, 15, 18]])

A multi-dimensional example:

>>> A = np.array([[1, 2, 3], [4, 5, 6]])

>>> A.shape
(2, 3)
>>> B = np.array([[1, 2, 3, 4]])
>>> B.shape
(1, 4)
>>> C = np.multiply.outer(A, B)
>>> C.shape; C
(2, 3, 1, 4)

384 Chapter 2. Universal functions (ufunc)

 NumPy Reference, Release 2.0.0.dev8464

array([[[[ 1, 2, 3, 4]],
[[ 2, 4, 6, 8]],
[[ 3, 6, 9, 12]]],
[[[ 4, 8, 12, 16]],
[[ 5, 10, 15, 20]],
[[ 6, 12, 18, 24]]]])

Warning: A reduce-like operation on an array with a data-type that has a range too small to handle the result
will silently wrap. One should use dtype to increase the size of the data-type over which reduction takes place.

2.7 Available ufuncs

There are currently more than 60 universal functions defined in numpy on one or more types, covering a wide variety
of operations. Some of these ufuncs are called automatically on arrays when the relevant infix notation is used (e.g.,
add(a, b) is called internally when a + b is written and a or b is an ndarray). Nevertheless, you may still want
to use the ufunc call in order to use the optional output argument(s) to place the output(s) in an object (or objects) of
your choice.
Recall that each ufunc operates element-by-element. Therefore, each ufunc will be described as if acting on a set of
scalar inputs to return a set of scalar outputs.
Note: The ufunc still returns its output(s) even if you use the optional output argument(s).

2.7. Available ufuncs 385

NumPy Reference, Release 2.0.0.dev8464

2.7.1 Math operations

add(x1) Add arguments element-wise.

subtract(x1) Subtract arguments, element-wise.
multiply(x1) Multiply arguments element-wise.
divide(x1) Divide arguments element-wise.
logaddexp(x1) Logarithm of the sum of exponentiations of the inputs.
logaddexp2(x1) Logarithm of the sum of exponentiations of the inputs in base-2.
true_divide(x1) Returns a true division of the inputs, element-wise.
floor_divide(x1) Return the largest integer smaller or equal to the division of the inputs.
negative() Returns an array with the negative of each element of the original array.
power(x1) Returns element-wise base array raised to power from second array.
remainder(x1) Return element-wise remainder of division.
mod(x1) Return element-wise remainder of division.
fmod(x1) Return the element-wise remainder of division.
absolute() Calculate the absolute value element-wise.
rint() Round elements of the array to the nearest integer.
sign() Returns an element-wise indication of the sign of a number.
conj() Return the complex conjugate, element-wise.
exp() Calculate the exponential of all elements in the input array.
exp2() Calculate 2**p for all p in the input array.
log() Natural logarithm, element-wise.
log2() Base-2 logarithm of x.
log10() Return the base 10 logarithm of the input array, element-wise.
expm1() Calculate exp(x) - 1 for all elements in the array.
log1p() Return the natural logarithm of one plus the input array, element-wise.
sqrt() Return the positive square-root of an array, element-wise.
square() Return the element-wise square of the input.
reciprocal() Return the reciprocal of the argument, element-wise.
ones_like() Returns an array of ones with the same shape and type as a given array.
Tip: The optional output arguments can be used to help you save memory for large calculations. If your arrays are
large, complicated expressions can take longer than absolutely necessary due to the creation and (later) destruction of
temporary calculation spaces. For example, the expression G = a * b + c is equivalent to t1 = A * B; G =
T1 + C; del t1. It will be more quickly executed as G = A * B; add(G, C, G) which is the same as G
= A * B; G += C.

2.7.2 Trigonometric functions

All trigonometric functions use radians when an angle is called for. The ratio of degrees to radians is 180 /.

386 Chapter 2. Universal functions (ufunc)

 NumPy Reference, Release 2.0.0.dev8464

sin() Trigonometric sine, element-wise.

cos() Cosine elementwise.
tan() Compute tangent element-wise.
arcsin() Inverse sine elementwise.
arccos() Trigonometric inverse cosine, element-wise.
arctan() Trigonometric inverse tangent, element-wise.
arctan2(x1) Elementwise arc tangent of x1/x2 choosing the quadrant correctly.
hypot(x1) Given the legs of a right triangle, return its hypotenuse.
sinh() Hyperbolic sine, element-wise.
cosh() Hyperbolic cosine, element-wise.
tanh() Compute hyperbolic tangent element-wise.
arcsinh() Inverse hyperbolic sine elementwise.
arccosh() Inverse hyperbolic cosine, elementwise.
arctanh() Inverse hyperbolic tangent elementwise.
deg2rad() Convert angles from degrees to radians.
rad2deg() Convert angles from radians to degrees.

2.7.3 Bit-twiddling functions

These function all require integer arguments and they manipulate the bit-pattern of those arguments.
bitwise_and(x1) Compute the bit-wise AND of two arrays element-wise.
bitwise_or(x1) Compute the bit-wise OR of two arrays element-wise.
bitwise_xor(x1) Compute the bit-wise XOR of two arrays element-wise.
invert() Compute bit-wise inversion, or bit-wise NOT, element-wise.
left_shift(x1) Shift the bits of an integer to the left.
right_shift(x1) Shift the bits of an integer to the right.

2.7.4 Comparison functions

greater(x1) Return (x1 > x2) element-wise.

greater_equal(x1) Return (x1 >= x2) element-wise.
less(x1) Return (x1 < x2) element-wise.
less_equal(x1) Return (x1 <= x2) element-wise.
not_equal(x1) Return (x1 != x2) element-wise.
equal(x1) Return (x1 == x2) element-wise.

Warning: Do not use the Python keywords and and or to combine logical array expressions. These keywords
will test the truth value of the entire array (not element-by-element as you might expect). Use the bitwise operators
& and | instead.

logical_and(x1) Compute the truth value of x1 AND x2 elementwise.

logical_or(x1) Compute the truth value of x1 OR x2 elementwise.
logical_xor(x1) Compute the truth value of x1 XOR x2 elementwise.
logical_not() Compute the truth value of NOT x elementwise.

Warning: The bit-wise operators & and | are the proper way to perform element-by-element array comparisons.
Be sure you understand the operator precedence: (a > 2) & (a < 5) is the proper syntax because a > 2
& a < 5 will result in an error due to the fact that 2 & a is evaluated first.

maximum(x1) Element-wise maximum of array elements.

2.7. Available ufuncs 387

NumPy Reference, Release 2.0.0.dev8464

Tip: The Python function max() will find the maximum over a one-dimensional array, but it will do so using a
slower sequence interface. The reduce method of the maximum ufunc is much faster. Also, the max() method will
not give answers you might expect for arrays with greater than one dimension. The reduce method of minimum also
allows you to compute a total minimum over an array.
minimum(x1) Element-wise minimum of array elements.

Warning: the behavior of maximum(a, b) is different than that of max(a, b). As a ufunc, maximum(a,
b) performs an element-by-element comparison of a and b and chooses each element of the result according to
which element in the two arrays is larger. In contrast, max(a, b) treats the objects a and b as a whole, looks at
the (total) truth value of a > b and uses it to return either a or b (as a whole). A similar difference exists between
minimum(a, b) and min(a, b).

2.7.5 Floating functions

Recall that all of these functions work element-by-element over an array, returning an array output. The description
details only a single operation.
isreal(x) Returns a bool array, where True if input element is real.
iscomplex(x) Returns a bool array, where True if input element is complex.
isfinite() Test element-wise for finite-ness (not infinity or not Not a Number).
isinf() Test element-wise for positive or negative infinity, return result as bool array.
isnan() Test element-wise for Not a Number (NaN), return result as a bool array.
signbit() Returns element-wise True where signbit is set (less than zero).
copysign(x1) Change the sign of x1 to that of x2, element-wise.
nextafter(x1) Return the next representable floating-point value after x1 in the direction of x2 element-wise.
modf(out1) Return the fractional and integral parts of an array, element-wise.
ldexp(x1) Compute y = x1 * 2**x2.
frexp(out1) Split the number, x, into a normalized fraction (y1) and exponent (y2)
fmod(x1) Return the element-wise remainder of division.
floor() Return the floor of the input, element-wise.
ceil() Return the ceiling of the input, element-wise.
trunc() Return the truncated value of the input, element-wise.

388 Chapter 2. Universal functions (ufunc)

 CHAPTER

THREE

ROUTINES

In this chapter routine docstrings are presented, grouped by functionality. Many docstrings contain example code,
which demonstrates basic usage of the routine. The examples assume that NumPy is imported with:

>>> import numpy as np

A convenient way to execute examples is the %doctest_mode mode of IPython, which allows for pasting of multi-
line examples and preserves indentation.

3.1 Array creation routines

See Also:
Array creation (in NumPy User Guide)

3.1.1 Ones and zeros

empty(shape[, dtype, order]) Return a new array of given shape and type, without initializing entries.
empty_like(a) Return a new array with the same shape and type as a given array.
eye(N[, M, k, dtype]) Return a 2-D array with ones on the diagonal and zeros elsewhere.
identity(n[, dtype]) Return the identity array.
ones(shape[, dtype, order]) Return a new array of given shape and type, filled with ones.
ones_like() Returns an array of ones with the same shape and type as a given array.
zeros(shape[, dtype, order]) Return a new array of given shape and type, filled with zeros.
zeros_like(a) Return an array of zeros with the same shape and type as a given array.
empty(shape, dtype=float, order=C)
Return a new array of given shape and type, without initializing entries.
Parameters
shape : int or tuple of int
Shape of the empty array
dtype : data-type, optional
Desired output data-type.
order : {C, F}, optional
Whether to store multi-dimensional data in C (row-major) or Fortran (column-major)
order in memory.

389
NumPy Reference, Release 2.0.0.dev8464

See Also:
empty_like, zeros, ones

Notes
empty, unlike zeros, does not set the array values to zero, and may therefore be marginally faster. On the other
hand, it requires the user to manually set all the values in the array, and should be used with caution.

Examples

>>> np.empty([2, 2])

array([[ -9.74499359e+001, 6.69583040e-309],
[ 2.13182611e-314, 3.06959433e-309]]) #random

>>> np.empty([2, 2], dtype=int)

array([[-1073741821, -1067949133],
[ 496041986, 19249760]]) #random

empty_like(a)
Return a new array with the same shape and type as a given array.
Parameters
a : array_like
The shape and data-type of a define the parameters of the returned array.
Returns
out : ndarray
Array of random data with the same shape and type as a.
See Also:

ones_like
Return an array of ones with shape and type of input.
zeros_like
Return an array of zeros with shape and type of input.
empty
Return a new uninitialized array.
ones
Return a new array setting values to one.
zeros
Return a new array setting values to zero.

Notes
This function does not initialize the returned array; to do that use zeros_like or ones_like instead. It may be
marginally faster than the functions that do set the array values.

Examples

>>> a = ([1,2,3], [4,5,6]) # a is array-like

>>> np.empty_like(a)
array([[-1073741821, -1073741821, 3], #random
[ 0, 0, -1073741821]])

390 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> a = np.array([[1., 2., 3.],[4.,5.,6.]])

>>> np.empty_like(a)
array([[ -2.00000715e+000, 1.48219694e-323, -2.00000572e+000], #random
[ 4.38791518e-305, -2.00000715e+000, 4.17269252e-309]])

eye(N, M=None, k=0, dtype=<type float>)

Return a 2-D array with ones on the diagonal and zeros elsewhere.
Parameters
N : int
Number of rows in the output.
M : int, optional
Number of columns in the output. If None, defaults to N.
k : int, optional
Index of the diagonal: 0 refers to the main diagonal, a positive value refers to an upper
diagonal, and a negative value to a lower diagonal.
dtype : dtype, optional
Data-type of the returned array.
Returns
I : ndarray (N,M)
An array where all elements are equal to zero, except for the k-th diagonal, whose values
are equal to one.
See Also:

diag
Return a diagonal 2-D array using a 1-D array specified by the user.

Examples

>>> np.eye(2, dtype=int)

array([[1, 0],
[0, 1]])
>>> np.eye(3, k=1)
array([[ 0., 1., 0.],
[ 0., 0., 1.],
[ 0., 0., 0.]])

identity(n, dtype=None)
Return the identity array.
The identity array is a square array with ones on the main diagonal.
Parameters
n : int
Number of rows (and columns) in n x n output.
dtype : data-type, optional
Data-type of the output. Defaults to float.
Returns
out : ndarray

3.1. Array creation routines 391

NumPy Reference, Release 2.0.0.dev8464

n x n array with its main diagonal set to one, and all other elements 0.

Examples

>>> np.identity(3)
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])

ones(shape, dtype=None, order=C)

Return a new array of given shape and type, filled with ones.
Please refer to the documentation for zeros.
See Also:
zeros

Examples

>>> np.ones(5)
array([ 1., 1., 1., 1., 1.])

>>> np.ones((5,), dtype=np.int)

array([1, 1, 1, 1, 1])

>>> np.ones((2, 1))

array([[ 1.],
[ 1.]])

>>> s = (2,2)
>>> np.ones(s)
array([[ 1., 1.],
[ 1., 1.]])

ones_like
Returns an array of ones with the same shape and type as a given array.
Equivalent to a.copy().fill(1).
Please refer to the documentation for zeros_like.
See Also:
zeros_like

Examples

>>> a = np.array([[1, 2, 3], [4, 5, 6]])

>>> np.ones_like(a)
array([[1, 1, 1],
[1, 1, 1]])

zeros(shape, dtype=float, order=C)

Return a new array of given shape and type, filled with zeros.
Parameters
shape : int or sequence of ints

392 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Shape of the new array, e.g., (2, 3) or 2.

dtype : data-type, optional
The desired data-type for the array, e.g., numpy.int8. Default is numpy.float64.
order : {C, F}, optional
Whether to store multidimensional data in C- or Fortran-contiguous (row- or column-
wise) order in memory.
Returns
out : ndarray
Array of zeros with the given shape, dtype, and order.
See Also:

zeros_like
Return an array of zeros with shape and type of input.
ones_like
Return an array of ones with shape and type of input.
empty_like
Return an empty array with shape and type of input.
ones
Return a new array setting values to one.
empty
Return a new uninitialized array.

Examples

>>> np.zeros(5)
array([ 0., 0., 0., 0., 0.])

>>> np.zeros((5,), dtype=numpy.int)

array([0, 0, 0, 0, 0])

>>> np.zeros((2, 1))

array([[ 0.],
[ 0.]])

>>> s = (2,2)
>>> np.zeros(s)
array([[ 0., 0.],
[ 0., 0.]])

>>> np.zeros((2,), dtype=[(x, i4), (y, i4)]) # custom dtype

array([(0, 0), (0, 0)],
dtype=[(x, <i4), (y, <i4)])

zeros_like(a)
Return an array of zeros with the same shape and type as a given array.
Equivalent to a.copy().fill(0).

3.1. Array creation routines 393

NumPy Reference, Release 2.0.0.dev8464

Parameters
a : array_like
The shape and data-type of a define the parameters of the returned array.
Returns
out : ndarray
Array of zeros with same shape and type as a.
See Also:

ones_like
Return an array of ones with shape and type of input.
empty_like
Return an empty array with shape and type of input.
zeros
Return a new array setting values to zero.
ones
Return a new array setting values to one.
empty
Return a new uninitialized array.

Examples

>>> x = np.arange(6)
>>> x = x.reshape((2, 3))
>>> x
array([[0, 1, 2],
[3, 4, 5]])
>>> np.zeros_like(x)
array([[0, 0, 0],
[0, 0, 0]])

>>> y = np.arange(3, dtype=np.float)

>>> y
array([ 0., 1., 2.])
>>> np.zeros_like(y)
array([ 0., 0., 0.])

394 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

3.1.2 From existing data

array(object[, dtype, copy, order, subok, Create an array.

ndmin])
asarray(a[, dtype, order]) Convert the input to an array.
asanyarray(a[, dtype, order]) Convert the input to a ndarray, but pass ndarray subclasses
through.
ascontiguousarray(a[, dtype]) Return a contiguous array in memory (C order).
asmatrix(data[, dtype]) Interpret the input as a matrix.
copy(a) Return an array copy of the given object.
frombuffer(buffer[, dtype, count, offset]) Interpret a buffer as a 1-dimensional array.
fromfile(file[, dtype, count, sep]) Construct an array from data in a text or binary file.
fromfunction(function, shape, **kwargs) Construct an array by executing a function over each coordinate.
fromiter(iterable, dtype[, count]) Create a new 1-dimensional array from an iterable object.
fromstring(string[, dtype, count, sep]) Return a new 1-D array initialized from raw binary or text data
in string.
loadtxt(fname[, dtype, comments, Load data from a text file.
delimiter, ...])
array(object, dtype=None, copy=True, order=None, subok=False, ndmin=0)
Create an array.
Parameters
object : array_like
An array, any object exposing the array interface, an object whose __array__ method
returns an array, or any (nested) sequence.
dtype : data-type, optional
The desired data-type for the array. If not given, then the type will be determined as the
minimum type required to hold the objects in the sequence. This argument can only be
used to upcast the array. For downcasting, use the .astype(t) method.
copy : bool, optional
If true (default), then the object is copied. Otherwise, a copy will only be made if
__array__ returns a copy, if obj is a nested sequence, or if a copy is needed to satisfy
any of the other requirements (dtype, order, etc.).
order : {C, F, A}, optional
Specify the order of the array. If order is C (default), then the array will be in C-
contiguous order (last-index varies the fastest). If order is F, then the returned array
will be in Fortran-contiguous order (first-index varies the fastest). If order is A, then
the returned array may be in any order (either C-, Fortran-contiguous, or even discon-
tiguous).
subok : bool, optional
If True, then sub-classes will be passed-through, otherwise the returned array will be
forced to be a base-class array (default).
ndmin : int, optional
Specifies the minimum number of dimensions that the resulting array should have. Ones
will be pre-pended to the shape as needed to meet this requirement.

3.1. Array creation routines 395

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.array([1, 2, 3])

array([1, 2, 3])

Upcasting:

>>> np.array([1, 2, 3.0])

array([ 1., 2., 3.])

More than one dimension:

>>> np.array([[1, 2], [3, 4]])

array([[1, 2],
[3, 4]])

Minimum dimensions 2:

>>> np.array([1, 2, 3], ndmin=2)

array([[1, 2, 3]])

Type provided:

>>> np.array([1, 2, 3], dtype=complex)

array([ 1.+0.j, 2.+0.j, 3.+0.j])

Data-type consisting of more than one element:

>>> x = np.array([(1,2),(3,4)],dtype=[(a,<i4),(b,<i4)])
>>> x[a]
array([1, 3])

Creating an array from sub-classes:

>>> np.array(np.mat(1 2; 3 4))

array([[1, 2],
[3, 4]])

>>> np.array(np.mat(1 2; 3 4), subok=True)

matrix([[1, 2],
[3, 4]])

asarray(a, dtype=None, order=None)

Convert the input to an array.
Parameters
a : array_like
Input data, in any form that can be converted to an array. This includes lists, lists of
tuples, tuples, tuples of tuples, tuples of lists and ndarrays.
dtype : data-type, optional
By default, the data-type is inferred from the input data.
order : {C, F}, optional

396 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Whether to use row-major (C) or column-major (F for FORTRAN) memory repre-

sentation. Defaults to C.
Returns
out : ndarray
Array interpretation of a. No copy is performed if the input is already an ndarray. If a
is a subclass of ndarray, a base class ndarray is returned.
See Also:

asanyarray
Similar function which passes through subclasses.
ascontiguousarray
Convert input to a contiguous array.
asfarray
Convert input to a floating point ndarray.
asfortranarray
Convert input to an ndarray with column-major memory order.
asarray_chkfinite
Similar function which checks input for NaNs and Infs.
fromiter
Create an array from an iterator.
fromfunction
Construct an array by executing a function on grid positions.

Examples
Convert a list into an array:

>>> a = [1, 2]
>>> np.asarray(a)
array([1, 2])

Existing arrays are not copied:

>>> a = np.array([1, 2])

>>> np.asarray(a) is a
True

If dtype is set, array is copied only if dtype does not match:

>>> a = np.array([1, 2], dtype=np.float32)

>>> np.asarray(a, dtype=np.float32) is a
True
>>> np.asarray(a, dtype=np.float64) is a
False

Contrary to asanyarray, ndarray subclasses are not passed through:

>>> issubclass(np.matrix, np.ndarray)

True
>>> a = np.matrix([[1, 2]])
>>> np.asarray(a) is a

3.1. Array creation routines 397

NumPy Reference, Release 2.0.0.dev8464

False
>>> np.asanyarray(a) is a
True

asanyarray(a, dtype=None, order=None)

Convert the input to a ndarray, but pass ndarray subclasses through.
Parameters
a : array_like
Input data, in any form that can be converted to an array. This includes scalars, lists,
lists of tuples, tuples, tuples of tuples, tuples of lists and ndarrays.
dtype : data-type, optional
By default, the data-type is inferred from the input data.
order : {C, F}, optional
Whether to use row-major (C) or column-major (F) memory representation. Defaults
to C.
Returns
out : ndarray or an ndarray subclass
Array interpretation of a. If a is an ndarray or a subclass of ndarray, it is returned as-is
and no copy is performed.
See Also:

asarray
Similar function which always returns ndarrays.
ascontiguousarray
Convert input to a contiguous array.
asfarray
Convert input to a floating point ndarray.
asfortranarray
Convert input to an ndarray with column-major memory order.
asarray_chkfinite
Similar function which checks input for NaNs and Infs.
fromiter
Create an array from an iterator.
fromfunction
Construct an array by executing a function on grid positions.

Examples
Convert a list into an array:

>>> a = [1, 2]
>>> np.asanyarray(a)
array([1, 2])

Instances of ndarray subclasses are passed through as-is:

398 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> a = np.matrix([1, 2])

>>> np.asanyarray(a) is a
True

ascontiguousarray(a, dtype=None)
Return a contiguous array in memory (C order).
Parameters
a : array_like
Input array.
dtype : str or dtype object, optional
Data-type of returned array.
Returns
out : ndarray
Contiguous array of same shape and content as a, with type dtype if specified.
See Also:

asfortranarray
Convert input to an ndarray with column-major memory order.
require
Return an ndarray that satisfies requirements.
ndarray.flags
Information about the memory layout of the array.

Examples

>>> x = np.arange(6).reshape(2,3)
>>> np.ascontiguousarray(x, dtype=np.float32)
array([[ 0., 1., 2.],
[ 3., 4., 5.]], dtype=float32)
>>> x.flags[C_CONTIGUOUS]
True

asmatrix(data, dtype=None)
Interpret the input as a matrix.
Unlike matrix, asmatrix does not make a copy if the input is already a matrix or an ndarray. Equivalent to
matrix(data, copy=False).
Parameters
data : array_like
Input data.
Returns
mat : matrix
data interpreted as a matrix.

Examples

>>> x = np.array([[1, 2], [3, 4]])

3.1. Array creation routines 399

NumPy Reference, Release 2.0.0.dev8464

>>> m = np.asmatrix(x)

>>> x[0,0] = 5

>>> m
matrix([[5, 2],
[3, 4]])

copy(a)
Return an array copy of the given object.
Parameters
a : array_like
Input data.
Returns
arr : ndarray
Array interpretation of a.

Notes
This is equivalent to

>>> np.array(a, copy=True)

Examples
Create an array x, with a reference y and a copy z:

>>> x = np.array([1, 2, 3])

>>> y = x
>>> z = np.copy(x)

Note that, when we modify x, y changes, but not z:

>>> x[0] = 10
>>> x[0] == y[0]
True
>>> x[0] == z[0]
False

frombuffer(buffer, dtype=float, count=-1, offset=0)

Interpret a buffer as a 1-dimensional array.
Parameters
buffer :
An object that exposes the buffer interface.
dtype : data-type, optional
Data type of the returned array.
count : int, optional
Number of items to read. -1 means all data in the buffer.
offset : int, optional

400 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Start reading the buffer from this offset.

Notes
If the buffer has data that is not in machine byte-order, this should be specified as part of the data-type, e.g.:

>>> dt = np.dtype(int)
>>> dt = dt.newbyteorder(>)
>>> np.frombuffer(buf, dtype=dt)

The data of the resulting array will not be byteswapped, but will be interpreted correctly.

Examples

>>> s = hello world

>>> np.frombuffer(s, dtype=S1, count=5, offset=6)
array([w, o, r, l, d],
dtype=|S1)

fromfile(file, dtype=float, count=-1, sep=)

Construct an array from data in a text or binary file.
A highly efficient way of reading binary data with a known data-type, as well as parsing simply formatted text
files. Data written using the tofile method can be read using this function.
Parameters
file : file or str
Open file object or filename.
dtype : data-type
Data type of the returned array. For binary files, it is used to determine the size and
byte-order of the items in the file.
count : int
Number of items to read. -1 means all items (i.e., the complete file).
sep : str
Separator between items if file is a text file. Empty () separator means the file should
be treated as binary. Spaces ( ) in the separator match zero or more whitespace char-
acters. A separator consisting only of spaces must match at least one whitespace.
See Also:
load, save, ndarray.tofile
loadtxt
More flexible way of loading data from a text file.

Notes
Do not rely on the combination of tofile and fromfile for data storage, as the binary files generated are are not
platform independent. In particular, no byte-order or data-type information is saved. Data can be stored in the
platform independent .npy format using save and load instead.

Examples
Construct an ndarray:

3.1. Array creation routines 401

NumPy Reference, Release 2.0.0.dev8464

>>> dt = np.dtype([(time, [(min, int), (sec, int)]),

... (temp, float)])
>>> x = np.zeros((1,), dtype=dt)
>>> x[time][min] = 10; x[temp] = 98.25
>>> x
array([((10, 0), 98.25)],
dtype=[(time, [(min, <i4), (sec, <i4)]), (temp, <f8)])

Save the raw data to disk:

>>> import os
>>> fname = os.tmpnam()
>>> x.tofile(fname)

Read the raw data from disk:

>>> np.fromfile(fname, dtype=dt)

array([((10, 0), 98.25)],
dtype=[(time, [(min, <i4), (sec, <i4)]), (temp, <f8)])

The recommended way to store and load data:

>>> np.save(fname, x)
>>> np.load(fname + .npy)
array([((10, 0), 98.25)],
dtype=[(time, [(min, <i4), (sec, <i4)]), (temp, <f8)])

fromfunction(function, shape, **kwargs)

Construct an array by executing a function over each coordinate.
The resulting array therefore has a value fn(x, y, z) at coordinate (x, y, z).
Parameters
function : callable
The function is called with N parameters, each of which represents the coordinates of
the array varying along a specific axis. For example, if shape were (2, 2), then the
parameters would be two arrays, [[0, 0], [1, 1]] and [[0, 1], [0, 1]].
function must be capable of operating on arrays, and should return a scalar value.
shape : (N,) tuple of ints
Shape of the output array, which also determines the shape of the coordinate arrays
passed to function.
dtype : data-type, optional
Data-type of the coordinate arrays passed to function. By default, dtype is float.
Returns
out : any
The result of the call to function is passed back directly. Therefore the type and shape
of out is completely determined by function.
See Also:
indices, meshgrid

402 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notes
Keywords other than shape and dtype are passed to function.

Examples

>>> np.fromfunction(lambda i, j: i == j, (3, 3), dtype=int)

array([[ True, False, False],
[False, True, False],
[False, False, True]], dtype=bool)

>>> np.fromfunction(lambda i, j: i + j, (3, 3), dtype=int)

array([[0, 1, 2],
[1, 2, 3],
[2, 3, 4]])

fromiter(iterable, dtype, count=-1)

Create a new 1-dimensional array from an iterable object.
Parameters
iterable : iterable object
An iterable object providing data for the array.
dtype : data-type
The data type of the returned array.
count : int, optional
The number of items to read from iterable. The default is -1, which means all data is
read.
Returns
out : ndarray
The output array.

Notes
Specify count to improve performance. It allows fromiter to pre-allocate the output array, instead of
resizing it on demand.

Examples

>>> iterable = (x*x for x in range(5))

>>> np.fromiter(iterable, np.float)
array([ 0., 1., 4., 9., 16.])

fromstring(string, dtype=float, count=-1, sep=)

Return a new 1-D array initialized from raw binary or text data in string.
Parameters
string : str
A string containing the data.
dtype : dtype, optional
The data type of the array. For binary input data, the data must be in exactly this format.
count : int, optional

3.1. Array creation routines 403

NumPy Reference, Release 2.0.0.dev8464

Read this number of dtype elements from the data. If this is negative, then the size will
be determined from the length of the data.
sep : str, optional
If provided and not empty, then the data will be interpreted as ASCII text with decimal
numbers. This argument is interpreted as the string separating numbers in the data.
Extra whitespace between elements is also ignored.
Returns
arr : array
The constructed array.
Raises
ValueError :
If the string is not the correct size to satisfy the requested dtype and count.

Examples

>>> np.fromstring(\x01\x02, dtype=np.uint8)

array([1, 2], dtype=uint8)
>>> np.fromstring(1 2, dtype=int, sep= )
array([1, 2])
>>> np.fromstring(1, 2, dtype=int, sep=,)
array([1, 2])
>>> np.fromstring(\x01\x02\x03\x04\x05, dtype=np.uint8, count=3)
array([1, 2, 3], dtype=uint8)

Invalid inputs:

>>> np.fromstring(\x01\x02\x03\x04\x05, dtype=np.int32)

Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: string size must be a multiple of element size
>>> np.fromstring(\x01\x02, dtype=np.uint8, count=3)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: string is smaller than requested size

loadtxt(fname, dtype=<type float>, comments=#, delimiter=None, converters=None, skiprows=0, usec-

ols=None, unpack=False)
Load data from a text file.
Each row in the text file must have the same number of values.
Parameters
fname : file or str
File or filename to read. If the filename extension is .gz or .bz2, the file is first
decompressed.
dtype : dtype, optional
Data type of the resulting array. If this is a record data-type, the resulting array will be
1-dimensional, and each row will be interpreted as an element of the array. In this case,
the number of columns used must match the number of fields in the data-type.
comments : str, optional
The character used to indicate the start of a comment.

404 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

delimiter : str, optional

The string used to separate values. By default, this is any whitespace.
converters : dict, optional
A dictionary mapping column number to a function that will convert that column to a
float. E.g., if column 0 is a date string: converters = {0: datestr2num}.
Converters can also be used to provide a default value for missing data: converters
= {3: lambda s: float(s or 0)}.
skiprows : int, optional
Skip the first skiprows lines.
usecols : sequence, optional
Which columns to read, with 0 being the first. For example, usecols = (1,4,5)
will extract the 2nd, 5th and 6th columns.
unpack : bool, optional
If True, the returned array is transposed, so that arguments may be unpacked using x,
y, z = loadtxt(...). Default is False.
Returns
out : ndarray
Data read from the text file.
See Also:
load, fromstring, fromregex
genfromtxt
Load data with missing values handled as specified.
scipy.io.loadmat
reads Matlab(R) data files

Notes
This function aims to be a fast reader for simply formatted files. The genfromtxt function provides more sophis-
ticated handling of, e.g., lines with missing values.

Examples

>>> from StringIO import StringIO # StringIO behaves like a file object
>>> c = StringIO("0 1\n2 3")
>>> np.loadtxt(c)
array([[ 0., 1.],
[ 2., 3.]])

>>> d = StringIO("M 21 72\nF 35 58")

>>> np.loadtxt(d, dtype={names: (gender, age, weight),
... formats: (S1, i4, f4)})
array([(M, 21, 72.0), (F, 35, 58.0)],
dtype=[(gender, |S1), (age, <i4), (weight, <f4)])

>>> c = StringIO("1,0,2\n3,0,4")
>>> x, y = np.loadtxt(c, delimiter=,, usecols=(0, 2), unpack=True)
>>> x

3.1. Array creation routines 405

NumPy Reference, Release 2.0.0.dev8464

array([ 1., 3.])

>>> y
array([ 2., 4.])

3.1.3 Creating record arrays (numpy.rec)

Note: numpy.rec is the preferred alias for numpy.core.records.

core.records.array(obj[, dtype, shape, ...]) Construct a record array from a wide-variety of objects.
core.records.fromarrays(arrayList[, dtype, create a record array from a (flat) list of arrays
...])
core.records.fromrecords(recList[, dtype, create a recarray from a list of records in text form
...])
core.records.fromstring(datastring[, dtype, create a (read-only) record array from binary data
...]) contained in
core.records.fromfile(fd[, dtype, shape, Create an array from binary file data
...])
array(obj, dtype=None, shape=None, offset=0, strides=None, formats=None, names=None, titles=None,
aligned=False, byteorder=None, copy=True)
Construct a record array from a wide-variety of objects.
fromarrays(arrayList, dtype=None, shape=None, formats=None, names=None, titles=None, aligned=False, by-
teorder=None)
create a record array from a (flat) list of arrays

>>> x1=np.array([1,2,3,4])
>>> x2=np.array([a,dd,xyz,12])
>>> x3=np.array([1.1,2,3,4])
>>> r = np.core.records.fromarrays([x1,x2,x3],names=a,b,c)
>>> print r[1]
(2, dd, 2.0)
>>> x1[1]=34
>>> r.a
array([1, 2, 3, 4])

fromrecords(recList, dtype=None, shape=None, formats=None, names=None, titles=None, aligned=False, by-

teorder=None)
create a recarray from a list of records in text form
The data in the same field can be heterogeneous, they will be promoted to the highest data type. This
method is intended for creating smaller record arrays. If used to create large array without formats
defined
r=fromrecords([(2,3.,abc)]*100000)
it can be slow.
If formats is None, then this will auto-detect formats. Use list of tuples rather than list of lists for
faster processing.

>>> r=np.core.records.fromrecords([(456,dbe,1.2),(2,de,1.3)],
... names=col1,col2,col3)
>>> print r[0]
(456, dbe, 1.2)
>>> r.col1
array([456, 2])

406 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> r.col2
chararray([dbe, de],
dtype=|S3)
>>> import cPickle
>>> print cPickle.loads(cPickle.dumps(r))
[(456, dbe, 1.2) (2, de, 1.3)]

fromstring(datastring, dtype=None, shape=None, offset=0, formats=None, names=None, titles=None,

aligned=False, byteorder=None)
create a (read-only) record array from binary data contained in a string
fromfile(fd, dtype=None, shape=None, offset=0, formats=None, names=None, titles=None, aligned=False, by-
teorder=None)
Create an array from binary file data
If file is a string then that file is opened, else it is assumed to be a file object.

>>> from tempfile import TemporaryFile

>>> a = np.empty(10,dtype=f8,i4,a5)
>>> a[5] = (0.5,10,abcde)
>>>
>>> fd=TemporaryFile()
>>> a = a.newbyteorder(<)
>>> a.tofile(fd)
>>>
>>> fd.seek(0)
>>> r=np.core.records.fromfile(fd, formats=f8,i4,a5, shape=10,
... byteorder=<)
>>> print r[5]
(0.5, 10, abcde)
>>> r.shape
(10,)

3.1.4 Creating character arrays (numpy.char)

Note: numpy.char is the preferred alias for numpy.core.defchararray.

core.defchararray.array(obj[, itemsize, Create a chararray.
...])
core.defchararray.asarray(obj[, Convert the input to a chararray, copying the data only if
itemsize, ...]) necessary.
array(obj, itemsize=None, copy=True, unicode=None, order=None)
Create a chararray.
Note: This class is provided for numarray backward-compatibility. New code (not concerned with numarray
compatibility) should use arrays of type string_ or unicode_ and use the free functions in numpy.char for
fast vectorized string operations instead.
Versus a regular Numpy array of type str or unicode, this class adds the following functionality:
1.values automatically have whitespace removed from the end when indexed
2.comparison operators automatically remove whitespace from the end when comparing values
3.vectorized string operations are provided as methods (e.g. str.endswith) and infix operators (e.g. +, *, %)

Parameters
obj : array of str or unicode-like

3.1. Array creation routines 407

NumPy Reference, Release 2.0.0.dev8464

itemsize : int, optional

itemsize is the number of characters per scalar in the resulting array. If itemsize is
None, and obj is an object array or a Python list, the itemsize will be automatically
determined. If itemsize is provided and obj is of type str or unicode, then the obj string
will be chunked into itemsize pieces.
copy : bool, optional
If true (default), then the object is copied. Otherwise, a copy will only be made if
__array__ returns a copy, if obj is a nested sequence, or if a copy is needed to satisfy
any of the other requirements (itemsize, unicode, order, etc.).
unicode : bool, optional
When true, the resulting chararray can contain Unicode characters, when false only
8-bit characters. If unicode is None and obj is one of the following:
a chararray,
an ndarray of type str or unicode
a Python str or unicode object,
then the unicode setting of the output array will be automatically determined.
order : {C, F, A}, optional
Specify the order of the array. If order is C (default), then the array will be in C-
contiguous order (last-index varies the fastest). If order is F, then the returned array
will be in Fortran-contiguous order (first-index varies the fastest). If order is A, then
the returned array may be in any order (either C-, Fortran-contiguous, or even discon-
tiguous).

asarray(obj, itemsize=None, unicode=None, order=None)

Convert the input to a chararray, copying the data only if necessary.
Versus a regular Numpy array of type str or unicode, this class adds the following functionality:
1.values automatically have whitespace removed from the end when indexed
2.comparison operators automatically remove whitespace from the end when comparing values
3.vectorized string operations are provided as methods (e.g. str.endswith) and infix operators (e.g. +, *, %)

Parameters
obj : array of str or unicode-like
itemsize : int, optional
itemsize is the number of characters per scalar in the resulting array. If itemsize is
None, and obj is an object array or a Python list, the itemsize will be automatically
determined. If itemsize is provided and obj is of type str or unicode, then the obj string
will be chunked into itemsize pieces.
unicode : bool, optional
When true, the resulting chararray can contain Unicode characters, when false only
8-bit characters. If unicode is None and obj is one of the following:
a chararray,
an ndarray of type str or unicode
a Python str or unicode object,

408 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

then the unicode setting of the output array will be automatically determined.
order : {C, F}, optional
Specify the order of the array. If order is C (default), then the array will be in C-
contiguous order (last-index varies the fastest). If order is F, then the returned array
will be in Fortran-contiguous order (first-index varies the fastest).

3.1.5 Numerical ranges

arange([dtype]) Return evenly spaced values within a given interval.

linspace(start, stop[, num, endpoint, Return evenly spaced numbers over a specified interval.
retstep])
logspace(start, stop[, num, endpoint, Return numbers spaced evenly on a log scale.
base])
meshgrid(x, y) Return coordinate matrices from two coordinate vectors.
mgrid nd_grid instance which returns a dense multi-dimensional
meshgrid.
ogrid nd_grid instance which returns an open multi-dimensional
meshgrid.
arange([start], stop, [step], dtype=None)
Return evenly spaced values within a given interval.
Values are generated within the half-open interval [start, stop) (in other words, the interval including
start but excluding stop). For integer arguments the function is equivalent to the Python built-in range function,
but returns a ndarray rather than a list.
Parameters
start : number, optional
Start of interval. The interval includes this value. The default start value is 0.
stop : number
End of interval. The interval does not include this value.
step : number, optional
Spacing between values. For any output out, this is the distance between two adjacent
values, out[i+1] - out[i]. The default step size is 1. If step is specified, start
must also be given.
dtype : dtype
The type of the output array. If dtype is not given, infer the data type from the other
input arguments.
Returns
out : ndarray
Array of evenly spaced values.
For floating point arguments, the length of the result is ceil((stop -
start)/step). Because of floating point overflow, this rule may result in the last
element of out being greater than stop.
See Also:

linspace
Evenly spaced numbers with careful handling of endpoints.

3.1. Array creation routines 409

NumPy Reference, Release 2.0.0.dev8464

ogrid
Arrays of evenly spaced numbers in N-dimensions
mgrid
Grid-shaped arrays of evenly spaced numbers in N-dimensions

Examples

>>> np.arange(3)
array([0, 1, 2])
>>> np.arange(3.0)
array([ 0., 1., 2.])
>>> np.arange(3,7)
array([3, 4, 5, 6])
>>> np.arange(3,7,2)
array([3, 5])

linspace(start, stop, num=50, endpoint=True, retstep=False)

Return evenly spaced numbers over a specified interval.
Returns num evenly spaced samples, calculated over the interval [start, stop ].
The endpoint of the interval can optionally be excluded.
Parameters
start : scalar
The starting value of the sequence.
stop : scalar
The end value of the sequence, unless endpoint is set to False. In that case, the sequence
consists of all but the last of num + 1 evenly spaced samples, so that stop is excluded.
Note that the step size changes when endpoint is False.
num : int, optional
Number of samples to generate. Default is 50.
endpoint : bool, optional
If True, stop is the last sample. Otherwise, it is not included. Default is True.
retstep : bool, optional
If True, return (samples, step), where step is the spacing between samples.
Returns
samples : ndarray
There are num equally spaced samples in the closed interval [start, stop] or
the half-open interval [start, stop) (depending on whether endpoint is True or
False).
step : float (only if retstep is True)
Size of spacing between samples.
See Also:

arange
Similiar to linspace, but uses a step size (instead of the number of samples).

410 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

logspace
Samples uniformly distributed in log space.

Examples

>>> np.linspace(2.0, 3.0, num=5)

array([ 2. , 2.25, 2.5 , 2.75, 3. ])
>>> np.linspace(2.0, 3.0, num=5, endpoint=False)
array([ 2. , 2.2, 2.4, 2.6, 2.8])
>>> np.linspace(2.0, 3.0, num=5, retstep=True)
(array([ 2. , 2.25, 2.5 , 2.75, 3. ]), 0.25)

Graphical illustration:

>>> import matplotlib.pyplot as plt

>>> N = 8
>>> y = np.zeros(N)
>>> x1 = np.linspace(0, 10, N, endpoint=True)
>>> x2 = np.linspace(0, 10, N, endpoint=False)
>>> plt.plot(x1, y, o)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.plot(x2, y + 0.5, o)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.ylim([-0.5, 1])
(-0.5, 1)
>>> plt.show()

1.0
0.8
0.6
0.4
0.2
0.0
0.2
0.4
0 2 4 6 8 10

logspace(start, stop, num=50, endpoint=True, base=10.0)

Return numbers spaced evenly on a log scale.
In linear space, the sequence starts at base ** start (base to the power of start) and ends with base **
stop (see endpoint below).
Parameters
start : float
base ** start is the starting value of the sequence.
stop : float

3.1. Array creation routines 411

NumPy Reference, Release 2.0.0.dev8464

base ** stop is the final value of the sequence, unless endpoint is False. In that
case, num + 1 values are spaced over the interval in log-space, of which all but the
last (a sequence of length num) are returned.
num : integer, optional
Number of samples to generate. Default is 50.
endpoint : boolean, optional
If true, stop is the last sample. Otherwise, it is not included. Default is True.
base : float, optional
The base of the log space. The step size between the elements in ln(samples) /
ln(base) (or log_base(samples)) is uniform. Default is 10.0.
Returns
samples : ndarray
num samples, equally spaced on a log scale.
See Also:

arange
Similiar to linspace, with the step size specified instead of the number of samples. Note that, when used
with a float endpoint, the endpoint may or may not be included.
linspace
Similar to logspace, but with the samples uniformly distributed in linear space, instead of log space.

Notes
Logspace is equivalent to the code

>>> y = np.linspace(start, stop, num=num, endpoint=endpoint)

...
>>> power(base, y)
...

Examples

>>> np.logspace(2.0, 3.0, num=4)

array([ 100. , 215.443469 , 464.15888336, 1000. ])
>>> np.logspace(2.0, 3.0, num=4, endpoint=False)
array([ 100. , 177.827941 , 316.22776602, 562.34132519])
>>> np.logspace(2.0, 3.0, num=4, base=2.0)
array([ 4. , 5.0396842 , 6.34960421, 8. ])

Graphical illustration:

>>> import matplotlib.pyplot as plt

>>> N = 10
>>> x1 = np.logspace(0.1, 1, N, endpoint=True)
>>> x2 = np.logspace(0.1, 1, N, endpoint=False)
>>> y = np.zeros(N)
>>> plt.plot(x1, y, o)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.plot(x2, y + 0.5, o)
[<matplotlib.lines.Line2D object at 0x...>]

412 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> plt.ylim([-0.5, 1])

(-0.5, 1)
>>> plt.show()

1.0
0.8
0.6
0.4
0.2
0.0
0.2
0.4
1 2 3 4 5 6 7 8 9 10

meshgrid(x, y)
Return coordinate matrices from two coordinate vectors.
Parameters
x, y : ndarray
Two 1-D arrays representing the x and y coordinates of a grid.
Returns
X, Y : ndarray
For vectors x, y with lengths Nx=len(x) and Ny=len(y), return X, Y where X and
Y are (Ny, Nx) shaped arrays with the elements of x and y repeated to fill the matrix
along the first dimension for x, the second for y.
See Also:

index_tricks.mgrid
Construct a multi-dimensional meshgrid using indexing notation.
index_tricks.ogrid
Construct an open multi-dimensional meshgrid using indexing notation.

Examples

>>> X, Y = np.meshgrid([1,2,3], [4,5,6,7])

>>> X
array([[1, 2, 3],
[1, 2, 3],
[1, 2, 3],
[1, 2, 3]])
>>> Y
array([[4, 4, 4],
[5, 5, 5],

3.1. Array creation routines 413

NumPy Reference, Release 2.0.0.dev8464

[6, 6, 6],
[7, 7, 7]])

meshgrid is very useful to evaluate functions on a grid.

>>> x = np.arange(-5, 5, 0.1)

>>> y = np.arange(-5, 5, 0.1)
>>> xx, yy = np.meshgrid(x, y)
>>> z = np.sin(xx**2+yy**2)/(xx**2+yy**2)

mgrid
nd_grid instance which returns a dense multi-dimensional meshgrid.
An instance of numpy.lib.index_tricks.nd_grid which returns an dense (or fleshed out) mesh-grid
when indexed, so that each returned argument has the same shape. The dimensions and number of the output
arrays are equal to the number of indexing dimensions. If the step length is not a complex number, then the stop
is not inclusive.
However, if the step length is a complex number (e.g. 5j), then the integer part of its magnitude is interpreted
as specifying the number of points to create between the start and stop values, where the stop value is inclusive.
Returns
mesh-grid ndarrays all of the same dimensions :
See Also:

numpy.lib.index_tricks.nd_grid
class of ogrid and mgrid objects
ogrid
like mgrid but returns open (not fleshed out) mesh grids
r_
array concatenator

Examples

>>> np.mgrid[0:5,0:5]
array([[[0, 0, 0, 0, 0],
[1, 1, 1, 1, 1],
[2, 2, 2, 2, 2],
[3, 3, 3, 3, 3],
[4, 4, 4, 4, 4]],
[[0, 1, 2, 3, 4],
[0, 1, 2, 3, 4],
[0, 1, 2, 3, 4],
[0, 1, 2, 3, 4],
[0, 1, 2, 3, 4]]])
>>> np.mgrid[-1:1:5j]
array([-1. , -0.5, 0. , 0.5, 1. ])

ogrid
nd_grid instance which returns an open multi-dimensional meshgrid.
An instance of numpy.lib.index_tricks.nd_grid which returns an open (i.e. not fleshed out) mesh-
grid when indexed, so that only one dimension of each returned array is greater than 1. The dimension and
number of the output arrays are equal to the number of indexing dimensions. If the step length is not a complex
number, then the stop is not inclusive.

414 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

However, if the step length is a complex number (e.g. 5j), then the integer part of its magnitude is interpreted
as specifying the number of points to create between the start and stop values, where the stop value is inclusive.
Returns
mesh-grid ndarrays with only one dimension :math:neq 1 :
See Also:

np.lib.index_tricks.nd_grid
class of ogrid and mgrid objects
mgrid
like ogrid but returns dense (or fleshed out) mesh grids
r_
array concatenator

Examples

>>> from numpy import ogrid

>>> ogrid[-1:1:5j]
array([-1. , -0.5, 0. , 0.5, 1. ])
>>> ogrid[0:5,0:5]
[array([[0],
[1],
[2],
[3],
[4]]), array([[0, 1, 2, 3, 4]])]

3.1.6 Building matrices

diag(v[, k]) Extract a diagonal or construct a diagonal array.

diagflat(v[, k]) Create a two-dimensional array with the flattened input as a diagonal.
tri(N[, M, k, dtype]) Construct an array filled with ones at and below the given diagonal.
tril(m[, k]) Lower triangle of an array.
triu(m[, k]) Upper triangle of an array.
vander(x[, N]) Generate a Van der Monde matrix.
diag(v, k=0)
Extract a diagonal or construct a diagonal array.
Parameters
v : array_like
If v is a 2-D array, return a copy of its k-th diagonal. If v is a 1-D array, return a 2-D
array with v on the k-th diagonal.
k : int, optional
Diagonal in question. The default is 0. Use k>0 for diagonals above the main diagonal,
and k for diagonals below the main diagonal.
Returns
out : ndarray
The extracted diagonal or constructed diagonal array.
See Also:

3.1. Array creation routines 415

NumPy Reference, Release 2.0.0.dev8464

diagonal
Return specified diagonals.
diagflat
Create a 2-D array with the flattened input as a diagonal.
trace
Sum along diagonals.
triu
Upper triangle of an array.
tril
Lower triange of an array.

Examples

>>> x = np.arange(9).reshape((3,3))
>>> x
array([[0, 1, 2],
[3, 4, 5],
[6, 7, 8]])

>>> np.diag(x)
array([0, 4, 8])
>>> np.diag(x, k=1)
array([1, 5])
>>> np.diag(x, k=-1)
array([3, 7])

>>> np.diag(np.diag(x))
array([[0, 0, 0],
[0, 4, 0],
[0, 0, 8]])

diagflat(v, k=0)
Create a two-dimensional array with the flattened input as a diagonal.
Parameters
v : array_like
Input data, which is flattened and set as the k-th diagonal of the output.
k : int, optional
Diagonal to set. The default is 0.
Returns
out : ndarray
The 2-D output array.
See Also:

diag
Matlab workalike for 1-D and 2-D arrays.
diagonal
Return specified diagonals.

416 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

trace
Sum along diagonals.

Examples

>>> np.diagflat([[1,2], [3,4]])

array([[1, 0, 0, 0],
[0, 2, 0, 0],
[0, 0, 3, 0],
[0, 0, 0, 4]])

>>> np.diagflat([1,2], 1)
array([[0, 1, 0],
[0, 0, 2],
[0, 0, 0]])

tri(N, M=None, k=0, dtype=<type float>)

Construct an array filled with ones at and below the given diagonal.
Parameters
N : int
Number of rows in the array.
M : int, optional
Number of columns in the array. By default, M is taken equal to N.
k : int, optional
The sub-diagonal below which the array is filled. k = 0 is the main diagonal, while k <
0 is below it, and k > 0 is above. The default is 0.
dtype : dtype, optional
Data type of the returned array. The default is float.
Returns
T : (N,M) ndarray
Array with a lower triangle filled with ones, in other words T[i,j] == 1 for i <=
j + k.

Examples

>>> np.tri(3, 5, 2, dtype=int)

array([[1, 1, 1, 0, 0],
[1, 1, 1, 1, 0],
[1, 1, 1, 1, 1]])

>>> np.tri(3, 5, -1)

array([[ 0., 0., 0., 0., 0.],
[ 1., 0., 0., 0., 0.],
[ 1., 1., 0., 0., 0.]])

tril(m, k=0)
Lower triangle of an array.
Return a copy of an array with elements above the k-th diagonal zeroed.

3.1. Array creation routines 417

NumPy Reference, Release 2.0.0.dev8464

Parameters
m : array_like, shape (M, N)
Input array.
k : int
Diagonal above which to zero elements. k = 0 is the main diagonal, k is below it and k
> 0 is above.
Returns
L : ndarray, shape (M, N)
Lower triangle of m, of same shape and data-type as m.
See Also:
triu

Examples

>>> np.tril([[1,2,3],[4,5,6],[7,8,9],[10,11,12]], -1)

array([[ 0, 0, 0],
[ 4, 0, 0],
[ 7, 8, 0],
[10, 11, 12]])

triu(m, k=0)
Upper triangle of an array.
Construct a copy of a matrix with elements below the k-th diagonal zeroed.
Please refer to the documentation for tril.
See Also:
tril

Examples

>>> np.triu([[1,2,3],[4,5,6],[7,8,9],[10,11,12]], -1)

array([[ 1, 2, 3],
[ 4, 5, 6],
[ 0, 8, 9],
[ 0, 0, 12]])

vander(x, N=None)
Generate a Van der Monde matrix.
The columns of the output matrix are decreasing powers of the input vector. Specifically, the i-th output column
is the input vector to the power of N - i - 1. Such a matrix with a geometric progression in each row is
named Van Der Monde, or Vandermonde matrix, from Alexandre-Theophile Vandermonde.
Parameters
x : array_like
1-D input array.
N : int, optional
Order of (number of columns in) the output. If N is not specified, a square array is
returned (N = len(x)).

418 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Returns
out : ndarray
Van der Monde matrix of order N. The first column is x^(N-1), the second x^(N-2)
and so forth.

References
[R276]

Examples

>>> x = np.array([1, 2, 3, 5])

>>> N = 3
>>> np.vander(x, N)
array([[ 1, 1, 1],
[ 4, 2, 1],
[ 9, 3, 1],
[25, 5, 1]])

>>> np.column_stack([x**(N-1-i) for i in range(N)])

array([[ 1, 1, 1],
[ 4, 2, 1],
[ 9, 3, 1],
[25, 5, 1]])

>>> x = np.array([1, 2, 3, 5])

>>> np.vander(x)
array([[ 1, 1, 1, 1],
[ 8, 4, 2, 1],
[ 27, 9, 3, 1],
[125, 25, 5, 1]])

The determinant of a square Vandermonde matrix is the product of the differences between the values of the
input vector:

>>> np.linalg.det(np.vander(x))
48.000000000000043
>>> (5-3)*(5-2)*(5-1)*(3-2)*(3-1)*(2-1)
48

3.1.7 The Matrix class

mat(data[, dtype]) Interpret the input as a matrix.

bmat(obj[, ldict, gdict]) Build a matrix object from a string, nested sequence, or array.
mat(data, dtype=None)
Interpret the input as a matrix.
Unlike matrix, asmatrix does not make a copy if the input is already a matrix or an ndarray. Equivalent to
matrix(data, copy=False).
Parameters
data : array_like
Input data.

3.1. Array creation routines 419

NumPy Reference, Release 2.0.0.dev8464

Returns
mat : matrix
data interpreted as a matrix.

Examples

>>> x = np.array([[1, 2], [3, 4]])

>>> m = np.asmatrix(x)

>>> x[0,0] = 5

>>> m
matrix([[5, 2],
[3, 4]])

bmat(obj, ldict=None, gdict=None)

Build a matrix object from a string, nested sequence, or array.
Parameters
obj : string, sequence or array
Input data. Variables names in the current scope may be referenced, even if obj is a
string.
Returns
out : matrix
Returns a matrix object, which is a specialized 2-D array.
See Also:
matrix

Examples

>>> A = np.mat(1 1; 1 1)
>>> B = np.mat(2 2; 2 2)
>>> C = np.mat(3 4; 5 6)
>>> D = np.mat(7 8; 9 0)

All the following expressions construct the same block matrix:

>>> np.bmat([[A, B], [C, D]])

matrix([[1, 1, 2, 2],
[1, 1, 2, 2],
[3, 4, 7, 8],
[5, 6, 9, 0]])
>>> np.bmat(np.r_[np.c_[A, B], np.c_[C, D]])
matrix([[1, 1, 2, 2],
[1, 1, 2, 2],
[3, 4, 7, 8],
[5, 6, 9, 0]])
>>> np.bmat(A,B; C,D)
matrix([[1, 1, 2, 2],
[1, 1, 2, 2],

420 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

[3, 4, 7, 8],
[5, 6, 9, 0]])

3.2 Array manipulation routines

3.2.1 Changing array shape

reshape(a, newshape[, order]) Gives a new shape to an array without changing its data.
ravel(a[, order]) Return a flattened array.
ndarray.flat A 1-D iterator over the array.
ndarray.flatten([order]) Return a copy of the array collapsed into one dimension.
reshape(a, newshape, order=C)
Gives a new shape to an array without changing its data.
Parameters
a : array_like
Array to be reshaped.
newshape : int or tuple of ints
The new shape should be compatible with the original shape. If an integer, then the
result will be a 1-D array of that length. One shape dimension can be -1. In this case,
the value is inferred from the length of the array and remaining dimensions.
order : {C, F}, optional
Determines whether the array data should be viewed as in C (row-major) order or FOR-
TRAN (column-major) order.
Returns
reshaped_array : ndarray
This will be a new view object if possible; otherwise, it will be a copy.
See Also:

ndarray.reshape
Equivalent method.

Notes
It is not always possible to change the shape of an array without copying the data. If you want an error to be
raise if the data is copied, you should assign the new shape to the shape attribute of the array:

>>> a = np.zeros((10, 2))

# A transpose make the array non-contiguous
>>> b = a.T
# Taking a view makes it possible to modify the shape without modiying the
# initial object.
>>> c = b.view()
>>> c.shape = (20)
AttributeError: incompatible shape for a non-contiguous array

3.2. Array manipulation routines 421

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.array([[1,2,3], [4,5,6]])

>>> np.reshape(a, 6)
array([1, 2, 3, 4, 5, 6])
>>> np.reshape(a, 6, order=F)
array([1, 4, 2, 5, 3, 6])

>>> np.reshape(a, (3,-1)) # the unspecified value is inferred to be 2

array([[1, 2],
[3, 4],
[5, 6]])

ravel(a, order=C)
Return a flattened array.
A 1-D array, containing the elements of the input, is returned. A copy is made only if needed.
Parameters
a : array_like
Input array. The elements in a are read in the order specified by order, and packed as a
1-D array.
order : {C,F}, optional
The elements of a are read in this order. It can be either C for row-major order, or F
for column-major order. By default, row-major order is used.
Returns
1d_array : ndarray
Output of the same dtype as a, and of shape (a.size(),).
See Also:

ndarray.flat
1-D iterator over an array.
ndarray.flatten
1-D array copy of the elements of an array in row-major order.

Notes
In row-major order, the row index varies the slowest, and the column index the quickest. This can be generalized
to multiple dimensions, where row-major order implies that the index along the first axis varies slowest, and the
index along the last quickest. The opposite holds for Fortran-, or column-major, mode.

Examples
If an array is in C-order (default), then ravel is equivalent to reshape(-1):

>>> x = np.array([[1, 2, 3], [4, 5, 6]])

>>> print x.reshape(-1)
[1 2 3 4 5 6]

>>> print np.ravel(x)

[1 2 3 4 5 6]

When flattening using Fortran-order, however, we see

422 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> print np.ravel(x, order=F)

[1 4 2 5 3 6]

flat
A 1-D iterator over the array.
This is a numpy.flatiter instance, which acts similarly to, but is not a subclass of, Pythons built-in iterator
object.
See Also:

flatten
Return a copy of the array collapsed into one dimension.

flatiter

Examples

>>> x = np.arange(1, 7).reshape(2, 3)

>>> x
array([[1, 2, 3],
[4, 5, 6]])
>>> x.flat[3]
4
>>> x.T
array([[1, 4],
[2, 5],
[3, 6]])
>>> x.T.flat[3]
5
>>> type(x.flat)
<type numpy.flatiter>

An assignment example:

>>> x.flat = 3; x
array([[3, 3, 3],
[3, 3, 3]])
>>> x.flat[[1,4]] = 1; x
array([[3, 1, 3],
[3, 1, 3]])

flatten(order=C)
Return a copy of the array collapsed into one dimension.
Parameters
order : {C, F}, optional
Whether to flatten in C (row-major) or Fortran (column-major) order. The default is
C.
Returns
y : ndarray
A copy of the input array, flattened to one dimension.
See Also:

3.2. Array manipulation routines 423

NumPy Reference, Release 2.0.0.dev8464

ravel
Return a flattened array.
flat
A 1-D flat iterator over the array.

Examples

>>> a = np.array([[1,2], [3,4]])

>>> a.flatten()
array([1, 2, 3, 4])
>>> a.flatten(F)
array([1, 3, 2, 4])

3.2.2 Transpose-like operations

rollaxis(a, axis[, start]) Roll the specified axis backwards, until it lies in a given position.
swapaxes(a, axis1, axis2) Interchange two axes of an array.
ndarray.T Same as self.transpose(), except that self is returned if self.ndim < 2.
transpose(a[, axes]) Permute the dimensions of an array.
rollaxis(a, axis, start=0)
Roll the specified axis backwards, until it lies in a given position.
Parameters
a : ndarray
Input array.
axis : int
The axis to roll backwards. The positions of the other axes do not change relative to one
another.
start : int, optional
The axis is rolled until it lies before this position.
Returns
res : ndarray
Output array.
See Also:

roll
Roll the elements of an array by a number of positions along a given axis.

Examples

>>> a = np.ones((3,4,5,6))
>>> np.rollaxis(a, 3, 1).shape
(3, 6, 4, 5)
>>> np.rollaxis(a, 2).shape
(5, 3, 4, 6)
>>> np.rollaxis(a, 1, 4).shape
(3, 5, 6, 4)

424 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

swapaxes(a, axis1, axis2)

Interchange two axes of an array.
Parameters
a : array_like
Input array.
axis1 : int
First axis.
axis2 : int
Second axis.
Returns
a_swapped : ndarray
If a is an ndarray, then a view of a is returned; otherwise a new array is created.

Examples

>>> x = np.array([[1,2,3]])
>>> np.swapaxes(x,0,1)
array([[1],
[2],
[3]])

>>> x = np.array([[[0,1],[2,3]],[[4,5],[6,7]]])
>>> x
array([[[0, 1],
[2, 3]],
[[4, 5],
[6, 7]]])

>>> np.swapaxes(x,0,2)
array([[[0, 4],
[2, 6]],
[[1, 5],
[3, 7]]])

T
Same as self.transpose(), except that self is returned if self.ndim < 2.

Examples

>>> x = np.array([[1.,2.],[3.,4.]])
>>> x
array([[ 1., 2.],
[ 3., 4.]])
>>> x.T
array([[ 1., 3.],
[ 2., 4.]])
>>> x = np.array([1.,2.,3.,4.])
>>> x
array([ 1., 2., 3., 4.])
>>> x.T
array([ 1., 2., 3., 4.])

3.2. Array manipulation routines 425

NumPy Reference, Release 2.0.0.dev8464

transpose(a, axes=None)
Permute the dimensions of an array.
Parameters
a : array_like
Input array.
axes : list of ints, optional
By default, reverse the dimensions, otherwise permute the axes according to the values
given.
Returns
p : ndarray
a with its axes permuted. A view is returned whenever possible.
See Also:
rollaxis

Examples

>>> x = np.arange(4).reshape((2,2))
>>> x
array([[0, 1],
[2, 3]])

>>> np.transpose(x)
array([[0, 2],
[1, 3]])

>>> x = np.ones((1, 2, 3))

>>> np.transpose(x, (1, 0, 2)).shape
(2, 1, 3)

3.2.3 Changing number of dimensions

atleast_1d(*arys) Convert inputs to arrays with at least one dimension.

atleast_2d(*arys) View inputs as arrays with at least two dimensions.
atleast_3d(*arys) View inputs as arrays with at least three dimensions.
broadcast Produce an object that mimics broadcasting.
broadcast_arrays(*args) Broadcast any number of arrays against each other.
expand_dims(a, axis) Expand the shape of an array.
squeeze(a) Remove single-dimensional entries from the shape of an array.
atleast_1d(*arys)
Convert inputs to arrays with at least one dimension.
Scalar inputs are converted to 1-dimensional arrays, whilst higher-dimensional inputs are preserved.
Parameters
array1, array2, ... : array_like
One or more input arrays.
Returns
ret : ndarray

426 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

An array, or sequence of arrays, each with a.ndim >= 1. Copies are made only if
necessary.
See Also:
atleast_2d, atleast_3d

Examples

>>> np.atleast_1d(1.0)
array([ 1.])

>>> x = np.arange(9.0).reshape(3,3)
>>> np.atleast_1d(x)
array([[ 0., 1., 2.],
[ 3., 4., 5.],
[ 6., 7., 8.]])
>>> np.atleast_1d(x) is x
True

>>> np.atleast_1d(1, [3, 4])

[array([1]), array([3, 4])]

atleast_2d(*arys)
View inputs as arrays with at least two dimensions.
Parameters
array1, array2, ... : array_like
One or more array-like sequences. Non-array inputs are converted to arrays. Arrays
that already have two or more dimensions are preserved.
Returns
res, res2, ... : ndarray
An array, or tuple of arrays, each with a.ndim >= 2. Copies are avoided where
possible, and views with two or more dimensions are returned.
See Also:
atleast_1d, atleast_3d

Examples

>>> np.atleast_2d(3.0)
array([[ 3.]])

>>> x = np.arange(3.0)
>>> np.atleast_2d(x)
array([[ 0., 1., 2.]])
>>> np.atleast_2d(x).base is x
True

>>> np.atleast_2d(1, [1, 2], [[1, 2]])

[array([[1]]), array([[1, 2]]), array([[1, 2]])]

atleast_3d(*arys)
View inputs as arrays with at least three dimensions.

3.2. Array manipulation routines 427

NumPy Reference, Release 2.0.0.dev8464

Parameters
array1, array2, ... : array_like
One or more array-like sequences. Non-array inputs are converted to arrays. Arrays
that already have three or more dimensions are preserved.
Returns
res1, res2, ... : ndarray
An array, or tuple of arrays, each with a.ndim >= 3. Copies are avoided where
possible, and views with three or more dimensions are returned. For example, a 1-D
array of shape N becomes a view of shape (1, N, 1). A 2-D array of shape (M, N)
becomes a view of shape (M, N, 1).
See Also:
atleast_1d, atleast_2d

Examples

>>> np.atleast_3d(3.0)
array([[[ 3.]]])

>>> x = np.arange(3.0)
>>> np.atleast_3d(x).shape
(1, 3, 1)

>>> x = np.arange(12.0).reshape(4,3)
>>> np.atleast_3d(x).shape
(4, 3, 1)
>>> np.atleast_3d(x).base is x
True

>>> for arr in np.atleast_3d([1, 2], [[1, 2]], [[[1, 2]]]):

... print arr, arr.shape
...
[[[1]
[2]]] (1, 2, 1)
[[[1]
[2]]] (1, 2, 1)
[[[1 2]]] (1, 1, 2)

class broadcast()
Produce an object that mimics broadcasting.
Parameters
in1, in2, ... : array_like
Input parameters.
Returns
b : broadcast object
Broadcast the input parameters against one another, and return an object that encapsu-
lates the result. Amongst others, it has shape and nd properties, and may be used as
an iterator.

428 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples
Manually adding two vectors, using broadcasting:

>>> x = np.array([[1], [2], [3]])

>>> y = np.array([4, 5, 6])
>>> b = np.broadcast(x, y)

>>> out = np.empty(b.shape)

>>> out.flat = [u+v for (u,v) in b]
>>> out
array([[ 5., 6., 7.],
[ 6., 7., 8.],
[ 7., 8., 9.]])

Compare against built-in broadcasting:

>>> x + y
array([[5, 6, 7],
[6, 7, 8],
[7, 8, 9]])

Methods
next x.next() -> the next value, or raise StopIteration
reset
next()
x.next() -> the next value, or raise StopIteration
reset()

broadcast_arrays(*args)
Broadcast any number of arrays against each other.
Parameters
*args : arrays
The arrays to broadcast.
Returns
broadcasted : list of arrays
These arrays are views on the original arrays. They are typically not contiguous. Fur-
thermore, more than one element of a broadcasted array may refer to a single memory
location. If you need to write to the arrays, make copies first.

Examples

>>> x = np.array([[1,2,3]])
>>> y = np.array([[1],[2],[3]])
>>> np.broadcast_arrays(x, y)
[array([[1, 2, 3],
[1, 2, 3],
[1, 2, 3]]), array([[1, 1, 1],
[2, 2, 2],
[3, 3, 3]])]

Here is a useful idiom for getting contiguous copies instead of non-contiguous views.

3.2. Array manipulation routines 429

NumPy Reference, Release 2.0.0.dev8464

>>> map(np.array, np.broadcast_arrays(x, y))

[array([[1, 2, 3],
[1, 2, 3],
[1, 2, 3]]), array([[1, 1, 1],
[2, 2, 2],
[3, 3, 3]])]

expand_dims(a, axis)
Expand the shape of an array.
Insert a new axis, corresponding to a given position in the array shape.
Parameters
a : array_like
Input array.
axis : int
Position (amongst axes) where new axis is to be inserted.
Returns
res : ndarray
Output array. The number of dimensions is one greater than that of the input array.
See Also:
doc.indexing, atleast_1d, atleast_2d, atleast_3d

Examples

>>> x = np.array([1,2])
>>> x.shape
(2,)

The following is equivalent to x[np.newaxis,:] or x[np.newaxis]:

>>> y = np.expand_dims(x, axis=0)

>>> y
array([[1, 2]])
>>> y.shape
(1, 2)

>>> y = np.expand_dims(x, axis=1) # Equivalent to x[:,newaxis]

>>> y
array([[1],
[2]])
>>> y.shape
(2, 1)

Note that some examples may use None instead of np.newaxis. These are the same objects:

>>> np.newaxis is None

True

squeeze(a)
Remove single-dimensional entries from the shape of an array.

430 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
a : array_like
Input data.
Returns
squeezed : ndarray
The input array, but with with all dimensions of length 1 removed. Whenever possible,
a view on a is returned.

Examples

>>> x = np.array([[[0], [1], [2]]])

>>> x.shape
(1, 3, 1)
>>> np.squeeze(x).shape
(3,)

3.2.4 Changing kind of array

asarray(a[, dtype, order]) Convert the input to an array.

asanyarray(a[, dtype, order]) Convert the input to a ndarray, but pass ndarray subclasses through.
asmatrix(data[, dtype]) Interpret the input as a matrix.
asfarray(a[, dtype]) Return an array converted to a float type.
asfortranarray(a[, dtype]) Return an array laid out in Fortran order in memory.
asscalar(a) Convert an array of size 1 to its scalar equivalent.
require(a[, dtype, requirements]) Return an ndarray of the provided type that satisfies requirements.
asarray(a, dtype=None, order=None)
Convert the input to an array.
Parameters
a : array_like
Input data, in any form that can be converted to an array. This includes lists, lists of
tuples, tuples, tuples of tuples, tuples of lists and ndarrays.
dtype : data-type, optional
By default, the data-type is inferred from the input data.
order : {C, F}, optional
Whether to use row-major (C) or column-major (F for FORTRAN) memory repre-
sentation. Defaults to C.
Returns
out : ndarray
Array interpretation of a. No copy is performed if the input is already an ndarray. If a
is a subclass of ndarray, a base class ndarray is returned.
See Also:

asanyarray
Similar function which passes through subclasses.
ascontiguousarray
Convert input to a contiguous array.

3.2. Array manipulation routines 431

NumPy Reference, Release 2.0.0.dev8464

asfarray
Convert input to a floating point ndarray.
asfortranarray
Convert input to an ndarray with column-major memory order.
asarray_chkfinite
Similar function which checks input for NaNs and Infs.
fromiter
Create an array from an iterator.
fromfunction
Construct an array by executing a function on grid positions.

Examples
Convert a list into an array:

>>> a = [1, 2]
>>> np.asarray(a)
array([1, 2])

Existing arrays are not copied:

>>> a = np.array([1, 2])

>>> np.asarray(a) is a
True

If dtype is set, array is copied only if dtype does not match:

>>> a = np.array([1, 2], dtype=np.float32)

>>> np.asarray(a, dtype=np.float32) is a
True
>>> np.asarray(a, dtype=np.float64) is a
False

Contrary to asanyarray, ndarray subclasses are not passed through:

>>> issubclass(np.matrix, np.ndarray)

True
>>> a = np.matrix([[1, 2]])
>>> np.asarray(a) is a
False
>>> np.asanyarray(a) is a
True

asanyarray(a, dtype=None, order=None)

Convert the input to a ndarray, but pass ndarray subclasses through.
Parameters
a : array_like
Input data, in any form that can be converted to an array. This includes scalars, lists,
lists of tuples, tuples, tuples of tuples, tuples of lists and ndarrays.
dtype : data-type, optional
By default, the data-type is inferred from the input data.
order : {C, F}, optional

432 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Whether to use row-major (C) or column-major (F) memory representation. Defaults

to C.
Returns
out : ndarray or an ndarray subclass
Array interpretation of a. If a is an ndarray or a subclass of ndarray, it is returned as-is
and no copy is performed.
See Also:

asarray
Similar function which always returns ndarrays.
ascontiguousarray
Convert input to a contiguous array.
asfarray
Convert input to a floating point ndarray.
asfortranarray
Convert input to an ndarray with column-major memory order.
asarray_chkfinite
Similar function which checks input for NaNs and Infs.
fromiter
Create an array from an iterator.
fromfunction
Construct an array by executing a function on grid positions.

Examples
Convert a list into an array:

>>> a = [1, 2]
>>> np.asanyarray(a)
array([1, 2])

Instances of ndarray subclasses are passed through as-is:

>>> a = np.matrix([1, 2])

>>> np.asanyarray(a) is a
True

asmatrix(data, dtype=None)
Interpret the input as a matrix.
Unlike matrix, asmatrix does not make a copy if the input is already a matrix or an ndarray. Equivalent to
matrix(data, copy=False).
Parameters
data : array_like
Input data.
Returns
mat : matrix
data interpreted as a matrix.

3.2. Array manipulation routines 433

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.array([[1, 2], [3, 4]])

>>> m = np.asmatrix(x)

>>> x[0,0] = 5

>>> m
matrix([[5, 2],
[3, 4]])

asfarray(a, dtype=<type numpy.float64>)

Return an array converted to a float type.
Parameters
a : array_like
The input array.
dtype : str or dtype object, optional
Float type code to coerce input array a. If dtype is one of the int dtypes, it is replaced
with float64.
Returns
out : ndarray
The input a as a float ndarray.

Examples

>>> np.asfarray([2, 3])

array([ 2., 3.])
>>> np.asfarray([2, 3], dtype=float)
array([ 2., 3.])
>>> np.asfarray([2, 3], dtype=int8)
array([ 2., 3.])

asfortranarray(a, dtype=None)
Return an array laid out in Fortran order in memory.
Parameters
a : array_like
Input array.
dtype : str or dtype object, optional
By default, the data-type is inferred from the input data.
Returns
out : ndarray
The input a in Fortran, or column-major, order.
See Also:

ascontiguousarray
Convert input to a contiguous (C order) array.

434 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

asanyarray
Convert input to an ndarray with either row or column-major memory order.
require
Return an ndarray that satisfies requirements.
ndarray.flags
Information about the memory layout of the array.

Examples

>>> x = np.arange(6).reshape(2,3)
>>> y = np.asfortranarray(x)
>>> x.flags[F_CONTIGUOUS]
False
>>> y.flags[F_CONTIGUOUS]
True

asscalar(a)
Convert an array of size 1 to its scalar equivalent.
Parameters
a : ndarray
Input array of size 1.
Returns
out : scalar
Scalar representation of a. The input data type is preserved.

Examples

>>> np.asscalar(np.array([24]))
24

require(a, dtype=None, requirements=None)

Return an ndarray of the provided type that satisfies requirements.
This function is useful to be sure that an array with the correct flags is returned for passing to compiled code
(perhaps through ctypes).
Parameters
a : array_like
The object to be converted to a type-and-requirement-satisfying array.
dtype : data-type
The required data-type, the default data-type is float64).
requirements : str or list of str
The requirements list can be any of the following
F_CONTIGUOUS (F) - ensure a Fortran-contiguous array
C_CONTIGUOUS (C) - ensure a C-contiguous array
ALIGNED (A) - ensure a data-type aligned array
WRITEABLE (W) - ensure a writable array
OWNDATA (O) - ensure an array that owns its own data

3.2. Array manipulation routines 435

NumPy Reference, Release 2.0.0.dev8464

See Also:

asarray
Convert input to an ndarray.
asanyarray
Convert to an ndarray, but pass through ndarray subclasses.
ascontiguousarray
Convert input to a contiguous array.
asfortranarray
Convert input to an ndarray with column-major memory order.
ndarray.flags
Information about the memory layout of the array.

Notes
The returned array will be guaranteed to have the listed requirements by making a copy if needed.

Examples

>>> x = np.arange(6).reshape(2,3)
>>> x.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : False
WRITEABLE : True
ALIGNED : True
UPDATEIFCOPY : False

>>> y = np.require(x, dtype=np.float32, requirements=[A, O, W, F])

>>> y.flags
C_CONTIGUOUS : False
F_CONTIGUOUS : True
OWNDATA : True
WRITEABLE : True
ALIGNED : True
UPDATEIFCOPY : False

3.2.5 Joining arrays

column_stack(tup) Stack 1-D arrays as columns into a 2-D array.

concatenate(a1, a2[, axis]) Join a sequence of arrays together.
dstack(tup) Stack arrays in sequence depth wise (along third axis).
hstack(tup) Stack arrays in sequence horizontally (column wise).
vstack(tup) Stack arrays in sequence vertically (row wise).
column_stack(tup)
Stack 1-D arrays as columns into a 2-D array.
Take a sequence of 1-D arrays and stack them as columns to make a single 2-D array. 2-D arrays are stacked
as-is, just like with hstack. 1-D arrays are turned into 2-D columns first.
Parameters
tup : sequence of 1-D or 2-D arrays.

436 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Arrays to stack. All of them must have the same first dimension.
Returns
stacked : 2-D array
The array formed by stacking the given arrays.
See Also:
hstack, vstack, concatenate

Notes
This function is equivalent to np.vstack(tup).T.

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.column_stack((a,b))
array([[1, 2],
[2, 3],
[3, 4]])

concatenate((a1, a2, ...), axis=0)

Join a sequence of arrays together.
Parameters
a1, a2, ... : sequence of array_like
The arrays must have the same shape, except in the dimension corresponding to axis
(the first, by default).
axis : int, optional
The axis along which the arrays will be joined. Default is 0.
Returns
res : ndarray
The concatenated array.
See Also:

ma.concatenate
Concatenate function that preserves input masks.
array_split
Split an array into multiple sub-arrays of equal or near-equal size.
split
Split array into a list of multiple sub-arrays of equal size.
hsplit
Split array into multiple sub-arrays horizontally (column wise)
vsplit
Split array into multiple sub-arrays vertically (row wise)
dsplit
Split array into multiple sub-arrays along the 3rd axis (depth).
hstack
Stack arrays in sequence horizontally (column wise)

3.2. Array manipulation routines 437

NumPy Reference, Release 2.0.0.dev8464

vstack
Stack arrays in sequence vertically (row wise)
dstack
Stack arrays in sequence depth wise (along third dimension)

Notes
When one or more of the arrays to be concatenated is a MaskedArray, this function will return a MaskedArray
object instead of an ndarray, but the input masks are not preserved. In cases where a MaskedArray is expected
as input, use the ma.concatenate function from the masked array module instead.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> b = np.array([[5, 6]])
>>> np.concatenate((a, b), axis=0)
array([[1, 2],
[3, 4],
[5, 6]])
>>> np.concatenate((a, b.T), axis=1)
array([[1, 2, 5],
[3, 4, 6]])

This function will not preserve masking of MaskedArray inputs.

>>> a = np.ma.arange(3)
>>> a[1] = np.ma.masked
>>> b = np.arange(2, 5)
>>> a
masked_array(data = [0 -- 2],
mask = [False True False],
fill_value = 999999)
>>> b
array([2, 3, 4])
>>> np.concatenate([a, b])
masked_array(data = [0 1 2 2 3 4],
mask = False,
fill_value = 999999)
>>> np.ma.concatenate([a, b])
masked_array(data = [0 -- 2 2 3 4],
mask = [False True False False False False],
fill_value = 999999)

dstack(tup)
Stack arrays in sequence depth wise (along third axis).
Takes a sequence of arrays and stack them along the third axis to make a single array. Rebuilds arrays divided
by dsplit. This is a simple way to stack 2D arrays (images) into a single 3D array for processing.
Parameters
tup : sequence of arrays
Arrays to stack. All of them must have the same shape along all but the third axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

438 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:

vstack
Stack along first axis.
hstack
Stack along second axis.
concatenate
Join arrays.
dsplit
Split array along third axis.

Notes
Equivalent to np.concatenate(tup, axis=2).

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.dstack((a,b))
array([[[1, 2],
[2, 3],
[3, 4]]])

>>> a = np.array([[1],[2],[3]])
>>> b = np.array([[2],[3],[4]])
>>> np.dstack((a,b))
array([[[1, 2]],
[[2, 3]],
[[3, 4]]])

hstack(tup)
Stack arrays in sequence horizontally (column wise).
Take a sequence of arrays and stack them horizontally to make a single array. Rebuild arrays divided by hsplit.
Parameters
tup : sequence of ndarrays
All arrays must have the same shape along all but the second axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.
See Also:

vstack
Stack arrays in sequence vertically (row wise).
dstack
Stack arrays in sequence depth wise (along third axis).
concatenate
Join a sequence of arrays together.

3.2. Array manipulation routines 439

NumPy Reference, Release 2.0.0.dev8464

hsplit
Split array along second axis.

Notes
Equivalent to np.concatenate(tup, axis=1)

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.hstack((a,b))
array([1, 2, 3, 2, 3, 4])
>>> a = np.array([[1],[2],[3]])
>>> b = np.array([[2],[3],[4]])
>>> np.hstack((a,b))
array([[1, 2],
[2, 3],
[3, 4]])

vstack(tup)
Stack arrays in sequence vertically (row wise).
Take a sequence of arrays and stack them vertically to make a single array. Rebuild arrays divided by vsplit.
Parameters
tup : sequence of ndarrays
Tuple containing arrays to be stacked. The arrays must have the same shape along all
but the first axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.
See Also:

hstack
Stack arrays in sequence horizontally (column wise).
dstack
Stack arrays in sequence depth wise (along third dimension).
concatenate
Join a sequence of arrays together.
vsplit
Split array into a list of multiple sub-arrays vertically.

Notes
Equivalent to np.concatenate(tup, axis=0)

Examples

>>> a = np.array([1, 2, 3])

>>> b = np.array([2, 3, 4])
>>> np.vstack((a,b))
array([[1, 2, 3],
[2, 3, 4]])

440 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> a = np.array([[1], [2], [3]])

>>> b = np.array([[2], [3], [4]])
>>> np.vstack((a,b))
array([[1],
[2],
[3],
[2],
[3],
[4]])

3.2.6 Splitting arrays

array_split(ary, indices_or_sections[, axis]) Split an array into multiple sub-arrays of equal or near-equal size.
dsplit(ary, indices_or_sections) Split array into multiple sub-arrays along the 3rd axis (depth).
hsplit(ary, indices_or_sections) Split an array into multiple sub-arrays horizontally (column-wise).
split(ary, indices_or_sections[, axis]) Split an array into multiple sub-arrays of equal size.
vsplit(ary, indices_or_sections) Split an array into multiple sub-arrays vertically (row-wise).
array_split(ary, indices_or_sections, axis=0)
Split an array into multiple sub-arrays of equal or near-equal size.
Please refer to the split documentation. The only difference between these functions is that array_split
allows indices_or_sections to be an integer that does not equally divide the axis.
See Also:

split
Split array into multiple sub-arrays of equal size.

Examples

>>> x = np.arange(8.0)
>>> np.array_split(x, 3)
[array([ 0., 1., 2.]), array([ 3., 4., 5.]), array([ 6., 7.])]

dsplit(ary, indices_or_sections)
Split array into multiple sub-arrays along the 3rd axis (depth).
Please refer to the split documentation. dsplit is equivalent to split with axis=2, the array is always split along
the third axis provided the array dimension is greater than or equal to 3.
See Also:

split
Split an array into multiple sub-arrays of equal size.

Examples

>>> x = np.arange(16.0).reshape(2, 2, 4)
>>> x
array([[[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.]],
[[ 8., 9., 10., 11.],
[ 12., 13., 14., 15.]]])
>>> np.dsplit(x, 2)

3.2. Array manipulation routines 441

NumPy Reference, Release 2.0.0.dev8464

[array([[[ 0., 1.],

[ 4., 5.]],
[[ 8., 9.],
[ 12., 13.]]]),
array([[[ 2., 3.],
[ 6., 7.]],
[[ 10., 11.],
[ 14., 15.]]])]
>>> np.dsplit(x, np.array([3, 6]))
[array([[[ 0., 1., 2.],
[ 4., 5., 6.]],
[[ 8., 9., 10.],
[ 12., 13., 14.]]]),
array([[[ 3.],
[ 7.]],
[[ 11.],
[ 15.]]]),
array([], dtype=float64)]

hsplit(ary, indices_or_sections)
Split an array into multiple sub-arrays horizontally (column-wise).
Please refer to the split documentation. hsplit is equivalent to split with axis=1, the array is always split along
the second axis regardless of the array dimension.
See Also:

split
Split an array into multiple sub-arrays of equal size.

Examples

>>> x = np.arange(16.0).reshape(4, 4)
>>> x
array([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.],
[ 12., 13., 14., 15.]])
>>> np.hsplit(x, 2)
[array([[ 0., 1.],
[ 4., 5.],
[ 8., 9.],
[ 12., 13.]]),
array([[ 2., 3.],
[ 6., 7.],
[ 10., 11.],
[ 14., 15.]])]
>>> np.hsplit(x, np.array([3, 6]))
[array([[ 0., 1., 2.],
[ 4., 5., 6.],
[ 8., 9., 10.],
[ 12., 13., 14.]]),
array([[ 3.],
[ 7.],
[ 11.],
[ 15.]]),
array([], dtype=float64)]

442 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

With a higher dimensional array the split is still along the second axis.

>>> x = np.arange(8.0).reshape(2, 2, 2)
>>> x
array([[[ 0., 1.],
[ 2., 3.]],
[[ 4., 5.],
[ 6., 7.]]])
>>> np.hsplit(x, 2)
[array([[[ 0., 1.]],
[[ 4., 5.]]]),
array([[[ 2., 3.]],
[[ 6., 7.]]])]

split(ary, indices_or_sections, axis=0)

Split an array into multiple sub-arrays of equal size.
Parameters
ary : ndarray
Array to be divided into sub-arrays.
indices_or_sections : int or 1-D array
If indices_or_sections is an integer, N, the array will be divided into N equal arrays
along axis. If such a split is not possible, an error is raised.
If indices_or_sections is a 1-D array of sorted integers, the entries indicate where along
axis the array is split. For example, [2, 3] would, for axis=0, result in
ary[:2]
ary[2:3]
ary[3:]
If an index exceeds the dimension of the array along axis, an empty sub-array is returned
correspondingly.
axis : int, optional
The axis along which to split, default is 0.
Returns
sub-arrays : list of ndarrays
A list of sub-arrays.
Raises
ValueError :
If indices_or_sections is given as an integer, but a split does not result in equal division.
See Also:

array_split
Split an array into multiple sub-arrays of equal or near-equal size. Does not raise an exception if an equal
division cannot be made.
hsplit
Split array into multiple sub-arrays horizontally (column-wise).
vsplit
Split array into multiple sub-arrays vertically (row wise).

3.2. Array manipulation routines 443

NumPy Reference, Release 2.0.0.dev8464

dsplit
Split array into multiple sub-arrays along the 3rd axis (depth).
concatenate
Join arrays together.
hstack
Stack arrays in sequence horizontally (column wise).
vstack
Stack arrays in sequence vertically (row wise).
dstack
Stack arrays in sequence depth wise (along third dimension).

Examples

>>> x = np.arange(9.0)
>>> np.split(x, 3)
[array([ 0., 1., 2.]), array([ 3., 4., 5.]), array([ 6., 7., 8.])]

>>> x = np.arange(8.0)
>>> np.split(x, [3, 5, 6, 10])
[array([ 0., 1., 2.]),
array([ 3., 4.]),
array([ 5.]),
array([ 6., 7.]),
array([], dtype=float64)]

vsplit(ary, indices_or_sections)
Split an array into multiple sub-arrays vertically (row-wise).
Please refer to the split documentation. vsplit is equivalent to split with axis=0 (default), the array is
always split along the first axis regardless of the array dimension.
See Also:

split
Split an array into multiple sub-arrays of equal size.

Examples

>>> x = np.arange(16.0).reshape(4, 4)
>>> x
array([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.],
[ 12., 13., 14., 15.]])
>>> np.vsplit(x, 2)
[array([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.]]),
array([[ 8., 9., 10., 11.],
[ 12., 13., 14., 15.]])]
>>> np.vsplit(x, np.array([3, 6]))
[array([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]]),

444 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

array([[ 12., 13., 14., 15.]]),

array([], dtype=float64)]

With a higher dimensional array the split is still along the first axis.

>>> x = np.arange(8.0).reshape(2, 2, 2)
>>> x
array([[[ 0., 1.],
[ 2., 3.]],
[[ 4., 5.],
[ 6., 7.]]])
>>> np.vsplit(x, 2)
[array([[[ 0., 1.],
[ 2., 3.]]]),
array([[[ 4., 5.],
[ 6., 7.]]])]

3.2.7 Tiling arrays

tile(A, reps) Construct an array by repeating A the number of times given by reps.
repeat(a, repeats[, axis]) Repeat elements of an array.
tile(A, reps)
Construct an array by repeating A the number of times given by reps.
If reps has length d, the result will have dimension of max(d, A.ndim).
If A.ndim < d, A is promoted to be d-dimensional by prepending new axes. So a shape (3,) array is promoted
to (1, 3) for 2-D replication, or shape (1, 1, 3) for 3-D replication. If this is not the desired behavior, promote A
to d-dimensions manually before calling this function.
If A.ndim > d, reps is promoted to A.ndim by pre-pending 1s to it. Thus for an A of shape (2, 3, 4, 5), a reps
of (2, 2) is treated as (1, 1, 2, 2).
Parameters
A : array_like
The input array.
reps : array_like
The number of repetitions of A along each axis.
Returns
c : ndarray
The tiled output array.
See Also:

repeat
Repeat elements of an array.

Examples

>>> a = np.array([0, 1, 2])

>>> np.tile(a, 2)
array([0, 1, 2, 0, 1, 2])

3.2. Array manipulation routines 445

NumPy Reference, Release 2.0.0.dev8464

>>> np.tile(a, (2, 2))

array([[0, 1, 2, 0, 1, 2],
[0, 1, 2, 0, 1, 2]])
>>> np.tile(a, (2, 1, 2))
array([[[0, 1, 2, 0, 1, 2]],
[[0, 1, 2, 0, 1, 2]]])

>>> b = np.array([[1, 2], [3, 4]])

>>> np.tile(b, 2)
array([[1, 2, 1, 2],
[3, 4, 3, 4]])
>>> np.tile(b, (2, 1))
array([[1, 2],
[3, 4],
[1, 2],
[3, 4]])

repeat(a, repeats, axis=None)

Repeat elements of an array.
Parameters
a : array_like
Input array.
repeats : {int, array of ints}
The number of repetitions for each element. repeats is broadcasted to fit the shape of
the given axis.
axis : int, optional
The axis along which to repeat values. By default, use the flattened input array, and
return a flat output array.
Returns
repeated_array : ndarray
Output array which has the same shape as a, except along the given axis.
See Also:

tile
Tile an array.

Examples

>>> x = np.array([[1,2],[3,4]])
>>> np.repeat(x, 2)
array([1, 1, 2, 2, 3, 3, 4, 4])
>>> np.repeat(x, 3, axis=1)
array([[1, 1, 1, 2, 2, 2],
[3, 3, 3, 4, 4, 4]])
>>> np.repeat(x, [1, 2], axis=0)
array([[1, 2],
[3, 4],
[3, 4]])

446 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

3.2.8 Adding and removing elements

delete(arr, obj[, axis]) Return a new array with sub-arrays along an axis deleted.
insert(arr, obj, values[, axis]) Insert values along the given axis before the given indices.
append(arr, values[, axis]) Append values to the end of an array.
resize(a, new_shape) Return a new array with the specified shape.
trim_zeros(filt[, trim]) Trim the leading and/or trailing zeros from a 1-D array or sequence.
unique(ar[, return_index, return_inverse]) Find the unique elements of an array.
delete(arr, obj, axis=None)
Return a new array with sub-arrays along an axis deleted.
Parameters
arr : array_like
Input array.
obj : slice, int or array of ints
Indicate which sub-arrays to remove.
axis : int, optional
The axis along which to delete the subarray defined by obj. If axis is None, obj is
applied to the flattened array.
Returns
out : ndarray
A copy of arr with the elements specified by obj removed. Note that delete does not
occur in-place. If axis is None, out is a flattened array.
See Also:

insert
Insert elements into an array.
append
Append elements at the end of an array.

Examples

>>> arr = np.array([[1,2,3,4], [5,6,7,8], [9,10,11,12]])

>>> arr
array([[ 1, 2, 3, 4],
[ 5, 6, 7, 8],
[ 9, 10, 11, 12]])
>>> np.delete(arr, 1, 0)
array([[ 1, 2, 3, 4],
[ 9, 10, 11, 12]])

>>> np.delete(arr, np.s_[::2], 1)

array([[ 2, 4],
[ 6, 8],
[10, 12]])
>>> np.delete(arr, [1,3,5], None)
array([ 1, 3, 5, 7, 8, 9, 10, 11, 12])

insert(arr, obj, values, axis=None)

Insert values along the given axis before the given indices.

3.2. Array manipulation routines 447

NumPy Reference, Release 2.0.0.dev8464

Parameters
arr : array_like
Input array.
obj : int, slice or sequence of ints
Object that defines the index or indices before which values is inserted.
values : array_like
Values to insert into arr. If the type of values is different from that of arr, values is
converted to the type of arr.
axis : int, optional
Axis along which to insert values. If axis is None then arr is flattened first.
Returns
out : ndarray
A copy of arr with values inserted. Note that insert does not occur in-place: a new array
is returned. If axis is None, out is a flattened array.
See Also:

append
Append elements at the end of an array.
delete
Delete elements from an array.

Examples

>>> a = np.array([[1, 1], [2, 2], [3, 3]])

>>> a
array([[1, 1],
[2, 2],
[3, 3]])
>>> np.insert(a, 1, 5)
array([1, 5, 1, 2, 2, 3, 3])
>>> np.insert(a, 1, 5, axis=1)
array([[1, 5, 1],
[2, 5, 2],
[3, 5, 3]])

>>> b = a.flatten()
>>> b
array([1, 1, 2, 2, 3, 3])
>>> np.insert(b, [2, 2], [5, 6])
array([1, 1, 5, 6, 2, 2, 3, 3])

>>> np.insert(b, slice(2, 4), [5, 6])

array([1, 1, 5, 2, 6, 2, 3, 3])

>>> np.insert(b, [2, 2], [7.13, False]) # type casting

array([1, 1, 7, 0, 2, 2, 3, 3])

448 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> x = np.arange(8).reshape(2, 4)
>>> idx = (1, 3)
>>> np.insert(x, idx, 999, axis=1)
array([[ 0, 999, 1, 2, 999, 3],
[ 4, 999, 5, 6, 999, 7]])

append(arr, values, axis=None)

Append values to the end of an array.
Parameters
arr : array_like
Values are appended to a copy of this array.
values : array_like
These values are appended to a copy of arr. It must be of the correct shape (the same
shape as arr, excluding axis). If axis is not specified, values can be any shape and will
be flattened before use.
axis : int, optional
The axis along which values are appended. If axis is not given, both arr and values are
flattened before use.
Returns
out : ndarray
A copy of arr with values appended to axis. Note that append does not occur in-place:
a new array is allocated and filled. If axis is None, out is a flattened array.
See Also:

insert
Insert elements into an array.
delete
Delete elements from an array.

Examples

>>> np.append([1, 2, 3], [[4, 5, 6], [7, 8, 9]])

array([1, 2, 3, 4, 5, 6, 7, 8, 9])

When axis is specified, values must have the correct shape.

>>> np.append([[1, 2, 3], [4, 5, 6]], [[7, 8, 9]], axis=0)

array([[1, 2, 3],
[4, 5, 6],
[7, 8, 9]])
>>> np.append([[1, 2, 3], [4, 5, 6]], [7, 8, 9], axis=0)
...
ValueError: arrays must have same number of dimensions

resize(a, new_shape)
Return a new array with the specified shape.
If the new array is larger than the original array, then the new array is filled with repeated copied of a. Note that
this behavior is different from a.resize(new_shape) which fills with zeros instead of repeated copies of a.

3.2. Array manipulation routines 449

NumPy Reference, Release 2.0.0.dev8464

Parameters
a : array_like
Array to be resized.
new_shape : {tuple, int}
Shape of resized array.
Returns
reshaped_array : ndarray
The new array is formed from the data in the old array, repeated if necessary to fill out
the required number of elements. The data are repeated in the order that the data are
stored in memory.
See Also:

ndarray.resize
resize an array in-place.

Examples

>>> a=np.array([[0,1],[2,3]])
>>> np.resize(a,(1,4))
array([[0, 1, 2, 3]])
>>> np.resize(a,(2,4))
array([[0, 1, 2, 3],
[0, 1, 2, 3]])

trim_zeros(filt, trim=fb)
Trim the leading and/or trailing zeros from a 1-D array or sequence.
Parameters
filt : 1-D array or sequence
Input array.
trim : str, optional
A string with f representing trim from front and b to trim from back. Default is fb,
trim zeros from both front and back of the array.
Returns
trimmed : 1-D array or sequence
The result of trimming the input. The input data type is preserved.

Examples

>>> a = np.array((0, 0, 0, 1, 2, 3, 0, 2, 1, 0))

>>> np.trim_zeros(a)
array([1, 2, 3, 0, 2, 1])

>>> np.trim_zeros(a, b)
array([0, 0, 0, 1, 2, 3, 0, 2, 1])

The input data type is preserved, list/tuple in means list/tuple out.

450 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.trim_zeros([0, 1, 2, 0])

[1, 2]

unique(ar, return_index=False, return_inverse=False)

Find the unique elements of an array.
Returns the sorted unique elements of an array. There are two optional outputs in addition to the unique elements:
the indices of the input array that give the unique values, and the indices of the unique array that reconstruct the
input array.
Parameters
ar : array_like
Input array. This will be flattened if it is not already 1-D.
return_index : bool, optional
If True, also return the indices of ar that result in the unique array.
return_inverse : bool, optional
If True, also return the indices of the unique array that can be used to reconstruct ar.
Returns
unique : ndarray
The sorted unique values.
unique_indices : ndarray, optional
The indices of the unique values in the (flattened) original array. Only provided if
return_index is True.
unique_inverse : ndarray, optional
The indices to reconstruct the (flattened) original array from the unique array. Only
provided if return_inverse is True.
See Also:

numpy.lib.arraysetops
Module with a number of other functions for performing set operations on arrays.

Examples

>>> np.unique([1, 1, 2, 2, 3, 3])

array([1, 2, 3])
>>> a = np.array([[1, 1], [2, 3]])
>>> np.unique(a)
array([1, 2, 3])

Return the indices of the original array that give the unique values:

>>> a = np.array([a, b, b, c, a])

>>> u, indices = np.unique(a, return_index=True)
>>> u
array([a, b, c],
dtype=|S1)
>>> indices
array([0, 1, 3])
>>> a[indices]

3.2. Array manipulation routines 451

NumPy Reference, Release 2.0.0.dev8464

array([a, b, c],
dtype=|S1)

Reconstruct the input array from the unique values:

>>> a = np.array([1, 2, 6, 4, 2, 3, 2])

>>> u, indices = np.unique(a, return_inverse=True)
>>> u
array([1, 2, 3, 4, 6])
>>> indices
array([0, 1, 4, 3, 1, 2, 1])
>>> u[indices]
array([1, 2, 6, 4, 2, 3, 2])

3.2.9 Rearranging elements

fliplr(m) Flip array in the left/right direction.

flipud(m) Flip array in the up/down direction.
reshape(a, newshape[, order]) Gives a new shape to an array without changing its data.
roll(a, shift[, axis]) Roll array elements along a given axis.
rot90(m[, k]) Rotate an array by 90 degrees in the counter-clockwise direction.
fliplr(m)
Flip array in the left/right direction.
Flip the entries in each row in the left/right direction. Columns are preserved, but appear in a different order
than before.
Parameters
m : array_like
Input array.
Returns
f : ndarray
A view of m with the columns reversed. Since a view is returned, this operation is O(1).
See Also:

flipud
Flip array in the up/down direction.
rot90
Rotate array counterclockwise.

Notes
Equivalent to A[:,::-1]. Does not require the array to be two-dimensional.

Examples

>>> A = np.diag([1.,2.,3.])
>>> A
array([[ 1., 0., 0.],
[ 0., 2., 0.],
[ 0., 0., 3.]])

452 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.fliplr(A)
array([[ 0., 0., 1.],
[ 0., 2., 0.],
[ 3., 0., 0.]])

>>> A = np.random.randn(2,3,5)
>>> np.all(np.fliplr(A)==A[:,::-1,...])
True

flipud(m)
Flip array in the up/down direction.
Flip the entries in each column in the up/down direction. Rows are preserved, but appear in a different order
than before.
Parameters
m : array_like
Input array.
Returns
out : array_like
A view of m with the rows reversed. Since a view is returned, this operation is O(1).
See Also:

fliplr
Flip array in the left/right direction.
rot90
Rotate array counterclockwise.

Notes
Equivalent to A[::-1,...]. Does not require the array to be two-dimensional.

Examples

>>> A = np.diag([1.0, 2, 3])

>>> A
array([[ 1., 0., 0.],
[ 0., 2., 0.],
[ 0., 0., 3.]])
>>> np.flipud(A)
array([[ 0., 0., 3.],
[ 0., 2., 0.],
[ 1., 0., 0.]])

>>> A = np.random.randn(2,3,5)
>>> np.all(np.flipud(A)==A[::-1,...])
True

>>> np.flipud([1,2])
array([2, 1])

reshape(a, newshape, order=C)

Gives a new shape to an array without changing its data.

3.2. Array manipulation routines 453

NumPy Reference, Release 2.0.0.dev8464

Parameters
a : array_like
Array to be reshaped.
newshape : int or tuple of ints
The new shape should be compatible with the original shape. If an integer, then the
result will be a 1-D array of that length. One shape dimension can be -1. In this case,
the value is inferred from the length of the array and remaining dimensions.
order : {C, F}, optional
Determines whether the array data should be viewed as in C (row-major) order or FOR-
TRAN (column-major) order.
Returns
reshaped_array : ndarray
This will be a new view object if possible; otherwise, it will be a copy.
See Also:

ndarray.reshape
Equivalent method.

Notes
It is not always possible to change the shape of an array without copying the data. If you want an error to be
raise if the data is copied, you should assign the new shape to the shape attribute of the array:

>>> a = np.zeros((10, 2))

# A transpose make the array non-contiguous
>>> b = a.T
# Taking a view makes it possible to modify the shape without modiying the
# initial object.
>>> c = b.view()
>>> c.shape = (20)
AttributeError: incompatible shape for a non-contiguous array

Examples

>>> a = np.array([[1,2,3], [4,5,6]])

>>> np.reshape(a, 6)
array([1, 2, 3, 4, 5, 6])
>>> np.reshape(a, 6, order=F)
array([1, 4, 2, 5, 3, 6])

>>> np.reshape(a, (3,-1)) # the unspecified value is inferred to be 2

array([[1, 2],
[3, 4],
[5, 6]])

roll(a, shift, axis=None)

Roll array elements along a given axis.
Elements that roll beyond the last position are re-introduced at the first.
Parameters
a : array_like

454 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Input array.
shift : int
The number of places by which elements are shifted.
axis : int, optional
The axis along which elements are shifted. By default, the array is flattened before
shifting, after which the original shape is restored.
Returns
res : ndarray
Output array, with the same shape as a.
See Also:

rollaxis
Roll the specified axis backwards, until it lies in a given position.

Examples

>>> x = np.arange(10)
>>> np.roll(x, 2)
array([8, 9, 0, 1, 2, 3, 4, 5, 6, 7])

>>> x2 = np.reshape(x, (2,5))

>>> x2
array([[0, 1, 2, 3, 4],
[5, 6, 7, 8, 9]])
>>> np.roll(x2, 1)
array([[9, 0, 1, 2, 3],
[4, 5, 6, 7, 8]])
>>> np.roll(x2, 1, axis=0)
array([[5, 6, 7, 8, 9],
[0, 1, 2, 3, 4]])
>>> np.roll(x2, 1, axis=1)
array([[4, 0, 1, 2, 3],
[9, 5, 6, 7, 8]])

rot90(m, k=1)
Rotate an array by 90 degrees in the counter-clockwise direction.
The first two dimensions are rotated; therefore, the array must be at least 2-D.
Parameters
m : array_like
Array of two or more dimensions.
k : integer
Number of times the array is rotated by 90 degrees.
Returns
y : ndarray
Rotated array.
See Also:

3.2. Array manipulation routines 455

NumPy Reference, Release 2.0.0.dev8464

fliplr
Flip an array horizontally.
flipud
Flip an array vertically.

Examples

>>> m = np.array([[1,2],[3,4]], int)

>>> m
array([[1, 2],
[3, 4]])
>>> np.rot90(m)
array([[2, 4],
[1, 3]])
>>> np.rot90(m, 2)
array([[4, 3],
[2, 1]])

3.3 Indexing routines

See Also:
Indexing

3.3.1 Generating index arrays

c_ Translates slice objects to concatenation along the second axis.

r_ Translates slice objects to concatenation along the first axis.
s_ A nicer way to build up index tuples for arrays.
nonzero(a) Return the indices of the elements that are non-zero.
where(condition, x) Return elements, either from x or y, depending on condition.
indices(dimensions[, dtype]) Return an array representing the indices of a grid.
ix_(*args) Construct an open mesh from multiple sequences.
ogrid nd_grid instance which returns an open multi-dimensional meshgrid.
unravel_index(x, dims) Convert a flat index to an index tuple for an array of given shape.
diag_indices(n[, ndim]) Return the indices to access the main diagonal of an array.
diag_indices_from(arr) Return the indices to access the main diagonal of an n-dimensional array.
mask_indices(n, mask_func[, k]) Return the indices to access (n, n) arrays, given a masking function.
tril_indices(n[, k]) Return the indices for the lower-triangle of an (n, n) array.
tril_indices_from(arr[, k]) Return the indices for the lower-triangle of an (n, n) array.
triu_indices(n[, k]) Return the indices for the upper-triangle of an (n, n) array.
triu_indices_from(arr[, k]) Return the indices for the upper-triangle of an (n, n) array.
c_
Translates slice objects to concatenation along the second axis.
This is short-hand for np.r_[-1,2,0, index expression], which is useful because of its common
occurrence. In particular, arrays will be stacked along their last axis after being upgraded to at least 2-D with
1s post-pended to the shape (column vectors made out of 1-D arrays).
For detailed documentation, see r_.

456 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.c_[np.array([[1,2,3]]), 0, 0, np.array([[4,5,6]])]

array([[1, 2, 3, 0, 0, 4, 5, 6]])

r_
Translates slice objects to concatenation along the first axis.
This is a simple way to build up arrays quickly. There are two use cases.
1.If the index expression contains comma separated arrays, then stack them along their first axis.
2.If the index expression contains slice notation or scalars then create a 1-D array with a range indicated by
the slice notation.
If slice notation is used, the syntax start:stop:step is equivalent to np.arange(start, stop,
step) inside of the brackets. However, if step is an imaginary number (i.e. 100j) then its integer
portion is interpreted as a number-of-points desired and the start and stop are inclusive. In other words
start:stop:stepj is interpreted as np.linspace(start, stop, step, endpoint=1) inside
of the brackets. After expansion of slice notation, all comma separated sequences are concatenated together.
Optional character strings placed as the first element of the index expression can be used to change the output.
The strings r or c result in matrix output. If the result is 1-D and r is specified a 1 x N (row) matrix is
produced. If the result is 1-D and c is specified, then a N x 1 (column) matrix is produced. If the result is 2-D
then both provide the same matrix result.
A string integer specifies which axis to stack multiple comma separated arrays along. A string of two comma-
separated integers allows indication of the minimum number of dimensions to force each entry into as the second
integer (the axis to concatenate along is still the first integer).
A string with three comma-separated integers allows specification of the axis to concatenate along, the minimum
number of dimensions to force the entries to, and which axis should contain the start of the arrays which are less
than the specified number of dimensions. In other words the third integer allows you to specify where the 1s
should be placed in the shape of the arrays that have their shapes upgraded. By default, they are placed in the
front of the shape tuple. The third argument allows you to specify where the start of the array should be instead.
Thus, a third argument of 0 would place the 1s at the end of the array shape. Negative integers specify where
in the new shape tuple the last dimension of upgraded arrays should be placed, so the default is -1.
Parameters
Not a function, so takes no parameters :
Returns
A concatenated ndarray or matrix. :
See Also:

concatenate
Join a sequence of arrays together.
c_
Translates slice objects to concatenation along the second axis.

Examples

>>> np.r_[np.array([1,2,3]), 0, 0, np.array([4,5,6])]

array([1, 2, 3, 0, 0, 4, 5, 6])
>>> np.r_[-1:1:6j, [0]*3, 5, 6]
array([-1. , -0.6, -0.2, 0.2, 0.6, 1. , 0. , 0. , 0. , 5. , 6. ])

String integers specify the axis to concatenate along or the minimum number of dimensions to force entries into.

3.3. Indexing routines 457

NumPy Reference, Release 2.0.0.dev8464

>>> a = np.array([[0, 1, 2], [3, 4, 5]])

>>> np.r_[-1, a, a] # concatenate along last axis
array([[0, 1, 2, 0, 1, 2],
[3, 4, 5, 3, 4, 5]])
>>> np.r_[0,2, [1,2,3], [4,5,6]] # concatenate along first axis, dim>=2
array([[1, 2, 3],
[4, 5, 6]])

>>> np.r_[0,2,0, [1,2,3], [4,5,6]]

array([[1],
[2],
[3],
[4],
[5],
[6]])
>>> np.r_[1,2,0, [1,2,3], [4,5,6]]
array([[1, 4],
[2, 5],
[3, 6]])

Using r or c as a first string argument creates a matrix.

>>> np.r_[r,[1,2,3], [4,5,6]]

matrix([[1, 2, 3, 4, 5, 6]])

s_
A nicer way to build up index tuples for arrays.
Note: Use one of the two predefined instances index_exp or s_ rather than directly using IndexExpression.
For any index combination, including slicing and axis insertion, a[indices] is the same as
a[np.index_exp[indices]] for any array a. However, np.index_exp[indices] can be used any-
where in Python code and returns a tuple of slice objects that can be used in the construction of complex index
expressions.
Parameters
maketuple : bool
If True, always returns a tuple.
See Also:

index_exp
Predefined instance that always returns a tuple: index_exp = IndexExpression(maketuple=True).
s_
Predefined instance without tuple conversion: s_ = IndexExpression(maketuple=False).

Notes
You can do all this with slice() plus a few special objects, but theres a lot to remember and this version is simpler
because it uses the standard array indexing syntax.

Examples

>>> np.s_[2::2]
slice(2, None, 2)

458 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.index_exp[2::2]
(slice(2, None, 2),)

>>> np.array([0, 1, 2, 3, 4])[np.s_[2::2]]

array([2, 4])

nonzero(a)
Return the indices of the elements that are non-zero.
Returns a tuple of arrays, one for each dimension of a, containing the indices of the non-zero elements in that
dimension. The corresponding non-zero values can be obtained with:

a[nonzero(a)]

To group the indices by element, rather than dimension, use:

transpose(nonzero(a))

The result of this is always a 2-D array, with a row for each non-zero element.
Parameters
a : array_like
Input array.
Returns
tuple_of_arrays : tuple
Indices of elements that are non-zero.
See Also:

flatnonzero
Return indices that are non-zero in the flattened version of the input array.
ndarray.nonzero
Equivalent ndarray method.

Examples

>>> x = np.eye(3)
>>> x
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])
>>> np.nonzero(x)
(array([0, 1, 2]), array([0, 1, 2]))

>>> x[np.nonzero(x)]
array([ 1., 1., 1.])
>>> np.transpose(np.nonzero(x))
array([[0, 0],
[1, 1],
[2, 2]])

A common use for nonzero is to find the indices of an array, where a condition is True. Given an array a, the
condition a > 3 is a boolean array and since False is interpreted as 0, np.nonzero(a > 3) yields the indices of the
a where the condition is true.

3.3. Indexing routines 459

NumPy Reference, Release 2.0.0.dev8464

>>> a = np.array([[1,2,3],[4,5,6],[7,8,9]])
>>> a > 3
array([[False, False, False],
[ True, True, True],
[ True, True, True]], dtype=bool)
>>> np.nonzero(a > 3)
(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

The nonzero method of the boolean array can also be called.

>>> (a > 3).nonzero()

(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

where(condition, [x, y])

Return elements, either from x or y, depending on condition.
If only condition is given, return condition.nonzero().
Parameters
condition : array_like, bool
When True, yield x, otherwise yield y.
x, y : array_like, optional
Values from which to choose. x and y need to have the same shape as condition.
Returns
out : ndarray or tuple of ndarrays
If both x and y are specified, the output array contains elements of x where condition is
True, and elements from y elsewhere.
If only condition is given, return the tuple condition.nonzero(), the indices
where condition is True.
See Also:
nonzero, choose

Notes
If x and y are given and input arrays are 1-D, where is equivalent to:

[xv if c else yv for (c,xv,yv) in zip(condition,x,y)]

Examples

>>> np.where([[True, False], [True, True]],

... [[1, 2], [3, 4]],
... [[9, 8], [7, 6]])
array([[1, 8],
[3, 4]])

>>> np.where([[0, 1], [1, 0]])

(array([0, 1]), array([1, 0]))

460 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> x = np.arange(9.).reshape(3, 3)
>>> np.where( x > 5 )
(array([2, 2, 2]), array([0, 1, 2]))
>>> x[np.where( x > 3.0 )] # Note: result is 1D.
array([ 4., 5., 6., 7., 8.])
>>> np.where(x < 5, x, -1) # Note: broadcasting.
array([[ 0., 1., 2.],
[ 3., 4., -1.],
[-1., -1., -1.]])

indices(dimensions, dtype=<type int>)

Return an array representing the indices of a grid.
Compute an array where the subarrays contain index values 0,1,... varying only along the corresponding axis.
Parameters
dimensions : sequence of ints
The shape of the grid.
dtype : dtype, optional
Data type of the result.
Returns
grid : ndarray
The array of grid indices, grid.shape = (len(dimensions),) +
tuple(dimensions).
See Also:
mgrid, meshgrid

Notes
The output shape is obtained by prepending the number of dimensions in front of the tuple of dimensions, i.e. if
dimensions is a tuple (r0, ..., rN-1) of length N, the output shape is (N,r0,...,rN-1).
The subarrays grid[k] contains the N-D array of indices along the k-th axis. Explicitly:

grid[k,i0,i1,...,iN-1] = ik

Examples

>>> grid = np.indices((2, 3))

>>> grid.shape
(2, 2, 3)
>>> grid[0] # row indices
array([[0, 0, 0],
[1, 1, 1]])
>>> grid[1] # column indices
array([[0, 1, 2],
[0, 1, 2]])

The indices can be used as an index into an array.

>>> x = np.arange(20).reshape(5, 4)
>>> row, col = np.indices((2, 3))
>>> x[row, col]

3.3. Indexing routines 461

NumPy Reference, Release 2.0.0.dev8464

array([[0, 1, 2],
[4, 5, 6]])

Note that it would be more straightforward in the above example to extract the required elements directly with
x[:2, :3].
ix_(*args)
Construct an open mesh from multiple sequences.
This function takes N 1-D sequences and returns N outputs with N dimensions each, such that the shape is 1 in
all but one dimension and the dimension with the non-unit shape value cycles through all N dimensions.
Using ix_ one can quickly construct index arrays that will index the cross product.
a[np.ix_([1,3],[2,5])] returns the array [[a[1,2] a[1,5]], [a[3,2] a[3,5]]].
Parameters
args : 1-D sequences
Returns
out : tuple of ndarrays
N arrays with N dimensions each, with N the number of input sequences. Together
these arrays form an open mesh.
See Also:
ogrid, mgrid, meshgrid

Examples

>>> a = np.arange(10).reshape(2, 5)
>>> a
array([[0, 1, 2, 3, 4],
[5, 6, 7, 8, 9]])
>>> ixgrid = np.ix_([0,1], [2,4])
>>> ixgrid
(array([[0],
[1]]), array([[2, 4]]))
>>> ixgrid[0].shape, ixgrid[1].shape
((2, 1), (1, 2))
>>> a[ixgrid]
array([[2, 4],
[7, 9]])

ogrid
nd_grid instance which returns an open multi-dimensional meshgrid.
An instance of numpy.lib.index_tricks.nd_grid which returns an open (i.e. not fleshed out) mesh-
grid when indexed, so that only one dimension of each returned array is greater than 1. The dimension and
number of the output arrays are equal to the number of indexing dimensions. If the step length is not a complex
number, then the stop is not inclusive.
However, if the step length is a complex number (e.g. 5j), then the integer part of its magnitude is interpreted
as specifying the number of points to create between the start and stop values, where the stop value is inclusive.
Returns
mesh-grid ndarrays with only one dimension :math:neq 1 :
See Also:

462 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

np.lib.index_tricks.nd_grid
class of ogrid and mgrid objects
mgrid
like ogrid but returns dense (or fleshed out) mesh grids
r_
array concatenator

Examples

>>> from numpy import ogrid

>>> ogrid[-1:1:5j]
array([-1. , -0.5, 0. , 0.5, 1. ])
>>> ogrid[0:5,0:5]
[array([[0],
[1],
[2],
[3],
[4]]), array([[0, 1, 2, 3, 4]])]

unravel_index(x, dims)
Convert a flat index to an index tuple for an array of given shape.
Parameters
x : int
Flattened index.
dims : tuple of ints
Input shape, the shape of an array into which indexing is required.
Returns
idx : tuple of ints
Tuple of the same shape as dims, containing the unraveled index.

Notes
In the Examples section, since arr.flat[x] == arr.max() it may be easier to use flattened indexing
than to re-map the index to a tuple.

Examples

>>> arr = np.arange(20).reshape(5, 4)

>>> arr
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11],
[12, 13, 14, 15],
[16, 17, 18, 19]])
>>> x = arr.argmax()
>>> x
19
>>> dims = arr.shape
>>> idx = np.unravel_index(x, dims)
>>> idx
(4, 3)

3.3. Indexing routines 463

NumPy Reference, Release 2.0.0.dev8464

>>> arr[idx] == arr.max()

True

diag_indices(n, ndim=2)
Return the indices to access the main diagonal of an array.
This returns a tuple of indices that can be used to access the main diagonal of an array a with a.ndim >= 2
dimensions and shape (n, n, ..., n). For a.ndim = 2 this is the usual diagonal, for a.ndim > 2 this is the
set of indices to access a[i, i, ..., i] for i = [0..n-1].
Parameters
n : int
The size, along each dimension, of the arrays for which the returned indices can be
used.
ndim : int, optional
The number of dimensions.
See Also:
diag_indices_from

Notes
New in version 1.4.0.

Examples
Create a set of indices to access the diagonal of a (4, 4) array:

>>> di = np.diag_indices(4)
>>> di
(array([0, 1, 2, 3]), array([0, 1, 2, 3]))
>>> a = np.arange(16).reshape(4, 4)
>>> a
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11],
[12, 13, 14, 15]])
>>> a[di] = 100
>>> a
array([[100, 1, 2, 3],
[ 4, 100, 6, 7],
[ 8, 9, 100, 11],
[ 12, 13, 14, 100]])

Now, we create indices to manipulate a 3-D array:

>>> d3 = np.diag_indices(2, 3)
>>> d3
(array([0, 1]), array([0, 1]), array([0, 1]))

And use it to set the diagonal of an array of zeros to 1:

>>> a = np.zeros((2, 2, 2), dtype=np.int)

>>> a[d3] = 1
>>> a
array([[[1, 0],

464 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

[0, 0]],
[[0, 0],
[0, 1]]])

diag_indices_from(arr)
Return the indices to access the main diagonal of an n-dimensional array.
See diag_indices for full details.
Parameters
arr : array, at least 2-D
See Also:
diag_indices

Notes
New in version 1.4.0.
mask_indices(n, mask_func, k=0)
Return the indices to access (n, n) arrays, given a masking function.
Assume mask_func is a function that, for a square array a of size (n, n) with a possible offset argument k,
when called as mask_func(a, k) returns a new array with zeros in certain locations (functions like triu or
tril do precisely this). Then this function returns the indices where the non-zero values would be located.
Parameters
n : int
The returned indices will be valid to access arrays of shape (n, n).
mask_func : callable
A function whose call signature is similar to that of triu, tril. That is, mask_func(x,
k) returns a boolean array, shaped like x. k is an optional argument to the function.
k : scalar
An optional argument which is passed through to mask_func. Functions like triu, tril
take a second argument that is interpreted as an offset.
Returns
indices : tuple of arrays.
The n arrays of indices corresponding to the locations where
mask_func(np.ones((n, n)), k) is True.
See Also:
triu, tril, triu_indices, tril_indices

Notes
New in version 1.4.0.

Examples
These are the indices that would allow you to access the upper triangular part of any 3x3 array:

>>> iu = np.mask_indices(3, np.triu)

For example, if a is a 3x3 array:

3.3. Indexing routines 465

NumPy Reference, Release 2.0.0.dev8464

>>> a = np.arange(9).reshape(3, 3)
>>> a
array([[0, 1, 2],
[3, 4, 5],
[6, 7, 8]])
>>> a[iu]
array([0, 1, 2, 4, 5, 8])

An offset can be passed also to the masking function. This gets us the indices starting on the first diagonal right
of the main one:

>>> iu1 = np.mask_indices(3, np.triu, 1)

with which we now extract only three elements:

>>> a[iu1]
array([1, 2, 5])

tril_indices(n, k=0)
Return the indices for the lower-triangle of an (n, n) array.
Parameters
n : int
Sets the size of the arrays for which the returned indices will be valid.
k : int, optional
Diagonal offset (see tril for details).
Returns
inds : tuple of arrays
The indices for the triangle. The returned tuple contains two arrays, each with the
indices along one dimension of the array.
See Also:

triu_indices
similar function, for upper-triangular.
mask_indices
generic function accepting an arbitrary mask function.

tril, triu

Notes
New in version 1.4.0.

Examples
Compute two different sets of indices to access 4x4 arrays, one for the lower triangular part starting at the main
diagonal, and one starting two diagonals further right:

>>> il1 = np.tril_indices(4)

>>> il2 = np.tril_indices(4, 2)

Here is how they can be used with a sample array:

466 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> a = np.arange(16).reshape(4, 4)
>>> a
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11],
[12, 13, 14, 15]])

Both for indexing:

>>> a[il1]
array([ 0, 4, 5, 8, 9, 10, 12, 13, 14, 15])

And for assigning values:

>>> a[il1] = -1
>>> a
array([[-1, 1, 2, 3],
[-1, -1, 6, 7],
[-1, -1, -1, 11],
[-1, -1, -1, -1]])

These cover almost the whole array (two diagonals right of the main one):

>>> a[il2] = -10

>>> a
array([[-10, -10, -10, 3],
[-10, -10, -10, -10],
[-10, -10, -10, -10],
[-10, -10, -10, -10]])

tril_indices_from(arr, k=0)
Return the indices for the lower-triangle of an (n, n) array.
See tril_indices for full details.
Parameters
n : int
Sets the size of the arrays for which the returned indices will be valid.
k : int, optional
Diagonal offset (see tril for details).
See Also:
tril_indices, tril

Notes
New in version 1.4.0.
triu_indices(n, k=0)
Return the indices for the upper-triangle of an (n, n) array.
Parameters
n : int
Sets the size of the arrays for which the returned indices will be valid.
k : int, optional

3.3. Indexing routines 467

NumPy Reference, Release 2.0.0.dev8464

Diagonal offset (see triu for details).

Returns
inds : tuple of arrays
The indices for the triangle. The returned tuple contains two arrays, each with the
indices along one dimension of the array.
See Also:

tril_indices
similar function, for lower-triangular.
mask_indices
generic function accepting an arbitrary mask function.

triu, tril

Notes
New in version 1.4.0.

Examples
Compute two different sets of indices to access 4x4 arrays, one for the upper triangular part starting at the main
diagonal, and one starting two diagonals further right:

>>> iu1 = np.triu_indices(4)

>>> iu2 = np.triu_indices(4, 2)

Here is how they can be used with a sample array:

>>> a = np.arange(16).reshape(4, 4)
>>> a
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11],
[12, 13, 14, 15]])

Both for indexing:

>>> a[iu1]
array([ 0, 1, 2, 3, 5, 6, 7, 10, 11, 15])

And for assigning values:

>>> a[iu1] = -1
>>> a
array([[-1, -1, -1, -1],
[ 4, -1, -1, -1],
[ 8, 9, -1, -1],
[12, 13, 14, -1]])

These cover only a small part of the whole array (two diagonals right of the main one):

>>> a[iu2] = -10

>>> a
array([[ -1, -1, -10, -10],

468 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

[ 4, -1, -1, -10],

[ 8, 9, -1, -1],
[ 12, 13, 14, -1]])

triu_indices_from(arr, k=0)
Return the indices for the upper-triangle of an (n, n) array.
See triu_indices for full details.
Parameters
n : int
Sets the size of the arrays for which the returned indices will be valid.
k : int, optional
Diagonal offset (see triu for details).
See Also:
triu_indices, triu

Notes
New in version 1.4.0.

3.3.2 Indexing-like operations

take(a, indices[, axis, out, mode]) Take elements from an array along an axis.
choose(a, choices[, out, mode]) Construct an array from an index array and a set of arrays to choose from.
compress(condition, a[, axis, out]) Return selected slices of an array along given axis.
diag(v[, k]) Extract a diagonal or construct a diagonal array.
diagonal(a[, offset, axis1, axis2]) Return specified diagonals.
select(condlist, choicelist[, default]) Return an array drawn from elements in choicelist, depending on conditions.
take(a, indices, axis=None, out=None, mode=raise)
Take elements from an array along an axis.
This function does the same thing as fancy indexing (indexing arrays using arrays); however, it can be easier
to use if you need elements along a given axis.
Parameters
a : array_like
The source array.
indices : array_like
The indices of the values to extract.
axis : int, optional
The axis over which to select values. By default, the flattened input array is used.
out : ndarray, optional
If provided, the result will be placed in this array. It should be of the appropriate shape
and dtype.
mode : {raise, wrap, clip}, optional
Specifies how out-of-bounds indices will behave.

3.3. Indexing routines 469

NumPy Reference, Release 2.0.0.dev8464

raise raise an error (default)

wrap wrap around
clip clip to the range
clip mode means that all indices that are too large are replaced by the index that ad-
dresses the last element along that axis. Note that this disables indexing with negative
numbers.
Returns
subarray : ndarray
The returned array has the same type as a.
See Also:

ndarray.take
equivalent method

Examples

>>> a = [4, 3, 5, 7, 6, 8]
>>> indices = [0, 1, 4]
>>> np.take(a, indices)
array([4, 3, 6])

In this example if a is an ndarray, fancy indexing can be used.

>>> a = np.array(a)
>>> a[indices]
array([4, 3, 6])

choose(a, choices, out=None, mode=raise)

Construct an array from an index array and a set of arrays to choose from.
First of all, if confused or uncertain, definitely look at the Examples - in its full generality, this function is less
simple than it might seem from the following code description (below ndi = numpy.lib.index_tricks):
np.choose(a,c) == np.array([c[a[I]][I] for I in ndi.ndindex(a.shape)]).
But this omits some subtleties. Here is a fully general summary:
Given an index array (a) of integers and a sequence of n arrays (choices), a and each choice array are first
broadcast, as necessary, to arrays of a common shape; calling these Ba and Bchoices[i], i = 0,...,n-1 we have that,
necessarily, Ba.shape == Bchoices[i].shape for each i. Then, a new array with shape Ba.shape is
created as follows:
if mode=raise (the default), then, first of all, each element of a (and thus Ba) must be in the range [0,
n-1]; now, suppose that i (in that range) is the value at the (j0, j1, ..., jm) position in Ba - then the value at
the same position in the new array is the value in Bchoices[i] at that same position;
if mode=wrap, values in a (and thus Ba) may be any (signed) integer; modular arithmetic is used to map
integers outside the range [0, n-1] back into that range; and then the new array is constructed as above;
if mode=clip, values in a (and thus Ba) may be any (signed) integer; negative integers are mapped to 0;
values greater than n-1 are mapped to n-1; and then the new array is constructed as above.

Parameters
a : int array

470 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

This array must contain integers in [0, n-1], where n is the number of choices, unless
mode=wrap or mode=clip, in which cases any integers are permissible.
choices : sequence of arrays
Choice arrays. a and all of the choices must be broadcastable to the same shape. If
choices is itself an array (not recommended), then its outermost dimension (i.e., the one
corresponding to choices.shape[0]) is taken as defining the sequence.
out : array, optional
If provided, the result will be inserted into this array. It should be of the appropriate
shape and dtype.
mode : {raise (default), wrap, clip}, optional
Specifies how indices outside [0, n-1] will be treated:
raise : an exception is raised
wrap : value becomes value mod n
clip : values < 0 are mapped to 0, values > n-1 are mapped to n-1
Returns
merged_array : array
The merged result.
Raises
ValueError: shape mismatch :
If a and each choice array are not all broadcastable to the same shape.

See Also:

ndarray.choose
equivalent method

Notes
To reduce the chance of misinterpretation, even though the following abuse is nominally supported, choices
should neither be, nor be thought of as, a single array, i.e., the outermost sequence-like container should be
either a list or a tuple.

Examples

>>> choices = [[0, 1, 2, 3], [10, 11, 12, 13],

... [20, 21, 22, 23], [30, 31, 32, 33]]
>>> np.choose([2, 3, 1, 0], choices
... # the first element of the result will be the first element of the
... # third (2+1) "array" in choices, namely, 20; the second element
... # will be the second element of the fourth (3+1) choice array, i.e.,
... # 31, etc.
... )
array([20, 31, 12, 3])
>>> np.choose([2, 4, 1, 0], choices, mode=clip) # 4 goes to 3 (4-1)
array([20, 31, 12, 3])
>>> # because there are 4 choice arrays
>>> np.choose([2, 4, 1, 0], choices, mode=wrap) # 4 goes to (4 mod 4)
array([20, 1, 12, 3])
>>> # i.e., 0

3.3. Indexing routines 471

NumPy Reference, Release 2.0.0.dev8464

A couple examples illustrating how choose broadcasts:

>>> a = [[1, 0, 1], [0, 1, 0], [1, 0, 1]]

>>> choices = [-10, 10]
>>> np.choose(a, choices)
array([[ 10, -10, 10],
[-10, 10, -10],
[ 10, -10, 10]])

>>> # With thanks to Anne Archibald

>>> a = np.array([0, 1]).reshape((2,1,1))
>>> c1 = np.array([1, 2, 3]).reshape((1,3,1))
>>> c2 = np.array([-1, -2, -3, -4, -5]).reshape((1,1,5))
>>> np.choose(a, (c1, c2)) # result is 2x3x5, res[0,:,:]=c1, res[1,:,:]=c2
array([[[ 1, 1, 1, 1, 1],
[ 2, 2, 2, 2, 2],
[ 3, 3, 3, 3, 3]],
[[-1, -2, -3, -4, -5],
[-1, -2, -3, -4, -5],
[-1, -2, -3, -4, -5]]])

compress(condition, a, axis=None, out=None)

Return selected slices of an array along given axis.
When working along a given axis, a slice along that axis is returned in output for each index where condition
evaluates to True. When working on a 1-D array, compress is equivalent to extract.
Parameters
condition : 1-D array of bools
Array that selects which entries to return. If len(condition) is less than the size of a
along the given axis, then output is truncated to the length of the condition array.
a : array_like
Array from which to extract a part.
axis : int, optional
Axis along which to take slices. If None (default), work on the flattened array.
out : ndarray, optional
Output array. Its type is preserved and it must be of the right shape to hold the output.
Returns
compressed_array : ndarray
A copy of a without the slices along axis for which condition is false.
See Also:
take, choose, diag, diagonal, select
ndarray.compress
Equivalent method.
numpy.doc.ufuncs
Section Output arguments

472 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.array([[1, 2], [3, 4], [5, 6]])

>>> a
array([[1, 2],
[3, 4],
[5, 6]])
>>> np.compress([0, 1], a, axis=0)
array([[3, 4]])
>>> np.compress([False, True, True], a, axis=0)
array([[3, 4],
[5, 6]])
>>> np.compress([False, True], a, axis=1)
array([[2],
[4],
[6]])

Working on the flattened array does not return slices along an axis but selects elements.

>>> np.compress([False, True], a)

array([2])

diag(v, k=0)
Extract a diagonal or construct a diagonal array.
Parameters
v : array_like
If v is a 2-D array, return a copy of its k-th diagonal. If v is a 1-D array, return a 2-D
array with v on the k-th diagonal.
k : int, optional
Diagonal in question. The default is 0. Use k>0 for diagonals above the main diagonal,
and k for diagonals below the main diagonal.
Returns
out : ndarray
The extracted diagonal or constructed diagonal array.
See Also:

diagonal
Return specified diagonals.
diagflat
Create a 2-D array with the flattened input as a diagonal.
trace
Sum along diagonals.
triu
Upper triangle of an array.
tril
Lower triange of an array.

3.3. Indexing routines 473

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.arange(9).reshape((3,3))
>>> x
array([[0, 1, 2],
[3, 4, 5],
[6, 7, 8]])

>>> np.diag(x)
array([0, 4, 8])
>>> np.diag(x, k=1)
array([1, 5])
>>> np.diag(x, k=-1)
array([3, 7])

>>> np.diag(np.diag(x))
array([[0, 0, 0],
[0, 4, 0],
[0, 0, 8]])

diagonal(a, offset=0, axis1=0, axis2=1)

Return specified diagonals.
If a is 2-D, returns the diagonal of a with the given offset, i.e., the collection of elements of the form a[i,i+offset].
If a has more than two dimensions, then the axes specified by axis1 and axis2 are used to determine the 2-D
subarray whose diagonal is returned. The shape of the resulting array can be determined by removing axis1 and
axis2 and appending an index to the right equal to the size of the resulting diagonals.
Parameters
a : array_like
Array from which the diagonals are taken.
offset : int, optional
Offset of the diagonal from the main diagonal. Can be both positive and negative.
Defaults to main diagonal (0).
axis1 : int, optional
Axis to be used as the first axis of the 2-D subarrays from which the diagonals should
be taken. Defaults to first axis (0).
axis2 : int, optional
Axis to be used as the second axis of the 2-D subarrays from which the diagonals should
be taken. Defaults to second axis (1).
Returns
array_of_diagonals : ndarray
If a is 2-D, a 1-D array containing the diagonal is returned. If a has larger dimensions,
then an array of diagonals is returned.
Raises
ValueError :
If the dimension of a is less than 2.
See Also:

474 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

diag
Matlab workalike for 1-D and 2-D arrays.
diagflat
Create diagonal arrays.
trace
Sum along diagonals.

Examples

>>> a = np.arange(4).reshape(2,2)
>>> a
array([[0, 1],
[2, 3]])
>>> a.diagonal()
array([0, 3])
>>> a.diagonal(1)
array([1])

>>> a = np.arange(8).reshape(2,2,2)
>>> a
array([[[0, 1],
[2, 3]],
[[4, 5],
[6, 7]]])
>>> a.diagonal(0,-2,-1)
array([[0, 3],
[4, 7]])

select(condlist, choicelist, default=0)

Return an array drawn from elements in choicelist, depending on conditions.
Parameters
condlist : list of bool ndarrays
The list of conditions which determine from which array in choicelist the output ele-
ments are taken. When multiple conditions are satisfied, the first one encountered in
condlist is used.
choicelist : list of ndarrays
The list of arrays from which the output elements are taken. It has to be of the same
length as condlist.
default : scalar, optional
The element inserted in output when all conditions evaluate to False.
Returns
output : ndarray
The output at position m is the m-th element of the array in choicelist where the m-th
element of the corresponding array in condlist is True.
See Also:

where
Return elements from one of two arrays depending on condition.

3.3. Indexing routines 475

NumPy Reference, Release 2.0.0.dev8464

take, choose, compress, diag, diagonal

Examples

>>> x = np.arange(10)
>>> condlist = [x<3, x>5]
>>> choicelist = [x, x**2]
>>> np.select(condlist, choicelist)
array([ 0, 1, 2, 0, 0, 0, 36, 49, 64, 81])

3.3.3 Inserting data into arrays

place(arr, mask, vals) Change elements of an array based on conditional and input values.
put(a, ind, v[, mode]) Replaces specified elements of an array with given values.
putmask(a, mask, values) Changes elements of an array based on conditional and input values.
fill_diagonal(a, val) Fill the main diagonal of the given array of any dimensionality.
place(arr, mask, vals)
Change elements of an array based on conditional and input values.
Similar to np.putmask(a, mask, vals), the difference is that place uses the first N elements of vals,
where N is the number of True values in mask, while putmask uses the elements where mask is True.
Note that extract does the exact opposite of place.
Parameters
a : array_like
Array to put data into.
mask : array_like
Boolean mask array. Must have the same size as a.
vals : 1-D sequence
Values to put into a. Only the first N elements are used, where N is the number of True
values in mask. If vals is smaller than N it will be repeated.
See Also:
putmask, put, take, extract

Examples

>>> x = np.arange(6).reshape(2, 3)
>>> np.place(x, x>2, [44, 55])
>>> x
array([[ 0, 1, 2],
[44, 55, 44]])

put(a, ind, v, mode=raise)

Replaces specified elements of an array with given values.
The indexing works on the flattened target array. put is roughly equivalent to:

a.flat[ind] = v

476 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
a : ndarray
Target array.
ind : array_like
Target indices, interpreted as integers.
v : array_like
Values to place in a at target indices. If v is shorter than ind it will be repeated as
necessary.
mode : {raise, wrap, clip}, optional
Specifies how out-of-bounds indices will behave.
raise raise an error (default)
wrap wrap around
clip clip to the range
clip mode means that all indices that are too large are replaced by the index that ad-
dresses the last element along that axis. Note that this disables indexing with negative
numbers.

See Also:
putmask, place

Examples

>>> a = np.arange(5)
>>> np.put(a, [0, 2], [-44, -55])
>>> a
array([-44, 1, -55, 3, 4])

>>> a = np.arange(5)
>>> np.put(a, 22, -5, mode=clip)
>>> a
array([ 0, 1, 2, 3, -5])

putmask(a, mask, values)

Changes elements of an array based on conditional and input values.
Sets a.flat[n] = values[n] for each n where mask.flat[n]==True.
If values is not the same size as a and mask then it will repeat. This gives behavior different from a[mask] =
values.
Parameters
a : array_like
Target array.
mask : array_like
Boolean mask array. It has to be the same shape as a.
values : array_like
Values to put into a where mask is True. If values is smaller than a it will be repeated.

3.3. Indexing routines 477

NumPy Reference, Release 2.0.0.dev8464

See Also:
place, put, take

Examples

>>> x = np.arange(6).reshape(2, 3)
>>> np.putmask(x, x>2, x**2)
>>> x
array([[ 0, 1, 2],
[ 9, 16, 25]])

If values is smaller than a it is repeated:

>>> x = np.arange(5)
>>> np.putmask(x, x>1, [-33, -44])
>>> x
array([ 0, 1, -33, -44, -33])

fill_diagonal(a, val)
Fill the main diagonal of the given array of any dimensionality.
For an array a with a.ndim > 2, the diagonal is the list of locations with indices a[i, i, ..., i] all
identical. This function modifies the input array in-place, it does not return a value.
Parameters
a : array, at least 2-D.
Array whose diagonal is to be filled, it gets modified in-place.
val : scalar
Value to be written on the diagonal, its type must be compatible with that of the array a.
See Also:
diag_indices, diag_indices_from

Notes
New in version 1.4.0. This functionality can be obtained via diag_indices, but internally this version uses a
much faster implementation that never constructs the indices and uses simple slicing.

Examples

>>> a = np.zeros((3, 3), int)

>>> np.fill_diagonal(a, 5)
>>> a
array([[5, 0, 0],
[0, 5, 0],
[0, 0, 5]])

The same function can operate on a 4-D array:

>>> a = np.zeros((3, 3, 3, 3), int)

>>> np.fill_diagonal(a, 4)

We only show a few blocks for clarity:

478 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> a[0, 0]
array([[4, 0, 0],
[0, 0, 0],
[0, 0, 0]])
>>> a[1, 1]
array([[0, 0, 0],
[0, 4, 0],
[0, 0, 0]])
>>> a[2, 2]
array([[0, 0, 0],
[0, 0, 0],
[0, 0, 4]])

3.3.4 Iterating over arrays

ndenumerate(arr) Multidimensional index iterator.

ndindex(*args) An N-dimensional iterator object to index arrays.
flatiter Flat iterator object to iterate over arrays.
class ndenumerate(arr)
Multidimensional index iterator.
Return an iterator yielding pairs of array coordinates and values.
Parameters
a : ndarray
Input array.
See Also:
ndindex, flatiter

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> for index, x in np.ndenumerate(a):
... print index, x
(0, 0) 1
(0, 1) 2
(1, 0) 3
(1, 1) 4

Methods
next() Standard iterator method, returns the index tuple and array value.
next()
Standard iterator method, returns the index tuple and array value.
Returns
coords : tuple of ints
The indices of the current iteration.
val : scalar
The array element of the current iteration.

3.3. Indexing routines 479

NumPy Reference, Release 2.0.0.dev8464

class ndindex(*args)
An N-dimensional iterator object to index arrays.
Given the shape of an array, an ndindex instance iterates over the N-dimensional index of the array. At each
iteration a tuple of indices is returned, the last dimension is iterated over first.
Parameters
*args : ints
The size of each dimension of the array.
See Also:
ndenumerate, flatiter

Examples

>>> for index in np.ndindex(3, 2, 1):

... print index
(0, 0, 0)
(0, 1, 0)
(1, 0, 0)
(1, 1, 0)
(2, 0, 0)
(2, 1, 0)

Methods
ndincr() Increment the multi-dimensional index by one.
next() Standard iterator method, updates the index and returns the index tuple.
ndincr()
Increment the multi-dimensional index by one.
ndincr takes care of the wrapping around of the axes. It is called by ndindex.next and not normally used
directly.
next()
Standard iterator method, updates the index and returns the index tuple.
Returns
val : tuple of ints
Returns a tuple containing the indices of the current iteration.
class flatiter()
Flat iterator object to iterate over arrays.
A flatiter iterator is returned by x.flat for any array x. It allows iterating over the array as if it were a 1-D
array, either in a for-loop or by calling its next method.
Iteration is done in C-contiguous style, with the last index varying the fastest. The iterator can also be indexed
using basic slicing or advanced indexing.
See Also:

ndarray.flat
Return a flat iterator over an array.
ndarray.flatten
Returns a flattened copy of an array.

480 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notes
A flatiter iterator can not be constructed directly from Python code by calling the flatiter constructor.

Examples

>>> x = np.arange(6).reshape(2, 3)
>>> fl = x.flat
>>> type(fl)
<type numpy.flatiter>
>>> for item in fl:
... print item
...
0
1
2
3
4
5

>>> fl[2:4]
array([2, 3])

Methods
copy() Get a copy of the iterator as a 1-D array.
next x.next() -> the next value, or raise StopIteration
copy()
Get a copy of the iterator as a 1-D array.

Examples

>>> x = np.arange(6).reshape(2, 3)
>>> x
array([[0, 1, 2],
[3, 4, 5]])
>>> fl = x.flat
>>> fl.copy()
array([0, 1, 2, 3, 4, 5])

next()
x.next() -> the next value, or raise StopIteration

3.4 Data type routines

can_cast(fromtype, totype) Returns True if cast between data types can occur without losing precision.
common_type(*arrays) Return a scalar type which is common to the input arrays.
obj2sctype(rep[, default])
can_cast(fromtype, totype)
Returns True if cast between data types can occur without losing precision.
Parameters
fromtype : dtype or dtype specifier

3.4. Data type routines 481

NumPy Reference, Release 2.0.0.dev8464

Data type to cast from.

totype : dtype or dtype specifier
Data type to cast to.
Returns
out : bool
True if cast can occur without losing precision.

Examples

>>> np.can_cast(np.int32, np.int64)

True
>>> np.can_cast(np.float64, np.complex)
True
>>> np.can_cast(np.complex, np.float)
False

>>> np.can_cast(i8, f8)

True
>>> np.can_cast(i8, f4)
False
>>> np.can_cast(i4, S4)
True

common_type(*arrays)
Return a scalar type which is common to the input arrays.
The return type will always be an inexact (i.e. floating point) scalar type, even if all the arrays are integer arrays.
If one of the inputs is an integer array, the minimum precision type that is returned is a 64-bit floating point
dtype.
All input arrays can be safely cast to the returned dtype without loss of information.
Parameters
array1, array2, ... : ndarrays
Input arrays.
Returns
out : data type code
Data type code.
See Also:
dtype, mintypecode

Examples

>>> np.common_type(np.arange(2, dtype=np.float32))

<type numpy.float32>
>>> np.common_type(np.arange(2, dtype=np.float32), np.arange(2))
<type numpy.float64>
>>> np.common_type(np.arange(4), np.array([45, 6.j]), np.array([45.0]))
<type numpy.complex128>

obj2sctype(rep, default=None)

482 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

3.4.1 Creating data types

dtype Create a data type object.

format_parser(formats, names, titles[, ...]) Class to convert formats, names, titles description to a dtype.
class dtype()
Create a data type object.
A numpy array is homogeneous, and contains elements described by a dtype object. A dtype object can be
constructed from different combinations of fundamental numeric types.
Parameters
obj :
Object to be converted to a data type object.
align : bool, optional
Add padding to the fields to match what a C compiler would output for a similar C-
struct. Can be True only if obj is a dictionary or a comma-separated string.
copy : bool, optional
Make a new copy of the data-type object. If False, the result may just be a reference
to a built-in data-type object.

Examples
Using array-scalar type:

>>> np.dtype(np.int16)
dtype(int16)

Record, one field name f1, containing int16:

>>> np.dtype([(f1, np.int16)])

dtype([(f1, <i2)])

Record, one field named f1, in itself containing a record with one field:

>>> np.dtype([(f1, [(f1, np.int16)])])

dtype([(f1, [(f1, <i2)])])

Record, two fields: the first field contains an unsigned int, the second an int32:

>>> np.dtype([(f1, np.uint), (f2, np.int32)])

dtype([(f1, <u4), (f2, <i4)])

Using array-protocol type strings:

>>> np.dtype([(a,f8),(b,S10)])
dtype([(a, <f8), (b, |S10)])

Using comma-separated field formats. The shape is (2,3):

>>> np.dtype("i4, (2,3)f8")

dtype([(f0, <i4), (f1, <f8, (2, 3))])

Using tuples. int is a fixed type, 3 the fields shape. void is a flexible type, here of size 10:

3.4. Data type routines 483

NumPy Reference, Release 2.0.0.dev8464

>>> np.dtype([(hello,(np.int,3)),(world,np.void,10)])
dtype([(hello, <i4, 3), (world, |V10)])

Subdivide int16 into 2 int8s, called x and y. 0 and 1 are the offsets in bytes:

>>> np.dtype((np.int16, {x:(np.int8,0), y:(np.int8,1)}))

dtype((<i2, [(x, |i1), (y, |i1)]))

Using dictionaries. Two fields named gender and age:

>>> np.dtype({names:[gender,age], formats:[S1,np.uint8]})

dtype([(gender, |S1), (age, |u1)])

Offsets in bytes, here 0 and 25:

>>> np.dtype({surname:(S25,0),age:(np.uint8,25)})
dtype([(surname, |S25), (age, |u1)])

Methods
newbyteorder([new_order]) Return a new dtype with a different byte order.
newbyteorder(new_order=S)
Return a new dtype with a different byte order.
Changes are also made in all fields and sub-arrays of the data type.
Parameters
new_order : string, optional
Byte order to force; a value from the byte order specifications below. The default value
(S) results in swapping the current byte order. new_order codes can be any of:

* S - swap dtype from current to opposite endian

* {<, L} - little endian
* {>, B} - big endian
* {=, N} - native order
* {|, I} - ignore (no change to byte order)

The code does a case-insensitive check on the first letter of new_order for these alterna-
tives. For example, any of > or B or b or brian are valid to specify big-endian.
Returns
new_dtype : dtype
New dtype object with the given change to the byte order.

Notes
Changes are also made in all fields and sub-arrays of the data type.

Examples

>>> import sys

>>> sys_is_le = sys.byteorder == little
>>> native_code = sys_is_le and < or >
>>> swapped_code = sys_is_le and > or <

484 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> native_dt = np.dtype(native_code+i2)

>>> swapped_dt = np.dtype(swapped_code+i2)
>>> native_dt.newbyteorder(S) == swapped_dt
True
>>> native_dt.newbyteorder() == swapped_dt
True
>>> native_dt == swapped_dt.newbyteorder(S)
True
>>> native_dt == swapped_dt.newbyteorder(=)
True
>>> native_dt == swapped_dt.newbyteorder(N)
True
>>> native_dt == native_dt.newbyteorder(|)
True
>>> np.dtype(<i2) == native_dt.newbyteorder(<)
True
>>> np.dtype(<i2) == native_dt.newbyteorder(L)
True
>>> np.dtype(>i2) == native_dt.newbyteorder(>)
True
>>> np.dtype(>i2) == native_dt.newbyteorder(B)
True

class format_parser(formats, names, titles, aligned=False, byteorder=None)

Class to convert formats, names, titles description to a dtype.
After constructing the format_parser object, the dtype attribute is the converted data-type: dtype =
format_parser(formats, names, titles).dtype
Parameters
formats : str or list of str
The format description, either specified as a string with comma-separated format de-
scriptions in the form f8, i4, a5, or a list of format description strings in the
form [f8, i4, a5].
names : str or list/tuple of str
The field names, either specified as a comma-separated string in the form col1,
col2, col3, or as a list or tuple of strings in the form [col1, col2,
col3]. An empty list can be used, in that case default field names (f0, f1, ...) are
used.
titles : sequence
Sequence of title strings. An empty list can be used to leave titles out.
aligned : bool, optional
If True, align the fields by padding as the C-compiler would. Default is False.
byteorder : str, optional
If specified, all the fields will be changed to the provided byte-order. Otherwise, the
default byte-order is used. For all available string specifiers, see dtype.newbyteorder.
See Also:
dtype, typename, sctype2char

3.4. Data type routines 485

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.format_parser([f8, i4, a5], [col1, col2, col3],

... [T1, T2, T3]).dtype
dtype([((T1, col1), <f8), ((T2, col2), <i4),
((T3, col3), |S5)])

names and/or titles can be empty lists. If titles is an empty list, titles will simply not appear. If names is empty,
default field names will be used.

>>> np.format_parser([f8, i4, a5], [col1, col2, col3],

... []).dtype
dtype([(col1, <f8), (col2, <i4), (col3, |S5)])
>>> np.format_parser([f8, i4, a5], [], []).dtype
dtype([(f0, <f8), (f1, <i4), (f2, |S5)])

Attributes
dtype dtype The converted data-type.

3.4.2 Data type information

finfo Machine limits for floating point types.

iinfo(type) Machine limits for integer types.
MachAr([float_conv, int_conv, ...]) Diagnosing machine parameters.
class finfo()
Machine limits for floating point types.
Parameters
dtype : floating point type, dtype, or instance
The kind of floating point data type to get information about.
See Also:

MachAr
The implementation of the tests that produce this information.
iinfo
The equivalent for integer data types.

Notes
For developers of NumPy: do not instantiate this at the module level. The initial calculation of these parameters
is expensive and negatively impacts import times. These objects are cached, so calling finfo() repeatedly
inside your functions is not a problem.

486 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Attributes
eps floating point number of the The smallest representable number such that 1.0 + eps !=
appropriate type 1.0.
epsneg floating point number of the The smallest representable number such that 1.0 - epsneg !=
appropriate type 1.0.
iexp int The number of bits in the exponent portion of the floating point
representation.
machar MachAr The object which calculated these parameters and holds more
detailed information.
machep int The exponent that yields eps.
max floating point number of the The largest representable number.
appropriate type
max- int The smallest positive power of the base (2) that causes overflow.
exp
min floating point number of the The smallest representable number, typically -max.
appropriate type
min- int The most negative power of the base (2) consistent with there being
exp no leading 0s in the mantissa.
negep int The exponent that yields epsneg.
nexp int The number of bits in the exponent including its sign and bias.
nmant int The number of bits in the mantissa.
preci- int The approximate number of decimal digits to which this kind of float
sion is precise.
resolu- floating point number of the The approximate decimal resolution of this type, i.e.
tion appropriate type 10**-precision.
tiny floating point number of the The smallest-magnitude usable number.
appropriate type
class iinfo(type)
Machine limits for integer types.
Parameters
type : integer type, dtype, or instance
The kind of integer data type to get information about.
See Also:

finfo
The equivalent for floating point data types.

Examples
With types:

>>> ii16 = np.iinfo(np.int16)

>>> ii16.min
-32768
>>> ii16.max
32767
>>> ii32 = np.iinfo(np.int32)
>>> ii32.min
-2147483648
>>> ii32.max
2147483647

With instances:

3.4. Data type routines 487

NumPy Reference, Release 2.0.0.dev8464

>>> ii32 = np.iinfo(np.int32(10))

>>> ii32.min
-2147483648
>>> ii32.max
2147483647

Attributes
min
max
min

max

class MachAr(float_conv=<type float>, int_conv=<type int>, float_to_float=<type float>,

float_to_str=<function <lambda> at 0x11840398>, title=Python floating point number)
Diagnosing machine parameters.
Parameters
float_conv : function, optional
Function that converts an integer or integer array to a float or float array. Default is float.
int_conv : function, optional
Function that converts a float or float array to an integer or integer array. Default is int.
float_to_float : function, optional
Function that converts a float array to float. Default is float. Note that this does not seem
to do anything useful in the current implementation.
float_to_str : function, optional
Function that converts a single float to a string. Default is lambda v:%24.16e
%v.
title : str, optional
Title that is printed in the string representation of MachAr.
See Also:

finfo
Machine limits for floating point types.
iinfo
Machine limits for integer types.

References
[R1]

488 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Attributes
ibeta int Radix in which numbers are represented.
it int Number of base-ibeta digits in the floating point mantissa M.
machep int Exponent of the smallest (most negative) power of ibeta that, added to 1.0, gives something
different from 1.0
eps float Floating-point number beta**machep (floating point precision)
negep int Exponent of the smallest power of ibeta that, substracted from 1.0, gives something different
from 1.0.
epsneg float Floating-point number beta**negep.
iexp int Number of bits in the exponent (including its sign and bias).
minexp int Smallest (most negative) power of ibeta consistent with there being no leading zeros in the
mantissa.
xmin float Floating point number beta**minexp (the smallest [in magnitude] usable floating value).
maxexp int Smallest (positive) power of ibeta that causes overflow.
xmax float (1-epsneg) * beta**maxexp (the largest [in magnitude] usable floating value).
irnd int In range(6), information on what kind of rounding is done in addition, and on how
underflow is handled.
ngrd int Number of guard digits used when truncating the product of two mantissas to fit the
representation.
epsilon float Same as eps.
tiny float Same as xmin.
huge float Same as xmax.
preci- float - int(-log10(eps))
sion
resolu- float - 10**(-precision)
tion

3.4.3 Data type testing

issctype(rep) Determines whether the given object represents a scalar data-type.

issubdtype(arg1, arg2) Returns True if first argument is a typecode lower/equal in type
hierarchy.
issubsctype(arg1, arg2) Determine if the first argument is a subclass of the second
argument.
issubclass_(arg1, arg2)
find_common_type(array_types, Determine common type following standard coercion rules.
scalar_types)
issctype(rep)
Determines whether the given object represents a scalar data-type.
Parameters
rep : any
If rep is an instance of a scalar dtype, True is returned. If not, False is returned.
Returns
out : bool
Boolean result of check whether rep is a scalar dtype.
See Also:
issubsctype, issubdtype, obj2sctype, sctype2char

3.4. Data type routines 489

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.issctype(np.int32)
True
>>> np.issctype(list)
False
>>> np.issctype(1.1)
False

issubdtype(arg1, arg2)
Returns True if first argument is a typecode lower/equal in type hierarchy.
Parameters
arg1, arg2 : dtype_like
dtype or string representing a typecode.
Returns
out : bool
See Also:
issubsctype, issubclass_
numpy.core.numerictypes
Overview of numpy type hierarchy.

Examples

>>> np.issubdtype(S1, str)

True
>>> np.issubdtype(np.float64, np.float32)
False

issubsctype(arg1, arg2)
Determine if the first argument is a subclass of the second argument.
Parameters
arg1, arg2 : dtype or dtype specifier
Data-types.
Returns
out : bool
The result.
See Also:
issctype, issubdtype, obj2sctype

Examples

>>> np.issubsctype(S8, str)

True
>>> np.issubsctype(np.array([1]), np.int)
True
>>> np.issubsctype(np.array([1]), np.float)
False

issubclass_(arg1, arg2)

490 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

find_common_type(array_types, scalar_types)
Determine common type following standard coercion rules.
Parameters
array_types : sequence
A list of dtypes or dtype convertible objects representing arrays.
scalar_types : sequence
A list of dtypes or dtype convertible objects representing scalars.
Returns
datatype : dtype
The common data type, which is the maximum of array_types ignoring scalar_types,
unless the maximum of scalar_types is of a different kind (dtype.kind). If the kind is
not understood, then None is returned.
See Also:
dtype, common_type, can_cast, mintypecode

Examples

>>> np.find_common_type([], [np.int64, np.float32, np.complex])

dtype(complex128)
>>> np.find_common_type([np.int64, np.float32], [])
dtype(float64)

The standard casting rules ensure that a scalar cannot up-cast an array unless the scalar is of a fundamentally
different kind of data (i.e. under a different hierarchy in the data type hierarchy) then the array:

>>> np.find_common_type([np.float32], [np.int64, np.float64])

dtype(float32)

Complex is of a different type, so it up-casts the float in the array_types argument:

>>> np.find_common_type([np.float32], [np.complex])

dtype(complex128)

Type specifier strings are convertible to dtypes and can therefore be used instead of dtypes:

>>> np.find_common_type([f4, f4, i4], [c8])

dtype(complex128)

3.4.4 Miscellaneous

typename(char) Return a description for the given data type code.

sctype2char(sctype) Return the string representation of a scalar dtype.
mintypecode(typechars[, typeset, Return the character for the minimum-size type to which given types can
default]) be safely cast.
typename(char)
Return a description for the given data type code.
Parameters
char : str

3.4. Data type routines 491

NumPy Reference, Release 2.0.0.dev8464

Data type code.

Returns
out : str
Description of the input data type code.
See Also:
dtype, typecodes

Examples

>>> typechars = [S1, ?, B, D, G, F, I, H, L, O, Q,

... S, U, V, b, d, g, f, i, h, l, q]
>>> for typechar in typechars:
... print typechar, : , np.typename(typechar)
...
S1 : character
? : bool
B : unsigned char
D : complex double precision
G : complex long double precision
F : complex single precision
I : unsigned integer
H : unsigned short
L : unsigned long integer
O : object
Q : unsigned long long integer
S : string
U : unicode
V : void
b : signed char
d : double precision
g : long precision
f : single precision
i : integer
h : short
l : long integer
q : long long integer

sctype2char(sctype)
Return the string representation of a scalar dtype.
Parameters
sctype : scalar dtype or object
If a scalar dtype, the corresponding string character is returned. If an object, sctype2char
tries to infer its scalar type and then return the corresponding string character.
Returns
typechar : str
The string character corresponding to the scalar type.
Raises
ValueError :
If sctype is an object for which the type can not be inferred.
See Also:

492 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

obj2sctype, issctype, issubsctype, mintypecode

Examples

>>> for sctype in [np.int32, np.float, np.complex, np.string_, np.ndarray]:

... print np.sctype2char(sctype)
l
d
D
S
O

>>> x = np.array([1., 2-1.j])

>>> np.sctype2char(x)
D
>>> np.sctype2char(list)
O

mintypecode(typechars, typeset=GDFgdf, default=d)

Return the character for the minimum-size type to which given types can be safely cast.
The returned type character must represent the smallest size dtype such that an array of the returned type can
handle the data from an array of all types in typechars (or if typechars is an array, then its dtype.char).
Parameters
typechars : list of str or array_like
If a list of strings, each string should represent a dtype. If array_like, the character
representation of the array dtype is used.
typeset : str or list of str, optional
The set of characters that the returned character is chosen from. The default set is
GDFgdf.
default : str, optional
The default character, this is returned if none of the characters in typechars matches a
character in typeset.
Returns
typechar : str
The character representing the minimum-size type that was found.
See Also:
dtype, sctype2char, maximum_sctype

Examples

>>> np.mintypecode([d, f, S])

d
>>> x = np.array([1.1, 2-3.j])
>>> np.mintypecode(x)
D

>>> np.mintypecode(abceh, default=G)

G

3.4. Data type routines 493

NumPy Reference, Release 2.0.0.dev8464

3.5 Input and output

3.5.1 NPZ files

load(file[, mmap_mode]) Load a pickled, .npy, or .npz binary file.

save(file, arr) Save an array to a binary file in NumPy .npy format.
savez(file, *args, **kwds) Save several arrays into a single, compressed file in .npz format.
load(file, mmap_mode=None)
Load a pickled, .npy, or .npz binary file.
Parameters
file : file-like object or string
The file to read. It must support seek() and read() methods. If the filename exten-
sion is .gz, the file is first decompressed.
mmap_mode: {None, r+, r, w+, c}, optional :
If not None, then memory-map the file, using the given mode (see numpy.memmap).
The mode has no effect for pickled or zipped files. A memory-mapped array is stored
on disk, and not directly loaded into memory. However, it can be accessed and sliced
like any ndarray. Memory mapping is especially useful for accessing small fragments
of large files without reading the entire file into memory.
Returns
result : array, tuple, dict, etc.
Data stored in the file.
Raises
IOError :
If the input file does not exist or cannot be read.
See Also:
save, savez, loadtxt
memmap
Create a memory-map to an array stored in a file on disk.

Notes
If the file contains pickle data, then whatever is stored in the pickle is returned.
If the file is a .npy file, then an array is returned.
If the file is a .npz file, then a dictionary-like object is returned, containing {filename: array}
key-value pairs, one for each file in the archive.

Examples
Store data to disk, and load it again:

>>> np.save(/tmp/123, np.array([[1, 2, 3], [4, 5, 6]]))

>>> np.load(/tmp/123.npy)
array([[1, 2, 3],
[4, 5, 6]])

Mem-map the stored array, and then access the second row directly from disk:

494 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> X = np.load(/tmp/123.npy, mmap_mode=r)

>>> X[1, :]
memmap([4, 5, 6])

save(file, arr)
Save an array to a binary file in NumPy .npy format.
Parameters
file : file or str
File or filename to which the data is saved. If file is a file-object, then the filename is
unchanged. If file is a string, a .npy extension will be appended to the file name if it
does not already have one.
arr : array_like
Array data to be saved.
See Also:

savez
Save several arrays into a .npz compressed archive

savetxt, load

Notes
For a description of the .npy format, see format.

Examples

>>> from tempfile import TemporaryFile

>>> outfile = TemporaryFile()

>>> x = np.arange(10)
>>> np.save(outfile, x)

>>> outfile.seek(0) # Only needed here to simulate closing & reopening file
>>> np.load(outfile)
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

savez(file, *args, **kwds)

Save several arrays into a single, compressed file in .npz format.
If arguments are passed in with no keywords, the corresponding variable names, in the .npz file, are arr_0,
arr_1, etc. If keyword arguments are given, the corresponding variable names, in the .npz file will match the
keyword names.
Parameters
file : str or file
Either the file name (string) or an open file (file-like object) where the data will be saved.
If file is a string, the .npz extension will be appended to the file name if it is not already
there.
*args : Arguments, optional

3.5. Input and output 495

NumPy Reference, Release 2.0.0.dev8464

Arrays to save to the file. Since it is not possible for Python to know the names of the
arrays outside savez, the arrays will be saved with names arr_0, arr_1, and so on.
These arguments can be any expression.
**kwds : Keyword arguments, optional
Arrays to save to the file. Arrays will be saved in the file with the keyword names.
Returns
None :
See Also:

save
Save a single array to a binary file in NumPy format.
savetxt
Save an array to a file as plain text.

Notes
The .npz file format is a zipped archive of files named after the variables they contain. Each file contains one
variable in .npy format. For a description of the .npy format, see format.
When opening the saved .npz file with load a NpzFile object is returned. This is a dictionary-like object which
can be queried for its list of arrays (with the .files attribute), and for the arrays themselves.

Examples

>>> from tempfile import TemporaryFile

>>> outfile = TemporaryFile()
>>> x = np.arange(10)
>>> y = np.sin(x)

Using savez with *args, the arrays are saved with default names.

>>> np.savez(outfile, x, y)
>>> outfile.seek(0) # Only needed here to simulate closing & reopening file
>>> npzfile = np.load(outfile)
>>> npzfile.files
[arr_1, arr_0]
>>> npzfile[arr_0]
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

Using savez with **kwds, the arrays are saved with the keyword names.

>>> outfile = TemporaryFile()

>>> np.savez(outfile, x=x, y=y)
>>> outfile.seek(0)
>>> npzfile = np.load(outfile)
>>> npzfile.files
[y, x]
>>> npzfile[x]
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

496 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

3.5.2 Text files

loadtxt(fname[, dtype, comments, Load data from a text file.

delimiter, ...])
savetxt(fname, X[, fmt, delimiter, Save an array to a text file.
newline])
genfromtxt(fname[, dtype, comments, ...]) Load data from a text file, with missing values handled as
specified.
fromregex(file, regexp, dtype) Construct an array from a text file, using regular expression
parsing.
fromstring(string[, dtype, count, sep]) Return a new 1-D array initialized from raw binary or text data
in string.
ndarray.tofile(fid[, sep, format]) Write array to a file as text or binary (default).
ndarray.tolist() Return the array as a (possibly nested) list.
loadtxt(fname, dtype=<type float>, comments=#, delimiter=None, converters=None, skiprows=0, usec-
ols=None, unpack=False)
Load data from a text file.
Each row in the text file must have the same number of values.
Parameters
fname : file or str
File or filename to read. If the filename extension is .gz or .bz2, the file is first
decompressed.
dtype : dtype, optional
Data type of the resulting array. If this is a record data-type, the resulting array will be
1-dimensional, and each row will be interpreted as an element of the array. In this case,
the number of columns used must match the number of fields in the data-type.
comments : str, optional
The character used to indicate the start of a comment.
delimiter : str, optional
The string used to separate values. By default, this is any whitespace.
converters : dict, optional
A dictionary mapping column number to a function that will convert that column to a
float. E.g., if column 0 is a date string: converters = {0: datestr2num}.
Converters can also be used to provide a default value for missing data: converters
= {3: lambda s: float(s or 0)}.
skiprows : int, optional
Skip the first skiprows lines.
usecols : sequence, optional
Which columns to read, with 0 being the first. For example, usecols = (1,4,5)
will extract the 2nd, 5th and 6th columns.
unpack : bool, optional
If True, the returned array is transposed, so that arguments may be unpacked using x,
y, z = loadtxt(...). Default is False.

3.5. Input and output 497

NumPy Reference, Release 2.0.0.dev8464

Returns
out : ndarray
Data read from the text file.
See Also:
load, fromstring, fromregex
genfromtxt
Load data with missing values handled as specified.
scipy.io.loadmat
reads Matlab(R) data files

Notes
This function aims to be a fast reader for simply formatted files. The genfromtxt function provides more sophis-
ticated handling of, e.g., lines with missing values.

Examples

>>> from StringIO import StringIO # StringIO behaves like a file object
>>> c = StringIO("0 1\n2 3")
>>> np.loadtxt(c)
array([[ 0., 1.],
[ 2., 3.]])

>>> d = StringIO("M 21 72\nF 35 58")

>>> np.loadtxt(d, dtype={names: (gender, age, weight),
... formats: (S1, i4, f4)})
array([(M, 21, 72.0), (F, 35, 58.0)],
dtype=[(gender, |S1), (age, <i4), (weight, <f4)])

>>> c = StringIO("1,0,2\n3,0,4")
>>> x, y = np.loadtxt(c, delimiter=,, usecols=(0, 2), unpack=True)
>>> x
array([ 1., 3.])
>>> y
array([ 2., 4.])

savetxt(fname, X, fmt=%.18e, delimiter= , newline=n)

Save an array to a text file.
Parameters
fname : filename or file handle
If the filename ends in .gz, the file is automatically saved in compressed gzip format.
loadtxt understands gzipped files transparently.
X : array_like
Data to be saved to a text file.
fmt : str or sequence of strs
A single format (%10.5f), a sequence of formats, or a multi-format string, e.g. Iteration
%d %10.5f, in which case delimiter is ignored.
delimiter : str

498 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Character separating columns.

newline : str
New in version 2.0. Character separating lines.
See Also:

save
Save an array to a binary file in NumPy .npy format
savez
Save several arrays into a .npz compressed archive

Notes
Further explanation of the fmt parameter (%[flag]width[.precision]specifier):
flags:
- : left justify
+ : Forces to preceed result with + or -.
0 : Left pad the number with zeros instead of space (see width).
width:
Minimum number of characters to be printed. The value is not truncated if it has more characters.
precision:

For integer specifiers (eg. d,i,o,x), the minimum number of digits.

For e, E and f specifiers, the number of digits to print after the decimal point.
For g and G, the maximum number of significant digits.
For s, the maximum number of characters.
specifiers:
c : character
d or i : signed decimal integer
e or E : scientific notation with e or E.
f : decimal floating point
g,G : use the shorter of e,E or f
o : signed octal
s : string of characters
u : unsigned decimal integer
x,X : unsigned hexadecimal integer
This explanation of fmt is not complete, for an exhaustive specification see [R268].

References
[R268]

3.5. Input and output 499

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = y = z = np.arange(0.0,5.0,1.0)
>>> np.savetxt(test.out, x, delimiter=,) # X is an array
>>> np.savetxt(test.out, (x,y,z)) # x,y,z equal sized 1D arrays
>>> np.savetxt(test.out, x, fmt=%1.4e) # use exponential notation

genfromtxt(fname, dtype=<type float>, comments=#, delimiter=None, skiprows=0, skip_header=0,

skip_footer=0, converters=None, missing=, missing_values=None, filling_values=None,
usecols=None, names=None, excludelist=None, deletechars=None, replace_space=_, au-
tostrip=False, case_sensitive=True, defaultfmt=f%i, unpack=None, usemask=False, loose=True,
invalid_raise=True)
Load data from a text file, with missing values handled as specified.
Each line past the first skiprows lines is split at the delimiter character, and characters following the comments
character are discarded.
Parameters
fname : file or str
File or filename to read. If the filename extension is gz or bz2, the file is first decom-
pressed.
dtype : dtype, optional
Data type of the resulting array. If None, the dtypes will be determined by the contents
of each column, individually.
comments : str, optional
The character used to indicate the start of a comment. All the characters occurring on a
line after a comment are discarded
delimiter : str, int, or sequence, optional
The string used to separate values. By default, any consecutive whitespaces act as
delimiter. An integer or sequence of integers can also be provided as width(s) of each
field.
skip_header : int, optional
The numbers of lines to skip at the beginning of the file.
skip_footer : int, optional
The numbers of lines to skip at the end of the file
converters : variable or None, optional
The set of functions that convert the data of a column to a value. The converters can also
be used to provide a default value for missing data: converters = {3: lambda
s: float(s or 0)}.
missing_values : variable or None, optional
The set of strings corresponding to missing data.
filling_values : variable or None, optional
The set of values to be used as default when the data are missing.
usecols : sequence or None, optional
Which columns to read, with 0 being the first. For example, usecols = (1, 4,
5) will extract the 2nd, 5th and 6th columns.

500 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

names : {None, True, str, sequence}, optional

If names is True, the field names are read from the first valid line after the first skiprows
lines. If names is a sequence or a single-string of comma-separated names, the names
will be used to define the field names in a structured dtype. If names is None, the names
of the dtype fields will be used, if any.
excludelist : sequence, optional
A list of names to exclude. This list is appended to the default list [return,file,print].
Excluded names are appended an underscore: for example, file would become file_.
deletechars : str, optional
A string combining invalid characters that must be deleted from the names.
defaultfmt : str, optional
A format used to define default field names, such as f%i or f_%02i.
autostrip : bool, optional
Whether to automatically strip white spaces from the variables.
replace_space : char, optional
Character(s) used in replacement of white spaces in the variables names. By default,
use a _.
case_sensitive : {True, False, upper, lower}, optional
If True, field names are case sensitive. If False or upper, field names are converted to
upper case. If lower, field names are converted to lower case.
unpack : bool, optional
If True, the returned array is transposed, so that arguments may be unpacked using x,
y, z = loadtxt(...)
usemask : bool, optional
If True, return a masked array. If False, return a regular array.
invalid_raise : bool, optional
If True, an exception is raised if an inconsistency is detected in the number of columns.
If False, a warning is emitted and the offending lines are skipped.
Returns
out : ndarray
Data read from the text file. If usemask is True, this is a masked array.
See Also:

numpy.loadtxt
equivalent function when no data is missing.

Notes
When spaces are used as delimiters, or when no delimiter has been given as input, there should not be any
missing data between two fields.
When the variables are named (either by a flexible dtype or with names, there must not be any header in
the file (else a ValueError exception is raised).

3.5. Input and output 501

NumPy Reference, Release 2.0.0.dev8464

Individual values are not stripped of spaces by default. When using a custom converter, make sure the
function does remove spaces.

Examples

>>> from StringIO import StringIO

>>> import numpy as np

Comma delimited file with mixed dtype

>>> s = StringIO("1,1.3,abcde")
>>> data = np.genfromtxt(s, dtype=[(myint,i8),(myfloat,f8),
... (mystring,S5)], delimiter=",")
>>> data
array((1, 1.3, abcde),
dtype=[(myint, <i8), (myfloat, <f8), (mystring, |S5)])

Using dtype = None

>>> s.seek(0) # needed for StringIO example only

>>> data = np.genfromtxt(s, dtype=None,
... names = [myint,myfloat,mystring], delimiter=",")
>>> data
array((1, 1.3, abcde),
dtype=[(myint, <i8), (myfloat, <f8), (mystring, |S5)])

Specifying dtype and names

>>> s.seek(0)
>>> data = np.genfromtxt(s, dtype="i8,f8,S5",
... names=[myint,myfloat,mystring], delimiter=",")
>>> data
array((1, 1.3, abcde),
dtype=[(myint, <i8), (myfloat, <f8), (mystring, |S5)])

An example with fixed-width columns

>>> s = StringIO("11.3abcde")
>>> data = np.genfromtxt(s, dtype=None, names=[intvar,fltvar,strvar],
... delimiter=[1,3,5])
>>> data
array((1, 1.3, abcde),
dtype=[(intvar, <i8), (fltvar, <f8), (strvar, |S5)])

fromregex(file, regexp, dtype)

Construct an array from a text file, using regular expression parsing.
The returned array is always a structured array, and is constructed from all matches of the regular expression in
the file. Groups in the regular expression are converted to fields of the structured array.
Parameters
file : str or file
File name or file object to read.
regexp : str or regexp

502 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Regular expression used to parse the file. Groups in the regular expression correspond
to fields in the dtype.
dtype : dtype or list of dtypes
Dtype for the structured array.
Returns
output : ndarray
The output array, containing the part of the content of file that was matched by regexp.
output is always a structured array.
Raises
TypeError :
When dtype is not a valid dtype for a structured array.
See Also:
fromstring, loadtxt

Notes
Dtypes for structured arrays can be specified in several forms, but all forms specify at least the data type and
field name. For details see doc.structured_arrays.

Examples

>>> f = open(test.dat, w)
>>> f.write("1312 foo\n1534 bar\n444 qux")
>>> f.close()

>>> regexp = r"(\d+)\s+(...)" # match [digits, whitespace, anything]

>>> output = np.fromregex(test.dat, regexp,
... [(num, np.int64), (key, S3)])
>>> output
array([(1312L, foo), (1534L, bar), (444L, qux)],
dtype=[(num, <i8), (key, |S3)])
>>> output[num]
array([1312, 1534, 444], dtype=int64)

fromstring(string, dtype=float, count=-1, sep=)

Return a new 1-D array initialized from raw binary or text data in string.
Parameters
string : str
A string containing the data.
dtype : dtype, optional
The data type of the array. For binary input data, the data must be in exactly this format.
count : int, optional
Read this number of dtype elements from the data. If this is negative, then the size will
be determined from the length of the data.
sep : str, optional

3.5. Input and output 503

NumPy Reference, Release 2.0.0.dev8464

If provided and not empty, then the data will be interpreted as ASCII text with decimal
numbers. This argument is interpreted as the string separating numbers in the data.
Extra whitespace between elements is also ignored.
Returns
arr : array
The constructed array.
Raises
ValueError :
If the string is not the correct size to satisfy the requested dtype and count.

Examples

>>> np.fromstring(\x01\x02, dtype=np.uint8)

array([1, 2], dtype=uint8)
>>> np.fromstring(1 2, dtype=int, sep= )
array([1, 2])
>>> np.fromstring(1, 2, dtype=int, sep=,)
array([1, 2])
>>> np.fromstring(\x01\x02\x03\x04\x05, dtype=np.uint8, count=3)
array([1, 2, 3], dtype=uint8)

Invalid inputs:

>>> np.fromstring(\x01\x02\x03\x04\x05, dtype=np.int32)

Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: string size must be a multiple of element size
>>> np.fromstring(\x01\x02, dtype=np.uint8, count=3)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: string is smaller than requested size

tofile(fid, sep="", format="%s")

Write array to a file as text or binary (default).
Data is always written in C order, independent of the order of a. The data produced by this method can be
recovered using the function fromfile().
Parameters
fid : file or str
An open file object, or a string containing a filename.
sep : str
Separator between array items for text output. If (empty), a binary file is written,
equivalent to file.write(a.tostring()).
format : str
Format string for text file output. Each entry in the array is formatted to text by first
converting it to the closest Python type, and then using format % item.

Notes
This is a convenience function for quick storage of array data. Information on endianness and precision is lost,
so this method is not a good choice for files intended to archive data or transport data between machines with

504 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

different endianness. Some of these problems can be overcome by outputting the data as text files, at the expense
of speed and file size.
tolist()
Return the array as a (possibly nested) list.
Return a copy of the array data as a (nested) Python list. Data items are converted to the nearest compatible
Python type.
Parameters
none :
Returns
y : list
The possibly nested list of array elements.

Notes
The array may be recreated, a = np.array(a.tolist()).

Examples

>>> a = np.array([1, 2])

>>> a.tolist()
[1, 2]
>>> a = np.array([[1, 2], [3, 4]])
>>> list(a)
[array([1, 2]), array([3, 4])]
>>> a.tolist()
[[1, 2], [3, 4]]

3.5.3 String formatting

array_repr(arr[, max_line_width, precision, ...]) Return the string representation of an array.

array_str(a[, max_line_width, precision, ...]) Return a string representation of the data in an array.
array_repr(arr, max_line_width=None, precision=None, suppress_small=None)
Return the string representation of an array.
Parameters
arr : ndarray
Input array.
max_line_width : int, optional
The maximum number of columns the string should span. Newline characters split the
string appropriately after array elements.
precision : int, optional
Floating point precision. Default is the current printing precision (usually 8), which can
be altered using set_printoptions.
suppress_small : bool, optional
Represent very small numbers as zero, default is False. Very small is defined by preci-
sion, if the precision is 8 then numbers smaller than 5e-9 are represented as zero.

3.5. Input and output 505

NumPy Reference, Release 2.0.0.dev8464

Returns
string : str
The string representation of an array.
See Also:
array_str, array2string, set_printoptions

Examples

>>> np.array_repr(np.array([1,2]))
array([1, 2])
>>> np.array_repr(np.ma.array([0.]))
MaskedArray([ 0.])
>>> np.array_repr(np.array([], np.int32))
array([], dtype=int32)

>>> x = np.array([1e-6, 4e-7, 2, 3])

>>> np.array_repr(x, precision=6, suppress_small=True)
array([ 0.000001, 0. , 2. , 3. ])

array_str(a, max_line_width=None, precision=None, suppress_small=None)

Return a string representation of the data in an array.
The data in the array is returned as a single string. This function is similar to array_repr, the difference is that
array_repr also returns information on the type of array and data type.
Parameters
a : ndarray
Input array.
max_line_width : int, optional
Inserts newlines if text is longer than max_line_width.
precision : int, optional
Floating point precision. Default is the current printing precision (usually 8), which can
be altered using set_printoptions.
suppress_small : bool, optional
Represent very small numbers as zero, default is False. Very small is defined by preci-
sion, if the precision is 8 then numbers smaller than 5e-9 are represented as zero.
See Also:
array2string, array_repr, set_printoptions

Examples

>>> np.array_str(np.arange(3))
[0 1 2]

3.5.4 Memory mapping files

memmap Create a memory-map to an array stored in a binary file on disk.

506 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

class memmap()
Create a memory-map to an array stored in a binary file on disk.
Memory-mapped files are used for accessing small segments of large files on disk, without reading the entire
file into memory. Numpys memmaps are array-like objects. This differs from Pythons mmap module, which
uses file-like objects.
Parameters
filename : str or file-like object
The file name or file object to be used as the array data buffer.
dtype : data-type, optional
The data-type used to interpret the file contents. Default is uint8.
mode : {r+, r, w+, c}, optional
The file is opened in this mode:
r Open existing file for reading only.
r+ Open existing file for reading and writing.
w+ Create or overwrite existing file for reading and writing.
c Copy-on-write: assignments affect data in memory, but changes are not saved to disk. The file on disk is
read-only.
Default is r+.
offset : int, optional
In the file, array data starts at this offset. Since offset is measured in bytes, it should be
a multiple of the byte-size of dtype. Requires shape=None. The default is 0.
shape : tuple, optional
The desired shape of the array. By default, the returned array will be 1-D with the
number of elements determined by file size and data-type.
order : {C, F}, optional
Specify the order of the ndarray memory layout: C (row-major) or Fortran (column-
major). This only has an effect if the shape is greater than 1-D. The default order is
C.

Notes
The memmap object can be used anywhere an ndarray is accepted. Given a memmap fp, isinstance(fp,
numpy.ndarray) returns True.
Memory-mapped arrays use the Python memory-map object which (prior to Python 2.5) does not allow files to
be larger than a certain size depending on the platform. This size is always < 2GB even on 64-bit systems.

Examples

>>> data = np.arange(12, dtype=float32)

>>> data.resize((3,4))

This example uses a temporary file so that doctest doesnt write files to your directory. You would use a normal
filename.

>>> from tempfile import mkdtemp

>>> import os.path as path
>>> filename = path.join(mkdtemp(), newfile.dat)

3.5. Input and output 507

NumPy Reference, Release 2.0.0.dev8464

Create a memmap with dtype and shape that matches our data:

>>> fp = np.memmap(filename, dtype=float32, mode=w+, shape=(3,4))

>>> fp
memmap([[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.]], dtype=float32)

Write data to memmap array:

>>> fp[:] = data[:]

>>> fp
memmap([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]], dtype=float32)

>>> fp.filename == path.abspath(filename)

True

Deletion flushes memory changes to disk before removing the object:

>>> del fp

Load the memmap and verify data was stored:

>>> newfp = np.memmap(filename, dtype=float32, mode=r, shape=(3,4))

>>> newfp
memmap([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]], dtype=float32)

Read-only memmap:

>>> fpr = np.memmap(filename, dtype=float32, mode=r, shape=(3,4))

>>> fpr.flags.writeable
False

Copy-on-write memmap:

>>> fpc = np.memmap(filename, dtype=float32, mode=c, shape=(3,4))

>>> fpc.flags.writeable
True

Its possible to assign to copy-on-write array, but values are only written into the memory copy of the array, and
not written to disk:

>>> fpc
memmap([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]], dtype=float32)
>>> fpc[0,:] = 0
>>> fpc
memmap([[ 0., 0., 0., 0.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]], dtype=float32)

508 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

File on disk is unchanged:

>>> fpr
memmap([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]], dtype=float32)

Offset into a memmap:

>>> fpo = np.memmap(filename, dtype=float32, mode=r, offset=16)

>>> fpo
memmap([ 4., 5., 6., 7., 8., 9., 10., 11.], dtype=float32)

Attributes
filename str Path to the mapped file.
offset int Offset position in the file.
mode str File mode.

Methods
close() Close the memmap file.
flush() Write any changes in the array to the file on disk.
close()
Close the memmap file. Does nothing.
flush()
Write any changes in the array to the file on disk.
For further information, see memmap.
Parameters
None :
See Also:
memmap

3.5.5 Text formatting options

set_printoptions([precision, threshold, ...]) Set printing options.

get_printoptions() Return the current print options.
set_string_function(f[, repr]) Set a Python function to be used when pretty printing arrays.
set_printoptions(precision=None, threshold=None, edgeitems=None, linewidth=None, suppress=None,
nanstr=None, infstr=None)
Set printing options.
These options determine the way floating point numbers, arrays and other NumPy objects are displayed.
Parameters
precision : int, optional
Number of digits of precision for floating point output (default 8).
threshold : int, optional
Total number of array elements which trigger summarization rather than full repr (de-
fault 1000).

3.5. Input and output 509

NumPy Reference, Release 2.0.0.dev8464

edgeitems : int, optional

Number of array items in summary at beginning and end of each dimension (default 3).
linewidth : int, optional
The number of characters per line for the purpose of inserting line breaks (default 75).
suppress : bool, optional
Whether or not suppress printing of small floating point values using scientific notation
(default False).
nanstr : str, optional
String representation of floating point not-a-number (default NaN).
infstr : str, optional
String representation of floating point infinity (default Inf).
See Also:
get_printoptions, set_string_function

Examples
Floating point precision can be set:

>>> np.set_printoptions(precision=4)
>>> print np.array([1.123456789])
[ 1.1235]

Long arrays can be summarised:

>>> np.set_printoptions(threshold=5)
>>> print np.arange(10)
[0 1 2 ..., 7 8 9]

Small results can be suppressed:

>>> eps = np.finfo(float).eps

>>> x = np.arange(4.)
>>> x**2 - (x + eps)**2
array([ -4.9304e-32, -4.4409e-16, 0.0000e+00, 0.0000e+00])
>>> np.set_printoptions(suppress=True)
>>> x**2 - (x + eps)**2
array([-0., -0., 0., 0.])

To put back the default options, you can use:

>>> np.set_printoptions(edgeitems=3,infstr=Inf,
... linewidth=75, nanstr=NaN, precision=8,
... suppress=False, threshold=1000)

get_printoptions()
Return the current print options.
Returns
print_opts : dict
Dictionary of current print options with keys

510 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

precision : int
threshold : int
edgeitems : int
linewidth : int
suppress : bool
nanstr : str
infstr : str
For a full description of these options, see set_printoptions.
See Also:
set_printoptions, set_string_function
set_string_function(f, repr=1)
Set a Python function to be used when pretty printing arrays.
Parameters
f : function or None
Function to be used to pretty print arrays. The function should expect a single array
argument and return a string of the representation of the array. If None, the function is
reset to the default NumPy function to print arrays.
repr : bool, optional
If True (default), the function for pretty printing (__repr__) is set, if False the func-
tion that returns the default string representation (__str__) is set.
See Also:
set_printoptions, get_printoptions

Examples

>>> def pprint(arr):

... return HA! - What are you going to do now?
...
>>> np.set_string_function(pprint)
>>> a = np.arange(10)
>>> a
HA! - What are you going to do now?
>>> print a
[0 1 2 3 4 5 6 7 8 9]

We can reset the function to the default:

>>> np.set_string_function(None)
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], l)

repr affects either pretty printing or normal string representation. Note that __repr__ is still affected by
setting __str__ because the width of each array element in the returned string becomes equal to the length of
the result of __str__().

3.5. Input and output 511

NumPy Reference, Release 2.0.0.dev8464

>>> x = np.arange(4)
>>> np.set_string_function(lambda x:random, repr=False)
>>> x.__str__()
random
>>> x.__repr__()
array([ 0, 1, 2, 3])

3.5.6 Base-n representations

binary_repr(num[, width]) Return the binary representation of the input number as a string.
base_repr(number[, base, padding]) Return a string representation of a number in the given base system.
binary_repr(num, width=None)
Return the binary representation of the input number as a string.
For negative numbers, if width is not given, a minus sign is added to the front. If width is given, the twos
complement of the number is returned, with respect to that width.
In a twos-complement system negative numbers are represented by the twos complement of the absolute
value. This is the most common method of representing signed integers on computers [R23]. A N-bit twos-
complement system can represent every integer in the range 2N 1 to +2N 1 1.
Parameters
num : int
Only an integer decimal number can be used.
width : int, optional
The length of the returned string if num is positive, the length of the twos complement
if num is negative.
Returns
bin : str
Binary representation of num or twos complement of num.
See Also:

base_repr
Return a string representation of a number in the given base system.

Notes
binary_repr is equivalent to using base_repr with base 2, but about 25x faster.

References
[R23]

Examples

>>> np.binary_repr(3)
11
>>> np.binary_repr(-3)
-11
>>> np.binary_repr(3, width=4)
0011

512 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

The twos complement is returned when the input number is negative and width is specified:

>>> np.binary_repr(-3, width=4)

1101

base_repr(number, base=2, padding=0)

Return a string representation of a number in the given base system.
Parameters
number : scalar
The value to convert. Only positive values are handled.
base : int, optional
Convert number to the base number system. The valid range is 2-36, the default value
is 2.
padding : int, optional
Number of zeros padded on the left. Default is 0 (no padding).
Returns
out : str
String representation of number in base system.
See Also:

binary_repr
Faster version of base_repr for base 2 that also handles negative numbers.

Examples

>>> np.base_repr(5)
101
>>> np.base_repr(6, 5)
11
>>> np.base_repr(7, base=5, padding=3)
00012

>>> np.base_repr(10, base=16)

A
>>> np.base_repr(32, base=16)
20

3.5.7 Data sources

DataSource([destpath]) A generic data source file (file, http, ftp, ...).

class DataSource(destpath=.)
A generic data source file (file, http, ftp, ...).
DataSources can be local files or remote files/URLs. The files may also be compressed or uncompressed.
DataSource hides some of the low-level details of downloading the file, allowing you to simply pass in a valid
file path (or URL) and obtain a file object.
Parameters
destpath : str or None, optional

3.5. Input and output 513

NumPy Reference, Release 2.0.0.dev8464

Path to the directory where the source file gets downloaded to for use. If destpath is
None, a temporary directory will be created. The default path is the current directory.

Notes
URLs require a scheme string (http://) to be used, without it they will fail:

>>> repos = DataSource()

>>> repos.exists(www.google.com/index.html)
False
>>> repos.exists(http://www.google.com/index.html)
True

Temporary directories are deleted when the DataSource is deleted.

Examples

>>> ds = DataSource(/home/guido)
>>> urlname = http://www.google.com/index.html
>>> gfile = ds.open(http://www.google.com/index.html) # remote file
>>> ds.abspath(urlname)
/home/guido/www.google.com/site/index.html

>>> ds = DataSource(None) # use with temporary file

>>> ds.open(/home/guido/foobar.txt)
<open file /home/guido.foobar.txt, mode r at 0x91d4430>
>>> ds.abspath(/home/guido/foobar.txt)
/tmp/tmpy4pgsP/home/guido/foobar.txt

Methods
abspath(path) Return absolute path of file in the DataSource directory.
exists(path) Test if path exists.
open(path[, mode]) Open and return file-like object.
abspath(path)
Return absolute path of file in the DataSource directory.
If path is an URL, then abspath will return either the location the file exists locally or the location it would
exist when opened using the open method.
Parameters
path : str
Can be a local file or a remote URL.
Returns
out : str
Complete path, including the DataSource destination directory.

Notes
The functionality is based on os.path.abspath.
exists(path)
Test if path exists.
Test if path exists as (and in this order):
a local file.

514 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

a remote URL that has been downloaded and stored locally in the DataSource directory.
a remote URL that has not been downloaded, but is valid and accessible.

Parameters
path : str
Can be a local file or a remote URL.
Returns
out : bool
True if path exists.

Notes
When path is an URL, exists will return True if its either stored locally in the DataSource directory, or is
a valid remote URL. DataSource does not discriminate between the two, the file is accessible if it exists in
either location.
open(path, mode=r)
Open and return file-like object.
If path is an URL, it will be downloaded, stored in the DataSource directory and opened from there.
Parameters
path : str
Local file path or URL to open.
mode : {r, w, a}, optional
Mode to open path. Mode r for reading, w for writing, a to append. Available
modes depend on the type of object specified by path. Default is r.
Returns
out : file object
File object.

3.6 Discrete Fourier Transform (numpy.fft)

3.6.1 Standard FFTs

fft(a[, n, axis]) Compute the one-dimensional discrete Fourier Transform.

ifft(a[, n, axis]) Compute the one-dimensional inverse discrete Fourier Transform.
fft2(a[, s, axes]) Compute the 2-dimensional discrete Fourier Transform
ifft2(a[, s, axes]) Compute the 2-dimensional inverse discrete Fourier Transform.
fftn(a[, s, axes]) Compute the N-dimensional discrete Fourier Transform.
ifftn(a[, s, axes]) Compute the N-dimensional inverse discrete Fourier Transform.
fft(a, n=None, axis=-1)
Compute the one-dimensional discrete Fourier Transform.
This function computes the one-dimensional n-point discrete Fourier Transform (DFT) with the efficient Fast
Fourier Transform (FFT) algorithm [CT].
Parameters
a : array_like

3.6. Discrete Fourier Transform (numpy.fft) 515

NumPy Reference, Release 2.0.0.dev8464

Input array, can be complex.

n : int, optional
Length of the transformed axis of the output. If n is smaller than the length of the input,
the input is cropped. If it is larger, the input is padded with zeros. If n is not given, the
length of the input (along the axis specified by axis) is used.
axis : int, optional
Axis over which to compute the FFT. If not given, the last axis is used.
Returns
out : complex ndarray
The truncated or zero-padded input, transformed along the axis indicated by axis, or the
last one if axis is not specified.
Raises
IndexError :
if axes is larger than the last axis of a.
See Also:

numpy.fft
for definition of the DFT and conventions used.
ifft
The inverse of fft.
fft2
The two-dimensional FFT.
fftn
The n-dimensional FFT.
rfftn
The n-dimensional FFT of real input.
fftfreq
Frequency bins for given FFT parameters.

Notes
FFT (Fast Fourier Transform) refers to a way the discrete Fourier Transform (DFT) can be calculated efficiently,
by using symmetries in the calculated terms. The symmetry is highest when n is a power of 2, and the transform
is therefore most efficient for these sizes.
The DFT is defined, with the conventions used in this implementation, in the documentation for the numpy.fft
module.

References
[CT]

Examples

>>> np.fft.fft(np.exp(2j * np.pi * np.arange(8) / 8))

array([ -3.44505240e-16 +1.14383329e-17j,
8.00000000e+00 -5.71092652e-15j,
2.33482938e-16 +1.22460635e-16j,
1.64863782e-15 +1.77635684e-15j,

516 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

9.95839695e-17 +2.33482938e-16j,
0.00000000e+00 +1.66837030e-15j,
1.14383329e-17 +1.22460635e-16j,
-1.64863782e-15 +1.77635684e-15j])

>>> import matplotlib.pyplot as plt

>>> t = np.arange(256)
>>> sp = np.fft.fft(np.sin(t))
>>> freq = np.fft.fftfreq(t.shape[-1])
>>> plt.plot(freq, sp.real, freq, sp.imag)
[<matplotlib.lines.Line2D object at 0x...>, <matplotlib.lines.Line2D object at 0x...>]
>>> plt.show()

In this example, real input has an FFT which is Hermitian, i.e., symmetric in the real part and anti-symmetric in
the imaginary part, as described in the numpy.fft documentation.

80
60
40
20
0
20
40
60
80
100
0.6 0.4 0.2 0.0 0.2 0.4 0.6

ifft(a, n=None, axis=-1)

Compute the one-dimensional inverse discrete Fourier Transform.
This function computes the inverse of the one-dimensional n-point discrete Fourier transform computed by fft. In
other words, ifft(fft(a)) == a to within numerical accuracy. For a general description of the algorithm
and definitions, see numpy.fft.
The input should be ordered in the same way as is returned by fft, i.e., a[0] should contain the zero fre-
quency term, a[1:n/2+1] should contain the positive-frequency terms, and a[n/2+1:] should contain the
negative-frequency terms, in order of decreasingly negative frequency. See numpy.fft for details.
Parameters
a : array_like
Input array, can be complex.
n : int, optional
Length of the transformed axis of the output. If n is smaller than the length of the input,
the input is cropped. If it is larger, the input is padded with zeros. If n is not given, the
length of the input (along the axis specified by axis) is used. See notes about padding
issues.

3.6. Discrete Fourier Transform (numpy.fft) 517

NumPy Reference, Release 2.0.0.dev8464

axis : int, optional

Axis over which to compute the inverse DFT. If not given, the last axis is used.
Returns
out : complex ndarray
The truncated or zero-padded input, transformed along the axis indicated by axis, or the
last one if axis is not specified.
Raises
IndexError :
If axes is larger than the last axis of a.
See Also:

numpy.fft
An introduction, with definitions and general explanations.
fft
The one-dimensional (forward) FFT, of which ifft is the inverse
ifft2
The two-dimensional inverse FFT.
ifftn
The n-dimensional inverse FFT.

Notes
If the input parameter n is larger than the size of the input, the input is padded by appending zeros at the end.
Even though this is the common approach, it might lead to surprising results. If a different padding is desired, it
must be performed before calling ifft.

Examples

>>> np.fft.ifft([0, 4, 0, 0])

array([ 1.+0.j, 0.+1.j, -1.+0.j, 0.-1.j])

Create and plot a band-limited signal with random phases:

>>> import matplotlib.pyplot as plt

>>> t = np.arange(400)
>>> n = np.zeros((400,), dtype=complex)
>>> n[40:60] = np.exp(1j*np.random.uniform(0, 2*np.pi, (20,)))
>>> s = np.fft.ifft(n)
>>> plt.plot(t, s.real, b-, t, s.imag, r--)
[<matplotlib.lines.Line2D object at 0x...>, <matplotlib.lines.Line2D object at 0x...>]
>>> plt.legend((real, imaginary))
<matplotlib.legend.Legend object at 0x...>
>>> plt.show()

518 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

0.03
real
0.02 imaginary
0.01
0.00
0.01
0.02
0.03
0 50 100 150 200 250 300 350 400

fft2(a, s=None, axes=(-2, -1))

Compute the 2-dimensional discrete Fourier Transform
This function computes the n-dimensional discrete Fourier Transform over any axes in an M-dimensional array
by means of the Fast Fourier Transform (FFT). By default, the transform is computed over the last two axes of
the input array, i.e., a 2-dimensional FFT.
Parameters
a : array_like
Input array, can be complex
s : sequence of ints, optional
Shape (length of each transformed axis) of the output (s[0] refers to axis 0, s[1] to axis
1, etc.). This corresponds to n for fft(x, n). Along each axis, if the given shape is smaller
than that of the input, the input is cropped. If it is larger, the input is padded with zeros.
if s is not given, the shape of the input (along the axes specified by axes) is used.
axes : sequence of ints, optional
Axes over which to compute the FFT. If not given, the last two axes are used. A repeated
index in axes means the transform over that axis is performed multiple times. A one-
element sequence means that a one-dimensional FFT is performed.
Returns
out : complex ndarray
The truncated or zero-padded input, transformed along the axes indicated by axes, or
the last two axes if axes is not given.
Raises
ValueError :
If s and axes have different length, or axes not given and len(s) != 2.
IndexError :
If an element of axes is larger than than the number of axes of a.
See Also:

3.6. Discrete Fourier Transform (numpy.fft) 519

NumPy Reference, Release 2.0.0.dev8464

numpy.fft
Overall view of discrete Fourier transforms, with definitions and conventions used.
ifft2
The inverse two-dimensional FFT.
fft
The one-dimensional FFT.
fftn
The n-dimensional FFT.
fftshift
Shifts zero-frequency terms to the center of the array. For two-dimensional input, swaps first and third
quadrants, and second and fourth quadrants.

Notes
fft2 is just fftn with a different default for axes.
The output, analogously to fft, contains the term for zero frequency in the low-order corner of the transformed
axes, the positive frequency terms in the first half of these axes, the term for the Nyquist frequency in the middle
of the axes and the negative frequency terms in the second half of the axes, in order of decreasingly negative
frequency.
See fftn for details and a plotting example, and numpy.fft for definitions and conventions used.

Examples

>>> a = np.mgrid[:5, :5][0]

>>> np.fft.fft2(a)
array([[ 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[ 5.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[ 10.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[ 15.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[ 20.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j]])

ifft2(a, s=None, axes=(-2, -1))

Compute the 2-dimensional inverse discrete Fourier Transform.
This function computes the inverse of the 2-dimensional discrete Fourier Transform over any number of axes
in an M-dimensional array by means of the Fast Fourier Transform (FFT). In other words, ifft2(fft2(a))
== a to within numerical accuracy. By default, the inverse transform is computed over the last two axes of the
input array.
The input, analogously to ifft, should be ordered in the same way as is returned by fft2, i.e. it should have the
term for zero frequency in the low-order corner of the two axes, the positive frequency terms in the first half of
these axes, the term for the Nyquist frequency in the middle of the axes and the negative frequency terms in the
second half of both axes, in order of decreasingly negative frequency.
Parameters
a : array_like
Input array, can be complex.
s : sequence of ints, optional
Shape (length of each axis) of the output (s[0] refers to axis 0, s[1] to axis 1, etc.).
This corresponds to n for ifft(x, n). Along each axis, if the given shape is smaller
than that of the input, the input is cropped. If it is larger, the input is padded with zeros.

520 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

if s is not given, the shape of the input (along the axes specified by axes) is used. See
notes for issue on ifft zero padding.
axes : sequence of ints, optional
Axes over which to compute the FFT. If not given, the last two axes are used. A repeated
index in axes means the transform over that axis is performed multiple times. A one-
element sequence means that a one-dimensional FFT is performed.
Returns
out : complex ndarray
The truncated or zero-padded input, transformed along the axes indicated by axes, or
the last two axes if axes is not given.
Raises
ValueError :
If s and axes have different length, or axes not given and len(s) != 2.
IndexError :
If an element of axes is larger than than the number of axes of a.
See Also:

numpy.fft
Overall view of discrete Fourier transforms, with definitions and conventions used.
fft2
The forward 2-dimensional FFT, of which ifft2 is the inverse.
ifftn
The inverse of the n-dimensional FFT.
fft
The one-dimensional FFT.
ifft
The one-dimensional inverse FFT.

Notes
ifft2 is just ifftn with a different default for axes.
See ifftn for details and a plotting example, and numpy.fft for definition and conventions used.
Zero-padding, analogously with ifft, is performed by appending zeros to the input along the specified dimension.
Although this is the common approach, it might lead to surprising results. If another form of zero padding is
desired, it must be performed before ifft2 is called.

Examples

>>> a = 4 * np.eye(4)
>>> np.fft.ifft2(a)
array([[ 1.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[ 0.+0.j, 0.+0.j, 0.+0.j, 1.+0.j],
[ 0.+0.j, 0.+0.j, 1.+0.j, 0.+0.j],
[ 0.+0.j, 1.+0.j, 0.+0.j, 0.+0.j]])

fftn(a, s=None, axes=None)

Compute the N-dimensional discrete Fourier Transform.

3.6. Discrete Fourier Transform (numpy.fft) 521

NumPy Reference, Release 2.0.0.dev8464

This function computes the N-dimensional discrete Fourier Transform over any number of axes in an M-
dimensional array by means of the Fast Fourier Transform (FFT).
Parameters
a : array_like
Input array, can be complex.
s : sequence of ints, optional
Shape (length of each transformed axis) of the output (s[0] refers to axis 0, s[1] to axis
1, etc.). This corresponds to n for fft(x, n). Along any axis, if the given shape is smaller
than that of the input, the input is cropped. If it is larger, the input is padded with zeros.
if s is not given, the shape of the input (along the axes specified by axes) is used.
axes : sequence of ints, optional
Axes over which to compute the FFT. If not given, the last len(s) axes are used, or
all axes if s is also not specified. Repeated indices in axes means that the transform over
that axis is performed multiple times.
Returns
out : complex ndarray
The truncated or zero-padded input, transformed along the axes indicated by axes, or
by a combination of s and a, as explained in the parameters section above.
Raises
ValueError :
If s and axes have different length.
IndexError :
If an element of axes is larger than than the number of axes of a.
See Also:

numpy.fft
Overall view of discrete Fourier transforms, with definitions and conventions used.
ifftn
The inverse of fftn, the inverse n-dimensional FFT.
fft
The one-dimensional FFT, with definitions and conventions used.
rfftn
The n-dimensional FFT of real input.
fft2
The two-dimensional FFT.
fftshift
Shifts zero-frequency terms to centre of array

Notes
The output, analogously to fft, contains the term for zero frequency in the low-order corner of all axes, the
positive frequency terms in the first half of all axes, the term for the Nyquist frequency in the middle of all axes
and the negative frequency terms in the second half of all axes, in order of decreasingly negative frequency.
See numpy.fft for details, definitions and conventions used.

522 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.mgrid[:3, :3, :3][0]

>>> np.fft.fftn(a, axes=(1, 2))
array([[[ 0.+0.j, 0.+0.j, 0.+0.j],
[ 0.+0.j, 0.+0.j, 0.+0.j],
[ 0.+0.j, 0.+0.j, 0.+0.j]],
[[ 9.+0.j, 0.+0.j, 0.+0.j],
[ 0.+0.j, 0.+0.j, 0.+0.j],
[ 0.+0.j, 0.+0.j, 0.+0.j]],
[[ 18.+0.j, 0.+0.j, 0.+0.j],
[ 0.+0.j, 0.+0.j, 0.+0.j],
[ 0.+0.j, 0.+0.j, 0.+0.j]]])
>>> np.fft.fftn(a, (2, 2), axes=(0, 1))
array([[[ 2.+0.j, 2.+0.j, 2.+0.j],
[ 0.+0.j, 0.+0.j, 0.+0.j]],
[[-2.+0.j, -2.+0.j, -2.+0.j],
[ 0.+0.j, 0.+0.j, 0.+0.j]]])

>>> import matplotlib.pyplot as plt

>>> [X, Y] = np.meshgrid(2 * np.pi * np.arange(200) / 12,
... 2 * np.pi * np.arange(200) / 34)
>>> S = np.sin(X) + np.cos(Y) + np.random.uniform(0, 1, X.shape)
>>> FS = np.fft.fftn(S)
>>> plt.imshow(np.log(np.abs(np.fft.fftshift(FS))**2))
<matplotlib.image.AxesImage object at 0x...>
>>> plt.show()

50

100

150

0 50 100 150

ifftn(a, s=None, axes=None)

Compute the N-dimensional inverse discrete Fourier Transform.
This function computes the inverse of the N-dimensional discrete Fourier Transform over any number of axes
in an M-dimensional array by means of the Fast Fourier Transform (FFT). In other words, ifftn(fftn(a))
== a to within numerical accuracy. For a description of the definitions and conventions used, see numpy.fft.
The input, analogously to ifft, should be ordered in the same way as is returned by fftn, i.e. it should have the
term for zero frequency in all axes in the low-order corner, the positive frequency terms in the first half of all
axes, the term for the Nyquist frequency in the middle of all axes and the negative frequency terms in the second

3.6. Discrete Fourier Transform (numpy.fft) 523

NumPy Reference, Release 2.0.0.dev8464

half of all axes, in order of decreasingly negative frequency.

Parameters
a : array_like
Input array, can be complex.
s : sequence of ints, optional
Shape (length of each transformed axis) of the output (s[0] refers to axis 0, s[1] to
axis 1, etc.). This corresponds to n for ifft(x, n). Along any axis, if the given
shape is smaller than that of the input, the input is cropped. If it is larger, the input is
padded with zeros. if s is not given, the shape of the input (along the axes specified by
axes) is used. See notes for issue on ifft zero padding.
axes : sequence of ints, optional
Axes over which to compute the IFFT. If not given, the last len(s) axes are used,
or all axes if s is also not specified. Repeated indices in axes means that the inverse
transform over that axis is performed multiple times.
Returns
out : complex ndarray
The truncated or zero-padded input, transformed along the axes indicated by axes, or
by a combination of s or a, as explained in the parameters section above.
Raises
ValueError :
If s and axes have different length.
IndexError :
If an element of axes is larger than than the number of axes of a.
See Also:

numpy.fft
Overall view of discrete Fourier transforms, with definitions and conventions used.
fftn
The forward n-dimensional FFT, of which ifftn is the inverse.
ifft
The one-dimensional inverse FFT.
ifft2
The two-dimensional inverse FFT.
ifftshift
Undoes fftshift, shifts zero-frequency terms to beginning of array.

Notes
See numpy.fft for definitions and conventions used.
Zero-padding, analogously with ifft, is performed by appending zeros to the input along the specified dimension.
Although this is the common approach, it might lead to surprising results. If another form of zero padding is
desired, it must be performed before ifftn is called.

524 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.eye(4)
>>> np.fft.ifftn(np.fft.fftn(a, axes=(0,)), axes=(1,))
array([[ 1.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[ 0.+0.j, 1.+0.j, 0.+0.j, 0.+0.j],
[ 0.+0.j, 0.+0.j, 1.+0.j, 0.+0.j],
[ 0.+0.j, 0.+0.j, 0.+0.j, 1.+0.j]])

Create and plot an image with band-limited frequency content:

>>> import matplotlib.pyplot as plt

>>> n = np.zeros((200,200), dtype=complex)
>>> n[60:80, 20:40] = np.exp(1j*np.random.uniform(0, 2*np.pi, (20, 20)))
>>> im = np.fft.ifftn(n).real
>>> plt.imshow(im)
<matplotlib.image.AxesImage object at 0x...>
>>> plt.show()

50

100

150

0 50 100 150

3.6.2 Real FFTs

rfft(a[, n, axis]) Compute the one-dimensional discrete Fourier Transform for real input.
irfft(a[, n, axis]) Compute the inverse of the n-point DFT for real input.
rfft2(a[, s, axes]) Compute the 2-dimensional FFT of a real array.
irfft2(a[, s, axes]) Compute the 2-dimensional inverse FFT of a real array.
rfftn(a[, s, axes]) Compute the N-dimensional discrete Fourier Transform for real input.
irfftn(a[, s, axes]) Compute the inverse of the N-dimensional FFT of real input.
rfft(a, n=None, axis=-1)
Compute the one-dimensional discrete Fourier Transform for real input.
This function computes the one-dimensional n-point discrete Fourier Transform (DFT) of a real-valued array by
means of an efficient algorithm called the Fast Fourier Transform (FFT).
Parameters
a : array_like

3.6. Discrete Fourier Transform (numpy.fft) 525

NumPy Reference, Release 2.0.0.dev8464

Input array
n : int, optional
Number of points along transformation axis in the input to use. If n is smaller than the
length of the input, the input is cropped. If it is larger, the input is padded with zeros. If
n is not given, the length of the input (along the axis specified by axis) is used.
axis : int, optional
Axis over which to compute the FFT. If not given, the last axis is used.
Returns
out : complex ndarray
The truncated or zero-padded input, transformed along the axis indicated by axis, or the
last one if axis is not specified. The length of the transformed axis is n/2+1.
Raises
IndexError :
If axis is larger than the last axis of a.
See Also:

numpy.fft
For definition of the DFT and conventions used.
irfft
The inverse of rfft.
fft
The one-dimensional FFT of general (complex) input.
fftn
The n-dimensional FFT.
rfftn
The n-dimensional FFT of real input.

Notes
When the DFT is computed for purely real input, the output is Hermite-symmetric, i.e. the negative frequency
terms are just the complex conjugates of the corresponding positive-frequency terms, and the negative-frequency
terms are therefore redundant. This function does not compute the negative frequency terms, and the length of
the transformed axis of the output is therefore n/2+1.
When A = rfft(a), A[0] contains the zero-frequency term, which must be purely real due to the Hermite
symmetry.
If n is even, A[-1] contains the term for frequencies n/2 and -n/2, and must also be purely real. If n is odd,
A[-1] contains the term for frequency A[(n-1)/2], and is complex in the general case.
If the input a contains an imaginary part, it is silently discarded.

Examples

>>> np.fft.fft([0, 1, 0, 0])

array([ 1.+0.j, 0.-1.j, -1.+0.j, 0.+1.j])
>>> np.fft.rfft([0, 1, 0, 0])
array([ 1.+0.j, 0.-1.j, -1.+0.j])

526 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notice how the final element of the fft output is the complex conjugate of the second element, for real input. For
rfft, this symmetry is exploited to compute only the non-negative frequency terms.
irfft(a, n=None, axis=-1)
Compute the inverse of the n-point DFT for real input.
This function computes the inverse of the one-dimensional n-point discrete Fourier Transform of real input
computed by rfft. In other words, irfft(rfft(a), len(a)) == a to within numerical accuracy. (See
Notes below for why len(a) is necessary here.)
The input is expected to be in the form returned by rfft, i.e. the real zero-frequency term followed by the complex
positive frequency terms in order of increasing frequency. Since the discrete Fourier Transform of real input is
Hermite-symmetric, the negative frequency terms are taken to be the complex conjugates of the corresponding
positive frequency terms.
Parameters
a : array_like
The input array.
n : int, optional
Length of the transformed axis of the output. For n output points, n/2+1 input points
are necessary. If the input is longer than this, it is cropped. If it is shorter than this, it is
padded with zeros. If n is not given, it is determined from the length of the input (along
the axis specified by axis).
axis : int, optional
Axis over which to compute the inverse FFT.
Returns
out : ndarray
The truncated or zero-padded input, transformed along the axis indicated by axis, or the
last one if axis is not specified. The length of the transformed axis is n, or, if n is not
given, 2*(m-1) where m is the length of the transformed axis of the input. To get an
odd number of output points, n must be specified.
Raises
IndexError :
If axis is larger than the last axis of a.
See Also:

numpy.fft
For definition of the DFT and conventions used.
rfft
The one-dimensional FFT of real input, of which irfft is inverse.
fft
The one-dimensional FFT.
irfft2
The inverse of the two-dimensional FFT of real input.
irfftn
The inverse of the n-dimensional FFT of real input.

3.6. Discrete Fourier Transform (numpy.fft) 527

NumPy Reference, Release 2.0.0.dev8464

Notes
Returns the real valued n-point inverse discrete Fourier transform of a, where a contains the non-negative fre-
quency terms of a Hermite-symmetric sequence. n is the length of the result, not the input.
If you specify an n such that a must be zero-padded or truncated, the extra/removed values will be
added/removed at high frequencies. One can thus resample a series to m points via Fourier interpolation by:
a_resamp = irfft(rfft(a), m).

Examples

>>> np.fft.ifft([1, -1j, -1, 1j])

array([ 0.+0.j, 1.+0.j, 0.+0.j, 0.+0.j])
>>> np.fft.irfft([1, -1j, -1])
array([ 0., 1., 0., 0.])

Notice how the last term in the input to the ordinary ifft is the complex conjugate of the second term, and the
output has zero imaginary part everywhere. When calling irfft, the negative frequencies are not specified, and
the output array is purely real.
rfft2(a, s=None, axes=(-2, -1))
Compute the 2-dimensional FFT of a real array.
Parameters
a : array
Input array, taken to be real.
s : sequence of ints, optional
Shape of the FFT.
axes : sequence of ints, optional
Axes over which to compute the FFT.
Returns
out : ndarray
The result of the real 2-D FFT.
See Also:

rfftn
Compute the N-dimensional discrete Fourier Transform for real input.

Notes
This is really just rfftn with different default behavior. For more details see rfftn.
irfft2(a, s=None, axes=(-2, -1))
Compute the 2-dimensional inverse FFT of a real array.
Parameters
a : array_like
The input array
s : sequence of ints, optional
Shape of the inverse FFT.
axes : sequence of ints, optional

528 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

The axes over which to compute the inverse fft. Default is the last two axes.
Returns
out : ndarray
The result of the inverse real 2-D FFT.
See Also:

irfftn
Compute the inverse of the N-dimensional FFT of real input.

Notes
This is really irfftn with different defaults. For more details see irfftn.
rfftn(a, s=None, axes=None)
Compute the N-dimensional discrete Fourier Transform for real input.
This function computes the N-dimensional discrete Fourier Transform over any number of axes in an M-
dimensional real array by means of the Fast Fourier Transform (FFT). By default, all axes are transformed,
with the real transform performed over the last axis, while the remaining transforms are complex.
Parameters
a : array_like
Input array, taken to be real.
s : sequence of ints, optional
Shape (length along each transformed axis) to use from the input. (s[0] refers to axis
0, s[1] to axis 1, etc.). The final element of s corresponds to n for rfft(x, n),
while for the remaining axes, it corresponds to n for fft(x, n). Along any axis, if
the given shape is smaller than that of the input, the input is cropped. If it is larger,
the input is padded with zeros. if s is not given, the shape of the input (along the axes
specified by axes) is used.
axes : sequence of ints, optional
Axes over which to compute the FFT. If not given, the last len(s) axes are used, or
all axes if s is also not specified.
Returns
out : complex ndarray
The truncated or zero-padded input, transformed along the axes indicated by axes, or by
a combination of s and a, as explained in the parameters section above. The length of
the last axis transformed will be s[-1]//2+1, while the remaining transformed axes
will have lengths according to s, or unchanged from the input.
Raises
ValueError :
If s and axes have different length.
IndexError :
If an element of axes is larger than than the number of axes of a.
See Also:

irfftn
The inverse of rfftn, i.e. the inverse of the n-dimensional FFT of real input.

3.6. Discrete Fourier Transform (numpy.fft) 529

NumPy Reference, Release 2.0.0.dev8464

fft
The one-dimensional FFT, with definitions and conventions used.
rfft
The one-dimensional FFT of real input.
fftn
The n-dimensional FFT.
rfft2
The two-dimensional FFT of real input.

Notes
The transform for real input is performed over the last transformation axis, as by rfft, then the transform over
the remaining axes is performed as by fftn. The order of the output is as for rfft for the final transformation axis,
and as for fftn for the remaining transformation axes.
See fft for details, definitions and conventions used.

Examples

>>> a = np.ones((2, 2, 2))

>>> np.fft.rfftn(a)
array([[[ 8.+0.j, 0.+0.j],
[ 0.+0.j, 0.+0.j]],
[[ 0.+0.j, 0.+0.j],
[ 0.+0.j, 0.+0.j]]])

>>> np.fft.rfftn(a, axes=(2, 0))

array([[[ 4.+0.j, 0.+0.j],
[ 4.+0.j, 0.+0.j]],
[[ 0.+0.j, 0.+0.j],
[ 0.+0.j, 0.+0.j]]])

irfftn(a, s=None, axes=None)

Compute the inverse of the N-dimensional FFT of real input.
This function computes the inverse of the N-dimensional discrete Fourier Transform for real input over any
number of axes in an M-dimensional array by means of the Fast Fourier Transform (FFT). In other words,
irfftn(rfftn(a), a.shape) == a to within numerical accuracy. (The a.shape is necessary like
len(a) is for irfft, and for the same reason.)
The input should be ordered in the same way as is returned by rfftn, i.e. as for irfft for the final transformation
axis, and as for ifftn along all the other axes.
Parameters
a : array_like
Input array.
s : sequence of ints, optional
Shape (length of each transformed axis) of the output (s[0] refers to axis 0, s[1] to
axis 1, etc.). s is also the number of input points used along this axis, except for the last
axis, where s[-1]//2+1 points of the input are used. Along any axis, if the shape
indicated by s is smaller than that of the input, the input is cropped. If it is larger, the
input is padded with zeros. If s is not given, the shape of the input (along the axes
specified by axes) is used.

530 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

axes : sequence of ints, optional

Axes over which to compute the inverse FFT. If not given, the last len(s) axes are used,
or all axes if s is also not specified. Repeated indices in axes means that the inverse
transform over that axis is performed multiple times.
Returns
out : ndarray
The truncated or zero-padded input, transformed along the axes indicated by axes, or
by a combination of s or a, as explained in the parameters section above. The length of
each transformed axis is as given by the corresponding element of s, or the length of the
input in every axis except for the last one if s is not given. In the final transformed axis
the length of the output when s is not given is 2*(m-1) where m is the length of the
final transformed axis of the input. To get an odd number of output points in the final
axis, s must be specified.
Raises
ValueError :
If s and axes have different length.
IndexError :
If an element of axes is larger than than the number of axes of a.
See Also:

rfftn
The forward n-dimensional FFT of real input, of which ifftn is the inverse.
fft
The one-dimensional FFT, with definitions and conventions used.
irfft
The inverse of the one-dimensional FFT of real input.
irfft2
The inverse of the two-dimensional FFT of real input.

Notes
See fft for definitions and conventions used.
See rfft for definitions and conventions used for real input.

Examples

>>> a = np.zeros((3, 2, 2))

>>> a[0, 0, 0] = 3 * 2 * 2
>>> np.fft.irfftn(a)
array([[[ 1., 1.],
[ 1., 1.]],
[[ 1., 1.],
[ 1., 1.]],
[[ 1., 1.],
[ 1., 1.]]])

3.6. Discrete Fourier Transform (numpy.fft) 531

NumPy Reference, Release 2.0.0.dev8464

3.6.3 Hermitian FFTs

hfft(a[, n, axis]) Compute the FFT of a signal whose spectrum has Hermitian symmetry.
ihfft(a[, n, axis]) Compute the inverse FFT of a signal whose spectrum has Hermitian symmetry.
hfft(a, n=None, axis=-1)
Compute the FFT of a signal whose spectrum has Hermitian symmetry.
Parameters
a : array_like
The input array.
n : int, optional
The length of the FFT.
axis : int, optional
The axis over which to compute the FFT, assuming Hermitian symmetry of the spec-
trum. Default is the last axis.
Returns
out : ndarray
The transformed input.
See Also:

rfft
Compute the one-dimensional FFT for real input.
ihfft
The inverse of hfft.

Notes
hfft/ihfft are a pair analogous to rfft/irfft, but for the opposite case: here the signal is real in the frequency domain
and has Hermite symmetry in the time domain. So here its hfft for which you must supply the length of the
result if it is to be odd: ihfft(hfft(a), len(a)) == a, within numerical accuracy.

Examples

>>> signal = np.array([[1, 1.j], [-1.j, 2]])

>>> np.conj(signal.T) - signal # check Hermitian symmetry
array([[ 0.-0.j, 0.+0.j],
[ 0.+0.j, 0.-0.j]])
>>> freq_spectrum = np.fft.hfft(signal)
>>> freq_spectrum
array([[ 1., 1.],
[ 2., -2.]])

ihfft(a, n=None, axis=-1)

Compute the inverse FFT of a signal whose spectrum has Hermitian symmetry.
Parameters
a : array_like
Input array.
n : int, optional
Length of the inverse FFT.

532 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

axis : int, optional

Axis over which to compute the inverse FFT, assuming Hermitian symmetry of the
spectrum. Default is the last axis.
Returns
out : ndarray
The transformed input.
See Also:
hfft, irfft

Notes
hfft/ihfft are a pair analogous to rfft/irfft, but for the opposite case: here the signal is real in the frequency domain
and has Hermite symmetry in the time domain. So here its hfft for which you must supply the length of the
result if it is to be odd: ihfft(hfft(a), len(a)) == a, within numerical accuracy.

3.6.4 Helper routines

fftfreq(n[, d]) Return the Discrete Fourier Transform sample frequencies.

fftshift(x[, axes]) Shift the zero-frequency component to the center of the spectrum.
ifftshift(x[, axes]) The inverse of fftshift.
fftfreq(n, d=1.0)
Return the Discrete Fourier Transform sample frequencies.
The returned float array contains the frequency bins in cycles/unit (with zero at the start) given a window length
n and a sample spacing d:

f = [0, 1, ..., n/2-1, -n/2, ..., -1] / (d*n) if n is even

f = [0, 1, ..., (n-1)/2, -(n-1)/2, ..., -1] / (d*n) if n is odd

Parameters
n : int
Window length.
d : scalar
Sample spacing.
Returns
out : ndarray
The array of length n, containing the sample frequencies.

Examples

>>> signal = np.array([-2, 8, 6, 4, 1, 0, 3, 5], dtype=float)

>>> fourier = np.fft.fft(signal)
>>> n = signal.size
>>> timestep = 0.1
>>> freq = np.fft.fftfreq(n, d=timestep)
>>> freq
array([ 0. , 1.25, 2.5 , 3.75, -5. , -3.75, -2.5 , -1.25])

3.6. Discrete Fourier Transform (numpy.fft) 533

NumPy Reference, Release 2.0.0.dev8464

fftshift(x, axes=None)
Shift the zero-frequency component to the center of the spectrum.
This function swaps half-spaces for all axes listed (defaults to all). Note that y[0] is the Nyquist component
only if len(x) is even.
Parameters
x : array_like
Input array.
axes : int or shape tuple, optional
Axes over which to shift. Default is None, which shifts all axes.
Returns
y : ndarray
The shifted array.
See Also:

ifftshift
The inverse of fftshift.

Examples

>>> freqs = np.fft.fftfreq(10, 0.1)

>>> freqs
array([ 0., 1., 2., 3., 4., -5., -4., -3., -2., -1.])
>>> np.fft.fftshift(freqs)
array([-5., -4., -3., -2., -1., 0., 1., 2., 3., 4.])

Shift the zero-frequency component only along the second axis:

>>> freqs = np.fft.fftfreq(9, d=1./9).reshape(3, 3)

>>> freqs
array([[ 0., 1., 2.],
[ 3., 4., -4.],
[-3., -2., -1.]])
>>> np.fft.fftshift(freqs, axes=(1,))
array([[ 2., 0., 1.],
[-4., 3., 4.],
[-1., -3., -2.]])

ifftshift(x, axes=None)
The inverse of fftshift.
Parameters
x : array_like
Input array.
axes : int or shape tuple, optional
Axes over which to calculate. Defaults to None, which shifts all axes.
Returns
y : ndarray
The shifted array.
See Also:

534 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

fftshift
Shift zero-frequency component to the center of the spectrum.

Examples

>>> freqs = np.fft.fftfreq(9, d=1./9).reshape(3, 3)

>>> freqs
array([[ 0., 1., 2.],
[ 3., 4., -4.],
[-3., -2., -1.]])
>>> np.fft.ifftshift(np.fft.fftshift(freqs))
array([[ 0., 1., 2.],
[ 3., 4., -4.],
[-3., -2., -1.]])

3.6.5 Background information

Fourier analysis is fundamentally a method for expressing a function as a sum of periodic components, and for recover-
ing the signal from those components. When both the function and its Fourier transform are replaced with discretized
counterparts, it is called the discrete Fourier transform (DFT). The DFT has become a mainstay of numerical com-
puting in part because of a very fast algorithm for computing it, called the Fast Fourier Transform (FFT), which was
known to Gauss (1805) and was brought to light in its current form by Cooley and Tukey [CT]. Press et al. [NR]
provide an accessible introduction to Fourier analysis and its applications.
Because the discrete Fourier transform separates its input into components that contribute at discrete frequencies, it
has a great number of applications in digital signal processing, e.g., for filtering, and in this context the discretized
input to the transform is customarily referred to as a signal, which exists in the time domain. The output is called a
spectrum or transform and exists in the frequency domain.
There are many ways to define the DFT, varying in the sign of the exponent, normalization, etc. In this implementation,
the DFT is defined as
n1  
X mk
Ak = am exp 2i k = 0, . . . , n 1.
m=0
n

The DFT is in general defined for complex inputs and outputs, and a single-frequency component at linear frequency
f is represented by a complex exponential am = exp{2i f mt}, where t is the sampling interval.
The values in the result follow so-called standard order: If A = fft(a, n), then A[0] contains the zero-
frequency term (the mean of the signal), which is always purely real for real inputs. Then A[1:n/2] contains
the positive-frequency terms, and A[n/2+1:] contains the negative-frequency terms, in order of decreasingly nega-
tive frequency. For an even number of input points, A[n/2] represents both positive and negative Nyquist frequency,
and is also purely real for real input. For an odd number of input points, A[(n-1)/2] contains the largest positive
frequency, while A[(n+1)/2] contains the largest negative frequency. The routine np.fft.fftfreq(A) returns
an array giving the frequencies of corresponding elements in the output. The routine np.fft.fftshift(A) shifts
transforms and their frequencies to put the zero-frequency components in the middle, and np.fft.ifftshift(A)
undoes that shift.
When the input a is a time-domain signal and A = fft(a), np.abs(A) is its amplitude spectrum and
np.abs(A)**2 is its power spectrum. The phase spectrum is obtained by np.angle(A).

3.6. Discrete Fourier Transform (numpy.fft) 535

NumPy Reference, Release 2.0.0.dev8464

The inverse DFT is defined as

n1  
1X mk
am = Ak exp 2i n = 0, . . . , n 1.
n n
k=0

It differs from the forward transform by the sign of the exponential argument and the normalization by 1/n.

Real and Hermitian transforms

When the input is purely real, its transform is Hermitian, i.e., the component at frequency fk is the complex conjugate
of the component at frequency fk , which means that for real inputs there is no information in the negative frequency
components that is not already available from the positive frequency components. The family of rfft functions is
designed to operate on real inputs, and exploits this symmetry by computing only the positive frequency components,
up to and including the Nyquist frequency. Thus, n input points produce n/2+1 complex output points. The inverses
of this family assumes the same symmetry of its input, and for an output of n points uses n/2+1 input points.
Correspondingly, when the spectrum is purely real, the signal is Hermitian. The hfft family of functions exploits this
symmetry by using n/2+1 complex points in the input (time) domain for n real points in the frequency domain.
In higher dimensions, FFTs are used, e.g., for image analysis and filtering. The computational efficiency of the FFT
means that it can also be a faster way to compute large convolutions, using the property that a convolution in the time
domain is equivalent to a point-by-point multiplication in the frequency domain.
In two dimensions, the DFT is defined as
M 1 N 1   
X X mk nl
Akl = amn exp 2i + k = 0, . . . , N 1; l = 0, . . . , M 1,
m=0 n=0
M N

which extends in the obvious way to higher dimensions, and the inverses in higher dimensions also extend in the same
way.

References

Examples

For examples, see the various functions.

3.7 Linear algebra (numpy.linalg)

3.7.1 Matrix and vector products

dot(a, b) Dot product of two arrays.

vdot(a, b) Return the dot product of two vectors.
inner(a, b) Inner product of two arrays.
outer(a, b) Compute the outer product of two vectors.
tensordot(a, b[, axes]) Compute tensor dot product along specified axes for arrays >= 1-D.
linalg.matrix_power(M, n) Raise a square matrix to the (integer) power n.
kron(a, b) Kronecker product of two arrays.

536 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

dot(a, b)
Dot product of two arrays.
For 2-D arrays it is equivalent to matrix multiplication, and for 1-D arrays to inner product of vectors (without
complex conjugation). For N dimensions it is a sum product over the last axis of a and the second-to-last of b:

dot(a, b)[i,j,k,m] = sum(a[i,j,:] * b[k,:,m])

Parameters
a : array_like
First argument.
b : array_like
Second argument.
Returns
output : ndarray
Returns the dot product of a and b. If a and b are both scalars or both 1-D arrays then a
scalar is returned; otherwise an array is returned.
Raises
ValueError :
If the last dimension of a is not the same size as the second-to-last dimension of b.

See Also:

vdot
Complex-conjugating dot product.
tensordot
Sum products over arbitrary axes.

Examples

>>> np.dot(3, 4)
12

Neither argument is complex-conjugated:

>>> np.dot([2j, 3j], [2j, 3j])

(-13+0j)

For 2-D arrays its the matrix product:

>>> a = [[1, 0], [0, 1]]

>>> b = [[4, 1], [2, 2]]
>>> np.dot(a, b)
array([[4, 1],
[2, 2]])

>>> a = np.arange(3*4*5*6).reshape((3,4,5,6))
>>> b = np.arange(3*4*5*6)[::-1].reshape((5,4,6,3))
>>> np.dot(a, b)[2,3,2,1,2,2]
499128

3.7. Linear algebra (numpy.linalg) 537

NumPy Reference, Release 2.0.0.dev8464

>>> sum(a[2,3,2,:] * b[1,2,:,2])

499128

vdot(a, b)
Return the dot product of two vectors.
The vdot(a, b) function handles complex numbers differently than dot(a, b). If the first argument is complex the
complex conjugate of the first argument is used for the calculation of the dot product.
For 2-D arrays it is equivalent to matrix multiplication, and for 1-D arrays to inner product of vectors (with
complex conjugation of a). For N dimensions it is a sum product over the last axis of a and the second-to-last
of b:

dot(a, b)[i,j,k,m] = sum(a[i,j,:] * b[k,:,m])

Parameters
a : array_like
If a is complex the complex conjugate is taken before calculation of the dot product.
b : array_like
Second argument to the dot product.
Returns
output : ndarray
Returns dot product of a and b. Can be an int, float, or complex depending on the types
of a and b.

See Also:

dot
Return the dot product without using the complex conjugate of the first argument.

Notes
The dot product is the summation of element wise multiplication.
n
X
ab= ai bi = a1 b1 + a2 b2 + + an bn
i=1

Examples

>>> a = np.array([1+2j,3+4j])
>>> b = np.array([5+6j,7+8j])
>>> np.vdot(a, b)
(70-8j)
>>> np.vdot(b, a)
(70+8j)
>>> a = np.array([[1, 4], [5, 6]])
>>> b = np.array([[4, 1], [2, 2]])
>>> np.vdot(a, b)
30
>>> np.vdot(b, a)
30

538 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

inner(a, b)
Inner product of two arrays.
Ordinary inner product of vectors for 1-D arrays (without complex conjugation), in higher dimensions a sum
product over the last axes.
Parameters
a, b : array_like
If a and b are nonscalar, their last dimensions of must match.
Returns
out : ndarray
out.shape = a.shape[:-1] + b.shape[:-1]
Raises
ValueError :
If the last dimension of a and b has different size.
See Also:

tensordot
Sum products over arbitrary axes.
dot
Generalised matrix product, using second last dimension of b.

Notes
For vectors (1-D arrays) it computes the ordinary inner-product:

np.inner(a, b) = sum(a[:]*b[:])

More generally, if ndim(a) = r > 0 and ndim(b) = s > 0:

np.inner(a, b) = np.tensordot(a, b, axes=(-1,-1))

or explicitly:

np.inner(a, b)[i0,...,ir-1,j0,...,js-1]
= sum(a[i0,...,ir-1,:]*b[j0,...,js-1,:])

In addition a or b may be scalars, in which case:

np.inner(a,b) = a*b

Examples
Ordinary inner product for vectors:

>>> a = np.array([1,2,3])
>>> b = np.array([0,1,0])
>>> np.inner(a, b)
2

A multidimensional example:

3.7. Linear algebra (numpy.linalg) 539

NumPy Reference, Release 2.0.0.dev8464

>>> a = np.arange(24).reshape((2,3,4))
>>> b = np.arange(4)
>>> np.inner(a, b)
array([[ 14, 38, 62],
[ 86, 110, 134]])

An example where b is a scalar:

>>> np.inner(np.eye(2), 7)
array([[ 7., 0.],
[ 0., 7.]])

outer(a, b)
Compute the outer product of two vectors.
Given two vectors, a = [a0, a1, ..., aM] and b = [b0, b1, ..., bN], the outer product [R63]
is:

[[a0*b0 a0*b1 ... a0*bN ]

[a1*b0 .
[ ... .
[aM*b0 aM*bN ]]

Parameters
a, b : array_like, shape (M,), (N,)
First and second input vectors. Inputs are flattened if they are not already 1-dimensional.
Returns
out : ndarray, shape (M, N)
out[i, j] = a[i] * b[j]

References
[R63]

Examples
Make a (very coarse) grid for computing a Mandelbrot set:

>>> rl = np.outer(np.ones((5,)), np.linspace(-2, 2, 5))

>>> rl
array([[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.]])
>>> im = np.outer(1j*np.linspace(2, -2, 5), np.ones((5,)))
>>> im
array([[ 0.+2.j, 0.+2.j, 0.+2.j, 0.+2.j, 0.+2.j],
[ 0.+1.j, 0.+1.j, 0.+1.j, 0.+1.j, 0.+1.j],
[ 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[ 0.-1.j, 0.-1.j, 0.-1.j, 0.-1.j, 0.-1.j],
[ 0.-2.j, 0.-2.j, 0.-2.j, 0.-2.j, 0.-2.j]])
>>> grid = rl + im
>>> grid
array([[-2.+2.j, -1.+2.j, 0.+2.j, 1.+2.j, 2.+2.j],

540 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

[-2.+1.j, -1.+1.j, 0.+1.j, 1.+1.j, 2.+1.j],

[-2.+0.j, -1.+0.j, 0.+0.j, 1.+0.j, 2.+0.j],
[-2.-1.j, -1.-1.j, 0.-1.j, 1.-1.j, 2.-1.j],
[-2.-2.j, -1.-2.j, 0.-2.j, 1.-2.j, 2.-2.j]])

An example using a vector of letters:

>>> x = np.array([a, b, c], dtype=object)

>>> np.outer(x, [1, 2, 3])
array([[a, aa, aaa],
[b, bb, bbb],
[c, cc, ccc]], dtype=object)

tensordot(a, b, axes=2)
Compute tensor dot product along specified axes for arrays >= 1-D.
Given two tensors (arrays of dimension greater than or equal to one), a and b, and an array_like object containing
two array_like objects, (a_axes, b_axes), sum the products of as and bs elements (components) over
the axes specified by a_axes and b_axes. The third argument can be a single non-negative integer_like
scalar, N; if it is such, then the last N dimensions of a and the first N dimensions of b are summed over.
Parameters
a, b : array_like, len(shape) >= 1
Tensors to dot.
axes : variable type
* integer_like scalar :
Number of axes to sum over (applies to both arrays); or
* array_like, shape = (2,), both elements array_like :
Axes to be summed over, first sequence applying to a, second to b.
See Also:
numpy.dot

Notes
When there is more than one axis to sum over - and they are not the last (first) axes of a (b) - the argument axes
should consist of two sequences of the same length, with the first axis to sum over given first in both sequences,
the second axis second, and so forth.

Examples
A traditional example:

>>> a = np.arange(60.).reshape(3,4,5)
>>> b = np.arange(24.).reshape(4,3,2)
>>> c = np.tensordot(a,b, axes=([1,0],[0,1]))
>>> c.shape
(5, 2)
>>> c
array([[ 4400., 4730.],
[ 4532., 4874.],
[ 4664., 5018.],
[ 4796., 5162.],
[ 4928., 5306.]])

3.7. Linear algebra (numpy.linalg) 541

NumPy Reference, Release 2.0.0.dev8464

>>> # A slower but equivalent way of computing the same...

>>> d = np.zeros((5,2))
>>> for i in range(5):
... for j in range(2):
... for k in range(3):
... for n in range(4):
... d[i,j] += a[k,n,i] * b[n,k,j]
>>> c == d
array([[ True, True],
[ True, True],
[ True, True],
[ True, True],
[ True, True]], dtype=bool)

An extended example taking advantage of the overloading of + and *:

>>> a = np.array(range(1, 9))

>>> a.shape = (2, 2, 2)
>>> A = np.array((a, b, c, d), dtype=object)
>>> A.shape = (2, 2)
>>> a; A
array([[[1, 2],
[3, 4]],
[[5, 6],
[7, 8]]])
array([[a, b],
[c, d]], dtype=object)

>>> np.tensordot(a, A) # third argument default is 2

array([abbcccdddd, aaaaabbbbbbcccccccdddddddd], dtype=object)

>>> np.tensordot(a, A, 1)
array([[[acc, bdd],
[aaacccc, bbbdddd]],
[[aaaaacccccc, bbbbbdddddd],
[aaaaaaacccccccc, bbbbbbbdddddddd]]], dtype=object)

>>> np.tensordot(a, A, 0) # "Left for reader" (result too long to incl.)

array([[[[[a, b],
[c, d]],
...

>>> np.tensordot(a, A, (0, 1))

array([[[abbbbb, cddddd],
[aabbbbbb, ccdddddd]],
[[aaabbbbbbb, cccddddddd],
[aaaabbbbbbbb, ccccdddddddd]]], dtype=object)

>>> np.tensordot(a, A, (2, 1))

array([[[abb, cdd],
[aaabbbb, cccdddd]],
[[aaaaabbbbbb, cccccdddddd],
[aaaaaaabbbbbbbb, cccccccdddddddd]]], dtype=object)

542 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.tensordot(a, A, ((0, 1), (0, 1)))

array([abbbcccccddddddd, aabbbbccccccdddddddd], dtype=object)

>>> np.tensordot(a, A, ((2, 1), (1, 0)))

array([acccbbdddd, aaaaacccccccbbbbbbdddddddd], dtype=object)

matrix_power(M, n)
Raise a square matrix to the (integer) power n.
For positive integers n, the power is computed by repeated matrix squarings and matrix multiplications. If n ==
0, the identity matrix of the same shape as M is returned. If n < 0, the inverse is computed and then raised to
the abs(n).
Parameters
M : ndarray or matrix object
Matrix to be powered. Must be square, i.e. M.shape == (m, m), with m a posi-
tive integer.
n : int
The exponent can be any integer or long integer, positive, negative, or zero.
Returns
M**n : ndarray or matrix object
The return value is the same shape and type as M; if the exponent is positive or zero
then the type of the elements is the same as those of M. If the exponent is negative the
elements are floating-point.
Raises
LinAlgError :
If the matrix is not numerically invertible.
See Also:

matrix
Provides an equivalent function as the exponentiation operator (**, not ^).

Examples

>>> from numpy import linalg as LA

>>> i = np.array([[0, 1], [-1, 0]]) # matrix equiv. of the imaginary unit
>>> LA.matrix_power(i, 3) # should = -i
array([[ 0, -1],
[ 1, 0]])
>>> LA.matrix_power(np.matrix(i), 3) # matrix arg returns matrix
matrix([[ 0, -1],
[ 1, 0]])
>>> LA.matrix_power(i, 0)
array([[1, 0],
[0, 1]])
>>> LA.matrix_power(i, -3) # should = 1/(-i) = i, but w/ f.p. elements
array([[ 0., 1.],
[-1., 0.]])

Somewhat more sophisticated example

3.7. Linear algebra (numpy.linalg) 543

NumPy Reference, Release 2.0.0.dev8464

>>> q = np.zeros((4, 4))

>>> q[0:2, 0:2] = -i
>>> q[2:4, 2:4] = i
>>> q # one of the three quarternion units not equal to 1
array([[ 0., -1., 0., 0.],
[ 1., 0., 0., 0.],
[ 0., 0., 0., 1.],
[ 0., 0., -1., 0.]])
>>> LA.matrix_power(q, 2) # = -np.eye(4)
array([[-1., 0., 0., 0.],
[ 0., -1., 0., 0.],
[ 0., 0., -1., 0.],
[ 0., 0., 0., -1.]])

kron(a, b)
Kronecker product of two arrays.
Computes the Kronecker product, a composite array made of blocks of the second array scaled by the first.
Parameters
a, b : array_like
Returns
out : ndarray
See Also:

outer
The outer product

Notes
The function assumes that the number of dimenensions of a and b are the same, if necessary prepending the
smallest with ones. If a.shape = (r0,r1,..,rN) and b.shape = (s0,s1,...,sN), the Kronecker product has shape
(r0*s0, r1*s1, ..., rN*SN). The elements are products of elements from a and b, organized explicitly by:

kron(a,b)[k0,k1,...,kN] = a[i0,i1,...,iN] * b[j0,j1,...,jN]

where:

kt = it * st + jt, t = 0,...,N

In the common 2-D case (N=1), the block structure can be visualized:

[[ a[0,0]*b, a[0,1]*b, ... , a[0,-1]*b ],

[ ... ... ],
[ a[-1,0]*b, a[-1,1]*b, ... , a[-1,-1]*b ]]

Examples

>>> np.kron([1,10,100], [5,6,7])

array([ 5, 6, 7, 50, 60, 70, 500, 600, 700])
>>> np.kron([5,6,7], [1,10,100])
array([ 5, 50, 500, 6, 60, 600, 7, 70, 700])

544 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.kron(np.eye(2), np.ones((2,2)))

array([[ 1., 1., 0., 0.],
[ 1., 1., 0., 0.],
[ 0., 0., 1., 1.],
[ 0., 0., 1., 1.]])

>>> a = np.arange(100).reshape((2,5,2,5))
>>> b = np.arange(24).reshape((2,3,4))
>>> c = np.kron(a,b)
>>> c.shape
(2, 10, 6, 20)
>>> I = (1,3,0,2)
>>> J = (0,2,1)
>>> J1 = (0,) + J # extend to ndim=4
>>> S1 = (1,) + b.shape
>>> K = tuple(np.array(I) * np.array(S1) + np.array(J1))
>>> c[K] == a[I]*b[J]
True

3.7.2 Decompositions

linalg.cholesky(a) Cholesky decomposition.

linalg.qr(a[, mode]) Compute the qr factorization of a matrix.
linalg.svd(a[, full_matrices, compute_uv]) Singular Value Decomposition.
cholesky(a)
Cholesky decomposition.
Return the Cholesky decomposition, L * L.H, of the square matrix a, where L is lower-triangular and .H is
the conjugate transpose operator (which is the ordinary transpose if a is real-valued). a must be Hermitian
(symmetric if real-valued) and positive-definite. Only L is actually returned.
Parameters
a : array_like, shape (M, M)
Hermitian (symmetric if all elements are real), positive-definite input matrix.
Returns
L : ndarray, or matrix object if a is, shape (M, M)
Lower-triangular Cholesky factor of a.
Raises
LinAlgError :
If the decomposition fails, for example, if a is not positive-definite.

Notes
The Cholesky decomposition is often used as a fast way of solving

Ax = b

(when A is both Hermitian/symmetric and positive-definite).

3.7. Linear algebra (numpy.linalg) 545

NumPy Reference, Release 2.0.0.dev8464

First, we solve for y in

Ly = b,

and then for x in

L.Hx = y.

Examples

>>> A = np.array([[1,-2j],[2j,5]])
>>> A
array([[ 1.+0.j, 0.-2.j],
[ 0.+2.j, 5.+0.j]])
>>> L = np.linalg.cholesky(A)
>>> L
array([[ 1.+0.j, 0.+0.j],
[ 0.+2.j, 1.+0.j]])
>>> np.dot(L, L.T.conj()) # verify that L * L.H = A
array([[ 1.+0.j, 0.-2.j],
[ 0.+2.j, 5.+0.j]])
>>> A = [[1,-2j],[2j,5]] # what happens if A is only array_like?
>>> np.linalg.cholesky(A) # an ndarray object is returned
array([[ 1.+0.j, 0.+0.j],
[ 0.+2.j, 1.+0.j]])
>>> # But a matrix object is returned if A is a matrix object
>>> LA.cholesky(np.matrix(A))
matrix([[ 1.+0.j, 0.+0.j],
[ 0.+2.j, 1.+0.j]])

qr(a, mode=full)
Compute the qr factorization of a matrix.
Factor the matrix a as qr, where q is orthonormal and r is upper-triangular.
Parameters
a : array_like
Matrix to be factored, of shape (M, N).
mode : {full, r, economic}, optional
Specifies the values to be returned. full is the default. Economic mode is slightly
faster then r mode if only r is needed.
Returns
q : ndarray of float or complex, optional
The orthonormal matrix, of shape (M, K). Only returned if mode=full.
r : ndarray of float or complex, optional
The upper-triangular matrix, of shape (K, N) with K = min(M, N). Only returned when
mode=full or mode=r.
a2 : ndarray of float or complex, optional
Array of shape (M, N), only returned when mode=economic. The diagonal and
the upper triangle of a2 contains r, while the rest of the matrix is undefined.

546 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Raises
LinAlgError :
If factoring fails.

Notes
This is an interface to the LAPACK routines dgeqrf, zgeqrf, dorgqr, and zungqr.
For more information on the qr factorization, see for example: http://en.wikipedia.org/wiki/QR_factorization
Subclasses of ndarray are preserved, so if a is of type matrix, all the return values will be matrices too.

Examples

>>> a = np.random.randn(9, 6)
>>> q, r = np.linalg.qr(a)
>>> np.allclose(a, np.dot(q, r)) # a does equal qr
True
>>> r2 = np.linalg.qr(a, mode=r)
>>> r3 = np.linalg.qr(a, mode=economic)
>>> np.allclose(r, r2) # mode=r returns the same r as mode=full
True
>>> # But only triu parts are guaranteed equal when mode=economic
>>> np.allclose(r, np.triu(r3[:6,:6], k=0))
True

Example illustrating a common use of qr: solving of least squares problems

What are the least-squares-best m and y0 in y = y0 + mx for the following data: {(0,1), (1,0), (1,2), (2,1)}.
(Graph the points and youll see that it should be y0 = 0, m = 1.) The answer is provided by solving the
over-determined matrix equation Ax = b, where:

A = array([[0, 1], [1, 1], [1, 1], [2, 1]])

x = array([[y0], [m]])
b = array([[1], [0], [2], [1]])

If A = qr such that q is orthonormal (which is always possible via Gram-Schmidt), then x = inv(r) *
(q.T) * b. (In numpy practice, however, we simply use lstsq.)

>>> A = np.array([[0, 1], [1, 1], [1, 1], [2, 1]])

>>> A
array([[0, 1],
[1, 1],
[1, 1],
[2, 1]])
>>> b = np.array([1, 0, 2, 1])
>>> q, r = LA.qr(A)
>>> p = np.dot(q.T, b)
>>> np.dot(LA.inv(r), p)
array([ 1.1e-16, 1.0e+00])

svd(a, full_matrices=1, compute_uv=1)

Singular Value Decomposition.
Factors the matrix a as u * np.diag(s) * v, where u and v are unitary and s is a 1-d array of as singular
values.
Parameters
a : array_like

3.7. Linear algebra (numpy.linalg) 547

NumPy Reference, Release 2.0.0.dev8464

A real or complex matrix of shape (M, N) .

full_matrices : bool, optional
If True (default), u and v have the shapes (M, M) and (N, N), respectively. Otherwise,
the shapes are (M, K) and (K, N), respectively, where K = min(M, N).
compute_uv : bool, optional
Whether or not to compute u and v in addition to s. True by default.
Returns
u : ndarray
Unitary matrix. The shape of u is (M, M) or (M, K) depending on value of
full_matrices.
s : ndarray
The singular values, sorted so that s[i] >= s[i+1]. s is a 1-d array of length
min(M, N).
v : ndarray
Unitary matrix of shape (N, N) or (K, N), depending on full_matrices.
Raises
LinAlgError :
If SVD computation does not converge.

Notes
The SVD is commonly written as a = U S V.H. The v returned by this function is V.H and u = U.
If U is a unitary matrix, it means that it satisfies U.H = inv(U).
The rows of v are the eigenvectors of a.H a. The columns of u are the eigenvectors of a a.H. For row i in v
and column i in u, the corresponding eigenvalue is s[i]**2.
If a is a matrix object (as opposed to an ndarray), then so are all the return values.

Examples

>>> a = np.random.randn(9, 6) + 1j*np.random.randn(9, 6)

Reconstruction based on full SVD:

>>> U, s, V = np.linalg.svd(a, full_matrices=True)

>>> U.shape, V.shape, s.shape
((9, 6), (6, 6), (6,))
>>> S = np.zeros((9, 6), dtype=complex)
>>> S[:6, :6] = np.diag(s)
>>> np.allclose(a, np.dot(U, np.dot(S, V)))
True

Reconstruction based on reduced SVD:

>>> U, s, V = np.linalg.svd(a, full_matrices=False)

>>> U.shape, V.shape, s.shape
((9, 6), (6, 6), (6,))
>>> S = np.diag(s)

548 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.allclose(a, np.dot(U, np.dot(S, V)))

True

3.7.3 Matrix eigenvalues

linalg.eig(a) Compute the eigenvalues and right eigenvectors of a square array.

linalg.eigh(a[, UPLO]) Return the eigenvalues and eigenvectors of a Hermitian or symmetric matrix.
linalg.eigvals(a) Compute the eigenvalues of a general matrix.
linalg.eigvalsh(a[, UPLO]) Compute the eigenvalues of a Hermitian or real symmetric matrix.
eig(a)
Compute the eigenvalues and right eigenvectors of a square array.
Parameters
a : array_like, shape (M, M)
A square array of real or complex elements.
Returns
w : ndarray, shape (M,)
The eigenvalues, each repeated according to its multiplicity. The eigenvalues are not
necessarily ordered, nor are they necessarily real for real arrays (though for real arrays
complex-valued eigenvalues should occur in conjugate pairs).
v : ndarray, shape (M, M)
The normalized (unit length) eigenvectors, such that the column v[:,i] is the eigen-
vector corresponding to the eigenvalue w[i].
Raises
LinAlgError :
If the eigenvalue computation does not converge.
See Also:

eigvalsh
eigenvalues of a symmetric or Hermitian (conjugate symmetric) array.
eigvals
eigenvalues of a non-symmetric array.

Notes
This is a simple interface to the LAPACK routines dgeev and zgeev which compute the eigenvalues and eigen-
vectors of, respectively, general real- and complex-valued square arrays.
The number w is an eigenvalue of a if there exists a vector v such that dot(a,v) = w * v. Thus, the arrays
a, w, and v satisfy the equations dot(a[i,:], v[i]) = w[i] * v[:,i] for i {0, ..., M 1}.
The array v of eigenvectors may not be of maximum rank, that is, some of the columns may be linearly depen-
dent, although round-off error may obscure that fact. If the eigenvalues are all different, then theoretically the
eigenvectors are linearly independent. Likewise, the (complex-valued) matrix of eigenvectors v is unitary if the
matrix a is normal, i.e., if dot(a, a.H) = dot(a.H, a), where a.H denotes the conjugate transpose of
a.
Finally, it is emphasized that v consists of the right (as in right-hand side) eigenvectors of a. A vector y satisfying
dot(y.T, a) = z * y.T for some number z is called a left eigenvector of a, and, in general, the left and
right eigenvectors of a matrix are not necessarily the (perhaps conjugate) transposes of each other.

3.7. Linear algebra (numpy.linalg) 549

NumPy Reference, Release 2.0.0.dev8464

References
G. Strang, Linear Algebra and Its Applications, 2nd Ed., Orlando, FL, Academic Press, Inc., 1980, Various pp.

Examples

>>> from numpy import linalg as LA

(Almost) trivial example with real e-values and e-vectors.

>>> w, v = LA.eig(np.diag((1, 2, 3)))

>>> w; v
array([ 1., 2., 3.])
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])

Real matrix possessing complex e-values and e-vectors; note that the e-values are complex conjugates of each
other.

>>> w, v = LA.eig(np.array([[1, -1], [1, 1]]))

>>> w; v
array([ 1. + 1.j, 1. - 1.j])
array([[ 0.70710678+0.j , 0.70710678+0.j ],
[ 0.00000000-0.70710678j, 0.00000000+0.70710678j]])

Complex-valued matrix with real e-values (but complex-valued e-vectors); note that a.conj().T = a, i.e., a is
Hermitian.

>>> a = np.array([[1, 1j], [-1j, 1]])

>>> w, v = LA.eig(a)
>>> w; v
array([ 2.00000000e+00+0.j, 5.98651912e-36+0.j]) # i.e., {2, 0}
array([[ 0.00000000+0.70710678j, 0.70710678+0.j ],
[ 0.70710678+0.j , 0.00000000+0.70710678j]])

Be careful about round-off error!

>>> a = np.array([[1 + 1e-9, 0], [0, 1 - 1e-9]])

>>> # Theor. e-values are 1 +/- 1e-9
>>> w, v = LA.eig(a)
>>> w; v
array([ 1., 1.])
array([[ 1., 0.],
[ 0., 1.]])

eigh(a, UPLO=L)
Return the eigenvalues and eigenvectors of a Hermitian or symmetric matrix.
Returns two objects, a 1-D array containing the eigenvalues of a, and a 2-D square array or matrix (depending
on the input type) of the corresponding eigenvectors (in columns).
Parameters
a : array_like, shape (M, M)
A complex Hermitian or real symmetric matrix.
UPLO : {L, U}, optional

550 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Specifies whether the calculation is done with the lower triangular part of a (L, default)
or the upper triangular part (U).
Returns
w : ndarray, shape (M,)
The eigenvalues, not necessarily ordered.
v : ndarray, or matrix object if a is, shape (M, M)
The column v[:, i] is the normalized eigenvector corresponding to the eigenvalue
w[i].
Raises
LinAlgError :
If the eigenvalue computation does not converge.
See Also:

eigvalsh
eigenvalues of symmetric or Hermitian arrays.
eig
eigenvalues and right eigenvectors for non-symmetric arrays.
eigvals
eigenvalues of non-symmetric arrays.

Notes
This is a simple interface to the LAPACK routines dsyevd and zheevd, which compute the eigenvalues and
eigenvectors of real symmetric and complex Hermitian arrays, respectively.
The eigenvalues of real symmetric or complex Hermitian matrices are always real. [R47] The array v of (column)
eigenvectors is unitary and a, w, and v satisfy the equations dot(a, v[:, i]) = w[i] * v[:, i].

References
[R47]

Examples

>>> from numpy import linalg as LA

>>> a = np.array([[1, -2j], [2j, 5]])
>>> a
array([[ 1.+0.j, 0.-2.j],
[ 0.+2.j, 5.+0.j]])
>>> w, v = LA.eigh(a)
>>> w; v
array([ 0.17157288, 5.82842712])
array([[-0.92387953+0.j , -0.38268343+0.j ],
[ 0.00000000+0.38268343j, 0.00000000-0.92387953j]])

>>> np.dot(a, v[:, 0]) - w[0] * v[:, 0] # verify 1st e-val/vec pair
array([2.77555756e-17 + 0.j, 0. + 1.38777878e-16j])
>>> np.dot(a, v[:, 1]) - w[1] * v[:, 1] # verify 2nd e-val/vec pair
array([ 0.+0.j, 0.+0.j])

3.7. Linear algebra (numpy.linalg) 551

NumPy Reference, Release 2.0.0.dev8464

>>> A = np.matrix(a) # what happens if input is a matrix object

>>> A
matrix([[ 1.+0.j, 0.-2.j],
[ 0.+2.j, 5.+0.j]])
>>> w, v = LA.eigh(A)
>>> w; v
array([ 0.17157288, 5.82842712])
matrix([[-0.92387953+0.j , -0.38268343+0.j ],
[ 0.00000000+0.38268343j, 0.00000000-0.92387953j]])

eigvals(a)
Compute the eigenvalues of a general matrix.
Main difference between eigvals and eig: the eigenvectors arent returned.
Parameters
a : array_like, shape (M, M)
A complex- or real-valued matrix whose eigenvalues will be computed.
Returns
w : ndarray, shape (M,)
The eigenvalues, each repeated according to its multiplicity. They are not necessarily
ordered, nor are they necessarily real for real matrices.
Raises
LinAlgError :
If the eigenvalue computation does not converge.
See Also:

eig
eigenvalues and right eigenvectors of general arrays
eigvalsh
eigenvalues of symmetric or Hermitian arrays.
eigh
eigenvalues and eigenvectors of symmetric/Hermitian arrays.

Notes
This is a simple interface to the LAPACK routines dgeev and zgeev that sets those routines flags to return only
the eigenvalues of general real and complex arrays, respectively.

Examples
Illustration, using the fact that the eigenvalues of a diagonal matrix are its diagonal elements, that multiplying
a matrix on the left by an orthogonal matrix, Q, and on the right by Q.T (the transpose of Q), preserves the
eigenvalues of the middle matrix. In other words, if Q is orthogonal, then Q * A * Q.T has the same
eigenvalues as A:

>>> from numpy import linalg as LA

>>> x = np.random.random()
>>> Q = np.array([[np.cos(x), -np.sin(x)], [np.sin(x), np.cos(x)]])
>>> LA.norm(Q[0, :]), LA.norm(Q[1, :]), np.dot(Q[0, :],Q[1, :])
(1.0, 1.0, 0.0)

552 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Now multiply a diagonal matrix by Q on one side and by Q.T on the other:

>>> D = np.diag((-1,1))
>>> LA.eigvals(D)
array([-1., 1.])
>>> A = np.dot(Q, D)
>>> A = np.dot(A, Q.T)
>>> LA.eigvals(A)
array([ 1., -1.])

eigvalsh(a, UPLO=L)
Compute the eigenvalues of a Hermitian or real symmetric matrix.
Main difference from eigh: the eigenvectors are not computed.
Parameters
a : array_like, shape (M, M)
A complex- or real-valued matrix whose eigenvalues are to be computed.
UPLO : {L, U}, optional
Specifies whether the calculation is done with the lower triangular part of a (L, default)
or the upper triangular part (U).
Returns
w : ndarray, shape (M,)
The eigenvalues, not necessarily ordered, each repeated according to its multiplicity.
Raises
LinAlgError :
If the eigenvalue computation does not converge.
See Also:

eigh
eigenvalues and eigenvectors of symmetric/Hermitian arrays.
eigvals
eigenvalues of general real or complex arrays.
eig
eigenvalues and right eigenvectors of general real or complex arrays.

Notes
This is a simple interface to the LAPACK routines dsyevd and zheevd that sets those routines flags to return
only the eigenvalues of real symmetric and complex Hermitian arrays, respectively.

Examples

>>> from numpy import linalg as LA

>>> a = np.array([[1, -2j], [2j, 5]])
>>> LA.eigvalsh(a)
array([ 0.17157288+0.j, 5.82842712+0.j])

3.7. Linear algebra (numpy.linalg) 553

NumPy Reference, Release 2.0.0.dev8464

3.7.4 Norms and other numbers

linalg.norm(x[, ord]) Matrix or vector norm.

linalg.cond(x[, p]) Compute the condition number of a matrix.
linalg.det(a) Compute the determinant of an array.
linalg.slogdet(a) Compute the sign and (natural) logarithm of the determinant of an array.
trace(a[, offset, axis1, axis2, dtype, out]) Return the sum along diagonals of the array.
norm(x, ord=None)
Matrix or vector norm.
This function is able to return one of seven different matrix norms, or one of an infinite number of vector norms
(described below), depending on the value of the ord parameter.
Parameters
x : array_like, shape (M,) or (M, N)
Input array.
ord : {non-zero int, inf, -inf, fro}, optional
Order of the norm (see table under Notes). inf means numpys inf object.
Returns
n : float
Norm of the matrix or vector.

Notes
For values of ord <= 0, the result is, strictly speaking, not a mathematical norm, but it may still be useful
for various numerical purposes.
The following norms can be calculated:
ord norm for matrices norm for vectors
None Frobenius norm 2-norm
fro Frobenius norm
inf max(sum(abs(x), axis=1)) max(abs(x))
-inf min(sum(abs(x), axis=1)) min(abs(x))
0 sum(x != 0)
1 max(sum(abs(x), axis=0)) as below
-1 min(sum(abs(x), axis=0)) as below
2 2-norm (largest sing. value) as below
-2 smallest singular value as below
other sum(abs(x)**ord)**(1./ord)
The Frobenius norm is given by [R48]:
||A||F = [ i,j abs(ai,j )2 ]1/2
P

References
[R48]

Examples

>>> from numpy import linalg as LA

>>> a = np.arange(9) - 4
>>> a
array([-4, -3, -2, -1, 0, 1, 2, 3, 4])

554 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> b = a.reshape((3, 3))

>>> b
array([[-4, -3, -2],
[-1, 0, 1],
[ 2, 3, 4]])

>>> LA.norm(a)
7.745966692414834
>>> LA.norm(b)
7.745966692414834
>>> LA.norm(b, fro)
7.745966692414834
>>> LA.norm(a, np.inf)
4
>>> LA.norm(b, np.inf)
9
>>> LA.norm(a, -np.inf)
0
>>> LA.norm(b, -np.inf)
2

>>> LA.norm(a, 1)
20
>>> LA.norm(b, 1)
7
>>> LA.norm(a, -1)
-4.6566128774142013e-010
>>> LA.norm(b, -1)
6
>>> LA.norm(a, 2)
7.745966692414834
>>> LA.norm(b, 2)
7.3484692283495345

>>> LA.norm(a, -2)

nan
>>> LA.norm(b, -2)
1.8570331885190563e-016
>>> LA.norm(a, 3)
5.8480354764257312
>>> LA.norm(a, -3)
nan

cond(x, p=None)
Compute the condition number of a matrix.
This function is capable of returning the condition number using one of seven different norms, depending on the
value of p (see Parameters below).
Parameters
x : array_like, shape (M, N)
The matrix whose condition number is sought.
p : {None, 1, -1, 2, -2, inf, -inf, fro}, optional
Order of the norm:

3.7. Linear algebra (numpy.linalg) 555

NumPy Reference, Release 2.0.0.dev8464

p norm for matrices

None 2-norm, computed directly using the SVD
fro Frobenius norm
inf max(sum(abs(x), axis=1))
-inf min(sum(abs(x), axis=1))
1 max(sum(abs(x), axis=0))
-1 min(sum(abs(x), axis=0))
2 2-norm (largest sing. value)
-2 smallest singular value
inf means the numpy.inf object, and the Frobenius norm is the root-of-sum-of-squares
norm.
Returns
c : {float, inf}
The condition number of the matrix. May be infinite.
See Also:
numpy.linalg.linalg.norm

Notes
The condition number of x is defined as the norm of x times the norm of the inverse of x [R46]; the norm can be
the usual L2-norm (root-of-sum-of-squares) or one of a number of other matrix norms.

References
[R46]

Examples

>>> from numpy import linalg as LA

>>> a = np.array([[1, 0, -1], [0, 1, 0], [1, 0, 1]])
>>> a
array([[ 1, 0, -1],
[ 0, 1, 0],
[ 1, 0, 1]])
>>> LA.cond(a)
1.4142135623730951
>>> LA.cond(a, fro)
3.1622776601683795
>>> LA.cond(a, np.inf)
2.0
>>> LA.cond(a, -np.inf)
1.0
>>> LA.cond(a, 1)
2.0
>>> LA.cond(a, -1)
1.0
>>> LA.cond(a, 2)
1.4142135623730951
>>> LA.cond(a, -2)
0.70710678118654746
>>> min(LA.svd(a, compute_uv=0))*min(LA.svd(LA.inv(a), compute_uv=0))
0.70710678118654746

det(a)
Compute the determinant of an array.

556 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
a : array_like, shape (M, M)
Input array.
Returns
det : ndarray
Determinant of a.
See Also:

slogdet
Another way to representing the determinant, more suitable for large matrices where underflow/overflow
may occur.

Notes
The determinant is computed via LU factorization using the LAPACK routine z/dgetrf.

Examples
The determinant of a 2-D array [[a, b], [c, d]] is ad - bc:

>>> a = np.array([[1, 2], [3, 4]])

>>> np.linalg.det(a)
-2.0

slogdet(a)
Compute the sign and (natural) logarithm of the determinant of an array.
If an array has a very small or very large determinant, than a call to det may overflow or underflow. This
routine is more robust against such issues, because it computes the logarithm of the determinant rather than the
determinant itself.
Parameters
a : array_like, shape (M, M)
Input array.
Returns
sign : float or complex
A number representing the sign of the determinant. For a real matrix, this is 1, 0, or -1.
For a complex matrix, this is a complex number with absolute value 1 (i.e., it is on the
unit circle), or else 0.
logdet : float
The natural log of the absolute value of the determinant.
If the determinant is zero, then sign will be 0 and logdet will be :
-Inf. In all cases, the determinant is equal to sign * np.exp(logdet). :
See Also:
det

Notes
The determinant is computed via LU factorization using the LAPACK routine z/dgetrf. New in version 2.0.0..

3.7. Linear algebra (numpy.linalg) 557

NumPy Reference, Release 2.0.0.dev8464

Examples
The determinant of a 2-D array [[a, b], [c, d]] is ad - bc:

>>> a = np.array([[1, 2], [3, 4]])

>>> (sign, logdet) = np.linalg.slogdet(a)
>>> (sign, logdet)
(-1, 0.69314718055994529)
>>> sign * np.exp(logdet)
-2.0

This routine succeeds where ordinary det does not:

>>> np.linalg.det(np.eye(500) * 0.1)

0.0
>>> np.linalg.slogdet(np.eye(500) * 0.1)
(1, -1151.2925464970228)

trace(a, offset=0, axis1=0, axis2=1, dtype=None, out=None)

Return the sum along diagonals of the array.
If a is 2-D, the sum along its diagonal with the given offset is returned, i.e., the sum of elements
a[i,i+offset] for all i.
If a has more than two dimensions, then the axes specified by axis1 and axis2 are used to determine the 2-D
sub-arrays whose traces are returned. The shape of the resulting array is the same as that of a with axis1 and
axis2 removed.
Parameters
a : array_like
Input array, from which the diagonals are taken.
offset : int, optional
Offset of the diagonal from the main diagonal. Can be both positive and negative.
Defaults to 0.
axis1, axis2 : int, optional
Axes to be used as the first and second axis of the 2-D sub-arrays from which the
diagonals should be taken. Defaults are the first two axes of a.
dtype : dtype, optional
Determines the data-type of the returned array and of the accumulator where the ele-
ments are summed. If dtype has the value None and a is of integer type of precision less
than the default integer precision, then the default integer precision is used. Otherwise,
the precision is the same as that of a.
out : ndarray, optional
Array into which the output is placed. Its type is preserved and it must be of the right
shape to hold the output.
Returns
sum_along_diagonals : ndarray
If a is 2-D, the sum along the diagonal is returned. If a has larger dimensions, then an
array of sums along diagonals is returned.

558 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:
diag, diagonal, diagflat

Examples

>>> np.trace(np.eye(3))
3.0
>>> a = np.arange(8).reshape((2,2,2))
>>> np.trace(a)
array([6, 8])

>>> a = np.arange(24).reshape((2,2,2,3))
>>> np.trace(a).shape
(2, 3)

3.7.5 Solving equations and inverting matrices

linalg.solve(a, b) Solve a linear matrix equation, or system of linear scalar equations.

linalg.tensorsolve(a, b[, axes]) Solve the tensor equation a x = b for x.
linalg.lstsq(a, b[, rcond]) Return the least-squares solution to a linear matrix equation.
linalg.inv(a) Compute the (multiplicative) inverse of a matrix.
linalg.pinv(a[, rcond]) Compute the (Moore-Penrose) pseudo-inverse of a matrix.
linalg.tensorinv(a[, ind]) Compute the inverse of an N-dimensional array.
solve(a, b)
Solve a linear matrix equation, or system of linear scalar equations.
Computes the exact solution, x, of the well-determined, i.e., full rank, linear matrix equation ax = b.
Parameters
a : array_like, shape (M, M)
Coefficient matrix.
b : array_like, shape (M,) or (M, N)
Ordinate or dependent variable values.
Returns
x : ndarray, shape (M,) or (M, N) depending on b
Solution to the system a x = b
Raises
LinAlgError :
If a is singular or not square.

Notes
solve is a wrapper for the LAPACK routines dgesv and zgesv, the former being used if a is real-valued, the latter
if it is complex-valued. The solution to the system of linear equations is computed using an LU decomposition
[R50] with partial pivoting and row interchanges.
a must be square and of full-rank, i.e., all rows (or, equivalently, columns) must be linearly independent; if either
is not true, use lstsq for the least-squares best solution of the system/equation.

3.7. Linear algebra (numpy.linalg) 559

NumPy Reference, Release 2.0.0.dev8464

References
[R50]

Examples
Solve the system of equations 3 * x0 + x1 = 9 and x0 + 2 * x1 = 8:

>>> a = np.array([[3,1], [1,2]])

>>> b = np.array([9,8])
>>> x = np.linalg.solve(a, b)
>>> x
array([ 2., 3.])

Check that the solution is correct:

>>> (np.dot(a, x) == b).all()

True

tensorsolve(a, b, axes=None)
Solve the tensor equation a x = b for x.
It is assumed that all indices of x are summed over in the product, together with the rightmost indices of a, as is
done in, for example, tensordot(a, x, axes=len(b.shape)).
Parameters
a : array_like
Coefficient tensor, of shape b.shape + Q. Q, a tuple, equals the shape of that
sub-tensor of a consisting of the appropriate number of its rightmost indices, and
must be such that
prod(Q) == prod(b.shape) (in which sense a is said to be square).
b : array_like
Right-hand tensor, which can be of any shape.
axes : tuple of ints, optional
Axes in a to reorder to the right, before inversion. If None (default), no reordering is
done.
Returns
x : ndarray, shape Q
Raises
LinAlgError :
If a is singular or not square (in the above sense).
See Also:
tensordot, tensorinv

Examples

>>> a = np.eye(2*3*4)
>>> a.shape = (2*3, 4, 2, 3, 4)
>>> b = np.random.randn(2*3, 4)
>>> x = np.linalg.tensorsolve(a, b)
>>> x.shape

560 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

(2, 3, 4)
>>> np.allclose(np.tensordot(a, x, axes=3), b)
True

lstsq(a, b, rcond=-1)
Return the least-squares solution to a linear matrix equation.
Solves the equation a x = b by computing a vector x that minimizes the norm || b - a x ||. The equation may
be under-, well-, or over- determined (i.e., the number of linearly independent rows of a can be less than, equal
to, or greater than its number of linearly independent columns). If a is square and of full rank, then x (but for
round-off error) is the exact solution of the equation.
Parameters
a : array_like, shape (M, N)
Coefficient matrix.
b : array_like, shape (M,) or (M, K)
Ordinate or dependent variable values. If b is two-dimensional, the least-squares
solution is calculated for each of the K columns of b.
rcond : float, optional
Cut-off ratio for small singular values of a. Singular values are set to zero if they are
smaller than rcond times the largest singular value of a.
Returns
x : ndarray, shape (N,) or (N, K)
Least-squares solution. The shape of x depends on the shape of b.
residues : ndarray, shape (), (1,), or (K,)
Sums of residues; squared Euclidean norm for each column in b - a*x. If the rank of
a is < N or > M, this is an empty array. If b is 1-dimensional, this is a (1,) shape array.
Otherwise the shape is (K,).
rank : int
Rank of matrix a.
s : ndarray, shape (min(M,N),)
Singular values of a.
Raises
LinAlgError :
If computation does not converge.

Notes
If b is a matrix, then all array results are returned as matrices.

Examples
Fit a line, y = mx + c, through some noisy data-points:

>>> x = np.array([0, 1, 2, 3])

>>> y = np.array([-1, 0.2, 0.9, 2.1])

3.7. Linear algebra (numpy.linalg) 561

NumPy Reference, Release 2.0.0.dev8464

By examining the coefficients, we see that the line should have a gradient of roughly 1 and cut the y-axis at,
more or less, -1.
We can rewrite the line equation as y = Ap, where A = [[x 1]] and p = [[m], [c]]. Now use lstsq
to solve for p:

>>> A = np.vstack([x, np.ones(len(x))]).T

>>> A
array([[ 0., 1.],
[ 1., 1.],
[ 2., 1.],
[ 3., 1.]])

>>> m, c = np.linalg.lstsq(A, y)[0]

>>> print m, c
1.0 -0.95

Plot the data along with the fitted line:

>>> import matplotlib.pyplot as plt

>>> plt.plot(x, y, o, label=Original data, markersize=10)
>>> plt.plot(x, m*x + c, r, label=Fitted line)
>>> plt.legend()
>>> plt.show()

2.5
Original data
2.0 Fitted line
1.5
1.0
0.5
0.0
0.5
1.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0

inv(a)
Compute the (multiplicative) inverse of a matrix.
Given a square matrix a, return the matrix ainv satisfying dot(a, ainv) = dot(ainv, a) =
eye(a.shape[0]).
Parameters
a : array_like, shape (M, M)
Matrix to be inverted.
Returns
ainv : ndarray or matrix, shape (M, M)
(Multiplicative) inverse of the matrix a.

562 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Raises
LinAlgError :
If a is singular or not square.

Examples

>>> from numpy import linalg as LA

>>> a = np.array([[1., 2.], [3., 4.]])
>>> ainv = LA.inv(a)
>>> np.allclose(np.dot(a, ainv), np.eye(2))
True
>>> np.allclose(np.dot(ainv, a), np.eye(2))
True

If a is a matrix object, then the return value is a matrix as well:

>>> ainv = LA.inv(np.matrix(a))

>>> ainv
matrix([[-2. , 1. ],
[ 1.5, -0.5]])

pinv(a, rcond=1.0000000000000001e-15)
Compute the (Moore-Penrose) pseudo-inverse of a matrix.
Calculate the generalized inverse of a matrix using its singular-value decomposition (SVD) and including all
large singular values.
Parameters
a : array_like, shape (M, N)
Matrix to be pseudo-inverted.
rcond : float
Cutoff for small singular values. Singular values smaller (in modulus) than rcond *
largest_singular_value (again, in modulus) are set to zero.
Returns
B : ndarray, shape (N, M)
The pseudo-inverse of a. If a is a matrix instance, then so is B.
Raises
LinAlgError :
If the SVD computation does not converge.

Notes
The pseudo-inverse of a matrix A, denoted A+ , is defined as: the matrix that solves [the least-squares prob-
lem] Ax = b, i.e., if x is said solution, then A+ is that matrix such that x = A+ b.
It can be shown that if Q1 QT2 = A is the singular value decomposition of A, then A+ = Q2 + QT1 , where
Q1,2 are orthogonal matrices, is a diagonal matrix consisting of As so-called singular values, (followed,
typically, by zeros), and then + is simply the diagonal matrix consisting of the reciprocals of As singular
values (again, followed by zeros). [R49]

References
[R49]

3.7. Linear algebra (numpy.linalg) 563

NumPy Reference, Release 2.0.0.dev8464

Examples
The following example checks that a * a+ * a == a and a+ * a * a+ == a+:

>>> a = np.random.randn(9, 6)
>>> B = np.linalg.pinv(a)
>>> np.allclose(a, np.dot(a, np.dot(B, a)))
True
>>> np.allclose(B, np.dot(B, np.dot(a, B)))
True

tensorinv(a, ind=2)
Compute the inverse of an N-dimensional array.
The result is an inverse for a relative to the tensordot operation tensordot(a, b, ind), i. e., up to
floating-point accuracy, tensordot(tensorinv(a), a, ind) is the identity tensor for the tensordot
operation.
Parameters
a : array_like
Tensor to invert. Its shape must be square, i. e., prod(a.shape[:ind]) ==
prod(a.shape[ind:]).
ind : int, optional
Number of first indices that are involved in the inverse sum. Must be a positive integer,
default is 2.
Returns
b : ndarray
as tensordot inverse, shape a.shape[:ind] + a.shape[ind:].
Raises
LinAlgError :
If a is singular or not square (in the above sense).
See Also:
tensordot, tensorsolve

Examples

>>> a = np.eye(4*6)
>>> a.shape = (4, 6, 8, 3)
>>> ainv = np.linalg.tensorinv(a, ind=2)
>>> ainv.shape
(8, 3, 4, 6)
>>> b = np.random.randn(4, 6)
>>> np.allclose(np.tensordot(ainv, b), np.linalg.tensorsolve(a, b))
True

>>> a = np.eye(4*6)
>>> a.shape = (24, 8, 3)
>>> ainv = np.linalg.tensorinv(a, ind=1)
>>> ainv.shape
(8, 3, 24)
>>> b = np.random.randn(24)

564 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.allclose(np.tensordot(ainv, b, 1), np.linalg.tensorsolve(a, b))

True

3.7.6 Exceptions

linalg.LinAlgError Generic Python-exception-derived object raised by linalg functions.

exception LinAlgError
Generic Python-exception-derived object raised by linalg functions.
General purpose exception class, derived from Pythons exception.Exception class, programmatically raised in
linalg functions when a Linear Algebra-related condition would prevent further correct execution of the function.
Parameters
None :

Examples

>>> from numpy import linalg as LA

>>> LA.inv(np.zeros((2,2)))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "...linalg.py", line 350,
in inv return wrap(solve(a, identity(a.shape[0], dtype=a.dtype)))
File "...linalg.py", line 249,
in solve
raise LinAlgError, Singular matrix
numpy.linalg.linalg.LinAlgError: Singular matrix

3.8 Random sampling (numpy.random)

3.8.1 Simple random data

rand(d0, d1, dn) Random values in a given shape.

randn(d1) Return a sample (or samples) from the standard normal distribution.
randint(low[, high, size]) Return random integers from low (inclusive) to high (exclusive).
random_integers(low[, high, size]) Return random integers between low and high, inclusive.
random_sample([size]) Return random floats in the half-open interval [0.0, 1.0).
bytes(length) Return random bytes.
rand(d0, d1, ..., dn)
Random values in a given shape.
Create an array of the given shape and propagate it with random samples from a uniform distribution over [0,
1).
Parameters
d0, d1, ..., dn : int
Shape of the output.
Returns
out : ndarray, shape (d0, d1, ..., dn)
Random values.

3.8. Random sampling (numpy.random) 565

NumPy Reference, Release 2.0.0.dev8464

See Also:
random

Notes
This is a convenience function. If you want an interface that takes a shape-tuple as the first argument, refer to
random.

Examples

>>> np.random.rand(3,2)
array([[ 0.14022471, 0.96360618], #random
[ 0.37601032, 0.25528411], #random
[ 0.49313049, 0.94909878]]) #random

randn([d1, ..., dn])

Return a sample (or samples) from the standard normal distribution.
If positive, int_like or int-convertible arguments are provided, randn generates an array of shape (d1, ...,
dn), filled with random floats sampled from a univariate normal (Gaussian) distribution of mean 0 and vari-
ance 1 (if any of the di are floats, they are first converted to integers by truncation). A single float randomly
sampled from the distribution is returned if no argument is provided.
This is a convenience function. If you want an interface that takes a tuple as the first argument, use
numpy.random.standard_normal instead.
Parameters
d1, ..., dn : n ints, optional
The dimensions of the returned array, should be all positive.
Returns
Z : ndarray or float
A (d1, ..., dn)-shaped array of floating-point samples from the standard normal
distribution, or a single such float if no parameters were supplied.
See Also:

random.standard_normal
Similar, but takes a tuple as its argument.

Notes
For random samples from N (, 2 ), use:
sigma * np.random.randn(...) + mu

Examples

>>> np.random.randn()
2.1923875335537315 #random

Two-by-four array of samples from N(3, 6.25):

>>> 2.5 * np.random.randn(2, 4) + 3

array([[-4.49401501, 4.00950034, -1.81814867, 7.29718677], #random
[ 0.39924804, 4.68456316, 4.99394529, 4.84057254]]) #random

566 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

randint(low, high=None, size=None)

Return random integers from low (inclusive) to high (exclusive).
Return random integers from the discrete uniform distribution in the half-open interval [low, high). If high
is None (the default), then results are from [0, low).
Parameters
low : int
Lowest (signed) integer to be drawn from the distribution (unless high=None, in
which case this parameter is the highest such integer).
high : int, optional
If provided, one above the largest (signed) integer to be drawn from the distribution (see
above for behavior if high=None).
size : int or tuple of ints, optional
Output shape. Default is None, in which case a single int is returned.
Returns
out : int or ndarray of ints
size-shaped array of random integers from the appropriate distribution, or a single such
random int if size not provided.
See Also:

random.random_integers
similar to randint, only for the closed interval [low, high], and 1 is the lowest value if high is omitted. In
particular, this other one is the one to use to generate uniformly distributed discrete non-integers.

Examples

>>> np.random.randint(2, size=10)

array([1, 0, 0, 0, 1, 1, 0, 0, 1, 0])
>>> np.random.randint(1, size=10)
array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0])

Generate a 2 x 4 array of ints between 0 and 4, inclusive:

>>> np.random.randint(5, size=(2, 4))

array([[4, 0, 2, 1],
[3, 2, 2, 0]])

random_integers(low, high=None, size=None)

Return random integers between low and high, inclusive.
Return random integers from the discrete uniform distribution in the closed interval [low, high]. If high is
None (the default), then results are from [1, low].
Parameters
low : int
Lowest (signed) integer to be drawn from the distribution (unless high=None, in
which case this parameter is the highest such integer).
high : int, optional
If provided, the largest (signed) integer to be drawn from the distribution (see above for
behavior if high=None).

3.8. Random sampling (numpy.random) 567

NumPy Reference, Release 2.0.0.dev8464

size : int or tuple of ints, optional

Output shape. Default is None, in which case a single int is returned.
Returns
out : int or ndarray of ints
size-shaped array of random integers from the appropriate distribution, or a single such
random int if size not provided.
See Also:

random.randint
Similar to random_integers, only for the half-open interval [low, high), and 0 is the lowest value if high is
omitted.

Notes
To sample from N evenly spaced floating-point numbers between a and b, use:

a + (b - a) * (np.random.random_integers(N) - 1) / (N - 1.)

Examples

>>> np.random.random_integers(5)
4
>>> type(np.random.random_integers(5))
<type int>
>>> np.random.random_integers(5, size=(3.,2.))
array([[5, 4],
[3, 3],
[4, 5]])

Choose five random numbers from the set of five evenly-spaced numbers between 0 and 2.5, inclusive (i.e., from
the set 0, 5/8, 10/8, 15/8, 20/8):

>>> 2.5 * (np.random.random_integers(5, size=(5,)) - 1) / 4.

array([ 0.625, 1.25 , 0.625, 0.625, 2.5 ])

Roll two six sided dice 1000 times and sum the results:

>>> d1 = np.random.random_integers(1, 6, 1000)

>>> d2 = np.random.random_integers(1, 6, 1000)
>>> dsums = d1 + d2

Display results as a histogram:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(dsums, 11, normed=True)
>>> plt.show()

568 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

0.20

0.15

0.10

0.05

0.00
2 4 6 8 10 12

random_sample(size=None)
Return random floats in the half-open interval [0.0, 1.0).
Results are from the continuous uniform distribution over the stated interval. To sample U nif [a, b), b > a
multiply the output of random_sample by (b-a) and add a:

(b - a) * random_sample() + a

Parameters
size : int or tuple of ints, optional
Defines the shape of the returned array of random floats. If None (the default), returns
a single float.
Returns
out : float or ndarray of floats
Array of random floats of shape size (unless size=None, in which case a single float
is returned).

Examples

>>> np.random.random_sample()
0.47108547995356098
>>> type(np.random.random_sample())
<type float>
>>> np.random.random_sample((5,))
array([ 0.30220482, 0.86820401, 0.1654503 , 0.11659149, 0.54323428])

Three-by-two array of random numbers from [-5, 0):

>>> 5 * np.random.random_sample((3, 2)) - 5

array([[-3.99149989, -0.52338984],
[-2.99091858, -0.79479508],
[-1.23204345, -1.75224494]])

bytes(length)
Return random bytes.

3.8. Random sampling (numpy.random) 569

NumPy Reference, Release 2.0.0.dev8464

Parameters
length : int
Number of random bytes.
Returns
out : str
String of length N.

Examples

>>> np.random.bytes(10)
eh\x85\x022SZ\xbf\xa4 #random

3.8.2 Permutations

shuffle(x) Modify a sequence in-place by shuffling its contents.

permutation(x) Randomly permute a sequence, or return a permuted range.
shuffle(x)
Modify a sequence in-place by shuffling its contents.
permutation(x)
Randomly permute a sequence, or return a permuted range.
Parameters
x : int or array_like
If x is an integer, randomly permute np.arange(x). If x is an array, make a copy
and shuffle the elements randomly.
Returns
out : ndarray
Permuted sequence or array range.

Examples

>>> np.random.permutation(10)
array([1, 7, 4, 3, 0, 9, 2, 5, 8, 6])

>>> np.random.permutation([1, 4, 9, 12, 15])

array([15, 1, 9, 4, 12])

3.8.3 Distributions

beta(a, b[, size]) The Beta distribution over [0, 1].

binomial(n, p[, size]) Draw samples from a binomial distribution.
chisquare(df[, size]) Draw samples from a chi-square distribution.
mtrand.dirichlet(alpha[, size]) Draw samples from the Dirichlet distribution.
exponential([scale, size]) Exponential distribution.
f(dfnum, dfden[, size]) Draw samples from a F distribution.
gamma(shape[, scale, size]) Draw samples from a Gamma distribution.
Continued on next page

570 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Table 3.1 continued from previous page

geometric(p[, size]) Draw samples from the geometric distribution.
gumbel([loc, scale, size]) Gumbel distribution.
hypergeometric(ngood, nbad, nsample[, size]) Draw samples from a Hypergeometric distribution.
laplace([loc, scale, size]) Draw samples from the Laplace or double exponential distribution with
logistic([loc, scale, size]) Draw samples from a Logistic distribution.
lognormal([mean, sigma, size]) Return samples drawn from a log-normal distribution.
logseries(p[, size]) Draw samples from a Logarithmic Series distribution.
multinomial(n, pvals[, size]) Draw samples from a multinomial distribution.
multivariate_normal(mean) Draw random samples from a multivariate normal distribution.
negative_binomial(n, p[, size]) Draw samples from a negative_binomial distribution.
noncentral_chisquare(df, nonc[, size]) Draw samples from a noncentral chi-square distribution.
noncentral_f(dfnum, dfden, nonc[, size]) Draw samples from the noncentral F distribution.
normal([loc, scale, size]) Draw random samples from a normal (Gaussian) distribution.
pareto(a[, size]) Draw samples from a Pareto distribution with specified shape.
poisson([lam, size]) Draw samples from a Poisson distribution.
power(a[, size]) Draws samples in [0, 1] from a power distribution with positive exponent a - 1
rayleigh([scale, size]) Draw samples from a Rayleigh distribution.
standard_cauchy([size]) Standard Cauchy distribution with mode = 0.
standard_exponential([size]) Draw samples from the standard exponential distribution.
standard_gamma(shape[, size]) Draw samples from a Standard Gamma distribution.
standard_normal([size]) Returns samples from a Standard Normal distribution (mean=0, stdev=1).
standard_t(df[, size]) Standard Students t distribution with df degrees of freedom.
triangular(left, mode, right[, size]) Draw samples from the triangular distribution.
uniform([low, high, size]) Draw samples from a uniform distribution.
vonmises([mu, kappa, size]) Draw samples from a von Mises distribution.
wald(mean, scale[, size]) Draw samples from a Wald, or Inverse Gaussian, distribution.
weibull(a[, size]) Weibull distribution.
zipf(a[, size]) Draw samples from a Zipf distribution.

beta(a, b, size=None)
The Beta distribution over [0, 1].
The Beta distribution is a special case of the Dirichlet distribution, and is related to the Gamma distribution. It
has the probability distribution function
1
f (x; a, b) = x1 (1 x)1 ,
B(, )

where the normalisation, B, is the beta function,

Z 1
B(, ) = t1 (1 t)1 dt.
0

It is often seen in Bayesian inference and order statistics.

Parameters
a : float
Alpha, non-negative.
b : float
Beta, non-negative.

3.8. Random sampling (numpy.random) 571

NumPy Reference, Release 2.0.0.dev8464

size : tuple of ints, optional

The number of samples to draw. The ouput is packed according to the size given.
Returns
out : ndarray
Array of the given shape, containing values drawn from a Beta distribution.
binomial(n, p, size=None)
Draw samples from a binomial distribution.
Samples are drawn from a Binomial distribution with specified parameters, n trials and p probability of success
where n an integer > 0 and p is in the interval [0,1]. (n may be input as a float, but it is truncated to an integer in
use)
Parameters
n : float (but truncated to an integer)
parameter, > 0.
p : float
parameter, >= 0 and <=1.
size : {tuple, int}
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : {ndarray, scalar}
where the values are all integers in [0, n].
See Also:

scipy.stats.distributions.binom
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Binomial distribution is
 
n N
P (N ) = p (1 p)nN ,
N

where n is the number of trials, p is the probability of success, and N is the number of successes.
When estimating the standard error of a proportion in a population by using a random sample, the normal
distribution works well unless the product p*n <=5, where p = population proportion estimate, and n = number
of samples, in which case the binomial distribution is used instead. For example, a sample of 15 people shows
4 who are left handed, and 11 who are right handed. Then p = 4/15 = 27%. 0.27*15 = 4, so the binomial
distribution should be used in this case.

References
[R69], [R70], [R71], [R72], [R73]

572 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples
Draw samples from the distribution:

>>> n, p = 10, .5 # number of trials, probability of each trial

>>> s = np.random.binomial(n, p, 1000)
# result of flipping a coin 10 times, tested 1000 times.

A real world example. A company drills 9 wild-cat oil exploration wells, each with an estimated probability of
success of 0.1. All nine wells fail. What is the probability of that happening?
Lets do 20,000 trials of the model, and count the number that generate zero positive results.

>>> sum(np.random.binomial(9,0.1,20000)==0)/20000.
answer = 0.38885, or 38%.

chisquare(df, size=None)
Draw samples from a chi-square distribution.
When df independent random variables, each with standard normal distributions (mean 0, variance 1), are
squared and summed, the resulting distribution is chi-square (see Notes). This distribution is often used in
hypothesis testing.
Parameters
df : int
Number of degrees of freedom.
size : tuple of ints, int, optional
Size of the returned array. By default, a scalar is returned.
Returns
output : ndarray
Samples drawn from the distribution, packed in a size-shaped array.
Raises
ValueError :
When df <= 0 or when an inappropriate size (e.g. size=-1) is given.

Notes
The variable obtained by summing the squares of df independent, standard normally distributed random vari-
ables:
df
X
Q= Xi2
i=0

is chi-square distributed, denoted

Q 2k .

The probability density function of the chi-squared distribution is

(1/2)k/2 k/21 x/2
p(x) = x e ,
(k/2)

3.8. Random sampling (numpy.random) 573

NumPy Reference, Release 2.0.0.dev8464

where is the gamma function,

Z
(x) = tx1 et dt.
0

References
[R74], [R75]

Examples

>>> np.random.chisquare(2,4)
array([ 1.89920014, 9.00867716, 3.13710533, 5.62318272])

dirichlet(alpha, size=None)
Draw samples from the Dirichlet distribution.
Draw size samples of dimension k from a Dirichlet distribution. A Dirichlet-distributed random variable can be
seen as a multivariate generalization of a Beta distribution. Dirichlet pdf is the conjugate prior of a multinomial
in Bayesian inference.
Parameters
alpha : array
Parameter of the distribution (k dimension for sample of dimension k).
size : array
Number of samples to draw.

Notes

k
Y
i 1
X x
i
i=1

Uses the following property for computation: for each dimension, draw a random sample y_i from a standard
gamma generator of shape alpha_i, then X = Pk1 y (y1 , . . . , yn ) is Dirichlet distributed.
i=1 i

References
[R235]
exponential(scale=1.0, size=None)
Exponential distribution.
Its probability density function is
1 1 x
f (x; ) = exp( ),

for x > 0 and 0 elsewhere. is the scale parameter, which is the inverse of the rate parameter = 1/. The
rate parameter is an alternative, widely used parameterization of the exponential distribution [R78].
The exponential distribution is a continuous analogue of the geometric distribution. It describes many common
situations, such as the size of raindrops measured over many rainstorms [R76], or the time between page requests
to Wikipedia [R77].

574 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
scale : float
The scale parameter, = 1/.
size : tuple of ints
Number of samples to draw. The output is shaped according to size.

References
[R76], [R77], [R78]
f(dfnum, dfden, size=None)
Draw samples from a F distribution.
Samples are drawn from an F distribution with specified parameters, dfnum (degrees of freedom in numerator)
and dfden (degrees of freedom in denominator), where both parameters should be greater than zero.
The random variate of the F distribution (also known as the Fisher distribution) is a continuous probability
distribution that arises in ANOVA tests, and is the ratio of two chi-square variates.
Parameters
dfnum : float
Degrees of freedom in numerator. Should be greater than zero.
dfden : float
Degrees of freedom in denominator. Should be greater than zero.
size : {tuple, int}, optional
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn. By default only one sample is returned.
Returns
samples : {ndarray, scalar}
Samples from the Fisher distribution.
See Also:

scipy.stats.distributions.f
probability density function, distribution or cumulative density function, etc.

Notes
The F statistic is used to compare in-group variances to between-group variances. Calculating the distribution
depends on the sampling, and so it is a function of the respective degrees of freedom in the problem. The
variable dfnum is the number of samples minus one, the between-groups degrees of freedom, while dfden is the
within-groups degrees of freedom, the sum of the number of samples in each group minus the number of groups.

References
[R79], [R80]

Examples
An example from Glantz[1], pp 47-40. Two groups, children of diabetics (25 people) and children from people
without diabetes (25 controls). Fasting blood glucose was measured, case group had a mean value of 86.1,
controls had a mean value of 82.2. Standard deviations were 2.09 and 2.49 respectively. Are these data consistent
with the null hypothesis that the parents diabetic status does not affect their childrens blood glucose levels?
Calculating the F statistic from the data gives a value of 36.01.

3.8. Random sampling (numpy.random) 575

NumPy Reference, Release 2.0.0.dev8464

Draw samples from the distribution:

>>> dfnum = 1. # between group degrees of freedom

>>> dfden = 48. # within groups degrees of freedom
>>> s = np.random.f(dfnum, dfden, 1000)

The lower bound for the top 1% of the samples is :

>>> sort(s)[-10]
7.61988120985

So there is about a 1% chance that the F statistic will exceed 7.62, the measured value is 36, so the null hypothesis
is rejected at the 1% level.
gamma(shape, scale=1.0, size=None)
Draw samples from a Gamma distribution.
Samples are drawn from a Gamma distribution with specified parameters, shape (sometimes designated k)
and scale (sometimes designated theta), where both parameters are > 0.
Parameters
shape : scalar > 0
The shape of the gamma distribution.
scale : scalar > 0, optional
The scale of the gamma distribution. Default is equal to 1.
size : shape_tuple, optional
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
out : ndarray, float
Returns one sample unless size parameter is specified.
See Also:

scipy.stats.distributions.gamma
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Gamma distribution is

ex/
p(x) = xk1 ,
k (k)

where k is the shape and the scale, and is the Gamma function.
The Gamma distribution is often used to model the times to failure of electronic components, and arises naturally
in processes for which the waiting times between Poisson distributed events are relevant.

References
[R81], [R82]

576 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples
geometric(p, size=None)
Draw samples from the geometric distribution.
Bernoulli trials are experiments with one of two outcomes: success or failure (an example of such an experiment
is flipping a coin). The geometric distribution models the number of trials that must be run in order to achieve
success. It is therefore supported on the positive integers, k = 1, 2, ....
The probability mass function of the geometric distribution is

f (k) = (1 p)k1 p

where p is the probability of success of an individual trial.

Parameters
p : float
The probability of success of an individual trial.
size : tuple of ints
Number of values to draw from the distribution. The output is shaped according to size.
Returns
out : ndarray
Samples from the geometric distribution, shaped according to size.

Examples
Draw ten thousand values from the geometric distribution, with the probability of an individual success equal to
0.35:

>>> z = np.random.geometric(p=0.35, size=10000)

How many trials succeeded after a single run?

>>> (z == 1).sum() / 10000.

0.34889999999999999 #random

gumbel(loc=0.0, scale=1.0, size=None)

Gumbel distribution.
Draw samples from a Gumbel distribution with specified location (or mean) and scale (or standard deviation).
The Gumbel (or Smallest Extreme Value (SEV) or the Smallest Extreme Value Type I) distribution is one of a
class of Generalized Extreme Value (GEV) distributions used in modeling extreme value problems. The Gumbel
is a special case of the Extreme Value Type I distribution for maximums from distributions with exponential-
like tails, it may be derived by considering a Gaussian process of measurements, and generating the pdf for the
maximum values from that set of measurements (see examples).
Parameters
loc : float
The location of the mode of the distribution.
scale : float
The scale parameter of the distribution.
size : tuple of ints

3.8. Random sampling (numpy.random) 577

NumPy Reference, Release 2.0.0.dev8464

Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
See Also:

scipy.stats.gumbel
probability density function, distribution or cumulative density function, etc.

weibull, scipy.stats.genextreme

Notes
The probability density for the Gumbel distribution is

e(x)/ e(x)/
p(x) = e ,

where is the mode, a location parameter, and is the scale parameter.

The Gumbel (named for German mathematician Emil Julius Gumbel) was used very early in the hydrology
literature, for modeling the occurrence of flood events. It is also used for modeling maximum wind speed and
rainfall rates. It is a fat-tailed distribution - the probability of an event in the tail of the distribution is larger
than if one used a Gaussian, hence the surprisingly frequent occurrence of 100-year floods. Floods were initially
modeled as a Gaussian process, which underestimated the frequency of extreme events.
It is one of a class of extreme value distributions, the Generalized Extreme Value (GEV) distributions, which
also includes the Weibull and Frechet.
2 2
The function has a mean of + 0.57721 and a variance of 6 .

References
[R83], [R84], [R85]

Examples
Draw samples from the distribution:

>>> mu, beta = 0, 0.1 # location and scale

>>> s = np.random.gumbel(mu, beta, 1000)

Display the histogram of the samples, along with the probability density function:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 30, normed=True)
>>> plt.plot(bins, (1/beta)*np.exp(-(bins - mu)/beta)
... * np.exp( -np.exp( -(bins - mu) /beta) ),
... linewidth=2, color=r)
>>> plt.show()

Show how an extreme value distribution can arise from a Gaussian process and compare to a Gaussian:

>>> means = []
>>> maxima = []
>>> for i in range(0,1000) :
... a = np.random.normal(mu, beta, 1000)
... means.append(a.mean())
... maxima.append(a.max())

578 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> count, bins, ignored = plt.hist(maxima, 30, normed=True)

>>> beta = np.std(maxima)*np.pi/np.sqrt(6)
>>> mu = np.mean(maxima) - 0.57721*beta
>>> plt.plot(bins, (1/beta)*np.exp(-(bins - mu)/beta)
... * np.exp(-np.exp(-(bins - mu)/beta)),
... linewidth=2, color=r)
>>> plt.plot(bins, 1/(beta * np.sqrt(2 * np.pi))
... * np.exp(-(bins - mu)**2 / (2 * beta**2)),
... linewidth=2, color=g)
>>> plt.show()

14
12
10
8
6
4
2
0
0.4 0.2 0.0 0.2 0.4 0.6 0.8

hypergeometric(ngood, nbad, nsample, size=None)

Draw samples from a Hypergeometric distribution.
Samples are drawn from a Hypergeometric distribution with specified parameters, ngood (ways to make a good
selection), nbad (ways to make a bad selection), and nsample = number of items sampled, which is less than or
equal to the sum ngood + nbad.
Parameters
ngood : float (but truncated to an integer)
parameter, > 0.
nbad : float
parameter, >= 0.
nsample : float
parameter, > 0 and <= ngood+nbad
size : {tuple, int}
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : {ndarray, scalar}
where the values are all integers in [0, n].
See Also:

3.8. Random sampling (numpy.random) 579

NumPy Reference, Release 2.0.0.dev8464

scipy.stats.distributions.hypergeom
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Hypergeometric distribution is
m N m



n nx
P (x) = N

,
n

where 0 x m and n + m N x n
for P(x) the probability of x successes, n = ngood, m = nbad, and N = number of samples.
Consider an urn with black and white marbles in it, ngood of them black and nbad are white. If you draw
nsample balls without replacement, then the Hypergeometric distribution describes the distribution of black
balls in the drawn sample.
Note that this distribution is very similar to the Binomial distribution, except that in this case, samples are drawn
without replacement, whereas in the Binomial case samples are drawn with replacement (or the sample space is
infinite). As the sample space becomes large, this distribution approaches the Binomial.

References
[R86], [R87], [R88]

Examples
Draw samples from the distribution:

>>> ngood, nbad, nsamp = 100, 2, 10

# number of good, number of bad, and number of samples
>>> s = np.random.hypergeometric(ngood, nbad, nsamp, 1000)
>>> hist(s)
# note that it is very unlikely to grab both bad items

Suppose you have an urn with 15 white and 15 black marbles. If you pull 15 marbles at random, how likely is
it that 12 or more of them are one color?

>>> s = np.random.hypergeometric(15, 15, 15, 100000)

>>> sum(s>=12)/100000. + sum(s<=3)/100000.
# answer = 0.003 ... pretty unlikely!

laplace(loc=0.0, scale=1.0, size=None)

Draw samples from the Laplace or double exponential distribution with specified location (or mean) and scale
(decay).
The Laplace distribution is similar to the Gaussian/normal distribution, but is sharper at the peak and has fatter
tails. It represents the difference between two independent, identically distributed exponential random variables.
Parameters
loc : float
The position, , of the distribution peak.
scale : float
, the exponential decay.

580 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notes
It has the probability density function
 
1 |x |
f (x; , ) = exp .
2

The first law of Laplace, from 1774, states that the frequency of an error can be expressed as an exponential
function of the absolute magnitude of the error, which leads to the Laplace distribution. For many problems
in Economics and Health sciences, this distribution seems to model the data better than the standard Gaussian
distribution

References
[R89], [R90], [R91], [R92]

Examples
Draw samples from the distribution

>>> loc, scale = 0., 1.

>>> s = np.random.laplace(loc, scale, 1000)

Display the histogram of the samples, along with the probability density function:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 30, normed=True)
>>> x = np.arange(-8., 8., .01)
>>> pdf = np.exp(-abs(x-loc/scale))/(2.*scale)
>>> plt.plot(x, pdf)

Plot Gaussian for comparison:

>>> g = (1/(scale * np.sqrt(2 * np.pi)) *

... np.exp( - (x - loc)**2 / (2 * scale**2) ))
>>> plt.plot(x,g)

0.5

0.4

0.3

0.2

0.1

0.0
8 6 4 2 0 2 4 6 8

3.8. Random sampling (numpy.random) 581

NumPy Reference, Release 2.0.0.dev8464

logistic(loc=0.0, scale=1.0, size=None)

Draw samples from a Logistic distribution.
Samples are drawn from a Logistic distribution with specified parameters, loc (location or mean, also median),
and scale (>0).
Parameters
loc : float
scale : float > 0.
size : {tuple, int}
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : {ndarray, scalar}
where the values are all integers in [0, n].
See Also:

scipy.stats.distributions.logistic
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Logistic distribution is

e(x)/s
P (x) = P (x) = ,
s(1 + e(x)/s )2

where = location and s = scale.

The Logistic distribution is used in Extreme Value problems where it can act as a mixture of Gumbel distribu-
tions, in Epidemiology, and by the World Chess Federation (FIDE) where it is used in the Elo ranking system,
assuming the performance of each player is a logistically distributed random variable.

References
[R93], [R94], [R95]

Examples
Draw samples from the distribution:

>>> loc, scale = 10, 1

>>> s = np.random.logistic(loc, scale, 10000)
>>> count, bins, ignored = plt.hist(s, bins=50)

# plot against distribution

>>> def logist(x, loc, scale):

... return exp((loc-x)/scale)/(scale*(1+exp((loc-x)/scale))**2)
>>> plt.plot(bins, logist(bins, loc, scale)*count.max()/\
... logist(bins, loc, scale).max())
>>> plt.show()

582 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

lognormal(mean=0.0, sigma=1.0, size=None)

Return samples drawn from a log-normal distribution.
Draw samples from a log-normal distribution with specified mean, standard deviation, and shape. Note that
the mean and standard deviation are not the values for the distribution itself, but of the underlying normal
distribution it is derived from.
Parameters
mean : float
Mean value of the underlying normal distribution
sigma : float, >0.
Standard deviation of the underlying normal distribution
size : tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
See Also:

scipy.stats.lognorm
probability density function, distribution, cumulative density function, etc.

Notes
A variable x has a log-normal distribution if log(x) is normally distributed.
The probability density function for the log-normal distribution is
1 (ln(x))2
p(x) = e( 22 )
x 2

where is the mean and is the standard deviation of the normally distributed logarithm of the variable.
A log-normal distribution results if a random variable is the product of a large number of independent,
identically-distributed variables in the same way that a normal distribution results if the variable is the sum of a
large number of independent, identically-distributed variables (see the last example). It is one of the so-called
fat-tailed distributions.
The log-normal distribution is commonly used to model the lifespan of units with fatigue-stress failure modes.
Since this includes most mechanical systems, the log-normal distribution has widespread application.
It is also commonly used to model oil field sizes, species abundance, and latent periods of infectious diseases.

References
[R96], [R97], [R98]

Examples
Draw samples from the distribution:

>>> mu, sigma = 3., 1. # mean and standard deviation

>>> s = np.random.lognormal(mu, sigma, 1000)

Display the histogram of the samples, along with the probability density function:

3.8. Random sampling (numpy.random) 583

NumPy Reference, Release 2.0.0.dev8464

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 100, normed=True, align=mid)

>>> x = np.linspace(min(bins), max(bins), 10000)

>>> pdf = (np.exp(-(np.log(x) - mu)**2 / (2 * sigma**2))
... / (x * sigma * np.sqrt(2 * np.pi)))

>>> plt.plot(x, pdf, linewidth=2, color=r)

>>> plt.axis(tight)
>>> plt.show()

Demonstrate that taking the products of random samples from a uniform distribution can be fit well by a log-
normal probability density function.

>>> # Generate a thousand samples: each is the product of 100 random

>>> # values, drawn from a normal distribution.
>>> b = []
>>> for i in range(1000):
... a = 10. + np.random.random(100)
... b.append(np.product(a))

>>> b = np.array(b) / np.min(b) # scale values to be positive

>>> count, bins, ignored = plt.hist(b, 100, normed=True, align=center)

>>> sigma = np.std(np.log(b))

>>> mu = np.mean(np.log(b))

>>> x = np.linspace(min(bins), max(bins), 10000)

>>> pdf = (np.exp(-(np.log(x) - mu)**2 / (2 * sigma**2))
... / (x * sigma * np.sqrt(2 * np.pi)))

>>> plt.plot(x, pdf, color=r, linewidth=2)

>>> plt.show()

584 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

0.035
0.030
0.025
0.020
0.015
0.010
0.005
0.000
50 100 150 200 250 300

logseries(p, size=None)
Draw samples from a Logarithmic Series distribution.
Samples are drawn from a Log Series distribution with specified parameter, p (probability, 0 < p < 1).
Parameters
loc : float
scale : float > 0.
size : {tuple, int}
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : {ndarray, scalar}
where the values are all integers in [0, n].
See Also:

scipy.stats.distributions.logser
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Log Series distribution is
pk
P (k) = ,
k ln(1 p)

where p = probability.
The Log Series distribution is frequently used to represent species richness and occurrence, first proposed by
Fisher, Corbet, and Williams in 1943 [2]. It may also be used to model the numbers of occupants seen in cars
[3].

References
[R99], [R100], [R101], [R102]

3.8. Random sampling (numpy.random) 585

NumPy Reference, Release 2.0.0.dev8464

Examples
Draw samples from the distribution:

>>> a = .6
>>> s = np.random.logseries(a, 10000)
>>> count, bins, ignored = plt.hist(s)

# plot against distribution

>>> def logseries(k, p):

... return -p**k/(k*log(1-p))
>>> plt.plot(bins, logseries(bins, a)*count.max()/
logseries(bins, a).max(), r)
>>> plt.show()

multinomial(n, pvals, size=None)

Draw samples from a multinomial distribution.
The multinomial distribution is a multivariate generalisation of the binomial distribution. Take an experiment
with one of p possible outcomes. An example of such an experiment is throwing a dice, where the outcome
can be 1 through 6. Each sample drawn from the distribution represents n such experiments. Its values, X_i =
[X_0, X_1, ..., X_p], represent the number of times the outcome was i.
Parameters
n : int
Number of experiments.
pvals : sequence of floats, length p
Probabilities of each of the p different outcomes. These should sum to 1 (however,
the last element is always assumed to account for the remaining probability, as long as
sum(pvals[:-1]) <= 1).
size : tuple of ints
Given a size of (M, N, K), then M*N*K samples are drawn, and the output shape
becomes (M, N, K, p), since each sample has shape (p,).

Examples
Throw a dice 20 times:

>>> np.random.multinomial(20, [1/6.]*6, size=1)

array([[4, 1, 7, 5, 2, 1]])

It landed 4 times on 1, once on 2, etc.

Now, throw the dice 20 times, and 20 times again:

>>> np.random.multinomial(20, [1/6.]*6, size=2)

array([[3, 4, 3, 3, 4, 3],
[2, 4, 3, 4, 0, 7]])

For the first run, we threw 3 times 1, 4 times 2, etc. For the second, we threw 2 times 1, 4 times 2, etc.
A loaded dice is more likely to land on number 6:

586 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.random.multinomial(100, [1/7.]*5)

array([13, 16, 13, 16, 42])

multivariate_normal(mean, cov, [size])

Draw random samples from a multivariate normal distribution.
The multivariate normal, multinormal or Gaussian distribution is a generalisation of the one-dimensional normal
distribution to higher dimensions.
Such a distribution is specified by its mean and covariance matrix, which are analogous to the mean (average or
centre) and variance (standard deviation squared or width) of the one-dimensional normal distribution.
Parameters
mean : (N,) ndarray
Mean of the N-dimensional distribution.
cov : (N,N) ndarray
Covariance matrix of the distribution.
size : tuple of ints, optional
Given a shape of, for example, (m,n,k), m*n*k samples are generated, and packed in an
m-by-n-by-k arrangement. Because each sample is N-dimensional, the output shape is
(m,n,k,N). If no shape is specified, a single sample is returned.
Returns
out : ndarray
The drawn samples, arranged according to size. If the shape given is (m,n,...), then the
shape of out is is (m,n,...,N).
In other words, each entry out[i,j,...,:] is an N-dimensional value drawn from
the distribution.

Notes
The mean is a coordinate in N-dimensional space, which represents the location where samples are most likely
to be generated. This is analogous to the peak of the bell curve for the one-dimensional or univariate normal
distribution.
Covariance indicates the level to which two variables vary together. From the multivariate normal distribution,
we draw N-dimensional samples, X = [x1 , x2 , ...xN ]. The covariance matrix element Cij is the covariance of
xi and xj . The element Cii is the variance of xi (i.e. its spread).
Instead of specifying the full covariance matrix, popular approximations include:
Spherical covariance (cov is a multiple of the identity matrix)
Diagonal covariance (cov has non-negative elements, and only on the diagonal)
This geometrical property can be seen in two dimensions by plotting generated data-points:

>>> mean = [0,0]

>>> cov = [[1,0],[0,100]] # diagonal covariance, points lie on x or y-axis

>>> import matplotlib.pyplot as plt

>>> x,y = np.random.multivariate_normal(mean,cov,5000).T
>>> plt.plot(x,y,x); plt.axis(equal); plt.show()

Note that the covariance matrix must be non-negative definite.

3.8. Random sampling (numpy.random) 587

NumPy Reference, Release 2.0.0.dev8464

References
[R236], [R237]

Examples

>>> mean = (1,2)

>>> cov = [[1,0],[1,0]]
>>> x = np.random.multivariate_normal(mean,cov,(3,3))
>>> x.shape
(3, 3, 2)

The following is probably true, given that 0.6 is roughly twice the standard deviation:

>>> print list( (x[0,0,:] - mean) < 0.6 )

[True, True]

negative_binomial(n, p, size=None)
Draw samples from a negative_binomial distribution.
Samples are drawn from a negative_Binomial distribution with specified parameters, n trials and p probability
of success where n is an integer > 0 and p is in the interval [0, 1].
Parameters
n : int
Parameter, > 0.
p : float
Parameter, >= 0 and <=1.
size : int or tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : int or ndarray of ints
Drawn samples.

Notes
The probability density for the Negative Binomial distribution is
 
N +n1 n
P (N ; n, p) = p (1 p)N ,
n1

where n 1 is the number of successes, p is the probability of success, and N + n 1 is the number of trials.
The negative binomial distribution gives the probability of n-1 successes and N failures in N+n-1 trials, and
success on the (N+n)th trial.
If one throws a die repeatedly until the third time a 1 appears, then the probability distribution of the number
of non-1s that appear before the third 1 is a negative binomial distribution.

References
[R238], [R239]

588 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples
Draw samples from the distribution:
A real world example. A company drills wild-cat oil exploration wells, each with an estimated probability
of success of 0.1. What is the probability of having one success for each successive well, that is what is the
probability of a single success after drilling 5 wells, after 6 wells, etc.?

>>> s = np.random.negative_binomial(1, 0.1, 100000)

>>> for i in range(1, 11):
... probability = sum(s<i) / 100000.
... print i, "wells drilled, probability of one success =", probability

noncentral_chisquare(df, nonc, size=None)

Draw samples from a noncentral chi-square distribution.
The noncentral 2 distribution is a generalisation of the 2 distribution.
Parameters
df : int
Degrees of freedom, should be >= 1.
nonc : float
Non-centrality, should be > 0.
size : int or tuple of ints
Shape of the output.

Notes
The probability density function for the noncentral Chi-square distribution is

X enonc/2 (nonc/2)i
P (x; df, nonc) = PYdf +2i (x),
i=0
i!

where Yq is the Chi-square with q degrees of freedom.

In Delhi (2007), it is noted that the noncentral chi-square is useful in bombing and coverage problems, the
probability of killing the point target given by the noncentral chi-squared distribution.

References
[R240], [R241]

Examples
Draw values from the distribution and plot the histogram

>>> import matplotlib.pyplot as plt

>>> values = plt.hist(np.random.noncentral_chisquare(3, 20, 100000),
... bins=200, normed=True)
>>> plt.show()

Draw values from a noncentral chisquare with very small noncentrality, and compare to a chisquare.

3.8. Random sampling (numpy.random) 589

NumPy Reference, Release 2.0.0.dev8464

>>> plt.figure()
>>> values = plt.hist(np.random.noncentral_chisquare(3, .0000001, 100000),
... bins=np.arange(0., 25, .1), normed=True)
>>> values2 = plt.hist(np.random.chisquare(3, 100000),
... bins=np.arange(0., 25, .1), normed=True)
>>> plt.plot(values[1][0:-1], values[0]-values2[0], ob)
>>> plt.show()

Demonstrate how large values of non-centrality lead to a more symmetric distribution.

>>> plt.figure()
>>> values = plt.hist(np.random.noncentral_chisquare(3, 20, 100000),
... bins=200, normed=True)
>>> plt.show()

0.05

0.04

0.03

0.02

0.01

0.00
0 10 20 30 40 50 60 70 80

0.25

0.20

0.15

0.10

0.05

0.00
0 5 10 15 20 25

590 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

0.05

0.04

0.03

0.02

0.01

0.00
0 10 20 30 40 50 60 70 80 90

noncentral_f(dfnum, dfden, nonc, size=None)

Draw samples from the noncentral F distribution.
Samples are drawn from an F distribution with specified parameters, dfnum (degrees of freedom in numerator)
and dfden (degrees of freedom in denominator), where both parameters > 1. nonc is the non-centrality parameter.
Parameters
dfnum : int
Parameter, should be > 1.
dfden : int
Parameter, should be > 1.
nonc : float
Parameter, should be >= 0.
size : int or tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : scalar or ndarray
Drawn samples.

Notes
When calculating the power of an experiment (power = probability of rejecting the null hypothesis when a
specific alternative is true) the non-central F statistic becomes important. When the null hypothesis is true, the
F statistic follows a central F distribution. When the null hypothesis is not true, then it follows a non-central F
statistic.

References
Weisstein, Eric W. Noncentral F-Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/NoncentralF-Distribution.html
Wikipedia, Noncentral F distribution, http://en.wikipedia.org/wiki/Noncentral_F-distribution

3.8. Random sampling (numpy.random) 591

NumPy Reference, Release 2.0.0.dev8464

Examples
In a study, testing for a specific alternative to the null hypothesis requires use of the Noncentral F distribution.
We need to calculate the area in the tail of the distribution that exceeds the value of the F distribution for the null
hypothesis. Well plot the two probability distributions for comparison.

>>> dfnum = 3 # between group deg of freedom

>>> dfden = 20 # within groups degrees of freedom
>>> nonc = 3.0
>>> nc_vals = np.random.noncentral_f(dfnum, dfden, nonc, 1000000)
>>> NF = np.histogram(nc_vals, bins=50, normed=True)
>>> c_vals = np.random.f(dfnum, dfden, 1000000)
>>> F = np.histogram(c_vals, bins=50, normed=True)
>>> plt.plot(F[1][1:], F[0])
>>> plt.plot(NF[1][1:], NF[0])
>>> plt.show()

normal(loc=0.0, scale=1.0, size=None)

Draw random samples from a normal (Gaussian) distribution.
The probability density function of the normal distribution, first derived by De Moivre and 200 years later by
both Gauss and Laplace independently [R243], is often called the bell curve because of its characteristic shape
(see the example below).
The normal distributions occurs often in nature. For example, it describes the commonly occurring distribution
of samples influenced by a large number of tiny, random disturbances, each with its own unique distribution
[R243].
Parameters
loc : float
Mean (centre) of the distribution.
scale : float
Standard deviation (spread or width) of the distribution.
size : tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
See Also:

scipy.stats.distributions.norm
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Gaussian distribution is
1 (x)2
p(x) = e 2 2 ,
2 2

where is the mean and the standard deviation. The square of the standard deviation, 2 , is called the
variance.
The function has its peak at the mean, and its spread increases with the standard deviation (the function
reaches 0.607 times its maximum at x + and x [R243]). This implies that numpy.random.normal is
more likely to return samples lying close to the mean, rather than those far away.

592 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

References
[R242], [R243]

Examples
Draw samples from the distribution:

>>> mu, sigma = 0, 0.1 # mean and standard deviation

>>> s = np.random.normal(mu, sigma, 1000)

Verify the mean and the variance:

>>> abs(mu - np.mean(s)) < 0.01

True

>>> abs(sigma - np.std(s, ddof=1)) < 0.01

True

Display the histogram of the samples, along with the probability density function:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 30, normed=True)
>>> plt.plot(bins, 1/(sigma * np.sqrt(2 * np.pi)) *
... np.exp( - (bins - mu)**2 / (2 * sigma**2) ),
... linewidth=2, color=r)
>>> plt.show()

0
0.4 0.3 0.2 0.1 0.0 0.1 0.2 0.3

pareto(a, size=None)
Draw samples from a Pareto distribution with specified shape.
This is a simplified version of the Generalized Pareto distribution (available in SciPy), with the scale set to one
and the location set to zero. Most authors default the location to one.
The Pareto distribution must be greater than zero, and is unbounded above. It is also known as the 80-20 rule.
In this distribution, 80 percent of the weights are in the lowest 20 percent of the range, while the other 20 percent
fill the remaining 80 percent of the range.

3.8. Random sampling (numpy.random) 593

NumPy Reference, Release 2.0.0.dev8464

Parameters
shape : float, > 0.
Shape of the distribution.
size : tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
See Also:

scipy.stats.distributions.genpareto.pdf
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Pareto distribution is
ama
p(x) =
xa+1

where a is the shape and m the location

The Pareto distribution, named after the Italian economist Vilfredo Pareto, is a power law probability distribution
useful in many real world problems. Outside the field of economics it is generally referred to as the Bradford
distribution. Pareto developed the distribution to describe the distribution of wealth in an economy. It has
also found use in insurance, web page access statistics, oil field sizes, and many other problems, including the
download frequency for projects in Sourceforge [1]. It is one of the so-called fat-tailed distributions.

References
[R244], [R245], [R246], [R247]

Examples
Draw samples from the distribution:

>>> a, m = 3., 1. # shape and mode

>>> s = np.random.pareto(a, 1000) + m

Display the histogram of the samples, along with the probability density function:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 100, normed=True, align=center)
>>> fit = a*m**a/bins**(a+1)
>>> plt.plot(bins, max(count)*fit/max(fit),linewidth=2, color=r)
>>> plt.show()

594 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

2.5

2.0

1.5

1.0

0.5

0.0
0 2 4 6 8 10 12 14 16 18

poisson(lam=1.0, size=None)
Draw samples from a Poisson distribution.
The Poisson distribution is the limit of the Binomial distribution for large N.
Parameters
lam : float
Expectation of interval, should be >= 0.
size : int or tuple of ints, optional
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.

Notes
The Poisson distribution
k e
f (k; ) =
k!

For events with an expected separation the Poisson distribution f (k; ) describes the probability of k events
occurring within the observed interval .

References
[R248], [R249]

Examples
Draw samples from the distribution:

>>> import numpy as np

>>> s = np.random.poisson(5, 10000)

Display histogram of the sample:

3.8. Random sampling (numpy.random) 595

NumPy Reference, Release 2.0.0.dev8464

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 14, normed=True)
>>> plt.show()

0.30
0.25
0.20
0.15
0.10
0.05
0.00
0 2 4 6 8 10 12 14 16 18

power(a, size=None)
Draws samples in [0, 1] from a power distribution with positive exponent a - 1.
Also known as the power function distribution.
Parameters
a : float
parameter, > 0
size : tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then
m * n * k samples are drawn.
Returns
samples : {ndarray, scalar}
The returned samples lie in [0, 1].
Raises
ValueError :
If a<1.

Notes
The probability density function is

P (x; a) = axa1 , 0 x 1, a > 0.

The power function distribution is just the inverse of the Pareto distribution. It may also be seen as a special
case of the Beta distribution.
It is used, for example, in modeling the over-reporting of insurance claims.

596 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

References
[R250], [R251]

Examples
rayleigh(scale=1.0, size=None)
Draw samples from a Rayleigh distribution.
The and Weibull distributions are generalizations of the Rayleigh.
Parameters
scale : scalar
Scale, also equals the mode. Should be >= 0.
size : int or tuple of ints, optional
Shape of the output. Default is None, in which case a single value is returned.

Notes
The probability density function for the Rayleigh distribution is
x x2
P (x; scale) = 2
e 2scale2
scale

The Rayleigh distribution arises if the wind speed and wind direction are both gaussian variables, then the
vector wind velocity forms a Rayleigh distribution. The Rayleigh distribution is used to model the expected
output from wind turbines.

References
..[1] Brighton Webs Ltd., Rayleigh Distribution,
http://www.brighton-webs.co.uk/distributions/rayleigh.asp
..[2] Wikipedia, Rayleigh distribution
http://en.wikipedia.org/wiki/Rayleigh_distribution

Examples
Draw values from the distribution and plot the histogram

>>> values = hist(np.random.rayleigh(3, 100000), bins=200, normed=True)

Wave heights tend to follow a Rayleigh distribution. If the mean wave height is 1 meter, what fraction of waves
are likely to be larger than 3 meters?

>>> meanvalue = 1
>>> modevalue = np.sqrt(2 / np.pi) * meanvalue
>>> s = np.random.rayleigh(modevalue, 1000000)

The percentage of waves larger than 3 meters is:

>>> 100.*sum(s>3)/1000000.
0.087300000000000003

standard_cauchy(size=None)
Standard Cauchy distribution with mode = 0.
Also known as the Lorentz distribution.

3.8. Random sampling (numpy.random) 597

NumPy Reference, Release 2.0.0.dev8464

Parameters
size : int or tuple of ints
Shape of the output.
Returns
samples : ndarray or scalar
The drawn samples.

Notes
The probability density function for the full Cauchy distribution is
1
P (x; x0 , ) =
1 + ( xx
 
0 2
)

and the Standard Cauchy distribution just sets x0 = 0 and = 1

The Cauchy distribution arises in the solution to the driven harmonic oscillator problem, and also describes
spectral line broadening. It also describes the distribution of values at which a line tilted at a random angle will
cut the x axis.
When studying hypothesis tests that assume normality, seeing how the tests perform on data from a Cauchy
distribution is a good indicator of their sensitivity to a heavy-tailed distribution, since the Cauchy looks very
much like a Gaussian distribution, but with heavier tails.

References
..[1] NIST/SEMATECH e-Handbook of Statistical Methods, Cauchy
Distribution, http://www.itl.nist.gov/div898/handbook/eda/section3/eda3663.htm
..[2] Weisstein, Eric W. Cauchy Distribution. From MathWorldA
Wolfram Web Resource. http://mathworld.wolfram.com/CauchyDistribution.html
..[3] Wikipedia, Cauchy distribution
http://en.wikipedia.org/wiki/Cauchy_distribution

Examples
Draw samples and plot the distribution:

>>> s = np.random.standard_cauchy(1000000)
>>> s = s[(s>-25) & (s<25)] # truncate distribution so it plots well
>>> plt.hist(s, bins=100)
>>> plt.show()

standard_exponential(size=None)
Draw samples from the standard exponential distribution.
standard_exponential is identical to the exponential distribution with a scale parameter of 1.
Parameters
size : int or tuple of ints
Shape of the output.
Returns
out : float or ndarray
Drawn samples.

598 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples
Output a 3x8000 array:

>>> n = np.random.standard_exponential((3, 8000))

standard_gamma(shape, size=None)
Draw samples from a Standard Gamma distribution.
Samples are drawn from a Gamma distribution with specified parameters, shape (sometimes designated k)
and scale=1.
Parameters
shape : float
Parameter, should be > 0.
size : int or tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : ndarray or scalar
The drawn samples.
See Also:

scipy.stats.distributions.gamma
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Gamma distribution is

ex/
p(x) = xk1 ,
k (k)

where k is the shape and the scale, and is the Gamma function.
The Gamma distribution is often used to model the times to failure of electronic components, and arises naturally
in processes for which the waiting times between Poisson distributed events are relevant.

References
[R253], [R254]

Examples
standard_normal(size=None)
Returns samples from a Standard Normal distribution (mean=0, stdev=1).
Parameters
size : int or tuple of ints, optional
Output shape. Default is None, in which case a single value is returned.
Returns
out : float or ndarray
Drawn samples.

3.8. Random sampling (numpy.random) 599

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> s = np.random.standard_normal(8000)
>>> s
array([ 0.6888893 , 0.78096262, -0.89086505, ..., 0.49876311, #random
-0.38672696, -0.4685006 ]) #random
>>> s.shape
(8000,)
>>> s = np.random.standard_normal(size=(3, 4, 2))
>>> s.shape
(3, 4, 2)

standard_t(df, size=None)
Standard Students t distribution with df degrees of freedom.
A special case of the hyperbolic distribution. As df gets large, the result resembles that of the standard normal
distribution (standard_normal).
Parameters
df : int
Degrees of freedom, should be > 0.
size : int or tuple of ints, optional
Output shape. Default is None, in which case a single value is returned.
Returns
samples : ndarray or scalar
Drawn samples.

Notes
The probability density function for the t distribution is

( df2+1 )  x2 (df +1)/2

P (x, df ) = df
1+
df ( 2 ) df

The t test is based on an assumption that the data come from a Normal distribution. The t test provides a way to
test whether the sample mean (that is the mean calculated from the data) is a good estimate of the true mean.
The derivation of the t-distribution was forst published in 1908 by William Gisset while working for the Guin-
ness Brewery in Dublin. Due to proprietary issues, he had to publish under a pseudonym, and so he used the
name Student.

References
[R255], [R256]

Examples
triangular(left, mode, right, size=None)
Draw samples from the triangular distribution.
The triangular distribution is a continuous probability distribution with lower limit left, peak at mode, and upper
limit right. Unlike the other distributions, these parameters directly define the shape of the pdf.
Parameters
left : scalar

600 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Lower limit.
mode : scalar
The value where the peak of the distribution occurs. The value should fulfill the condi-
tion left <= mode <= right.
right : scalar
Upper limit, should be larger than left.
size : int or tuple of ints, optional
Output shape. Default is None, in which case a single value is returned.
Returns
samples : ndarray or scalar
The returned samples all lie in the interval [left, right].

Notes
The probability density function for the Triangular distribution is
2(xl)
(rl)(ml) for l x m,

2(mx)
P (x; l, m, r) = (rl)(rm) for m x r,

0 otherwise.

The triangular distribution is often used in ill-defined problems where the underlying distribution is not known,
but some knowledge of the limits and mode exists. Often it is used in simulations.

References
..[1] Wikipedia, Triangular distribution
http://en.wikipedia.org/wiki/Triangular_distribution

Examples
Draw values from the distribution and plot the histogram:

>>> import matplotlib.pyplot as plt

>>> h = plt.hist(np.random.triangular(-3, 0, 8, 100000), bins=200,
... normed=True)
>>> plt.show()

3.8. Random sampling (numpy.random) 601

NumPy Reference, Release 2.0.0.dev8464

0.20

0.15

0.10

0.05

0.00
4 2 0 2 4 6 8

uniform(low=0.0, high=1.0, size=1)

Draw samples from a uniform distribution.
Samples are uniformly distributed over the half-open interval [low, high) (includes low, but excludes high).
In other words, any value within the given interval is equally likely to be drawn by uniform.
Parameters
low : float, optional
Lower boundary of the output interval. All values generated will be greater than or
equal to low. The default value is 0.
high : float
Upper boundary of the output interval. All values generated will be less than high. The
default value is 1.0.
size : tuple of ints, int, optional
Shape of output. If the given size is, for example, (m,n,k), m*n*k samples are generated.
If no shape is specified, a single sample is returned.
Returns
out : ndarray
Drawn samples, with shape size.
See Also:

randint
Discrete uniform distribution, yielding integers.
random_integers
Discrete uniform distribution over the closed interval [low, high].
random_sample
Floats uniformly distributed over [0, 1).
random
Alias for random_sample.

602 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

rand
Convenience function that accepts dimensions as input, e.g., rand(2,2) would generate a 2-by-2 array
of floats, uniformly distributed over [0, 1).

Notes
The probability density function of the uniform distribution is
1
p(x) =
ba

anywhere within the interval [a, b), and zero elsewhere.

Examples
Draw samples from the distribution:

>>> s = np.random.uniform(-1,0,1000)

All values are within the given interval:

>>> np.all(s >= -1)

True

>>> np.all(s < 0)

True

Display the histogram of the samples, along with the probability density function:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 15, normed=True)
>>> plt.plot(bins, np.ones_like(bins), linewidth=2, color=r)
>>> plt.show()

1.2
1.0
0.8
0.6
0.4
0.2
0.0
1.0 0.8 0.6 0.4 0.2 0.0

vonmises(mu=0.0, kappa=1.0, size=None)

Draw samples from a von Mises distribution.

3.8. Random sampling (numpy.random) 603

NumPy Reference, Release 2.0.0.dev8464

Samples are drawn from a von Mises distribution with specified mode (mu) and dispersion (kappa), on the
interval [-pi, pi].
The von Mises distribution (also known as the circular normal distribution) is a continuous probability distribu-
tion on the circle. It may be thought of as the circular analogue of the normal distribution.
Parameters
mu : float
Mode (center) of the distribution.
kappa : float, >= 0.
Dispersion of the distribution.
size : {tuple, int}
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : {ndarray, scalar}
The returned samples live on the unit circle [-pi, pi].
See Also:

scipy.stats.distributions.vonmises
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the von Mises distribution is

ecos(x)
p(x) = ,
2I0 ()

where is the mode and the dispersion, and I0 () is the modified Bessel function of order 0.
The von Mises, named for Richard Edler von Mises, born in Austria-Hungary, in what is now the Ukraine.
He fled to the United States in 1939 and became a professor at Harvard. He worked in probability theory,
aerodynamics, fluid mechanics, and philosophy of science.

References
[R257], [R258], [R259]

Examples
wald(mean, scale, size=None)
Draw samples from a Wald, or Inverse Gaussian, distribution.
As the scale approaches infinity, the distribution becomes more like a Gaussian.
Some references claim that the Wald is an Inverse Gaussian with mean=1, but this is by no means universal.
The Inverse Gaussian distribution was first studied in relationship to Brownian motion. In 1956 M.C.K. Tweedie
used the name Inverse Gaussian because there is an inverse relationship between the time to cover a unit distance
and distance covered in unit time.
Parameters
mean : scalar

604 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Distribution mean, should be > 0.

scale : scalar
Scale parameter, should be >= 0.
size : int or tuple of ints, optional
Output shape. Default is None, in which case a single value is returned.
Returns
samples : ndarray or scalar
Drawn sample, all greater than zero.

Notes
The probability density function for the Wald distribution is
r
scale scale(xmean)2

P (x; mean, scale) = 3

e 2mean2 x
2x

As noted above the Inverse Gaussian distribution first arise from attempts to model Brownian Motion. It is also a
competitor to the Weibull for use in reliability modeling and modeling stock returns and interest rate processes.

References
..[1] Brighton Webs Ltd., Wald Distribution,
http://www.brighton-webs.co.uk/distributions/wald.asp
..[2] Chhikara, Raj S., and Folks, J. Leroy, The Inverse Gaussian
Distribution: Theory : Methodology, and Applications, CRC Press, 1988.
..[3] Wikipedia, Wald distribution
http://en.wikipedia.org/wiki/Wald_distribution

Examples
Draw values from the distribution and plot the histogram:

>>> import matplotlib.pyplot as plt

>>> h = plt.hist(np.random.wald(3, 2, 100000), bins=200, normed=True)
>>> plt.show()

3.8. Random sampling (numpy.random) 605

NumPy Reference, Release 2.0.0.dev8464

0.45
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
0 10 20 30 40 50 60 70 80

weibull(a, size=None)
Weibull distribution.
Draw samples from a 1-parameter Weibull distribution with the given shape parameter.

X = (ln(U ))1/a

Here, U is drawn from the uniform distribution over (0,1].

The more common 2-parameter Weibull, including a scale parameter is just X = (ln(U ))1/a .
The Weibull (or Type III asymptotic extreme value distribution for smallest values, SEV Type III, or Rosin-
Rammler distribution) is one of a class of Generalized Extreme Value (GEV) distributions used in modeling
extreme value problems. This class includes the Gumbel and Frechet distributions.
Parameters
a : float
Shape of the distribution.
size : tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
See Also:

scipy.stats.distributions.weibull
probability density function, distribution or cumulative density function, etc.

gumbel, scipy.stats.distributions.genextreme

Notes
The probability density for the Weibull distribution is
a x a1 (x/)a
p(x) = ( ) e ,

606 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

where a is the shape and the scale.

The function has its peak (the mode) at ( a1
a )
1/a
.
When a = 1, the Weibull distribution reduces to the exponential distribution.

References
[R260], [R261], [R262]

Examples
Draw samples from the distribution:

>>> a = 5. # shape
>>> s = np.random.weibull(a, 1000)

Display the histogram of the samples, along with the probability density function:

>>> import matplotlib.pyplot as plt

>>> x = np.arange(1,100.)/50.
>>> def weib(x,n,a):
... return (a / n) * (x / n)**(a - 1) * np.exp(-(x / n)**a)

>>> count, bins, ignored = plt.hist(np.random.weibull(5.,1000))

>>> x = np.arange(1,100.)/50.
>>> scale = count.max()/weib(x, 1., 5.).max()
>>> plt.plot(x, weib(x, 1., 5.)*scale)
>>> plt.show()

250

200

150

100

50

0
0.0 0.5 1.0 1.5 2.0

zipf(a, size=None)
Draw samples from a Zipf distribution.
Samples are drawn from a Zipf distribution with specified parameter (a), where a > 1.
The zipf distribution (also known as the zeta distribution) is a continuous probability distribution that satisfies
Zipfs law, where the frequency of an item is inversely proportional to its rank in a frequency table.
Parameters
a : float

3.8. Random sampling (numpy.random) 607

NumPy Reference, Release 2.0.0.dev8464

parameter, > 1.
size : {tuple, int}
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : {ndarray, scalar}
The returned samples are greater than or equal to one.
See Also:

scipy.stats.distributions.zipf
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Zipf distribution is

xa
p(x) = ,
(a)

where is the Riemann Zeta function.

Named after the American linguist George Kingsley Zipf, who noted that the frequency of any word in a sample
of a language is inversely proportional to its rank in the frequency table.

References
[R263], [R264], [R265], [R266]

Examples

3.8.4 Random generator

mtrand.RandomState Container for the Mersenne Twister pseudo-random number generator.

seed([seed]) Seed the generator.
get_state() Return a tuple representing the internal state of the generator.
set_state(state) Set the internal state of the generator from a tuple.
class RandomState()
Container for the Mersenne Twister pseudo-random number generator.
RandomState exposes a number of methods for generating random numbers drawn from a variety of probability
distributions. In addition to the distribution-specific arguments, each method takes a keyword argument size that
defaults to None. If size is None, then a single value is generated and returned. If size is an integer, then a
1-D array filled with generated values is returned. If size is a tuple, then an array with that shape is filled and
returned.
Parameters
seed : int or array_like, optional
Random seed initializing the pseudo-random number generator. Can be an integer, an
array (or other sequence) of integers of any length, or None (the default). If seed is
None, then RandomState will try to read data from /dev/urandom (or the Windows
analogue) if available or seed from the clock otherwise.

608 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notes
The Python stdlib module random also contains a Mersenne Twister pseudo-random number generator with a
number of methods that are similar to the ones available in RandomState. RandomState, besides being NumPy-
aware, has the advantage that it provides a much larger number of probability distributions to choose from.

Methods

beta(a, b[, size]) The Beta distribution over [0, 1].

binomial(n, p[, size]) Draw samples from a binomial distribution.
bytes(length) Return random bytes.
chisquare(df[, size]) Draw samples from a chi-square distribution.
dirichlet(alpha[, size]) Draw samples from the Dirichlet distribution.
exponential([scale, size]) Exponential distribution.
f(dfnum, dfden[, size]) Draw samples from a F distribution.
gamma(shape[, scale, size]) Draw samples from a Gamma distribution.
geometric(p[, size]) Draw samples from the geometric distribution.
get_state() Return a tuple representing the internal state of the generato
gumbel([loc, scale, size]) Gumbel distribution.
hypergeometric(ngood, nbad, nsample[, size]) Draw samples from a Hypergeometric distribution.
laplace([loc, scale, size]) Draw samples from the Laplace or double exponential distribution
logistic([loc, scale, size]) Draw samples from a Logistic distribution.
lognormal([mean, sigma, size]) Return samples drawn from a log-normal distribution.
logseries(p[, size]) Draw samples from a Logarithmic Series distribution.
multinomial(n, pvals[, size]) Draw samples from a multinomial distribution.
multivariate_normal(mean) Draw random samples from a multivariate normal distribution
negative_binomial(n, p[, size]) Draw samples from a negative_binomial distribution.
noncentral_chisquare(df, nonc[, size]) Draw samples from a noncentral chi-square distribution.
noncentral_f(dfnum, dfden, nonc[, size]) Draw samples from the noncentral F distribution.
normal([loc, scale, size]) Draw random samples from a normal (Gaussian) distribution
pareto(a[, size]) Draw samples from a Pareto distribution with specified shap
permutation(x) Randomly permute a sequence, or return a permuted range
poisson([lam, size]) Draw samples from a Poisson distribution.
power(a[, size]) Draws samples in [0, 1] from a power distribution with positive exponent
rand(d0, d1, dn) Random values in a given shape.
randint(low[, high, size]) Return random integers from low (inclusive) to high (exclusiv
randn(d1) Return a sample (or samples) from the standard normal distribu
random_integers(low[, high, size]) Return random integers between low and high, inclusive.
random_sample([size]) Return random floats in the half-open interval [0.0, 1.0).
rayleigh([scale, size]) Draw samples from a Rayleigh distribution.
seed([seed]) Seed the generator.
set_state(state) Set the internal state of the generator from a tuple.
shuffle(x) Modify a sequence in-place by shuffling its contents.
standard_cauchy([size]) Standard Cauchy distribution with mode = 0.
standard_exponential([size]) Draw samples from the standard exponential distribution.
standard_gamma(shape[, size]) Draw samples from a Standard Gamma distribution.
standard_normal([size]) Returns samples from a Standard Normal distribution (mean=0, stdev
standard_t(df[, size]) Standard Students t distribution with df degrees of freedom
tomaxint([size]) Uniformly sample discrete random integers x such that 0 <= x <= sys.m
triangular(left, mode, right[, size]) Draw samples from the triangular distribution.
uniform([low, high, size]) Draw samples from a uniform distribution.
vonmises([mu, kappa, size]) Draw samples from a von Mises distribution.
wald(mean, scale[, size]) Draw samples from a Wald, or Inverse Gaussian, distributio
weibull(a[, size]) Weibull distribution.
Continued on ne

3.8. Random sampling (numpy.random) 609

NumPy Reference, Release 2.0.0.dev8464

Table 3.2 continued from previous page

zipf(a[, size]) Draw samples from a Zipf distribution.

beta(a, b, size=None)
The Beta distribution over [0, 1].
The Beta distribution is a special case of the Dirichlet distribution, and is related to the Gamma distribution.
It has the probability distribution function
1
f (x; a, b) = x1 (1 x)1 ,
B(, )

where the normalisation, B, is the beta function,

Z 1
B(, ) = t1 (1 t)1 dt.
0

It is often seen in Bayesian inference and order statistics.

Parameters
a : float
Alpha, non-negative.
b : float
Beta, non-negative.
size : tuple of ints, optional
The number of samples to draw. The ouput is packed according to the size given.
Returns
out : ndarray
Array of the given shape, containing values drawn from a Beta distribution.
binomial(n, p, size=None)
Draw samples from a binomial distribution.
Samples are drawn from a Binomial distribution with specified parameters, n trials and p probability of
success where n an integer > 0 and p is in the interval [0,1]. (n may be input as a float, but it is truncated
to an integer in use)
Parameters
n : float (but truncated to an integer)
parameter, > 0.
p : float
parameter, >= 0 and <=1.
size : {tuple, int}
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : {ndarray, scalar}

610 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

where the values are all integers in [0, n].

See Also:

scipy.stats.distributions.binom
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Binomial distribution is
 
n N
P (N ) = p (1 p)nN ,
N

where n is the number of trials, p is the probability of success, and N is the number of successes.
When estimating the standard error of a proportion in a population by using a random sample, the normal
distribution works well unless the product p*n <=5, where p = population proportion estimate, and n =
number of samples, in which case the binomial distribution is used instead. For example, a sample of 15
people shows 4 who are left handed, and 11 who are right handed. Then p = 4/15 = 27%. 0.27*15 = 4, so
the binomial distribution should be used in this case.

References
[R169], [R170], [R171], [R172], [R173]

Examples
Draw samples from the distribution:

>>> n, p = 10, .5 # number of trials, probability of each trial

>>> s = np.random.binomial(n, p, 1000)
# result of flipping a coin 10 times, tested 1000 times.

A real world example. A company drills 9 wild-cat oil exploration wells, each with an estimated probability
of success of 0.1. All nine wells fail. What is the probability of that happening?
Lets do 20,000 trials of the model, and count the number that generate zero positive results.

>>> sum(np.random.binomial(9,0.1,20000)==0)/20000.
answer = 0.38885, or 38%.

bytes(length)
Return random bytes.
Parameters
length : int
Number of random bytes.
Returns
out : str
String of length N.

3.8. Random sampling (numpy.random) 611

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.random.bytes(10)
eh\x85\x022SZ\xbf\xa4 #random

chisquare(df, size=None)
Draw samples from a chi-square distribution.
When df independent random variables, each with standard normal distributions (mean 0, variance 1), are
squared and summed, the resulting distribution is chi-square (see Notes). This distribution is often used in
hypothesis testing.
Parameters
df : int
Number of degrees of freedom.
size : tuple of ints, int, optional
Size of the returned array. By default, a scalar is returned.
Returns
output : ndarray
Samples drawn from the distribution, packed in a size-shaped array.
Raises
ValueError :
When df <= 0 or when an inappropriate size (e.g. size=-1) is given.

Notes
The variable obtained by summing the squares of df independent, standard normally distributed random
variables:
df
X
Q= Xi2
i=0

is chi-square distributed, denoted

Q 2k .

The probability density function of the chi-squared distribution is

(1/2)k/2 k/21 x/2

p(x) = x e ,
(k/2)

where is the gamma function,

Z
(x) = tx1 et dt.
0

References
[R174], [R175]

612 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.random.chisquare(2,4)
array([ 1.89920014, 9.00867716, 3.13710533, 5.62318272])

dirichlet(alpha, size=None)
Draw samples from the Dirichlet distribution.
Draw size samples of dimension k from a Dirichlet distribution. A Dirichlet-distributed random variable
can be seen as a multivariate generalization of a Beta distribution. Dirichlet pdf is the conjugate prior of a
multinomial in Bayesian inference.
Parameters
alpha : array
Parameter of the distribution (k dimension for sample of dimension k).
size : array
Number of samples to draw.

Notes

k
Y
i 1
X x
i
i=1

Uses the following property for computation: for each dimension, draw a random sample y_i from a
standard gamma generator of shape alpha_i, then X = Pk1 y (y1 , . . . , yn ) is Dirichlet distributed.
i=1 i

References
[R176]
exponential(scale=1.0, size=None)
Exponential distribution.
Its probability density function is
1 1 x
f (x; ) = exp( ),

for x > 0 and 0 elsewhere. is the scale parameter, which is the inverse of the rate parameter = 1/.
The rate parameter is an alternative, widely used parameterization of the exponential distribution [R179].
The exponential distribution is a continuous analogue of the geometric distribution. It describes many
common situations, such as the size of raindrops measured over many rainstorms [R177], or the time
between page requests to Wikipedia [R178].
Parameters
scale : float
The scale parameter, = 1/.
size : tuple of ints
Number of samples to draw. The output is shaped according to size.

3.8. Random sampling (numpy.random) 613

NumPy Reference, Release 2.0.0.dev8464

References
[R177], [R178], [R179]
f(dfnum, dfden, size=None)
Draw samples from a F distribution.
Samples are drawn from an F distribution with specified parameters, dfnum (degrees of freedom in nu-
merator) and dfden (degrees of freedom in denominator), where both parameters should be greater than
zero.
The random variate of the F distribution (also known as the Fisher distribution) is a continuous probability
distribution that arises in ANOVA tests, and is the ratio of two chi-square variates.
Parameters
dfnum : float
Degrees of freedom in numerator. Should be greater than zero.
dfden : float
Degrees of freedom in denominator. Should be greater than zero.
size : {tuple, int}, optional
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn. By default only one sample is returned.
Returns
samples : {ndarray, scalar}
Samples from the Fisher distribution.
See Also:

scipy.stats.distributions.f
probability density function, distribution or cumulative density function, etc.

Notes
The F statistic is used to compare in-group variances to between-group variances. Calculating the distribu-
tion depends on the sampling, and so it is a function of the respective degrees of freedom in the problem.
The variable dfnum is the number of samples minus one, the between-groups degrees of freedom, while
dfden is the within-groups degrees of freedom, the sum of the number of samples in each group minus the
number of groups.

References
[R180], [R181]

Examples
An example from Glantz[1], pp 47-40. Two groups, children of diabetics (25 people) and children from
people without diabetes (25 controls). Fasting blood glucose was measured, case group had a mean value
of 86.1, controls had a mean value of 82.2. Standard deviations were 2.09 and 2.49 respectively. Are
these data consistent with the null hypothesis that the parents diabetic status does not affect their childrens
blood glucose levels? Calculating the F statistic from the data gives a value of 36.01.
Draw samples from the distribution:

614 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> dfnum = 1. # between group degrees of freedom

>>> dfden = 48. # within groups degrees of freedom
>>> s = np.random.f(dfnum, dfden, 1000)

The lower bound for the top 1% of the samples is :

>>> sort(s)[-10]
7.61988120985

So there is about a 1% chance that the F statistic will exceed 7.62, the measured value is 36, so the null
hypothesis is rejected at the 1% level.
gamma(shape, scale=1.0, size=None)
Draw samples from a Gamma distribution.
Samples are drawn from a Gamma distribution with specified parameters, shape (sometimes designated
k) and scale (sometimes designated theta), where both parameters are > 0.
Parameters
shape : scalar > 0
The shape of the gamma distribution.
scale : scalar > 0, optional
The scale of the gamma distribution. Default is equal to 1.
size : shape_tuple, optional
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
out : ndarray, float
Returns one sample unless size parameter is specified.
See Also:

scipy.stats.distributions.gamma
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Gamma distribution is

ex/
p(x) = xk1 ,
k (k)

where k is the shape and the scale, and is the Gamma function.
The Gamma distribution is often used to model the times to failure of electronic components, and arises
naturally in processes for which the waiting times between Poisson distributed events are relevant.

References
[R182], [R183]

3.8. Random sampling (numpy.random) 615

NumPy Reference, Release 2.0.0.dev8464

Examples
geometric(p, size=None)
Draw samples from the geometric distribution.
Bernoulli trials are experiments with one of two outcomes: success or failure (an example of such an
experiment is flipping a coin). The geometric distribution models the number of trials that must be run in
order to achieve success. It is therefore supported on the positive integers, k = 1, 2, ....
The probability mass function of the geometric distribution is

f (k) = (1 p)k1 p

where p is the probability of success of an individual trial.

Parameters
p : float
The probability of success of an individual trial.
size : tuple of ints
Number of values to draw from the distribution. The output is shaped according to size.
Returns
out : ndarray
Samples from the geometric distribution, shaped according to size.

Examples
Draw ten thousand values from the geometric distribution, with the probability of an individual success
equal to 0.35:

>>> z = np.random.geometric(p=0.35, size=10000)

How many trials succeeded after a single run?

>>> (z == 1).sum() / 10000.

0.34889999999999999 #random

get_state()
Return a tuple representing the internal state of the generator.
For more details, see set_state.
Returns
out : tuple(str, ndarray of 624 uints, int, int, float)
The returned tuple has the following items:
1. the string MT19937.
2. a 1-D array of 624 unsigned integer keys.
3. an integer pos.
4. an integer has_gauss.
5. a float cached_gaussian.
See Also:
set_state

616 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notes
set_state and get_state are not needed to work with any of the random distributions in NumPy. If the
internal state is manually altered, the user should know exactly what he/she is doing.
gumbel(loc=0.0, scale=1.0, size=None)
Gumbel distribution.
Draw samples from a Gumbel distribution with specified location (or mean) and scale (or standard devia-
tion).
The Gumbel (or Smallest Extreme Value (SEV) or the Smallest Extreme Value Type I) distribution is one
of a class of Generalized Extreme Value (GEV) distributions used in modeling extreme value problems.
The Gumbel is a special case of the Extreme Value Type I distribution for maximums from distributions
with exponential-like tails, it may be derived by considering a Gaussian process of measurements, and
generating the pdf for the maximum values from that set of measurements (see examples).
Parameters
loc : float
The location of the mode of the distribution.
scale : float
The scale parameter of the distribution.
size : tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
See Also:

scipy.stats.gumbel
probability density function, distribution or cumulative density function, etc.

weibull, scipy.stats.genextreme

Notes
The probability density for the Gumbel distribution is

e(x)/ e(x)/
p(x) = e ,

where is the mode, a location parameter, and is the scale parameter.

The Gumbel (named for German mathematician Emil Julius Gumbel) was used very early in the hydrology
literature, for modeling the occurrence of flood events. It is also used for modeling maximum wind speed
and rainfall rates. It is a fat-tailed distribution - the probability of an event in the tail of the distribution is
larger than if one used a Gaussian, hence the surprisingly frequent occurrence of 100-year floods. Floods
were initially modeled as a Gaussian process, which underestimated the frequency of extreme events.
It is one of a class of extreme value distributions, the Generalized Extreme Value (GEV) distributions,
which also includes the Weibull and Frechet.
2 2
The function has a mean of + 0.57721 and a variance of 6 .

References
[R184], [R185], [R186]

3.8. Random sampling (numpy.random) 617

NumPy Reference, Release 2.0.0.dev8464

Examples
Draw samples from the distribution:

>>> mu, beta = 0, 0.1 # location and scale

>>> s = np.random.gumbel(mu, beta, 1000)

Display the histogram of the samples, along with the probability density function:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 30, normed=True)
>>> plt.plot(bins, (1/beta)*np.exp(-(bins - mu)/beta)
... * np.exp( -np.exp( -(bins - mu) /beta) ),
... linewidth=2, color=r)
>>> plt.show()

Show how an extreme value distribution can arise from a Gaussian process and compare to a Gaussian:

>>> means = []
>>> maxima = []
>>> for i in range(0,1000) :
... a = np.random.normal(mu, beta, 1000)
... means.append(a.mean())
... maxima.append(a.max())
>>> count, bins, ignored = plt.hist(maxima, 30, normed=True)
>>> beta = np.std(maxima)*np.pi/np.sqrt(6)
>>> mu = np.mean(maxima) - 0.57721*beta
>>> plt.plot(bins, (1/beta)*np.exp(-(bins - mu)/beta)
... * np.exp(-np.exp(-(bins - mu)/beta)),
... linewidth=2, color=r)
>>> plt.plot(bins, 1/(beta * np.sqrt(2 * np.pi))
... * np.exp(-(bins - mu)**2 / (2 * beta**2)),
... linewidth=2, color=g)
>>> plt.show()

14
12
10
8
6
4
2
0
0.4 0.2 0.0 0.2 0.4 0.6 0.8

hypergeometric(ngood, nbad, nsample, size=None)

Draw samples from a Hypergeometric distribution.

618 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Samples are drawn from a Hypergeometric distribution with specified parameters, ngood (ways to make a
good selection), nbad (ways to make a bad selection), and nsample = number of items sampled, which is
less than or equal to the sum ngood + nbad.
Parameters
ngood : float (but truncated to an integer)
parameter, > 0.
nbad : float
parameter, >= 0.
nsample : float
parameter, > 0 and <= ngood+nbad
size : {tuple, int}
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : {ndarray, scalar}
where the values are all integers in [0, n].
See Also:

scipy.stats.distributions.hypergeom
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Hypergeometric distribution is
m N m



n nx
P (x) = N

,
n

where 0 x m and n + m N x n
for P(x) the probability of x successes, n = ngood, m = nbad, and N = number of samples.
Consider an urn with black and white marbles in it, ngood of them black and nbad are white. If you
draw nsample balls without replacement, then the Hypergeometric distribution describes the distribution
of black balls in the drawn sample.
Note that this distribution is very similar to the Binomial distribution, except that in this case, samples
are drawn without replacement, whereas in the Binomial case samples are drawn with replacement (or the
sample space is infinite). As the sample space becomes large, this distribution approaches the Binomial.

References
[R187], [R188], [R189]

Examples
Draw samples from the distribution:

3.8. Random sampling (numpy.random) 619

NumPy Reference, Release 2.0.0.dev8464

>>> ngood, nbad, nsamp = 100, 2, 10

# number of good, number of bad, and number of samples
>>> s = np.random.hypergeometric(ngood, nbad, nsamp, 1000)
>>> hist(s)
# note that it is very unlikely to grab both bad items

Suppose you have an urn with 15 white and 15 black marbles. If you pull 15 marbles at random, how
likely is it that 12 or more of them are one color?

>>> s = np.random.hypergeometric(15, 15, 15, 100000)

>>> sum(s>=12)/100000. + sum(s<=3)/100000.
# answer = 0.003 ... pretty unlikely!

laplace(loc=0.0, scale=1.0, size=None)

Draw samples from the Laplace or double exponential distribution with specified location (or mean) and
scale (decay).
The Laplace distribution is similar to the Gaussian/normal distribution, but is sharper at the peak and
has fatter tails. It represents the difference between two independent, identically distributed exponential
random variables.
Parameters
loc : float
The position, , of the distribution peak.
scale : float
, the exponential decay.

Notes
It has the probability density function
 
1 |x |
f (x; , ) = exp .
2

The first law of Laplace, from 1774, states that the frequency of an error can be expressed as an exponential
function of the absolute magnitude of the error, which leads to the Laplace distribution. For many prob-
lems in Economics and Health sciences, this distribution seems to model the data better than the standard
Gaussian distribution

References
[R190], [R191], [R192], [R193]

Examples
Draw samples from the distribution

>>> loc, scale = 0., 1.

>>> s = np.random.laplace(loc, scale, 1000)

Display the histogram of the samples, along with the probability density function:

620 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 30, normed=True)
>>> x = np.arange(-8., 8., .01)
>>> pdf = np.exp(-abs(x-loc/scale))/(2.*scale)
>>> plt.plot(x, pdf)

Plot Gaussian for comparison:

>>> g = (1/(scale * np.sqrt(2 * np.pi)) *

... np.exp( - (x - loc)**2 / (2 * scale**2) ))
>>> plt.plot(x,g)

0.5

0.4

0.3

0.2

0.1

0.0
8 6 4 2 0 2 4 6 8

logistic(loc=0.0, scale=1.0, size=None)

Draw samples from a Logistic distribution.
Samples are drawn from a Logistic distribution with specified parameters, loc (location or mean, also
median), and scale (>0).
Parameters
loc : float
scale : float > 0.
size : {tuple, int}
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : {ndarray, scalar}
where the values are all integers in [0, n].
See Also:

scipy.stats.distributions.logistic
probability density function, distribution or cumulative density function, etc.

3.8. Random sampling (numpy.random) 621

NumPy Reference, Release 2.0.0.dev8464

Notes
The probability density for the Logistic distribution is

e(x)/s
P (x) = P (x) = ,
s(1 + e(x)/s )2

where = location and s = scale.

The Logistic distribution is used in Extreme Value problems where it can act as a mixture of Gumbel
distributions, in Epidemiology, and by the World Chess Federation (FIDE) where it is used in the Elo
ranking system, assuming the performance of each player is a logistically distributed random variable.

References
[R194], [R195], [R196]

Examples
Draw samples from the distribution:

>>> loc, scale = 10, 1

>>> s = np.random.logistic(loc, scale, 10000)
>>> count, bins, ignored = plt.hist(s, bins=50)

# plot against distribution

>>> def logist(x, loc, scale):

... return exp((loc-x)/scale)/(scale*(1+exp((loc-x)/scale))**2)
>>> plt.plot(bins, logist(bins, loc, scale)*count.max()/\
... logist(bins, loc, scale).max())
>>> plt.show()

lognormal(mean=0.0, sigma=1.0, size=None)

Return samples drawn from a log-normal distribution.
Draw samples from a log-normal distribution with specified mean, standard deviation, and shape. Note
that the mean and standard deviation are not the values for the distribution itself, but of the underlying
normal distribution it is derived from.
Parameters
mean : float
Mean value of the underlying normal distribution
sigma : float, >0.
Standard deviation of the underlying normal distribution
size : tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
See Also:

scipy.stats.lognorm
probability density function, distribution, cumulative density function, etc.

622 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notes
A variable x has a log-normal distribution if log(x) is normally distributed.
The probability density function for the log-normal distribution is
1 (ln(x))2
p(x) = e( 22 )
x 2

where is the mean and is the standard deviation of the normally distributed logarithm of the variable.
A log-normal distribution results if a random variable is the product of a large number of independent,
identically-distributed variables in the same way that a normal distribution results if the variable is the sum
of a large number of independent, identically-distributed variables (see the last example). It is one of the
so-called fat-tailed distributions.
The log-normal distribution is commonly used to model the lifespan of units with fatigue-stress failure
modes. Since this includes most mechanical systems, the log-normal distribution has widespread applica-
tion.
It is also commonly used to model oil field sizes, species abundance, and latent periods of infectious
diseases.

References
[R197], [R198], [R199]

Examples
Draw samples from the distribution:

>>> mu, sigma = 3., 1. # mean and standard deviation

>>> s = np.random.lognormal(mu, sigma, 1000)

Display the histogram of the samples, along with the probability density function:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 100, normed=True, align=mid)

>>> x = np.linspace(min(bins), max(bins), 10000)

>>> pdf = (np.exp(-(np.log(x) - mu)**2 / (2 * sigma**2))
... / (x * sigma * np.sqrt(2 * np.pi)))

>>> plt.plot(x, pdf, linewidth=2, color=r)

>>> plt.axis(tight)
>>> plt.show()

Demonstrate that taking the products of random samples from a uniform distribution can be fit well by a
log-normal probability density function.

>>> # Generate a thousand samples: each is the product of 100 random

>>> # values, drawn from a normal distribution.
>>> b = []
>>> for i in range(1000):
... a = 10. + np.random.random(100)
... b.append(np.product(a))

3.8. Random sampling (numpy.random) 623

NumPy Reference, Release 2.0.0.dev8464

>>> b = np.array(b) / np.min(b) # scale values to be positive

>>> count, bins, ignored = plt.hist(b, 100, normed=True, align=center)

>>> sigma = np.std(np.log(b))

>>> mu = np.mean(np.log(b))

>>> x = np.linspace(min(bins), max(bins), 10000)

>>> pdf = (np.exp(-(np.log(x) - mu)**2 / (2 * sigma**2))
... / (x * sigma * np.sqrt(2 * np.pi)))

>>> plt.plot(x, pdf, color=r, linewidth=2)

>>> plt.show()

0.035
0.030
0.025
0.020
0.015
0.010
0.005
0.000
50 100 150 200 250 300

logseries(p, size=None)
Draw samples from a Logarithmic Series distribution.
Samples are drawn from a Log Series distribution with specified parameter, p (probability, 0 < p < 1).
Parameters
loc : float
scale : float > 0.
size : {tuple, int}
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : {ndarray, scalar}
where the values are all integers in [0, n].
See Also:

624 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

scipy.stats.distributions.logser
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Log Series distribution is

pk
P (k) = ,
k ln(1 p)

where p = probability.
The Log Series distribution is frequently used to represent species richness and occurrence, first proposed
by Fisher, Corbet, and Williams in 1943 [2]. It may also be used to model the numbers of occupants seen
in cars [3].

References
[R200], [R201], [R202], [R203]

Examples
Draw samples from the distribution:

>>> a = .6
>>> s = np.random.logseries(a, 10000)
>>> count, bins, ignored = plt.hist(s)

# plot against distribution

>>> def logseries(k, p):

... return -p**k/(k*log(1-p))
>>> plt.plot(bins, logseries(bins, a)*count.max()/
logseries(bins, a).max(), r)
>>> plt.show()

multinomial(n, pvals, size=None)

Draw samples from a multinomial distribution.
The multinomial distribution is a multivariate generalisation of the binomial distribution. Take an experi-
ment with one of p possible outcomes. An example of such an experiment is throwing a dice, where the
outcome can be 1 through 6. Each sample drawn from the distribution represents n such experiments. Its
values, X_i = [X_0, X_1, ..., X_p], represent the number of times the outcome was i.
Parameters
n : int
Number of experiments.
pvals : sequence of floats, length p
Probabilities of each of the p different outcomes. These should sum to 1 (however,
the last element is always assumed to account for the remaining probability, as long as
sum(pvals[:-1]) <= 1).
size : tuple of ints
Given a size of (M, N, K), then M*N*K samples are drawn, and the output shape
becomes (M, N, K, p), since each sample has shape (p,).

3.8. Random sampling (numpy.random) 625

NumPy Reference, Release 2.0.0.dev8464

Examples
Throw a dice 20 times:

>>> np.random.multinomial(20, [1/6.]*6, size=1)

array([[4, 1, 7, 5, 2, 1]])

It landed 4 times on 1, once on 2, etc.

Now, throw the dice 20 times, and 20 times again:

>>> np.random.multinomial(20, [1/6.]*6, size=2)

array([[3, 4, 3, 3, 4, 3],
[2, 4, 3, 4, 0, 7]])

For the first run, we threw 3 times 1, 4 times 2, etc. For the second, we threw 2 times 1, 4 times 2, etc.
A loaded dice is more likely to land on number 6:

>>> np.random.multinomial(100, [1/7.]*5)

array([13, 16, 13, 16, 42])

multivariate_normal(mean, cov, [size])

Draw random samples from a multivariate normal distribution.
The multivariate normal, multinormal or Gaussian distribution is a generalisation of the one-dimensional
normal distribution to higher dimensions.
Such a distribution is specified by its mean and covariance matrix, which are analogous to the mean
(average or centre) and variance (standard deviation squared or width) of the one-dimensional normal
distribution.
Parameters
mean : (N,) ndarray
Mean of the N-dimensional distribution.
cov : (N,N) ndarray
Covariance matrix of the distribution.
size : tuple of ints, optional
Given a shape of, for example, (m,n,k), m*n*k samples are generated, and packed in an
m-by-n-by-k arrangement. Because each sample is N-dimensional, the output shape is
(m,n,k,N). If no shape is specified, a single sample is returned.
Returns
out : ndarray
The drawn samples, arranged according to size. If the shape given is (m,n,...), then the
shape of out is is (m,n,...,N).
In other words, each entry out[i,j,...,:] is an N-dimensional value drawn from
the distribution.

Notes
The mean is a coordinate in N-dimensional space, which represents the location where samples are most
likely to be generated. This is analogous to the peak of the bell curve for the one-dimensional or univariate
normal distribution.

626 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Covariance indicates the level to which two variables vary together. From the multivariate normal distri-
bution, we draw N-dimensional samples, X = [x1 , x2 , ...xN ]. The covariance matrix element Cij is the
covariance of xi and xj . The element Cii is the variance of xi (i.e. its spread).
Instead of specifying the full covariance matrix, popular approximations include:
Spherical covariance (cov is a multiple of the identity matrix)
Diagonal covariance (cov has non-negative elements, and only on the diagonal)
This geometrical property can be seen in two dimensions by plotting generated data-points:

>>> mean = [0,0]

>>> cov = [[1,0],[0,100]] # diagonal covariance, points lie on x or y-axis

>>> import matplotlib.pyplot as plt

>>> x,y = np.random.multivariate_normal(mean,cov,5000).T
>>> plt.plot(x,y,x); plt.axis(equal); plt.show()

Note that the covariance matrix must be non-negative definite.

References
[R204], [R205]

Examples

>>> mean = (1,2)

>>> cov = [[1,0],[1,0]]
>>> x = np.random.multivariate_normal(mean,cov,(3,3))
>>> x.shape
(3, 3, 2)

The following is probably true, given that 0.6 is roughly twice the standard deviation:

>>> print list( (x[0,0,:] - mean) < 0.6 )

[True, True]

negative_binomial(n, p, size=None)
Draw samples from a negative_binomial distribution.
Samples are drawn from a negative_Binomial distribution with specified parameters, n trials and p proba-
bility of success where n is an integer > 0 and p is in the interval [0, 1].
Parameters
n : int
Parameter, > 0.
p : float
Parameter, >= 0 and <=1.
size : int or tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : int or ndarray of ints
Drawn samples.

3.8. Random sampling (numpy.random) 627

NumPy Reference, Release 2.0.0.dev8464

Notes
The probability density for the Negative Binomial distribution is
 
N +n1 n
P (N ; n, p) = p (1 p)N ,
n1

where n 1 is the number of successes, p is the probability of success, and N + n 1 is the number of
trials.
The negative binomial distribution gives the probability of n-1 successes and N failures in N+n-1 trials,
and success on the (N+n)th trial.
If one throws a die repeatedly until the third time a 1 appears, then the probability distribution of the
number of non-1s that appear before the third 1 is a negative binomial distribution.

References
[R206], [R207]

Examples
Draw samples from the distribution:
A real world example. A company drills wild-cat oil exploration wells, each with an estimated probability
of success of 0.1. What is the probability of having one success for each successive well, that is what is
the probability of a single success after drilling 5 wells, after 6 wells, etc.?

>>> s = np.random.negative_binomial(1, 0.1, 100000)

>>> for i in range(1, 11):
... probability = sum(s<i) / 100000.
... print i, "wells drilled, probability of one success =", probability

noncentral_chisquare(df, nonc, size=None)

Draw samples from a noncentral chi-square distribution.
The noncentral 2 distribution is a generalisation of the 2 distribution.
Parameters
df : int
Degrees of freedom, should be >= 1.
nonc : float
Non-centrality, should be > 0.
size : int or tuple of ints
Shape of the output.

Notes
The probability density function for the noncentral Chi-square distribution is

X enonc/2 (nonc/2)i
P (x; df, nonc) = PYdf +2i (x),
i=0
i!

where Yq is the Chi-square with q degrees of freedom.

628 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

In Delhi (2007), it is noted that the noncentral chi-square is useful in bombing and coverage problems, the
probability of killing the point target given by the noncentral chi-squared distribution.

References
[R208], [R209]

Examples
Draw values from the distribution and plot the histogram

>>> import matplotlib.pyplot as plt

>>> values = plt.hist(np.random.noncentral_chisquare(3, 20, 100000),
... bins=200, normed=True)
>>> plt.show()

Draw values from a noncentral chisquare with very small noncentrality, and compare to a chisquare.

>>> plt.figure()
>>> values = plt.hist(np.random.noncentral_chisquare(3, .0000001, 100000),
... bins=np.arange(0., 25, .1), normed=True)
>>> values2 = plt.hist(np.random.chisquare(3, 100000),
... bins=np.arange(0., 25, .1), normed=True)
>>> plt.plot(values[1][0:-1], values[0]-values2[0], ob)
>>> plt.show()

Demonstrate how large values of non-centrality lead to a more symmetric distribution.

>>> plt.figure()
>>> values = plt.hist(np.random.noncentral_chisquare(3, 20, 100000),
... bins=200, normed=True)
>>> plt.show()

0.05

0.04

0.03

0.02

0.01

0.00
0 10 20 30 40 50 60 70 80

3.8. Random sampling (numpy.random) 629

NumPy Reference, Release 2.0.0.dev8464

0.25

0.20

0.15

0.10

0.05

0.00
0 5 10 15 20 25

0.05

0.04

0.03

0.02

0.01

0.00
0 10 20 30 40 50 60 70 80 90

noncentral_f(dfnum, dfden, nonc, size=None)

Draw samples from the noncentral F distribution.
Samples are drawn from an F distribution with specified parameters, dfnum (degrees of freedom in numer-
ator) and dfden (degrees of freedom in denominator), where both parameters > 1. nonc is the non-centrality
parameter.
Parameters
dfnum : int
Parameter, should be > 1.
dfden : int
Parameter, should be > 1.
nonc : float
Parameter, should be >= 0.
size : int or tuple of ints

630 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : scalar or ndarray
Drawn samples.

Notes
When calculating the power of an experiment (power = probability of rejecting the null hypothesis when
a specific alternative is true) the non-central F statistic becomes important. When the null hypothesis is
true, the F statistic follows a central F distribution. When the null hypothesis is not true, then it follows a
non-central F statistic.

References
Weisstein, Eric W. Noncentral F-Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/NoncentralF-Distribution.html
Wikipedia, Noncentral F distribution, http://en.wikipedia.org/wiki/Noncentral_F-distribution

Examples
In a study, testing for a specific alternative to the null hypothesis requires use of the Noncentral F distribu-
tion. We need to calculate the area in the tail of the distribution that exceeds the value of the F distribution
for the null hypothesis. Well plot the two probability distributions for comparison.

>>> dfnum = 3 # between group deg of freedom

>>> dfden = 20 # within groups degrees of freedom
>>> nonc = 3.0
>>> nc_vals = np.random.noncentral_f(dfnum, dfden, nonc, 1000000)
>>> NF = np.histogram(nc_vals, bins=50, normed=True)
>>> c_vals = np.random.f(dfnum, dfden, 1000000)
>>> F = np.histogram(c_vals, bins=50, normed=True)
>>> plt.plot(F[1][1:], F[0])
>>> plt.plot(NF[1][1:], NF[0])
>>> plt.show()

normal(loc=0.0, scale=1.0, size=None)

Draw random samples from a normal (Gaussian) distribution.
The probability density function of the normal distribution, first derived by De Moivre and 200 years later
by both Gauss and Laplace independently [R211], is often called the bell curve because of its characteristic
shape (see the example below).
The normal distributions occurs often in nature. For example, it describes the commonly occurring dis-
tribution of samples influenced by a large number of tiny, random disturbances, each with its own unique
distribution [R211].
Parameters
loc : float
Mean (centre) of the distribution.
scale : float
Standard deviation (spread or width) of the distribution.
size : tuple of ints

3.8. Random sampling (numpy.random) 631

NumPy Reference, Release 2.0.0.dev8464

Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
See Also:

scipy.stats.distributions.norm
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Gaussian distribution is
1 (x)2
p(x) = e 2 2 ,
2 2

where is the mean and the standard deviation. The square of the standard deviation, 2 , is called the
variance.
The function has its peak at the mean, and its spread increases with the standard deviation (the function
reaches 0.607 times its maximum at x+ and x [R211]). This implies that numpy.random.normal
is more likely to return samples lying close to the mean, rather than those far away.

References
[R210], [R211]

Examples
Draw samples from the distribution:

>>> mu, sigma = 0, 0.1 # mean and standard deviation

>>> s = np.random.normal(mu, sigma, 1000)

Verify the mean and the variance:

>>> abs(mu - np.mean(s)) < 0.01

True

>>> abs(sigma - np.std(s, ddof=1)) < 0.01

True

Display the histogram of the samples, along with the probability density function:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 30, normed=True)
>>> plt.plot(bins, 1/(sigma * np.sqrt(2 * np.pi)) *
... np.exp( - (bins - mu)**2 / (2 * sigma**2) ),
... linewidth=2, color=r)
>>> plt.show()

632 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

0
0.4 0.3 0.2 0.1 0.0 0.1 0.2 0.3

pareto(a, size=None)
Draw samples from a Pareto distribution with specified shape.
This is a simplified version of the Generalized Pareto distribution (available in SciPy), with the scale set to
one and the location set to zero. Most authors default the location to one.
The Pareto distribution must be greater than zero, and is unbounded above. It is also known as the 80-20
rule. In this distribution, 80 percent of the weights are in the lowest 20 percent of the range, while the
other 20 percent fill the remaining 80 percent of the range.
Parameters
shape : float, > 0.
Shape of the distribution.
size : tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
See Also:

scipy.stats.distributions.genpareto.pdf
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Pareto distribution is
ama
p(x) =
xa+1

where a is the shape and m the location

The Pareto distribution, named after the Italian economist Vilfredo Pareto, is a power law probability
distribution useful in many real world problems. Outside the field of economics it is generally referred to
as the Bradford distribution. Pareto developed the distribution to describe the distribution of wealth in an
economy. It has also found use in insurance, web page access statistics, oil field sizes, and many other
problems, including the download frequency for projects in Sourceforge [1]. It is one of the so-called
fat-tailed distributions.

3.8. Random sampling (numpy.random) 633

NumPy Reference, Release 2.0.0.dev8464

References
[R212], [R213], [R214], [R215]

Examples
Draw samples from the distribution:

>>> a, m = 3., 1. # shape and mode

>>> s = np.random.pareto(a, 1000) + m

Display the histogram of the samples, along with the probability density function:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 100, normed=True, align=center)
>>> fit = a*m**a/bins**(a+1)
>>> plt.plot(bins, max(count)*fit/max(fit),linewidth=2, color=r)
>>> plt.show()

2.5

2.0

1.5

1.0

0.5

0.0
0 2 4 6 8 10 12 14 16 18

permutation(x)
Randomly permute a sequence, or return a permuted range.
Parameters
x : int or array_like
If x is an integer, randomly permute np.arange(x). If x is an array, make a copy and
shuffle the elements randomly.
Returns
out : ndarray
Permuted sequence or array range.

Examples

>>> np.random.permutation(10)
array([1, 7, 4, 3, 0, 9, 2, 5, 8, 6])

634 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.random.permutation([1, 4, 9, 12, 15])

array([15, 1, 9, 4, 12])

poisson(lam=1.0, size=None)
Draw samples from a Poisson distribution.
The Poisson distribution is the limit of the Binomial distribution for large N.
Parameters
lam : float
Expectation of interval, should be >= 0.
size : int or tuple of ints, optional
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.

Notes
The Poisson distribution
k e
f (k; ) =
k!

For events with an expected separation the Poisson distribution f (k; ) describes the probability of k
events occurring within the observed interval .

References
[R216], [R217]

Examples
Draw samples from the distribution:

>>> import numpy as np

>>> s = np.random.poisson(5, 10000)

Display histogram of the sample:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 14, normed=True)
>>> plt.show()

3.8. Random sampling (numpy.random) 635

NumPy Reference, Release 2.0.0.dev8464

0.30
0.25
0.20
0.15
0.10
0.05
0.00
0 2 4 6 8 10 12 14 16 18

power(a, size=None)
Draws samples in [0, 1] from a power distribution with positive exponent a - 1.
Also known as the power function distribution.
Parameters
a : float
parameter, > 0
size : tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then
m * n * k samples are drawn.
Returns
samples : {ndarray, scalar}
The returned samples lie in [0, 1].
Raises
ValueError :
If a<1.

Notes
The probability density function is

P (x; a) = axa1 , 0 x 1, a > 0.

The power function distribution is just the inverse of the Pareto distribution. It may also be seen as a special
case of the Beta distribution.
It is used, for example, in modeling the over-reporting of insurance claims.

References
[R218], [R219]

636 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples
rand(d0, d1, ..., dn)
Random values in a given shape.
Create an array of the given shape and propagate it with random samples from a uniform distribution over
[0, 1).
Parameters
d0, d1, ..., dn : int
Shape of the output.
Returns
out : ndarray, shape (d0, d1, ..., dn)
Random values.
See Also:
random

Notes
This is a convenience function. If you want an interface that takes a shape-tuple as the first argument, refer
to random.

Examples

>>> np.random.rand(3,2)
array([[ 0.14022471, 0.96360618], #random
[ 0.37601032, 0.25528411], #random
[ 0.49313049, 0.94909878]]) #random

randint(low, high=None, size=None)

Return random integers from low (inclusive) to high (exclusive).
Return random integers from the discrete uniform distribution in the half-open interval [low, high). If
high is None (the default), then results are from [0, low).
Parameters
low : int
Lowest (signed) integer to be drawn from the distribution (unless high=None, in
which case this parameter is the highest such integer).
high : int, optional
If provided, one above the largest (signed) integer to be drawn from the distribution (see
above for behavior if high=None).
size : int or tuple of ints, optional
Output shape. Default is None, in which case a single int is returned.
Returns
out : int or ndarray of ints
size-shaped array of random integers from the appropriate distribution, or a single such
random int if size not provided.
See Also:

3.8. Random sampling (numpy.random) 637

NumPy Reference, Release 2.0.0.dev8464

random.random_integers
similar to randint, only for the closed interval [low, high], and 1 is the lowest value if high is omitted.
In particular, this other one is the one to use to generate uniformly distributed discrete non-integers.

Examples

>>> np.random.randint(2, size=10)

array([1, 0, 0, 0, 1, 1, 0, 0, 1, 0])
>>> np.random.randint(1, size=10)
array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0])

Generate a 2 x 4 array of ints between 0 and 4, inclusive:

>>> np.random.randint(5, size=(2, 4))

array([[4, 0, 2, 1],
[3, 2, 2, 0]])

randn([d1, ..., dn])

Return a sample (or samples) from the standard normal distribution.
If positive, int_like or int-convertible arguments are provided, randn generates an array of shape (d1,
..., dn), filled with random floats sampled from a univariate normal (Gaussian) distribution of mean
0 and variance 1 (if any of the di are floats, they are first converted to integers by truncation). A single
float randomly sampled from the distribution is returned if no argument is provided.
This is a convenience function. If you want an interface that takes a tuple as the first argument, use
numpy.random.standard_normal instead.
Parameters
d1, ..., dn : n ints, optional
The dimensions of the returned array, should be all positive.
Returns
Z : ndarray or float
A (d1, ..., dn)-shaped array of floating-point samples from the standard normal
distribution, or a single such float if no parameters were supplied.
See Also:

random.standard_normal
Similar, but takes a tuple as its argument.

Notes
For random samples from N (, 2 ), use:
sigma * np.random.randn(...) + mu

Examples

>>> np.random.randn()
2.1923875335537315 #random

Two-by-four array of samples from N(3, 6.25):

638 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> 2.5 * np.random.randn(2, 4) + 3

array([[-4.49401501, 4.00950034, -1.81814867, 7.29718677], #random
[ 0.39924804, 4.68456316, 4.99394529, 4.84057254]]) #random

random_integers(low, high=None, size=None)

Return random integers between low and high, inclusive.
Return random integers from the discrete uniform distribution in the closed interval [low, high]. If high
is None (the default), then results are from [1, low].
Parameters
low : int
Lowest (signed) integer to be drawn from the distribution (unless high=None, in
which case this parameter is the highest such integer).
high : int, optional
If provided, the largest (signed) integer to be drawn from the distribution (see above for
behavior if high=None).
size : int or tuple of ints, optional
Output shape. Default is None, in which case a single int is returned.
Returns
out : int or ndarray of ints
size-shaped array of random integers from the appropriate distribution, or a single such
random int if size not provided.
See Also:

random.randint
Similar to random_integers, only for the half-open interval [low, high), and 0 is the lowest value if
high is omitted.

Notes
To sample from N evenly spaced floating-point numbers between a and b, use:

a + (b - a) * (np.random.random_integers(N) - 1) / (N - 1.)

Examples

>>> np.random.random_integers(5)
4
>>> type(np.random.random_integers(5))
<type int>
>>> np.random.random_integers(5, size=(3.,2.))
array([[5, 4],
[3, 3],
[4, 5]])

Choose five random numbers from the set of five evenly-spaced numbers between 0 and 2.5, inclusive (i.e.,
from the set 0, 5/8, 10/8, 15/8, 20/8):

>>> 2.5 * (np.random.random_integers(5, size=(5,)) - 1) / 4.

array([ 0.625, 1.25 , 0.625, 0.625, 2.5 ])

3.8. Random sampling (numpy.random) 639

NumPy Reference, Release 2.0.0.dev8464

Roll two six sided dice 1000 times and sum the results:

>>> d1 = np.random.random_integers(1, 6, 1000)

>>> d2 = np.random.random_integers(1, 6, 1000)
>>> dsums = d1 + d2

Display results as a histogram:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(dsums, 11, normed=True)
>>> plt.show()

0.20

0.15

0.10

0.05

0.00
2 4 6 8 10 12

random_sample(size=None)
Return random floats in the half-open interval [0.0, 1.0).
Results are from the continuous uniform distribution over the stated interval. To sample U nif [a, b), b >
a multiply the output of random_sample by (b-a) and add a:

(b - a) * random_sample() + a

Parameters
size : int or tuple of ints, optional
Defines the shape of the returned array of random floats. If None (the default), returns
a single float.
Returns
out : float or ndarray of floats
Array of random floats of shape size (unless size=None, in which case a single float
is returned).

Examples

>>> np.random.random_sample()
0.47108547995356098
>>> type(np.random.random_sample())
<type float>

640 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.random.random_sample((5,))
array([ 0.30220482, 0.86820401, 0.1654503 , 0.11659149, 0.54323428])

Three-by-two array of random numbers from [-5, 0):

>>> 5 * np.random.random_sample((3, 2)) - 5

array([[-3.99149989, -0.52338984],
[-2.99091858, -0.79479508],
[-1.23204345, -1.75224494]])

rayleigh(scale=1.0, size=None)
Draw samples from a Rayleigh distribution.
The and Weibull distributions are generalizations of the Rayleigh.
Parameters
scale : scalar
Scale, also equals the mode. Should be >= 0.
size : int or tuple of ints, optional
Shape of the output. Default is None, in which case a single value is returned.

Notes
The probability density function for the Rayleigh distribution is
x x2
P (x; scale) = e 2scale2
scale2

The Rayleigh distribution arises if the wind speed and wind direction are both gaussian variables, then
the vector wind velocity forms a Rayleigh distribution. The Rayleigh distribution is used to model the
expected output from wind turbines.

References
..[1] Brighton Webs Ltd., Rayleigh Distribution,
http://www.brighton-webs.co.uk/distributions/rayleigh.asp
..[2] Wikipedia, Rayleigh distribution
http://en.wikipedia.org/wiki/Rayleigh_distribution

Examples
Draw values from the distribution and plot the histogram

>>> values = hist(np.random.rayleigh(3, 100000), bins=200, normed=True)

Wave heights tend to follow a Rayleigh distribution. If the mean wave height is 1 meter, what fraction of
waves are likely to be larger than 3 meters?

>>> meanvalue = 1
>>> modevalue = np.sqrt(2 / np.pi) * meanvalue
>>> s = np.random.rayleigh(modevalue, 1000000)

The percentage of waves larger than 3 meters is:

3.8. Random sampling (numpy.random) 641

NumPy Reference, Release 2.0.0.dev8464

>>> 100.*sum(s>3)/1000000.
0.087300000000000003

seed(seed=None)
Seed the generator.
This method is called when RandomState is initialized. It can be called again to re-seed the generator. For
details, see RandomState.
Parameters
seed : int or array_like, optional
Seed for RandomState.
See Also:
RandomState
set_state(state)
Set the internal state of the generator from a tuple.
For use if one has reason to manually (re-)set the internal state of the Mersenne Twister[R220] pseudo-
random number generating algorithm.
Parameters
state : tuple(str, ndarray of 624 uints, int, int, float)
The state tuple has the following items:
1. the string MT19937, specifying the Mersenne Twister algorithm.
2. a 1-D array of 624 unsigned integers keys.
3. an integer pos.
4. an integer has_gauss.
5. a float cached_gaussian.
Returns
out : None
Returns None on success.
See Also:
get_state

Notes
set_state and get_state are not needed to work with any of the random distributions in NumPy. If the
internal state is manually altered, the user should know exactly what he/she is doing.
For backwards compatibility, the form (str, array of 624 uints, int) is also accepted although it is missing
some information about the cached Gaussian value: state = (MT19937, keys, pos).

References
[R220]
shuffle(x)
Modify a sequence in-place by shuffling its contents.

642 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

standard_cauchy(size=None)
Standard Cauchy distribution with mode = 0.
Also known as the Lorentz distribution.
Parameters
size : int or tuple of ints
Shape of the output.
Returns
samples : ndarray or scalar
The drawn samples.

Notes
The probability density function for the full Cauchy distribution is
1
P (x; x0 , ) =
1 + ( xx
 
0 2
)

and the Standard Cauchy distribution just sets x0 = 0 and = 1

The Cauchy distribution arises in the solution to the driven harmonic oscillator problem, and also describes
spectral line broadening. It also describes the distribution of values at which a line tilted at a random angle
will cut the x axis.
When studying hypothesis tests that assume normality, seeing how the tests perform on data from a Cauchy
distribution is a good indicator of their sensitivity to a heavy-tailed distribution, since the Cauchy looks
very much like a Gaussian distribution, but with heavier tails.

References
..[1] NIST/SEMATECH e-Handbook of Statistical Methods, Cauchy
Distribution, http://www.itl.nist.gov/div898/handbook/eda/section3/eda3663.htm
..[2] Weisstein, Eric W. Cauchy Distribution. From MathWorldA
Wolfram Web Resource. http://mathworld.wolfram.com/CauchyDistribution.html
..[3] Wikipedia, Cauchy distribution
http://en.wikipedia.org/wiki/Cauchy_distribution

Examples
Draw samples and plot the distribution:

>>> s = np.random.standard_cauchy(1000000)
>>> s = s[(s>-25) & (s<25)] # truncate distribution so it plots well
>>> plt.hist(s, bins=100)
>>> plt.show()

standard_exponential(size=None)
Draw samples from the standard exponential distribution.
standard_exponential is identical to the exponential distribution with a scale parameter of 1.
Parameters
size : int or tuple of ints
Shape of the output.

3.8. Random sampling (numpy.random) 643

NumPy Reference, Release 2.0.0.dev8464

Returns
out : float or ndarray
Drawn samples.

Examples
Output a 3x8000 array:

>>> n = np.random.standard_exponential((3, 8000))

standard_gamma(shape, size=None)
Draw samples from a Standard Gamma distribution.
Samples are drawn from a Gamma distribution with specified parameters, shape (sometimes designated
k) and scale=1.
Parameters
shape : float
Parameter, should be > 0.
size : int or tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : ndarray or scalar
The drawn samples.
See Also:

scipy.stats.distributions.gamma
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Gamma distribution is

ex/
p(x) = xk1 ,
k (k)

where k is the shape and the scale, and is the Gamma function.
The Gamma distribution is often used to model the times to failure of electronic components, and arises
naturally in processes for which the waiting times between Poisson distributed events are relevant.

References
[R221], [R222]

Examples
standard_normal(size=None)
Returns samples from a Standard Normal distribution (mean=0, stdev=1).
Parameters
size : int or tuple of ints, optional
Output shape. Default is None, in which case a single value is returned.

644 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Returns
out : float or ndarray
Drawn samples.

Examples

>>> s = np.random.standard_normal(8000)
>>> s
array([ 0.6888893 , 0.78096262, -0.89086505, ..., 0.49876311, #random
-0.38672696, -0.4685006 ]) #random
>>> s.shape
(8000,)
>>> s = np.random.standard_normal(size=(3, 4, 2))
>>> s.shape
(3, 4, 2)

standard_t(df, size=None)
Standard Students t distribution with df degrees of freedom.
A special case of the hyperbolic distribution. As df gets large, the result resembles that of the standard
normal distribution (standard_normal).
Parameters
df : int
Degrees of freedom, should be > 0.
size : int or tuple of ints, optional
Output shape. Default is None, in which case a single value is returned.
Returns
samples : ndarray or scalar
Drawn samples.

Notes
The probability density function for the t distribution is

( df2+1 )  x2 (df +1)/2

P (x, df ) = df
1+
df ( 2 ) df

The t test is based on an assumption that the data come from a Normal distribution. The t test provides a
way to test whether the sample mean (that is the mean calculated from the data) is a good estimate of the
true mean.
The derivation of the t-distribution was forst published in 1908 by William Gisset while working for the
Guinness Brewery in Dublin. Due to proprietary issues, he had to publish under a pseudonym, and so he
used the name Student.

References
[R223], [R224]

Examples
tomaxint(size=None)
Uniformly sample discrete random integers x such that 0 <= x <= sys.maxint.

3.8. Random sampling (numpy.random) 645

NumPy Reference, Release 2.0.0.dev8464

Parameters
size : tuple of ints, int, optional
Shape of output. If the given size is, for example, (m,n,k), m*n*k samples are generated.
If no shape is specified, a single sample is returned.
Returns
out : ndarray
Drawn samples, with shape size.
See Also:

randint
Uniform sampling over a given half-open interval of integers.
random_integers
Uniform sampling over a given closed interval of integers.

triangular(left, mode, right, size=None)

Draw samples from the triangular distribution.
The triangular distribution is a continuous probability distribution with lower limit left, peak at mode, and
upper limit right. Unlike the other distributions, these parameters directly define the shape of the pdf.
Parameters
left : scalar
Lower limit.
mode : scalar
The value where the peak of the distribution occurs. The value should fulfill the condi-
tion left <= mode <= right.
right : scalar
Upper limit, should be larger than left.
size : int or tuple of ints, optional
Output shape. Default is None, in which case a single value is returned.
Returns
samples : ndarray or scalar
The returned samples all lie in the interval [left, right].

Notes
The probability density function for the Triangular distribution is
2(xl)
(rl)(ml) for l x m,

2(mx)
P (x; l, m, r) = (rl)(rm) for m x r,

0 otherwise.

The triangular distribution is often used in ill-defined problems where the underlying distribution is not
known, but some knowledge of the limits and mode exists. Often it is used in simulations.

646 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

References
..[1] Wikipedia, Triangular distribution
http://en.wikipedia.org/wiki/Triangular_distribution

Examples
Draw values from the distribution and plot the histogram:

>>> import matplotlib.pyplot as plt

>>> h = plt.hist(np.random.triangular(-3, 0, 8, 100000), bins=200,
... normed=True)
>>> plt.show()

0.20

0.15

0.10

0.05

0.00
4 2 0 2 4 6 8

uniform(low=0.0, high=1.0, size=1)

Draw samples from a uniform distribution.
Samples are uniformly distributed over the half-open interval [low, high) (includes low, but excludes
high). In other words, any value within the given interval is equally likely to be drawn by uniform.
Parameters
low : float, optional
Lower boundary of the output interval. All values generated will be greater than or
equal to low. The default value is 0.
high : float
Upper boundary of the output interval. All values generated will be less than high. The
default value is 1.0.
size : tuple of ints, int, optional
Shape of output. If the given size is, for example, (m,n,k), m*n*k samples are generated.
If no shape is specified, a single sample is returned.
Returns
out : ndarray
Drawn samples, with shape size.
See Also:

3.8. Random sampling (numpy.random) 647

NumPy Reference, Release 2.0.0.dev8464

randint
Discrete uniform distribution, yielding integers.
random_integers
Discrete uniform distribution over the closed interval [low, high].
random_sample
Floats uniformly distributed over [0, 1).
random
Alias for random_sample.
rand
Convenience function that accepts dimensions as input, e.g., rand(2,2) would generate a 2-by-2
array of floats, uniformly distributed over [0, 1).

Notes
The probability density function of the uniform distribution is
1
p(x) =
ba

anywhere within the interval [a, b), and zero elsewhere.

Examples
Draw samples from the distribution:

>>> s = np.random.uniform(-1,0,1000)

All values are within the given interval:

>>> np.all(s >= -1)

True

>>> np.all(s < 0)

True

Display the histogram of the samples, along with the probability density function:

>>> import matplotlib.pyplot as plt

>>> count, bins, ignored = plt.hist(s, 15, normed=True)
>>> plt.plot(bins, np.ones_like(bins), linewidth=2, color=r)
>>> plt.show()

648 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

1.2
1.0
0.8
0.6
0.4
0.2
0.0
1.0 0.8 0.6 0.4 0.2 0.0

vonmises(mu=0.0, kappa=1.0, size=None)

Draw samples from a von Mises distribution.
Samples are drawn from a von Mises distribution with specified mode (mu) and dispersion (kappa), on the
interval [-pi, pi].
The von Mises distribution (also known as the circular normal distribution) is a continuous probability
distribution on the circle. It may be thought of as the circular analogue of the normal distribution.
Parameters
mu : float
Mode (center) of the distribution.
kappa : float, >= 0.
Dispersion of the distribution.
size : {tuple, int}
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : {ndarray, scalar}
The returned samples live on the unit circle [-pi, pi].
See Also:

scipy.stats.distributions.vonmises
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the von Mises distribution is

ecos(x)
p(x) = ,
2I0 ()

3.8. Random sampling (numpy.random) 649

NumPy Reference, Release 2.0.0.dev8464

where is the mode and the dispersion, and I0 () is the modified Bessel function of order 0.
The von Mises, named for Richard Edler von Mises, born in Austria-Hungary, in what is now the Ukraine.
He fled to the United States in 1939 and became a professor at Harvard. He worked in probability theory,
aerodynamics, fluid mechanics, and philosophy of science.

References
[R225], [R226], [R227]

Examples
wald(mean, scale, size=None)
Draw samples from a Wald, or Inverse Gaussian, distribution.
As the scale approaches infinity, the distribution becomes more like a Gaussian.
Some references claim that the Wald is an Inverse Gaussian with mean=1, but this is by no means universal.
The Inverse Gaussian distribution was first studied in relationship to Brownian motion. In 1956 M.C.K.
Tweedie used the name Inverse Gaussian because there is an inverse relationship between the time to cover
a unit distance and distance covered in unit time.
Parameters
mean : scalar
Distribution mean, should be > 0.
scale : scalar
Scale parameter, should be >= 0.
size : int or tuple of ints, optional
Output shape. Default is None, in which case a single value is returned.
Returns
samples : ndarray or scalar
Drawn sample, all greater than zero.

Notes
The probability density function for the Wald distribution is
r
scale scale(xmean)2

P (x; mean, scale) = 3

e 2mean2 x
2x

As noted above the Inverse Gaussian distribution first arise from attempts to model Brownian Motion. It
is also a competitor to the Weibull for use in reliability modeling and modeling stock returns and interest
rate processes.

References
..[1] Brighton Webs Ltd., Wald Distribution,
http://www.brighton-webs.co.uk/distributions/wald.asp
..[2] Chhikara, Raj S., and Folks, J. Leroy, The Inverse Gaussian
Distribution: Theory : Methodology, and Applications, CRC Press, 1988.
..[3] Wikipedia, Wald distribution
http://en.wikipedia.org/wiki/Wald_distribution

650 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples
Draw values from the distribution and plot the histogram:

>>> import matplotlib.pyplot as plt

>>> h = plt.hist(np.random.wald(3, 2, 100000), bins=200, normed=True)
>>> plt.show()

0.45
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
0 10 20 30 40 50 60 70 80

weibull(a, size=None)
Weibull distribution.
Draw samples from a 1-parameter Weibull distribution with the given shape parameter.

X = (ln(U ))1/a

Here, U is drawn from the uniform distribution over (0,1].

The more common 2-parameter Weibull, including a scale parameter is just X = (ln(U ))1/a .
The Weibull (or Type III asymptotic extreme value distribution for smallest values, SEV Type III, or
Rosin-Rammler distribution) is one of a class of Generalized Extreme Value (GEV) distributions used in
modeling extreme value problems. This class includes the Gumbel and Frechet distributions.
Parameters
a : float
Shape of the distribution.
size : tuple of ints
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
See Also:

scipy.stats.distributions.weibull
probability density function, distribution or cumulative density function, etc.

gumbel, scipy.stats.distributions.genextreme

3.8. Random sampling (numpy.random) 651

NumPy Reference, Release 2.0.0.dev8464

Notes
The probability density for the Weibull distribution is
a x a1 (x/)a
p(x) = ( ) e ,

where a is the shape and the scale.

The function has its peak (the mode) at ( a1
a )
1/a
.
When a = 1, the Weibull distribution reduces to the exponential distribution.

References
[R228], [R229], [R230]

Examples
Draw samples from the distribution:

>>> a = 5. # shape
>>> s = np.random.weibull(a, 1000)

Display the histogram of the samples, along with the probability density function:

>>> import matplotlib.pyplot as plt

>>> x = np.arange(1,100.)/50.
>>> def weib(x,n,a):
... return (a / n) * (x / n)**(a - 1) * np.exp(-(x / n)**a)

>>> count, bins, ignored = plt.hist(np.random.weibull(5.,1000))

>>> x = np.arange(1,100.)/50.
>>> scale = count.max()/weib(x, 1., 5.).max()
>>> plt.plot(x, weib(x, 1., 5.)*scale)
>>> plt.show()

250

200

150

100

50

0
0.0 0.5 1.0 1.5 2.0

652 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

zipf(a, size=None)
Draw samples from a Zipf distribution.
Samples are drawn from a Zipf distribution with specified parameter (a), where a > 1.
The zipf distribution (also known as the zeta distribution) is a continuous probability distribution that
satisfies Zipfs law, where the frequency of an item is inversely proportional to its rank in a frequency
table.
Parameters
a : float
parameter, > 1.
size : {tuple, int}
Output shape. If the given shape is, e.g., (m, n, k), then m * n * k samples are
drawn.
Returns
samples : {ndarray, scalar}
The returned samples are greater than or equal to one.
See Also:

scipy.stats.distributions.zipf
probability density function, distribution or cumulative density function, etc.

Notes
The probability density for the Zipf distribution is

xa
p(x) = ,
(a)

where is the Riemann Zeta function.

Named after the American linguist George Kingsley Zipf, who noted that the frequency of any word in a
sample of a language is inversely proportional to its rank in the frequency table.

References
[R231], [R232], [R233], [R234]

Examples
seed(seed=None)
Seed the generator.
This method is called when RandomState is initialized. It can be called again to re-seed the generator. For
details, see RandomState.
Parameters
seed : int or array_like, optional
Seed for RandomState.
See Also:
RandomState

3.8. Random sampling (numpy.random) 653

NumPy Reference, Release 2.0.0.dev8464

get_state()
Return a tuple representing the internal state of the generator.
For more details, see set_state.
Returns
out : tuple(str, ndarray of 624 uints, int, int, float)
The returned tuple has the following items:
1. the string MT19937.
2. a 1-D array of 624 unsigned integer keys.
3. an integer pos.
4. an integer has_gauss.
5. a float cached_gaussian.
See Also:
set_state

Notes
set_state and get_state are not needed to work with any of the random distributions in NumPy. If the internal
state is manually altered, the user should know exactly what he/she is doing.
set_state(state)
Set the internal state of the generator from a tuple.
For use if one has reason to manually (re-)set the internal state of the Mersenne Twister[R252] pseudo-random
number generating algorithm.
Parameters
state : tuple(str, ndarray of 624 uints, int, int, float)
The state tuple has the following items:
1. the string MT19937, specifying the Mersenne Twister algorithm.
2. a 1-D array of 624 unsigned integers keys.
3. an integer pos.
4. an integer has_gauss.
5. a float cached_gaussian.
Returns
out : None
Returns None on success.
See Also:
get_state

Notes
set_state and get_state are not needed to work with any of the random distributions in NumPy. If the internal
state is manually altered, the user should know exactly what he/she is doing.
For backwards compatibility, the form (str, array of 624 uints, int) is also accepted although it is missing some
information about the cached Gaussian value: state = (MT19937, keys, pos).

654 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

References
[R252]

3.9 Sorting and searching

3.9.1 Sorting

sort(a[, axis, kind, order]) Return a sorted copy of an array.

lexsort(keys[, axis]) Perform an indirect sort using a sequence of keys.
argsort(a[, axis, kind, order]) Returns the indices that would sort an array.
ndarray.sort([axis, kind, order]) Sort an array, in-place.
msort(a) Return a copy of an array sorted along the first axis.
sort_complex(a) Sort a complex array using the real part first, then the imaginary part.
sort(a, axis=-1, kind=quicksort, order=None)
Return a sorted copy of an array.
Parameters
a : array_like
Array to be sorted.
axis : int or None, optional
Axis along which to sort. If None, the array is flattened before sorting. The default is
-1, which sorts along the last axis.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm. Default is quicksort.
order : list, optional
When a is a structured array, this argument specifies which fields to compare first, sec-
ond, and so on. This list does not need to include all of the fields.
Returns
sorted_array : ndarray
Array of the same type and shape as a.
See Also:

ndarray.sort
Method to sort an array in-place.
argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in a sorted array.

3.9. Sorting and searching 655

NumPy Reference, Release 2.0.0.dev8464

Notes
The various sorting algorithms are characterized by their average speed, worst case performance, work space
size, and whether they are stable. A stable sort keeps items with the same key in the same relative order. The
three available algorithms have the following properties:
kind speed worst case work space stable
quicksort 1 O(n^2) 0 no
mergesort 2 O(n*log(n)) ~n/2 yes
heapsort 3 O(n*log(n)) 0 no
All the sort algorithms make temporary copies of the data when sorting along any but the last axis. Consequently,
sorting along the last axis is faster and uses less space than sorting along any other axis.
The sort order for complex numbers is lexicographic. If both the real and imaginary parts are non-nan then the
order is determined by the real parts except when they are equal, in which case the order is determined by the
imaginary parts.
Previous to numpy 1.4.0 sorting real and complex arrays containing nan values led to undefined behaviour. In
numpy versions >= 1.4.0 nan values are sorted to the end. The extended sort order is:
Real: [R, nan]
Complex: [R + Rj, R + nanj, nan + Rj, nan + nanj]
where R is a non-nan real value. Complex values with the same nan placements are sorted according to the
non-nan part if it exists. Non-nan values are sorted as before.

Examples

>>> a = np.array([[1,4],[3,1]])
>>> np.sort(a) # sort along the last axis
array([[1, 4],
[1, 3]])
>>> np.sort(a, axis=None) # sort the flattened array
array([1, 1, 3, 4])
>>> np.sort(a, axis=0) # sort along the first axis
array([[1, 1],
[3, 4]])

Use the order keyword to specify a field to use when sorting a structured array:

>>> dtype = [(name, S10), (height, float), (age, int)]

>>> values = [(Arthur, 1.8, 41), (Lancelot, 1.9, 38),
... (Galahad, 1.7, 38)]
>>> a = np.array(values, dtype=dtype) # create a structured array
>>> np.sort(a, order=height)
array([(Galahad, 1.7, 38), (Arthur, 1.8, 41),
(Lancelot, 1.8999999999999999, 38)],
dtype=[(name, |S10), (height, <f8), (age, <i4)])

Sort by age, then height if ages are equal:

>>> np.sort(a, order=[age, height])

array([(Galahad, 1.7, 38), (Lancelot, 1.8999999999999999, 38),
(Arthur, 1.8, 41)],
dtype=[(name, |S10), (height, <f8), (age, <i4)])

lexsort(keys, axis=-1)
Perform an indirect sort using a sequence of keys.

656 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Given multiple sorting keys, which can be interpreted as columns in a spreadsheet, lexsort returns an array of
integer indices that describes the sort order by multiple columns. The last key in the sequence is used for the
primary sort order, the second-to-last key for the secondary sort order, and so on. The keys argument must be
a sequence of objects that can be converted to arrays of the same shape. If a 2D array is provided for the keys
argument, its rows are interpreted as the sorting keys and sorting is according to the last row, second last row
etc.
Parameters
keys : (k,N) array or tuple containing k (N,)-shaped sequences
The k different columns to be sorted. The last column (or row if keys is a 2D array) is
the primary sort key.
axis : int, optional
Axis to be indirectly sorted. By default, sort over the last axis.
Returns
indices : (N,) ndarray of ints
Array of indices that sort the keys along the specified axis.
See Also:

argsort
Indirect sort.
ndarray.sort
In-place sort.
sort
Return a sorted copy of an array.

Examples
Sort names: first by surname, then by name.

>>> surnames = (Hertz, Galilei, Hertz)

>>> first_names = (Heinrich, Galileo, Gustav)
>>> ind = np.lexsort((first_names, surnames))
>>> ind
array([1, 2, 0])

>>> [surnames[i] + ", " + first_names[i] for i in ind]

[Galilei, Galileo, Hertz, Gustav, Hertz, Heinrich]

Sort two columns of numbers:

>>> a = [1,5,1,4,3,4,4] # First column

>>> b = [9,4,0,4,0,2,1] # Second column
>>> ind = np.lexsort((b,a)) # Sort by a, then by b
>>> print ind
[2 0 4 6 5 3 1]

>>> [(a[i],b[i]) for i in ind]

[(1, 0), (1, 9), (3, 0), (4, 1), (4, 2), (4, 4), (5, 4)]

Note that sorting is first according to the elements of a. Secondary sorting is according to the elements of b.
A normal argsort would have yielded:

3.9. Sorting and searching 657

NumPy Reference, Release 2.0.0.dev8464

>>> [(a[i],b[i]) for i in np.argsort(a)]

[(1, 9), (1, 0), (3, 0), (4, 4), (4, 2), (4, 1), (5, 4)]

Structured arrays are sorted lexically by argsort:

>>> x = np.array([(1,9), (5,4), (1,0), (4,4), (3,0), (4,2), (4,1)],

... dtype=np.dtype([(x, int), (y, int)]))

>>> np.argsort(x) # or np.argsort(x, order=(x, y))

array([2, 0, 4, 6, 5, 3, 1])

argsort(a, axis=-1, kind=quicksort, order=None)

Returns the indices that would sort an array.
Perform an indirect sort along the given axis using the algorithm specified by the kind keyword. It returns an
array of indices of the same shape as a that index data along the given axis in sorted order.
Parameters
a : array_like
Array to sort.
axis : int or None, optional
Axis along which to sort. The default is -1 (the last axis). If None, the flattened array is
used.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm.
order : list, optional
When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
Returns
index_array : ndarray, int
Array of indices that sort a along the specified axis. In other words,
a[index_array] yields a sorted a.
See Also:

sort
Describes sorting algorithms used.
lexsort
Indirect stable sort with multiple keys.
ndarray.sort
Inplace sort.

Notes
See sort for notes on the different sorting algorithms.
As of NumPy 1.4.0 argsort works with real/complex arrays containing nan values. The enhanced sort order is
documented in sort.

658 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples
One dimensional array:

>>> x = np.array([3, 1, 2])

>>> np.argsort(x)
array([1, 2, 0])

Two-dimensional array:

>>> x = np.array([[0, 3], [2, 2]])

>>> x
array([[0, 3],
[2, 2]])

>>> np.argsort(x, axis=0)

array([[0, 1],
[1, 0]])

>>> np.argsort(x, axis=1)

array([[0, 1],
[0, 1]])

Sorting with keys:

>>> x = np.array([(1, 0), (0, 1)], dtype=[(x, <i4), (y, <i4)])

>>> x
array([(1, 0), (0, 1)],
dtype=[(x, <i4), (y, <i4)])

>>> np.argsort(x, order=(x,y))

array([1, 0])

>>> np.argsort(x, order=(y,x))

array([0, 1])

sort(axis=-1, kind=quicksort, order=None)

Sort an array, in-place.
Parameters
axis : int, optional
Axis along which to sort. Default is -1, which means sort along the last axis.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm. Default is quicksort.
order : list, optional
When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
See Also:

numpy.sort
Return a sorted copy of an array.

3.9. Sorting and searching 659

NumPy Reference, Release 2.0.0.dev8464

argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in sorted array.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = np.array([[1,4], [3,1]])

>>> a.sort(axis=1)
>>> a
array([[1, 4],
[1, 3]])
>>> a.sort(axis=0)
>>> a
array([[1, 3],
[1, 4]])

Use the order keyword to specify a field to use when sorting a structured array:

>>> a = np.array([(a, 2), (c, 1)], dtype=[(x, S1), (y, int)])

>>> a.sort(order=y)
>>> a
array([(c, 1), (a, 2)],
dtype=[(x, |S1), (y, <i4)])

msort(a)
Return a copy of an array sorted along the first axis.
Parameters
a : array_like
Array to be sorted.
Returns
sorted_array : ndarray
Array of the same type and shape as a.
See Also:
sort

Notes
np.msort(a) is equivalent to np.sort(a, axis=0).
sort_complex(a)
Sort a complex array using the real part first, then the imaginary part.
Parameters
a : array_like
Input array

660 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Returns
out : complex ndarray
Always returns a sorted complex array.

Examples

>>> np.sort_complex([5, 3, 6, 2, 1])

array([ 1.+0.j, 2.+0.j, 3.+0.j, 5.+0.j, 6.+0.j])

>>> np.sort_complex([1 + 2j, 2 - 1j, 3 - 2j, 3 - 3j, 3 + 5j])

array([ 1.+2.j, 2.-1.j, 3.-3.j, 3.-2.j, 3.+5.j])

3.9.2 Searching

argmax(a[, axis]) Indices of the maximum values along an axis.

nanargmax(a[, axis]) Return indices of the maximum values over an axis, ignoring NaNs.
argmin(a[, axis]) Return the indices of the minimum values along an axis.
nanargmin(a[, axis]) Return indices of the minimum values over an axis, ignoring NaNs.
argwhere(a) Find the indices of array elements that are non-zero, grouped by element.
nonzero(a) Return the indices of the elements that are non-zero.
flatnonzero(a) Return indices that are non-zero in the flattened version of a.
where(condition, x) Return elements, either from x or y, depending on condition.
searchsorted(a, v[, side]) Find indices where elements should be inserted to maintain order.
extract(condition, arr) Return the elements of an array that satisfy some condition.
argmax(a, axis=None)
Indices of the maximum values along an axis.
Parameters
a : array_like
Input array.
axis : int, optional
By default, the index is into the flattened array, otherwise along the specified axis.
Returns
index_array : ndarray of ints
Array of indices into the array. It has the same shape as a.shape with the dimension
along axis removed.
See Also:
ndarray.argmax, argmin
amax
The maximum value along a given axis.
unravel_index
Convert a flat index into an index tuple.

Notes
In case of multiple occurrences of the maximum values, the indices corresponding to the first occurrence are
returned.

3.9. Sorting and searching 661

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.arange(6).reshape(2,3)
>>> a
array([[0, 1, 2],
[3, 4, 5]])
>>> np.argmax(a)
5
>>> np.argmax(a, axis=0)
array([1, 1, 1])
>>> np.argmax(a, axis=1)
array([2, 2])

>>> b = np.arange(6)
>>> b[1] = 5
>>> b
array([0, 5, 2, 3, 4, 5])
>>> np.argmax(b) # Only the first occurrence is returned.
1

nanargmax(a, axis=None)
Return indices of the maximum values over an axis, ignoring NaNs.
Parameters
a : array_like
Input data.
axis : int, optional
Axis along which to operate. By default flattened input is used.
Returns
index_array : ndarray
An array of indices or a single index value.
See Also:
argmax, nanargmin

Examples

>>> a = np.array([[np.nan, 4], [2, 3]])

>>> np.argmax(a)
0
>>> np.nanargmax(a)
1
>>> np.nanargmax(a, axis=0)
array([1, 0])
>>> np.nanargmax(a, axis=1)
array([1, 1])

argmin(a, axis=None)
Return the indices of the minimum values along an axis.
See Also:

argmax
Similar function. Please refer to numpy.argmax for detailed documentation.

662 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

nanargmin(a, axis=None)
Return indices of the minimum values over an axis, ignoring NaNs.
Parameters
a : array_like
Input data.
axis : int, optional
Axis along which to operate. By default flattened input is used.
Returns
index_array : ndarray
An array of indices or a single index value.
See Also:
argmin, nanargmax

Examples

>>> a = np.array([[np.nan, 4], [2, 3]])

>>> np.argmin(a)
0
>>> np.nanargmin(a)
2
>>> np.nanargmin(a, axis=0)
array([1, 1])
>>> np.nanargmin(a, axis=1)
array([1, 0])

argwhere(a)
Find the indices of array elements that are non-zero, grouped by element.
Parameters
a : array_like
Input data.
Returns
index_array : ndarray
Indices of elements that are non-zero. Indices are grouped by element.
See Also:
where, nonzero

Notes
np.argwhere(a) is the same as np.transpose(np.nonzero(a)).
The output of argwhere is not suitable for indexing arrays. For this purpose use where(a) instead.

Examples

>>> x = np.arange(6).reshape(2,3)
>>> x
array([[0, 1, 2],
[3, 4, 5]])
>>> np.argwhere(x>1)

3.9. Sorting and searching 663

NumPy Reference, Release 2.0.0.dev8464

array([[0, 2],
[1, 0],
[1, 1],
[1, 2]])

nonzero(a)
Return the indices of the elements that are non-zero.
Returns a tuple of arrays, one for each dimension of a, containing the indices of the non-zero elements in that
dimension. The corresponding non-zero values can be obtained with:

a[nonzero(a)]

To group the indices by element, rather than dimension, use:

transpose(nonzero(a))

The result of this is always a 2-D array, with a row for each non-zero element.
Parameters
a : array_like
Input array.
Returns
tuple_of_arrays : tuple
Indices of elements that are non-zero.
See Also:

flatnonzero
Return indices that are non-zero in the flattened version of the input array.
ndarray.nonzero
Equivalent ndarray method.

Examples

>>> x = np.eye(3)
>>> x
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])
>>> np.nonzero(x)
(array([0, 1, 2]), array([0, 1, 2]))

>>> x[np.nonzero(x)]
array([ 1., 1., 1.])
>>> np.transpose(np.nonzero(x))
array([[0, 0],
[1, 1],
[2, 2]])

A common use for nonzero is to find the indices of an array, where a condition is True. Given an array a, the
condition a > 3 is a boolean array and since False is interpreted as 0, np.nonzero(a > 3) yields the indices of the
a where the condition is true.

664 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> a = np.array([[1,2,3],[4,5,6],[7,8,9]])
>>> a > 3
array([[False, False, False],
[ True, True, True],
[ True, True, True]], dtype=bool)
>>> np.nonzero(a > 3)
(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

The nonzero method of the boolean array can also be called.

>>> (a > 3).nonzero()

(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

flatnonzero(a)
Return indices that are non-zero in the flattened version of a.
This is equivalent to a.ravel().nonzero()[0].
Parameters
a : ndarray
Input array.
Returns
res : ndarray
Output array, containing the indices of the elements of a.ravel() that are non-zero.
See Also:

nonzero
Return the indices of the non-zero elements of the input array.
ravel
Return a 1-D array containing the elements of the input array.

Examples

>>> x = np.arange(-2, 3)
>>> x
array([-2, -1, 0, 1, 2])
>>> np.flatnonzero(x)
array([0, 1, 3, 4])

Use the indices of the non-zero elements as an index array to extract these elements:

>>> x.ravel()[np.flatnonzero(x)]
array([-2, -1, 1, 2])

where(condition, [x, y])

Return elements, either from x or y, depending on condition.
If only condition is given, return condition.nonzero().
Parameters
condition : array_like, bool
When True, yield x, otherwise yield y.
x, y : array_like, optional

3.9. Sorting and searching 665

NumPy Reference, Release 2.0.0.dev8464

Values from which to choose. x and y need to have the same shape as condition.
Returns
out : ndarray or tuple of ndarrays
If both x and y are specified, the output array contains elements of x where condition is
True, and elements from y elsewhere.
If only condition is given, return the tuple condition.nonzero(), the indices
where condition is True.
See Also:
nonzero, choose

Notes
If x and y are given and input arrays are 1-D, where is equivalent to:

[xv if c else yv for (c,xv,yv) in zip(condition,x,y)]

Examples

>>> np.where([[True, False], [True, True]],

... [[1, 2], [3, 4]],
... [[9, 8], [7, 6]])
array([[1, 8],
[3, 4]])

>>> np.where([[0, 1], [1, 0]])

(array([0, 1]), array([1, 0]))

>>> x = np.arange(9.).reshape(3, 3)
>>> np.where( x > 5 )
(array([2, 2, 2]), array([0, 1, 2]))
>>> x[np.where( x > 3.0 )] # Note: result is 1D.
array([ 4., 5., 6., 7., 8.])
>>> np.where(x < 5, x, -1) # Note: broadcasting.
array([[ 0., 1., 2.],
[ 3., 4., -1.],
[-1., -1., -1.]])

searchsorted(a, v, side=left)
Find indices where elements should be inserted to maintain order.
Find the indices into a sorted array a such that, if the corresponding elements in v were inserted before the
indices, the order of a would be preserved.
Parameters
a : 1-D array_like
Input array, sorted in ascending order.
v : array_like
Values to insert into a.
side : {left, right}, optional

666 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

If left, the index of the first suitable location found is given. If right, return the last
such index. If there is no suitable index, return either 0 or N (where N is the length of
a).
Returns
indices : array of ints
Array of insertion points with the same shape as v.
See Also:

sort
Return a sorted copy of an array.
histogram
Produce histogram from 1-D data.

Notes
Binary search is used to find the required insertion points.
As of Numpy 1.4.0 searchsorted works with real/complex arrays containing nan values. The enhanced sort
order is documented in sort.

Examples

>>> np.searchsorted([1,2,3,4,5], 3)
2
>>> np.searchsorted([1,2,3,4,5], 3, side=right)
3
>>> np.searchsorted([1,2,3,4,5], [-10, 10, 2, 3])
array([0, 5, 1, 2])

extract(condition, arr)
Return the elements of an array that satisfy some condition.
This is equivalent to np.compress(ravel(condition), ravel(arr)). If condition is boolean
np.extract is equivalent to arr[condition].
Parameters
condition : array_like
An array whose nonzero or True entries indicate the elements of arr to extract.
arr : array_like
Input array of the same size as condition.
See Also:
take, put, putmask, compress

Examples

>>> arr = np.arange(12).reshape((3, 4))

>>> arr
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> condition = np.mod(arr, 3)==0
>>> condition

3.9. Sorting and searching 667

NumPy Reference, Release 2.0.0.dev8464

array([[ True, False, False, True],

[False, False, True, False],
[False, True, False, False]], dtype=bool)
>>> np.extract(condition, arr)
array([0, 3, 6, 9])

If condition is boolean:

>>> arr[condition]
array([0, 3, 6, 9])

3.10 Logic functions

3.10.1 Truth value testing

all(a[, axis, out]) Test whether all array elements along a given axis evaluate to True.
any(a[, axis, out]) Test whether any array element along a given axis evaluates to True.
all(a, axis=None, out=None)
Test whether all array elements along a given axis evaluate to True.
Parameters
a : array_like
Input array or object that can be converted to an array.
axis : int, optional
Axis along which a logical AND is performed. The default (axis = None) is to perform
a logical AND over a flattened input array. axis may be negative, in which case it counts
from the last to the first axis.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output and the type is preserved. See doc.ufuncs (Section Output arguments)
for more details.
Returns
all : ndarray, bool
A new boolean or array is returned unless out is specified, in which case a reference to
out is returned.
See Also:

ndarray.all
equivalent method
any
Test whether any array element along a given axis evaluates to True.

Notes
Not a Number (NaN), positive infinity and negative infinity evaluate to True because these are not equal to zero.

668 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.all([[True,False],[True,True]])
False

>>> np.all([[True,False],[True,True]], axis=0)

array([ True, False], dtype=bool)

>>> np.all([-1, 4, 5])

True

>>> np.all([1.0, np.nan])

True

>>> o=np.array([False])
>>> z=np.all([-1, 4, 5], out=o)
>>> id(z), id(o), z
(28293632, 28293632, array([ True], dtype=bool))

any(a, axis=None, out=None)

Test whether any array element along a given axis evaluates to True.
Returns single boolean unless axis is not None
Parameters
a : array_like
Input array or object that can be converted to an array.
axis : int, optional
Axis along which a logical OR is performed. The default (axis = None) is to perform
a logical OR over a flattened input array. axis may be negative, in which case it counts
from the last to the first axis.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output and the type is preserved. See doc.ufuncs (Section Output arguments)
for details.
Returns
any : bool, ndarray
A new boolean or ndarray is returned unless out is specified, in which case a reference
to out is returned.
See Also:

ndarray.any
equivalent method
all
Test whether all array elements along a given axis evaluate to True.

Notes
Not a Number (NaN), positive infinity and negative infinity evaluate to True because these are not equal to zero.

3.10. Logic functions 669

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.any([[True, False], [True, True]])

True

>>> np.any([[True, False], [False, False]], axis=0)

array([ True, False], dtype=bool)

>>> np.any([-1, 0, 5])

True

>>> np.any(np.nan)
True

>>> o=np.array([False])
>>> z=np.any([-1, 4, 5], out=o)
>>> z, o
(array([ True], dtype=bool), array([ True], dtype=bool))
>>> # Check now that z is a reference to o
>>> z is o
True
>>> id(z), id(o) # identity of z and o
(191614240, 191614240)

3.10.2 Array contents

isfinite() Test element-wise for finite-ness (not infinity or not Not a Number).
isinf() Test element-wise for positive or negative infinity, return result as bool array.
isnan() Test element-wise for Not a Number (NaN), return result as a bool array.
isneginf(x[, y]) Test element-wise for negative infinity, return result as bool array.
isposinf(x[, y]) Test element-wise for positive infinity, return result as bool array.
isfinite
Test element-wise for finite-ness (not infinity or not Not a Number).
The result is returned as a boolean array.
Parameters
x : array_like
Input values.
out : ndarray, optional
Array into which the output is placed. Its type is preserved and it must be of the right
shape to hold the output. See doc.ufuncs.
Returns
y : ndarray, bool
For scalar input, the result is a new boolean with value True if the input is finite; oth-
erwise the value is False (input is either positive infinity, negative infinity or Not a
Number).
For array input, the result is a boolean array with the same dimensions as the input and
the values are True if the corresponding element of the input is finite; otherwise the
values are False (element is either positive infinity, negative infinity or Not a Number).

670 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:
isinf, isneginf, isposinf, isnan

Notes
Not a Number, positive infinity and negative infinity are considered to be non-finite.
Numpy uses the IEEE Standard for Binary Floating-Point for Arithmetic (IEEE 754). This means that Not a
Number is not equivalent to infinity. Also that positive infinity is not equivalent to negative infinity. But infinity
is equivalent to positive infinity. Errors result if the second argument is also supplied when x is a scalar input,
or if first and second arguments have different shapes.

Examples

>>> np.isfinite(1)
True
>>> np.isfinite(0)
True
>>> np.isfinite(np.nan)
False
>>> np.isfinite(np.inf)
False
>>> np.isfinite(np.NINF)
False
>>> np.isfinite([np.log(-1.),1.,np.log(0)])
array([False, True, False], dtype=bool)

>>> x = np.array([-np.inf, 0., np.inf])

>>> y = np.array([2, 2, 2])
>>> np.isfinite(x, y)
array([0, 1, 0])
>>> y
array([0, 1, 0])

isinf
Test element-wise for positive or negative infinity, return result as bool array.
Parameters
x : array_like
Input values
y : array_like, optional
An array with the same shape as x to store the result.
Returns
y : {ndarray, bool}
For scalar input, the result is a new boolean with value True if the input is positive or
negative infinity; otherwise the value is False.
For array input, the result is a boolean array with the same dimensions as the input
and the values are True if the corresponding element of the input is positive or negative
infinity; otherwise the values are False. If a second argument is supplied the result is
stored there. If the type of that array is a numeric type the result is represented as zeros
and ones, if the type is boolean then as False and True. The return value y is then a
reference to that array.

3.10. Logic functions 671

NumPy Reference, Release 2.0.0.dev8464

See Also:
isneginf, isposinf, isnan, isfinite

Notes
Numpy uses the IEEE Standard for Binary Floating-Point for Arithmetic (IEEE 754).
Errors result if second argument is also supplied with scalar input or if first and second arguments have different
shapes.

Examples

>>> np.isinf(np.inf)
True
>>> np.isinf(np.nan)
False
>>> np.isinf(np.NINF)
True
>>> np.isinf([np.inf, -np.inf, 1.0, np.nan])
array([ True, True, False, False], dtype=bool)

>>> x = np.array([-np.inf, 0., np.inf])

>>> y = np.array([2, 2, 2])
>>> np.isinf(x, y)
array([1, 0, 1])
>>> y
array([1, 0, 1])

isnan
Test element-wise for Not a Number (NaN), return result as a bool array.
Parameters
x : array_like
Input array.
Returns
y : {ndarray, bool}
For scalar input, the result is a new boolean with value True if the input is NaN; other-
wise the value is False.
For array input, the result is a boolean array with the same dimensions as the input
and the values are True if the corresponding element of the input is NaN; otherwise the
values are False.
See Also:
isinf, isneginf, isposinf, isfinite

Notes
Numpy uses the IEEE Standard for Binary Floating-Point for Arithmetic (IEEE 754). This means that Not a
Number is not equivalent to infinity.

Examples

>>> np.isnan(np.nan)
True

672 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.isnan(np.inf)
False
>>> np.isnan([np.log(-1.),1.,np.log(0)])
array([ True, False, False], dtype=bool)

isneginf(x, y=None)
Test element-wise for negative infinity, return result as bool array.
Parameters
x : array_like
The input array.
y : array_like, optional
A boolean array with the same shape and type as x to store the result.
Returns
y : ndarray
A boolean array with the same dimensions as the input. If second argument is not sup-
plied then a numpy boolean array is returned with values True where the corresponding
element of the input is negative infinity and values False where the element of the input
is not negative infinity.
If a second argument is supplied the result is stored there. If the type of that array is a
numeric type the result is represented as zeros and ones, if the type is boolean then as
False and True. The return value y is then a reference to that array.
See Also:
isinf, isposinf, isnan, isfinite

Notes
Numpy uses the IEEE Standard for Binary Floating-Point for Arithmetic (IEEE 754).
Errors result if the second argument is also supplied when x is a scalar input, or if first and second arguments
have different shapes.

Examples

>>> np.isneginf(np.NINF)
array(True, dtype=bool)
>>> np.isneginf(np.inf)
array(False, dtype=bool)
>>> np.isneginf(np.PINF)
array(False, dtype=bool)
>>> np.isneginf([-np.inf, 0., np.inf])
array([ True, False, False], dtype=bool)

>>> x = np.array([-np.inf, 0., np.inf])

>>> y = np.array([2, 2, 2])
>>> np.isneginf(x, y)
array([1, 0, 0])
>>> y
array([1, 0, 0])

isposinf(x, y=None)
Test element-wise for positive infinity, return result as bool array.

3.10. Logic functions 673

NumPy Reference, Release 2.0.0.dev8464

Parameters
x : array_like
The input array.
y : array_like, optional
A boolean array with the same shape as x to store the result.
Returns
y : ndarray
A boolean array with the same dimensions as the input. If second argument is not
supplied then a boolean array is returned with values True where the corresponding
element of the input is positive infinity and values False where the element of the input
is not positive infinity.
If a second argument is supplied the result is stored there. If the type of that array is a
numeric type the result is represented as zeros and ones, if the type is boolean then as
False and True. The return value y is then a reference to that array.
See Also:
isinf, isneginf, isfinite, isnan

Notes
Numpy uses the IEEE Standard for Binary Floating-Point for Arithmetic (IEEE 754).
Errors result if the second argument is also supplied when x is a scalar input, or if first and second arguments
have different shapes.

Examples

>>> np.isposinf(np.PINF)
array(True, dtype=bool)
>>> np.isposinf(np.inf)
array(True, dtype=bool)
>>> np.isposinf(np.NINF)
array(False, dtype=bool)
>>> np.isposinf([-np.inf, 0., np.inf])
array([False, False, True], dtype=bool)

>>> x = np.array([-np.inf, 0., np.inf])

>>> y = np.array([2, 2, 2])
>>> np.isposinf(x, y)
array([0, 0, 1])
>>> y
array([0, 0, 1])

3.10.3 Array type testing

iscomplex(x) Returns a bool array, where True if input element is complex.

iscomplexobj(x) Return True if x is a complex type or an array of complex numbers.
isfortran(a) Returns True if array is arranged in Fortran-order in memory
isreal(x) Returns a bool array, where True if input element is real.
isrealobj(x) Return True if x is a not complex type or an array of complex numbers.
isscalar(num) Returns True if the type of num is a scalar type.

674 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

iscomplex(x)
Returns a bool array, where True if input element is complex.
What is tested is whether the input has a non-zero imaginary part, not if the input type is complex.
Parameters
x : array_like
Input array.
Returns
out : ndarray of bools
Output array.
See Also:
isreal
iscomplexobj
Return True if x is a complex type or an array of complex numbers.

Examples

>>> np.iscomplex([1+1j, 1+0j, 4.5, 3, 2, 2j])

array([ True, False, False, False, False, True], dtype=bool)

iscomplexobj(x)
Return True if x is a complex type or an array of complex numbers.
The type of the input is checked, not the value. So even if the input has an imaginary part equal to zero,
iscomplexobj evaluates to True if the data type is complex.
Parameters
x : any
The input can be of any type and shape.
Returns
y : bool
The return value, True if x is of a complex type.
See Also:
isrealobj, iscomplex

Examples

>>> np.iscomplexobj(1)
False
>>> np.iscomplexobj(1+0j)
True
>>> np.iscomplexobj([3, 1+0j, True])
True

isfortran(a)
Returns True if array is arranged in Fortran-order in memory and dimension > 1.
Parameters
a : ndarray
Input array.

3.10. Logic functions 675

NumPy Reference, Release 2.0.0.dev8464

Examples
np.array allows to specify whether the array is written in C-contiguous order (last index varies the fastest), or
FORTRAN-contiguous order in memory (first index varies the fastest).

>>> a = np.array([[1, 2, 3], [4, 5, 6]], order=C)

>>> a
array([[1, 2, 3],
[4, 5, 6]])
>>> np.isfortran(a)
False

>>> b = np.array([[1, 2, 3], [4, 5, 6]], order=FORTRAN)

>>> b
array([[1, 2, 3],
[4, 5, 6]])
>>> np.isfortran(b)
True

The transpose of a C-ordered array is a FORTRAN-ordered array.

>>> a = np.array([[1, 2, 3], [4, 5, 6]], order=C)

>>> a
array([[1, 2, 3],
[4, 5, 6]])
>>> np.isfortran(a)
False
>>> b = a.T
>>> b
array([[1, 4],
[2, 5],
[3, 6]])
>>> np.isfortran(b)
True

1-D arrays always evaluate as False.

>>> np.isfortran(np.array([1, 2], order=FORTRAN))

False

isreal(x)
Returns a bool array, where True if input element is real.
If element has complex type with zero complex part, the return value for that element is True.
Parameters
x : array_like
Input array.
Returns
out : ndarray, bool
Boolean array of same shape as x.
See Also:
iscomplex
isrealobj
Return True if x is not a complex type.

676 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.isreal([1+1j, 1+0j, 4.5, 3, 2, 2j])

array([False, True, True, True, True, False], dtype=bool)

isrealobj(x)
Return True if x is a not complex type or an array of complex numbers.
The type of the input is checked, not the value. So even if the input has an imaginary part equal to zero, isrealobj
evaluates to False if the data type is complex.
Parameters
x : any
The input can be of any type and shape.
Returns
y : bool
The return value, False if x is of a complex type.
See Also:
iscomplexobj, isreal

Examples

>>> np.isrealobj(1)
True
>>> np.isrealobj(1+0j)
False
>>> np.isrealobj([3, 1+0j, True])
False

isscalar(num)
Returns True if the type of num is a scalar type.
Parameters
num : any
Input argument, can be of any type and shape.
Returns
val : bool
True if num is a scalar type, False if it is not.

Examples

>>> np.isscalar(3.1)
True
>>> np.isscalar([3.1])
False
>>> np.isscalar(False)
True

3.10. Logic functions 677

NumPy Reference, Release 2.0.0.dev8464

3.10.4 Logical operations

logical_and(x1) Compute the truth value of x1 AND x2 elementwise.

logical_or(x1) Compute the truth value of x1 OR x2 elementwise.
logical_not() Compute the truth value of NOT x elementwise.
logical_xor(x1) Compute the truth value of x1 XOR x2 elementwise.
logical_and
Compute the truth value of x1 AND x2 elementwise.
Parameters
x1, x2 : array_like
Input arrays. x1 and x2 must be of the same shape.
Returns
y : {ndarray, bool}
Boolean result with the same shape as x1 and x2 of the logical AND operation on cor-
responding elements of x1 and x2.
See Also:
logical_or, logical_not, logical_xor, bitwise_and

Examples

>>> np.logical_and(True, False)

False
>>> np.logical_and([True, False], [False, False])
array([False, False], dtype=bool)

>>> x = np.arange(5)
>>> np.logical_and(x>1, x<4)
array([False, False, True, True, False], dtype=bool)

logical_or
Compute the truth value of x1 OR x2 elementwise.
Parameters
x1, x2 : array_like
Logical OR is applied to the elements of x1 and x2. They have to be of the same shape.
Returns
y : {ndarray, bool}
Boolean result with the same shape as x1 and x2 of the logical OR operation on elements
of x1 and x2.
See Also:
logical_and, logical_not, logical_xor, bitwise_or

Examples

>>> np.logical_or(True, False)

True
>>> np.logical_or([True, False], [False, False])
array([ True, False], dtype=bool)

678 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> x = np.arange(5)
>>> np.logical_or(x < 1, x > 3)
array([ True, False, False, False, True], dtype=bool)

logical_not
Compute the truth value of NOT x elementwise.
Parameters
x : array_like
Logical NOT is applied to the elements of x.
Returns
y : {ndarray, bool}
Boolean result with the same shape as x of the NOT operation on elements of x.
See Also:
logical_and, logical_or, logical_xor

Examples

>>> np.logical_not(3)
False
>>> np.logical_not([True, False, 0, 1])
array([False, True, True, False], dtype=bool)

>>> x = np.arange(5)
>>> np.logical_not(x<3)
array([False, False, False, True, True], dtype=bool)

logical_xor
Compute the truth value of x1 XOR x2 elementwise.
Parameters
x1, x2 : array_like
Logical XOR is applied to the elements of x1 and x2. They have to be of the same
shape.
Returns
y : {ndarray, bool}
Boolean result with the same shape as x1 and x2 of the logical XOR operation on ele-
ments of x1 and x2.
See Also:
logical_and, logical_or, logical_not, bitwise_xor

Examples

>>> np.logical_xor(True, False)

True
>>> np.logical_xor([True, True, False, False], [True, False, True, False])
array([False, True, True, False], dtype=bool)

3.10. Logic functions 679

NumPy Reference, Release 2.0.0.dev8464

>>> x = np.arange(5)
>>> np.logical_xor(x < 1, x > 3)
array([ True, False, False, False, True], dtype=bool)

3.10.5 Comparison

allclose(a, b[, rtol, atol]) Returns True if two arrays are element-wise equal within a tolerance.
array_equal(a1, a2) True if two arrays have the same shape and elements, False otherwise.
array_equiv(a1, a2) Returns True if input arrays are shape consistent and all elements equal.
allclose(a, b, rtol=1.0000000000000001e-05, atol=1e-08)
Returns True if two arrays are element-wise equal within a tolerance.
The tolerance values are positive, typically very small numbers. The relative difference (rtol * abs(b)) and the
absolute difference atol are added together to compare against the absolute difference between a and b.
Parameters
a, b : array_like
Input arrays to compare.
rtol : float
The relative tolerance parameter (see Notes).
atol : float
The absolute tolerance parameter (see Notes).
Returns
y : bool
Returns True if the two arrays are equal within the given tolerance; False otherwise. If
either array contains NaN, then False is returned.
See Also:
all, any, alltrue, sometrue

Notes
If the following equation is element-wise True, then allclose returns True.
absolute(a - b) <= (atol + rtol * absolute(b))
The above equation is not symmetric in a and b, so that allclose(a, b) might be different from allclose(b, a) in
some rare cases.

Examples

>>> np.allclose([1e10,1e-7], [1.00001e10,1e-8])

False
>>> np.allclose([1e10,1e-8], [1.00001e10,1e-9])
True
>>> np.allclose([1e10,1e-8], [1.0001e10,1e-9])
False
>>> np.allclose([1.0, np.nan], [1.0, np.nan])
False

680 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

array_equal(a1, a2)
True if two arrays have the same shape and elements, False otherwise.
Parameters
a1, a2 : array_like
Input arrays.
Returns
b : bool
Returns True if the arrays are equal.
See Also:

allclose
Returns True if two arrays are element-wise equal within a tolerance.
array_equiv
Returns True if input arrays are shape consistent and all elements equal.

Examples

>>> np.array_equal([1, 2], [1, 2])

True
>>> np.array_equal(np.array([1, 2]), np.array([1, 2]))
True
>>> np.array_equal([1, 2], [1, 2, 3])
False
>>> np.array_equal([1, 2], [1, 4])
False

array_equiv(a1, a2)
Returns True if input arrays are shape consistent and all elements equal.
Shape consistent means they are either the same shape, or one input array can be broadcasted to create the same
shape as the other one.
Parameters
a1, a2 : array_like
Input arrays.
Returns
out : bool
True if equivalent, False otherwise.

Examples

>>> np.array_equiv([1, 2], [1, 2])

True
>>> np.array_equiv([1, 2], [1, 3])
False

Showing the shape equivalence:

>>> np.array_equiv([1, 2], [[1, 2], [1, 2]])

True
>>> np.array_equiv([1, 2], [[1, 2, 1, 2], [1, 2, 1, 2]])
False

3.10. Logic functions 681

NumPy Reference, Release 2.0.0.dev8464

>>> np.array_equiv([1, 2], [[1, 2], [1, 3]])

False

greater(x1) Return (x1 > x2) element-wise.

greater_equal(x1) Return (x1 >= x2) element-wise.
less(x1) Return (x1 < x2) element-wise.
less_equal(x1) Return (x1 <= x2) element-wise.
equal(x1) Return (x1 == x2) element-wise.
not_equal(x1) Return (x1 != x2) element-wise.
greater
Return (x1 > x2) element-wise.
Parameters
x1, x2 : array_like
Input arrays.
Returns
Out : {ndarray, bool}
Output array of bools, or a single bool if x1 and x2 are scalars.
See Also:
greater_equal, less, less_equal, equal, not_equal

Examples

>>> np.greater([4,2],[2,2])
array([ True, False], dtype=bool)

If the inputs are ndarrays, then np.greater is equivalent to >.

>>> a = np.array([4,2])
>>> b = np.array([2,2])
>>> a > b
array([ True, False], dtype=bool)

greater_equal
Return (x1 >= x2) element-wise.
Parameters
x1, x2 : array_like
Input arrays.
Returns
out : {ndarray, bool}
Output array of bools, or a single bool if x1 and x2 are scalars.
See Also:
greater, less, less_equal, equal, not_equal

Examples

>>> np.greater_equal([4, 2, 1], [2, 2, 2])

array([ True, True, False], dtype=bool)

682 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

less
Return (x1 < x2) element-wise.
Parameters
x1, x2 : array_like
Input arrays.
Returns
Out : ndarray of bools
Output array of bools, or a single bool if x1 and x2 are scalars.
See Also:
less_equal, greater, greater_equal, equal, not_equal

Examples

>>> np.less([1, 2], [2, 2])

array([ True, False], dtype=bool)

less_equal
Return (x1 <= x2) element-wise.
Parameters
x1, x2 : array_like
Input arrays.
Returns
Out : {ndarray, bool}
Output array of bools, or a single bool if x1 and x2 are scalars.
See Also:
less, greater_equal, greater, equal, not_equal

Examples

>>> np.less_equal([1, 2, 3], [2, 2, 2])

array([ True, True, False], dtype=bool)

equal
Return (x1 == x2) element-wise.
Parameters
x1, x2 : array_like
Input arrays of the same shape.
Returns
out : {ndarray, bool}
Output array of bools, or a single bool if x1 and x2 are scalars.
See Also:
not_equal, greater_equal, less_equal, greater, less

3.10. Logic functions 683

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.equal([0, 1, 3], np.arange(3))

array([ True, True, False], dtype=bool)

What is compared are values, not types. So an int (1) and an array of length one can evaluate as True:

>>> np.equal(1, np.ones(1))

array([ True], dtype=bool)

not_equal
Return (x1 != x2) element-wise.
Parameters
x1, x2 : array_like
Input arrays.
out : ndarray, optional
A placeholder the same shape as x1 to store the result. See doc.ufuncs (Section Output
arguments) for more details.
Returns
not_equal : ndarray bool, scalar bool
For each element in x1, x2, return True if x1 is not equal to x2 and False otherwise.
See Also:
equal, greater, greater_equal, less, less_equal

Examples

>>> np.not_equal([1.,2.], [1., 3.])

array([False, True], dtype=bool)
>>> np.not_equal([1, 2], [[1, 3],[1, 4]])
array([[False, True],
[False, True]], dtype=bool)

3.11 Binary operations

3.11.1 Elementwise bit operations

bitwise_and(x1) Compute the bit-wise AND of two arrays element-wise.

bitwise_or(x1) Compute the bit-wise OR of two arrays element-wise.
bitwise_xor(x1) Compute the bit-wise XOR of two arrays element-wise.
invert() Compute bit-wise inversion, or bit-wise NOT, element-wise.
left_shift(x1) Shift the bits of an integer to the left.
right_shift(x1) Shift the bits of an integer to the right.
bitwise_and
Compute the bit-wise AND of two arrays element-wise.
Computes the bit-wise AND of the underlying binary representation of the integers in the input arrays. This
ufunc implements the C/Python operator &.

684 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
x1, x2 : array_like
Only integer types are handled (including booleans).
Returns
out : array_like
Result.
See Also:
logical_and, bitwise_or, bitwise_xor
binary_repr
Return the binary representation of the input number as a string.

Examples
The number 13 is represented by 00001101. Likewise, 17 is represented by 00010001. The bit-wise AND
of 13 and 17 is therefore 000000001, or 1:

>>> np.bitwise_and(13, 17)

1

>>> np.bitwise_and(14, 13)

12
>>> np.binary_repr(12)
1100
>>> np.bitwise_and([14,3], 13)
array([12, 1])

>>> np.bitwise_and([11,7], [4,25])

array([0, 1])
>>> np.bitwise_and(np.array([2,5,255]), np.array([3,14,16]))
array([ 2, 4, 16])
>>> np.bitwise_and([True, True], [False, True])
array([False, True], dtype=bool)

bitwise_or
Compute the bit-wise OR of two arrays element-wise.
Computes the bit-wise OR of the underlying binary representation of the integers in the input arrays. This ufunc
implements the C/Python operator |.
Parameters
x1, x2 : array_like
Only integer types are handled (including booleans).
out : ndarray, optional
Array into which the output is placed. Its type is preserved and it must be of the right
shape to hold the output. See doc.ufuncs.
Returns
out : array_like
Result.

3.11. Binary operations 685

NumPy Reference, Release 2.0.0.dev8464

See Also:
logical_or, bitwise_and, bitwise_xor
binary_repr
Return the binary representation of the input number as a string.

Examples
The number 13 has the binaray representation 00001101. Likewise, 16 is represented by 00010000. The
bit-wise OR of 13 and 16 is then 000111011, or 29:

>>> np.bitwise_or(13, 16)

29
>>> np.binary_repr(29)
11101

>>> np.bitwise_or(32, 2)
34
>>> np.bitwise_or([33, 4], 1)
array([33, 5])
>>> np.bitwise_or([33, 4], [1, 2])
array([33, 6])

>>> np.bitwise_or(np.array([2, 5, 255]), np.array([4, 4, 4]))

array([ 6, 5, 255])
>>> np.array([2, 5, 255]) | np.array([4, 4, 4])
array([ 6, 5, 255])
>>> np.bitwise_or(np.array([2, 5, 255, 2147483647L], dtype=np.int32),
... np.array([4, 4, 4, 2147483647L], dtype=np.int32))
array([ 6, 5, 255, 2147483647])
>>> np.bitwise_or([True, True], [False, True])
array([ True, True], dtype=bool)

bitwise_xor
Compute the bit-wise XOR of two arrays element-wise.
Computes the bit-wise XOR of the underlying binary representation of the integers in the input arrays. This
ufunc implements the C/Python operator ^.
Parameters
x1, x2 : array_like
Only integer types are handled (including booleans).
Returns
out : array_like
Result.
See Also:
logical_xor, bitwise_and, bitwise_or
binary_repr
Return the binary representation of the input number as a string.

686 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples
The number 13 is represented by 00001101. Likewise, 17 is represented by 00010001. The bit-wise XOR
of 13 and 17 is therefore 00011100, or 28:

>>> np.bitwise_xor(13, 17)

28
>>> np.binary_repr(28)
11100

>>> np.bitwise_xor(31, 5)
26
>>> np.bitwise_xor([31,3], 5)
array([26, 6])

>>> np.bitwise_xor([31,3], [5,6])

array([26, 5])
>>> np.bitwise_xor([True, True], [False, True])
array([ True, False], dtype=bool)

invert
Compute bit-wise inversion, or bit-wise NOT, element-wise.
Computes the bit-wise NOT of the underlying binary representation of the integers in the input arrays. This
ufunc implements the C/Python operator ~.
For signed integer inputs, the twos complement is returned. In a twos-complement system negative numbers
are represented by the twos complement of the absolute value. This is the most common method of representing
signed integers on computers [R41]. A N-bit twos-complement system can represent every integer in the range
2N 1 to +2N 1 1.
Parameters
x1 : array_like
Only integer types are handled (including booleans).
Returns
out : array_like
Result.
See Also:
bitwise_and, bitwise_or, bitwise_xor, logical_not
binary_repr
Return the binary representation of the input number as a string.

Notes
bitwise_not is an alias for invert:

>>> np.bitwise_not is np.invert

True

References
[R41]

3.11. Binary operations 687

NumPy Reference, Release 2.0.0.dev8464

Examples
Weve seen that 13 is represented by 00001101. The invert or bit-wise NOT of 13 is then:

>>> np.invert(np.array([13], dtype=uint8))

array([242], dtype=uint8)
>>> np.binary_repr(x, width=8)
00001101
>>> np.binary_repr(242, width=8)
11110010

The result depends on the bit-width:

>>> np.invert(np.array([13], dtype=uint16))

array([65522], dtype=uint16)
>>> np.binary_repr(x, width=16)
0000000000001101
>>> np.binary_repr(65522, width=16)
1111111111110010

When using signed integer types the result is the twos complement of the result for the unsigned type:

>>> np.invert(np.array([13], dtype=int8))

array([-14], dtype=int8)
>>> np.binary_repr(-14, width=8)
11110010

Booleans are accepted as well:

>>> np.invert(array([True, False]))

array([False, True], dtype=bool)

left_shift
Shift the bits of an integer to the left.
Bits are shifted to the left by appending x2 0s at the right of x1. Since the internal representation of numbers is
in binary format, this operation is equivalent to multiplying x1 by 2**x2.
Parameters
x1 : array_like of integer type
Input values.
x2 : array_like of integer type
Number of zeros to append to x1. Has to be non-negative.
Returns
out : array of integer type
Return x1 with bits shifted x2 times to the left.
See Also:

right_shift
Shift the bits of an integer to the right.
binary_repr
Return the binary representation of the input number as a string.

688 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.binary_repr(5)
101
>>> np.left_shift(5, 2)
20
>>> np.binary_repr(20)
10100

>>> np.left_shift(5, [1,2,3])

array([10, 20, 40])

right_shift
Shift the bits of an integer to the right.
Bits are shifted to the right by removing x2 bits at the right of x1. Since the internal representation of numbers
is in binary format, this operation is equivalent to dividing x1 by 2**x2.
Parameters
x1 : array_like, int
Input values.
x2 : array_like, int
Number of bits to remove at the right of x1.
Returns
out : ndarray, int
Return x1 with bits shifted x2 times to the right.
See Also:

left_shift
Shift the bits of an integer to the left.
binary_repr
Return the binary representation of the input number as a string.

Examples

>>> np.binary_repr(10)
1010
>>> np.right_shift(10, 1)
5
>>> np.binary_repr(5)
101

>>> np.right_shift(10, [1,2,3])

array([5, 2, 1])

3.11.2 Bit packing

packbits(myarray[, axis]) Packs the elements of a binary-valued array into bits in a uint8 array.
unpackbits(myarray[, axis]) Unpacks elements of a uint8 array into a binary-valued output array.

3.11. Binary operations 689

NumPy Reference, Release 2.0.0.dev8464

packbits(myarray, axis=None)
Packs the elements of a binary-valued array into bits in a uint8 array.
The result is padded to full bytes by inserting zero bits at the end.
Parameters
myarray : array_like
An integer type array whose elements should be packed to bits.
axis : int, optional
The dimension over which bit-packing is done. None implies packing the flattened
array.
Returns
packed : ndarray
Array of type uint8 whose elements represent bits corresponding to the logical (0 or
nonzero) value of the input elements. The shape of packed has the same number of
dimensions as the input (unless axis is None, in which case the output is 1-D).
See Also:

unpackbits
Unpacks elements of a uint8 array into a binary-valued output array.

Examples

>>> a = np.array([[[1,0,1],
... [0,1,0]],
... [[1,1,0],
... [0,0,1]]])
>>> b = np.packbits(a, axis=-1)
>>> b
array([[[160],[64]],[[192],[32]]], dtype=uint8)

Note that in binary 160 = 1010 0000, 64 = 0100 0000, 192 = 1100 0000, and 32 = 0010 0000.
unpackbits(myarray, axis=None)
Unpacks elements of a uint8 array into a binary-valued output array.
Each element of myarray represents a bit-field that should be unpacked into a binary-valued output array. The
shape of the output array is either 1-D (if axis is None) or the same shape as the input array with unpacking done
along the axis specified.
Parameters
myarray : ndarray, uint8 type
Input array.
axis : int, optional
Unpacks along this axis.
Returns
unpacked : ndarray, uint8 type
The elements are binary-valued (0 or 1).
See Also:

690 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

packbits
Packs the elements of a binary-valued array into bits in a uint8 array.

Examples

>>> a = np.array([[2], [7], [23]], dtype=np.uint8)

>>> a
array([[ 2],
[ 7],
[23]], dtype=uint8)
>>> b = np.unpackbits(a, axis=1)
>>> b
array([[0, 0, 0, 0, 0, 0, 1, 0],
[0, 0, 0, 0, 0, 1, 1, 1],
[0, 0, 0, 1, 0, 1, 1, 1]], dtype=uint8)

3.11.3 Output formatting

binary_repr(num[, width]) Return the binary representation of the input number as a string.
binary_repr(num, width=None)
Return the binary representation of the input number as a string.
For negative numbers, if width is not given, a minus sign is added to the front. If width is given, the twos
complement of the number is returned, with respect to that width.
In a twos-complement system negative numbers are represented by the twos complement of the absolute
value. This is the most common method of representing signed integers on computers [R23]. A N-bit twos-
complement system can represent every integer in the range 2N 1 to +2N 1 1.
Parameters
num : int
Only an integer decimal number can be used.
width : int, optional
The length of the returned string if num is positive, the length of the twos complement
if num is negative.
Returns
bin : str
Binary representation of num or twos complement of num.
See Also:

base_repr
Return a string representation of a number in the given base system.

Notes
binary_repr is equivalent to using base_repr with base 2, but about 25x faster.

References
[R23]

3.11. Binary operations 691

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.binary_repr(3)
11
>>> np.binary_repr(-3)
-11
>>> np.binary_repr(3, width=4)
0011

The twos complement is returned when the input number is negative and width is specified:

>>> np.binary_repr(-3, width=4)

1101

3.12 Statistics

3.12.1 Extremal values

amin(a[, axis, out]) Return the minimum of an array or minimum along an axis.
amax(a[, axis, out]) Return the maximum of an array or maximum along an axis.
nanmax(a[, axis]) Return the maximum of an array or maximum along an axis ignoring any NaNs.
nanmin(a[, axis]) Return the minimum of an array or minimum along an axis ignoring any NaNs.
ptp(a[, axis, out]) Range of values (maximum - minimum) along an axis.
amin(a, axis=None, out=None)
Return the minimum of an array or minimum along an axis.
Parameters
a : array_like
Input data.
axis : int, optional
Axis along which to operate. By default a flattened input is used.
out : ndarray, optional
Alternative output array in which to place the result. Must be of the same shape and
buffer length as the expected output. See doc.ufuncs (Section Output arguments) for
more details.
Returns
amin : ndarray
A new array or a scalar array with the result.
See Also:

nanmin
nan values are ignored instead of being propagated
fmin
same behavior as the C99 fmin function
argmin
Return the indices of the minimum values.

692 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

amax, nanmax, fmax

Notes
NaN values are propagated, that is if at least one item is nan, the corresponding min value will be nan as well.
To ignore NaN values (matlab behavior), please use nanmin.

Examples

>>> a = np.arange(4).reshape((2,2))
>>> a
array([[0, 1],
[2, 3]])
>>> np.amin(a) # Minimum of the flattened array
0
>>> np.amin(a, axis=0) # Minima along the first axis
array([0, 1])
>>> np.amin(a, axis=1) # Minima along the second axis
array([0, 2])

>>> b = np.arange(5, dtype=np.float)

>>> b[2] = np.NaN
>>> np.amin(b)
nan
>>> np.nanmin(b)
0.0

amax(a, axis=None, out=None)

Return the maximum of an array or maximum along an axis.
Parameters
a : array_like
Input data.
axis : int, optional
Axis along which to operate. By default flattened input is used.
out : ndarray, optional
Alternative output array in which to place the result. Must be of the same shape and
buffer length as the expected output. See doc.ufuncs (Section Output arguments) for
more details.
Returns
amax : ndarray
A new array or a scalar array with the result.
See Also:

nanmax
nan values are ignored instead of being propagated
fmax
same behavior as the C99 fmax function
argmax
Indices of the maximum values.

3.12. Statistics 693

NumPy Reference, Release 2.0.0.dev8464

Notes
NaN values are propagated, that is if at least one item is nan, the corresponding max value will be nan as well.
To ignore NaN values (matlab behavior), please use nanmax.

Examples

>>> a = np.arange(4).reshape((2,2))
>>> a
array([[0, 1],
[2, 3]])
>>> np.amax(a)
3
>>> np.amax(a, axis=0)
array([2, 3])
>>> np.amax(a, axis=1)
array([1, 3])

>>> b = np.arange(5, dtype=np.float)

>>> b[2] = np.NaN
>>> np.amax(b)
nan
>>> np.nanmax(b)
4.0

nanmax(a, axis=None)
Return the maximum of an array or maximum along an axis ignoring any NaNs.
Parameters
a : array_like
Array containing numbers whose maximum is desired. If a is not an array, a conversion
is attempted.
axis : int, optional
Axis along which the maximum is computed. The default is to compute the maximum
of the flattened array.
Returns
nanmax : ndarray
An array with the same shape as a, with the specified axis removed. If a is a 0-d array,
or if axis is None, a ndarray scalar is returned. The the same dtype as a is returned.
See Also:

numpy.amax
Maximum across array including any Not a Numbers.
numpy.nanmin
Minimum across array ignoring any Not a Numbers.
isnan
Shows which elements are Not a Number (NaN).
isfinite
Shows which elements are not: Not a Number, positive and negative infinity

694 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notes
Numpy uses the IEEE Standard for Binary Floating-Point for Arithmetic (IEEE 754). This means that Not a
Number is not equivalent to infinity. Positive infinity is treated as a very large number and negative infinity is
treated as a very small (i.e. negative) number.
If the input has a integer type, an integer type is returned unless the input contains NaNs and infinity.

Examples

>>> a = np.array([[1, 2], [3, np.nan]])

>>> np.nanmax(a)
3.0
>>> np.nanmax(a, axis=0)
array([ 3., 2.])
>>> np.nanmax(a, axis=1)
array([ 2., 3.])

When positive infinity and negative infinity are present:

>>> np.nanmax([1, 2, np.nan, np.NINF])

2.0
>>> np.nanmax([1, 2, np.nan, np.inf])
inf

nanmin(a, axis=None)
Return the minimum of an array or minimum along an axis ignoring any NaNs.
Parameters
a : array_like
Array containing numbers whose minimum is desired.
axis : int, optional
Axis along which the minimum is computed.The default is to compute the minimum of
the flattened array.
Returns
nanmin : ndarray
A new array or a scalar array with the result.
See Also:

numpy.amin
Minimum across array including any Not a Numbers.
numpy.nanmax
Maximum across array ignoring any Not a Numbers.
isnan
Shows which elements are Not a Number (NaN).
isfinite
Shows which elements are not: Not a Number, positive and negative infinity

Notes
Numpy uses the IEEE Standard for Binary Floating-Point for Arithmetic (IEEE 754). This means that Not a
Number is not equivalent to infinity. Positive infinity is treated as a very large number and negative infinity is
treated as a very small (i.e. negative) number.

3.12. Statistics 695

NumPy Reference, Release 2.0.0.dev8464

If the input has a integer type, an integer type is returned unless the input contains NaNs and infinity.

Examples

>>> a = np.array([[1, 2], [3, np.nan]])

>>> np.nanmin(a)
1.0
>>> np.nanmin(a, axis=0)
array([ 1., 2.])
>>> np.nanmin(a, axis=1)
array([ 1., 3.])

When positive infinity and negative infinity are present:

>>> np.nanmin([1, 2, np.nan, np.inf])

1.0
>>> np.nanmin([1, 2, np.nan, np.NINF])
-inf

ptp(a, axis=None, out=None)

Range of values (maximum - minimum) along an axis.
The name of the function comes from the acronym for peak to peak.
Parameters
a : array_like
Input values.
axis : int, optional
Axis along which to find the peaks. By default, flatten the array.
out : array_like
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output, but the type of the output values will be cast if
necessary.
Returns
ptp : ndarray
A new array holding the result, unless out was specified, in which case a reference to
out is returned.

Examples

>>> x = np.arange(4).reshape((2,2))
>>> x
array([[0, 1],
[2, 3]])

>>> np.ptp(x, axis=0)

array([2, 2])

>>> np.ptp(x, axis=1)

array([1, 1])

696 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

3.12.2 Averages and variances

average(a[, axis, weights, returned]) Compute the weighted average along the specified axis.
mean(a[, axis, dtype, out]) Compute the arithmetic mean along the specified axis.
median(a[, axis, out, overwrite_input]) Compute the median along the specified axis.
std(a[, axis, dtype, out, ddof]) Compute the standard deviation along the specified axis.
var(a[, axis, dtype, out, ddof]) Compute the variance along the specified axis.
average(a, axis=None, weights=None, returned=False)
Compute the weighted average along the specified axis.
Parameters
a : array_like
Array containing data to be averaged. If a is not an array, a conversion is attempted.
axis : int, optional
Axis along which to average a. If None, averaging is done over the flattened array.
weights : array_like, optional
An array of weights associated with the values in a. Each value in a contributes to the
average according to its associated weight. The weights array can either be 1-D (in
which case its length must be the size of a along the given axis) or of the same shape as
a. If weights=None, then all data in a are assumed to have a weight equal to one.
returned : bool, optional
Default is False. If True, the tuple (average, sum_of_weights) is returned, otherwise
only the average is returned. If weights=None, sum_of_weights is equivalent to the
number of elements over which the average is taken.
Returns
average, [sum_of_weights] : {array_type, double}
Return the average along the specified axis. When returned is True, return a tuple with
the average as the first element and the sum of the weights as the second element.
The return type is Float if a is of integer type, otherwise it is of the same type as a.
sum_of_weights is of the same type as average.
Raises
ZeroDivisionError :
When all weights along axis are zero. See numpy.ma.average for a version robust
to this type of error.
TypeError :
When the length of 1D weights is not the same as the shape of a along axis.
See Also:
mean
ma.average
average for masked arrays

Examples

>>> data = range(1,5)

>>> data
[1, 2, 3, 4]

3.12. Statistics 697

NumPy Reference, Release 2.0.0.dev8464

>>> np.average(data)
2.5
>>> np.average(range(1,11), weights=range(10,0,-1))
4.0

>>> data = np.arange(6).reshape((3,2))

>>> data
array([[0, 1],
[2, 3],
[4, 5]])
>>> np.average(data, axis=1, weights=[1./4, 3./4])
array([ 0.75, 2.75, 4.75])
>>> np.average(data, weights=[1./4, 3./4])
...
TypeError: Axis must be specified when shapes of a and weights differ.

mean(a, axis=None, dtype=None, out=None)

Compute the arithmetic mean along the specified axis.
Returns the average of the array elements. The average is taken over the flattened array by default, otherwise
over the specified axis. float64 intermediate and return values are used for integer inputs.
Parameters
a : array_like
Array containing numbers whose mean is desired. If a is not an array, a conversion is
attempted.
axis : int, optional
Axis along which the means are computed. The default is to compute the mean of the
flattened array.
dtype : dtype, optional
Type to use in computing the mean. For integer inputs, the default is float64; for floating
point, inputs it is the same as the input dtype.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary. See doc.ufuncs for details.
Returns
m : ndarray, see dtype parameter above
If out=None, returns a new array containing the mean values, otherwise a reference to
the output array is returned.
See Also:

average
Weighted average

Notes
The arithmetic mean is the sum of the elements along the axis divided by the number of elements.
Note that for floating-point input, the mean is computed using the same precision the input has. Depending on
the input data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying
a higher-accuracy accumulator using the dtype keyword can alleviate this issue.

698 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> np.mean(a)
2.5
>>> np.mean(a, axis=0)
array([ 2., 3.])
>>> np.mean(a, axis=1)
array([ 1.5, 3.5])

In single precision, mean can be inaccurate:

>>> a = np.zeros((2, 512*512), dtype=np.float32)

>>> a[0, :] = 1.0
>>> a[1, :] = 0.1
>>> np.mean(a)
0.546875

Computing the mean in float64 is more accurate:

>>> np.mean(a, dtype=np.float64)

0.55000000074505806

median(a, axis=None, out=None, overwrite_input=False)

Compute the median along the specified axis.
Returns the median of the array elements.
Parameters
a : array_like
Input array or object that can be converted to an array.
axis : {None, int}, optional
Axis along which the medians are computed. The default (axis=None) is to compute
the median along a flattened version of the array.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output, but the type (of the output) will be cast if necessary.
overwrite_input : {False, True}, optional
If True, then allow use of memory of input array (a) for calculations. The input array
will be modified by the call to median. This will save memory when you do not need to
preserve the contents of the input array. Treat the input as undefined, but it will probably
be fully or partially sorted. Default is False. Note that, if overwrite_input is True and
the input is not already an ndarray, an error will be raised.
Returns
median : ndarray
A new array holding the result (unless out is specified, in which case that array is re-
turned instead). If the input contains integers, or floats of smaller precision than 64,
then the output data-type is float64. Otherwise, the output data-type is the same as that
of the input.
See Also:
mean, percentile

3.12. Statistics 699

NumPy Reference, Release 2.0.0.dev8464

Notes
Given a vector V of length N, the median of V is the middle value of a sorted copy of V, V_sorted - i.e.,
V_sorted[(N-1)/2], when N is odd. When N is even, it is the average of the two middle values of
V_sorted.

Examples

>>> a = np.array([[10, 7, 4], [3, 2, 1]])

>>> a
array([[10, 7, 4],
[ 3, 2, 1]])
>>> np.median(a)
3.5
>>> np.median(a, axis=0)
array([ 6.5, 4.5, 2.5])
>>> np.median(a, axis=1)
array([ 7., 2.])
>>> m = np.median(a, axis=0)
>>> out = np.zeros_like(m)
>>> np.median(a, axis=0, out=m)
array([ 6.5, 4.5, 2.5])
>>> m
array([ 6.5, 4.5, 2.5])
>>> b = a.copy()
>>> np.median(b, axis=1, overwrite_input=True)
array([ 7., 2.])
>>> assert not np.all(a==b)
>>> b = a.copy()
>>> np.median(b, axis=None, overwrite_input=True)
3.5
>>> assert not np.all(a==b)

std(a, axis=None, dtype=None, out=None, ddof=0)

Compute the standard deviation along the specified axis.
Returns the standard deviation, a measure of the spread of a distribution, of the array elements. The standard
deviation is computed for the flattened array by default, otherwise over the specified axis.
Parameters
a : array_like
Calculate the standard deviation of these values.
axis : int, optional
Axis along which the standard deviation is computed. The default is to compute the
standard deviation of the flattened array.
dtype : dtype, optional
Type to use in computing the standard deviation. For arrays of integer type the default
is float64, for arrays of float types it is the same as the array type.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type (of the calculated values) will be cast if necessary.
ddof : int, optional

700 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Means Delta Degrees of Freedom. The divisor used in calculations is N - ddof,

where N represents the number of elements. By default ddof is zero.
Returns
standard_deviation : ndarray, see dtype parameter above.
If out is None, return a new array containing the standard deviation, otherwise return a
reference to the output array.
See Also:
var, mean
numpy.doc.ufuncs
Section Output arguments

Notes
The standard deviation is the square root of the average of the squared deviations from the mean, i.e., std =
sqrt(mean(abs(x - x.mean())**2)).
The average squared deviation is normally calculated as x.sum() / N, where N = len(x). If, however,
ddof is specified, the divisor N - ddof is used instead. In standard statistical practice, ddof=1 provides an
unbiased estimator of the variance of the infinite population. ddof=0 provides a maximum likelihood estimate
of the variance for normally distributed variables. The standard deviation computed in this function is the
square root of the estimated variance, so even with ddof=1, it will not be an unbiased estimate of the standard
deviation per se.
Note that, for complex numbers, std takes the absolute value before squaring, so that the result is always real
and nonnegative.
For floating-point input, the std is computed using the same precision the input has. Depending on the input
data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a higher-
accuracy accumulator using the dtype keyword can alleviate this issue.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> np.std(a)
1.1180339887498949
>>> np.std(a, axis=0)
array([ 1., 1.])
>>> np.std(a, axis=1)
array([ 0.5, 0.5])

In single precision, std() can be inaccurate:

>>> a = np.zeros((2,512*512), dtype=np.float32)

>>> a[0,:] = 1.0
>>> a[1,:] = 0.1
>>> np.std(a)
0.45172946707416706

Computing the standard deviation in float64 is more accurate:

>>> np.std(a, dtype=np.float64)

0.44999999925552653

var(a, axis=None, dtype=None, out=None, ddof=0)

Compute the variance along the specified axis.

3.12. Statistics 701

NumPy Reference, Release 2.0.0.dev8464

Returns the variance of the array elements, a measure of the spread of a distribution. The variance is computed
for the flattened array by default, otherwise over the specified axis.
Parameters
a : array_like
Array containing numbers whose variance is desired. If a is not an array, a conversion
is attempted.
axis : int, optional
Axis along which the variance is computed. The default is to compute the variance of
the flattened array.
dtype : dtype, optional
Type to use in computing the variance. For arrays of integer type the default is float32;
for arrays of float types it is the same as the array type.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type is cast if necessary.
ddof : int, optional
Delta Degrees of Freedom: the divisor used in calculation is N - ddof, where N
represents the number of elements. By default ddof is zero.
Returns
variance : ndarray, see dtype parameter above
If out=None, returns a new array containing the variance; otherwise a reference to the
output array is returned.
See Also:

std
Standard deviation
mean
Average
numpy.doc.ufuncs
Section Output arguments

Notes
The variance is the average of the squared deviations from the mean, i.e., var = mean(abs(x -
x.mean())**2).
The mean is normally calculated as x.sum() / N, where N = len(x). If, however, ddof is specified, the
divisor N - ddof is used instead. In standard statistical practice, ddof=1 provides an unbiased estimator of
the variance of the infinite population. ddof=0 provides a maximum likelihood estimate of the variance for
normally distributed variables.
Note that for complex numbers, the absolute value is taken before squaring, so that the result is always real and
nonnegative.
For floating-point input, the variance is computed using the same precision the input has. Depending on the
input data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a
higher-accuracy accumulator using the dtype keyword can alleviate this issue.

702 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.array([[1,2],[3,4]])
>>> np.var(a)
1.25
>>> np.var(a,0)
array([ 1., 1.])
>>> np.var(a,1)
array([ 0.25, 0.25])

In single precision, var() can be inaccurate:

>>> a = np.zeros((2,512*512), dtype=np.float32)

>>> a[0,:] = 1.0
>>> a[1,:] = 0.1
>>> np.var(a)
0.20405951142311096

Computing the standard deviation in float64 is more accurate:

>>> np.var(a, dtype=np.float64)

0.20249999932997387
>>> ((1-0.55)**2 + (0.1-0.55)**2)/2
0.20250000000000001

3.12.3 Correlating

corrcoef(x[, y, rowvar, bias]) Return correlation coefficients.

correlate(a, v[, mode, old_behavior]) Discrete, linear correlation of two 1-dimensional sequences.
cov(m[, y, rowvar, bias]) Estimate a covariance matrix, given data.
corrcoef(x, y=None, rowvar=1, bias=0)
Return correlation coefficients.
Please refer to the documentation for cov for more detail. The relationship between the correlation coefficient
matrix, P, and the covariance matrix, C, is
Cij
Pij = p
Cii Cjj

The values of P are between -1 and 1.

Parameters
m : array_like
A 1-D or 2-D array containing multiple variables and observations. Each row of m
represents a variable, and each column a single observation of all those variables. Also
see rowvar below.
y : array_like, optional
An additional set of variables and observations. y has the same shape as m.
rowvar : int, optional
If rowvar is non-zero (default), then each row represents a variable, with observations
in the columns. Otherwise, the relationship is transposed: each column represents a
variable, while the rows contain observations.

3.12. Statistics 703

NumPy Reference, Release 2.0.0.dev8464

bias : int, optional

Default normalization is by (N-1), where N is the number of observations given (un-
biased estimate). If bias is 1, then normalization is by N.
Returns
out : ndarray
The correlation coefficient matrix of the variables.
See Also:

cov
Covariance matrix

correlate(a, v, mode=valid, old_behavior=True)

Discrete, linear correlation of two 1-dimensional sequences.
This function is equivalent to

>>> np.convolve(a, v[::-1], mode=mode)

...

where v[::-1] is the reverse of v.

Parameters
a, v : array_like
Input sequences.
mode : {valid, same, full}, optional
Refer to the convolve docstring. Note that the default is valid, unlike convolve, which
uses full.
old_behavior : bool
If True, uses the old, numeric behavior (correlate(a,v) == correlate(v, a), and the conju-
gate is not taken for complex arrays). If False, uses the conventional signal processing
definition (see note).
See Also:

convolve
Discrete, linear convolution of two one-dimensional sequences.
acorrelate
Discrete correlation following the usual signal processing definition for complex arrays, and without as-
suming that correlate(a, b) == correlate(b, a).

Notes
If old_behavior is False, this function computes the correlation as generally defined in signal processing texts:

z[k] = sum_n a[n] * conj(v[n+k])

with a and v sequences being zero-padded where necessary and conj being the conjugate.

704 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.correlate([1, 2, 3], [0, 1, 0.5])

array([ 3.5])
>>> np.correlate([1, 2, 3], [0, 1, 0.5], "same")
array([ 2. , 3.5, 3. ])
>>> np.correlate([1, 2, 3], [0, 1, 0.5], "full")
array([ 0.5, 2. , 3.5, 3. , 0. ])

cov(m, y=None, rowvar=1, bias=0)

Estimate a covariance matrix, given data.
Covariance indicates the level to which two variables vary together. If we examine N-dimensional samples,
X = [x1 , x2 , ...xN ]T , then the covariance matrix element Cij is the covariance of xi and xj . The element Cii
is the variance of xi .
Parameters
m : array_like
A 1-D or 2-D array containing multiple variables and observations. Each row of m
represents a variable, and each column a single observation of all those variables. Also
see rowvar below.
y : array_like, optional
An additional set of variables and observations. y has the same form as that of m.
rowvar : int, optional
If rowvar is non-zero (default), then each row represents a variable, with observations
in the columns. Otherwise, the relationship is transposed: each column represents a
variable, while the rows contain observations.
bias : int, optional
Default normalization is by (N-1), where N is the number of observations given (un-
biased estimate). If bias is 1, then normalization is by N.
Returns
out : ndarray
The covariance matrix of the variables.
See Also:

corrcoef
Normalized covariance matrix

Examples
Consider two variables, x0 and x1 , which correlate perfectly, but in opposite directions:

>>> x = np.array([[0, 2], [1, 1], [2, 0]]).T

>>> x
array([[0, 1, 2],
[2, 1, 0]])

Note how x0 increases while x1 decreases. The covariance matrix shows this clearly:

3.12. Statistics 705

NumPy Reference, Release 2.0.0.dev8464

>>> np.cov(x)
array([[ 1., -1.],
[-1., 1.]])

Note that element C0,1 , which shows the correlation between x0 and x1 , is negative.
Further, note how x and y are combined:

>>> x = [-2.1, -1, 4.3]

>>> y = [3, 1.1, 0.12]
>>> X = np.vstack((x,y))
>>> print np.cov(X)
[[ 11.71 -4.286 ]
[ -4.286 2.14413333]]
>>> print np.cov(x, y)
[[ 11.71 -4.286 ]
[ -4.286 2.14413333]]
>>> print np.cov(x)
11.71

3.12.4 Histograms

histogram(a[, bins, range, normed, Compute the histogram of a set of data.

weights])
histogram2d(x, y[, bins, range, normed, Compute the bi-dimensional histogram of two data samples.
weights])
histogramdd(sample[, bins, range, normed, Compute the multidimensional histogram of some data.
...])
bincount(x[, weights]) Count number of occurrences of each value in array of
non-negative ints.
digitize(x, bins) Return the indices of the bins to which each value in input array
belongs.
histogram(a, bins=10, range=None, normed=False, weights=None)
Compute the histogram of a set of data.
Parameters
a : array_like
Input data. The histogram is computed over the flattened array.
bins : int or sequence of scalars, optional
If bins is an int, it defines the number of equal-width bins in the given range (10, by
default). If bins is a sequence, it defines the bin edges, including the rightmost edge,
allowing for non-uniform bin widths.
range : (float, float), optional
The lower and upper range of the bins. If not provided, range is simply (a.min(),
a.max()). Values outside the range are ignored.
normed : bool, optional
If False, the result will contain the number of samples in each bin. If True, the result is
the value of the probability density function at the bin, normalized such that the integral
over the range is 1. Note that the sum of the histogram values will not be equal to 1
unless bins of unity width are chosen; it is not a probability mass function.

706 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

weights : array_like, optional

An array of weights, of the same shape as a. Each value in a only contributes its
associated weight towards the bin count (instead of 1). If normed is True, the weights
are normalized, so that the integral of the density over the range remains 1
Returns
hist : array
The values of the histogram. See normed and weights for a description of the possible
semantics.
bin_edges : array of dtype float
Return the bin edges (length(hist)+1).
See Also:
histogramdd, bincount, searchsorted

Notes
All but the last (righthand-most) bin is half-open. In other words, if bins is:

[1, 2, 3, 4]

then the first bin is [1, 2) (including 1, but excluding 2) and the second [2, 3). The last bin, however, is
[3, 4], which includes 4.

Examples

>>> np.histogram([1, 2, 1], bins=[0, 1, 2, 3])

(array([0, 2, 1]), array([0, 1, 2, 3]))
>>> np.histogram(np.arange(4), bins=np.arange(5), normed=True)
(array([ 0.25, 0.25, 0.25, 0.25]), array([0, 1, 2, 3, 4]))
>>> np.histogram([[1, 2, 1], [1, 0, 1]], bins=[0,1,2,3])
(array([1, 4, 1]), array([0, 1, 2, 3]))

>>> a = np.arange(5)
>>> hist, bin_edges = np.histogram(a, normed=True)
>>> hist
array([ 0.5, 0. , 0.5, 0. , 0. , 0.5, 0. , 0.5, 0. , 0.5])
>>> hist.sum()
2.4999999999999996
>>> np.sum(hist*np.diff(bin_edges))
1.0

histogram2d(x, y, bins=10, range=None, normed=False, weights=None)

Compute the bi-dimensional histogram of two data samples.
Parameters
x : array_like, shape(N,)
A sequence of values to be histogrammed along the first dimension.
y : array_like, shape(M,)
A sequence of values to be histogrammed along the second dimension.
bins : int or [int, int] or array_like or [array, array], optional
The bin specification:

3.12. Statistics 707

NumPy Reference, Release 2.0.0.dev8464

If int, the number of bins for the two dimensions (nx=ny=bins).

If [int, int], the number of bins in each dimension (nx, ny = bins).
If array_like, the bin edges for the two dimensions (x_edges=y_edges=bins).
If [array, array], the bin edges in each dimension (x_edges, y_edges = bins).
range : array_like, shape(2,2), optional
The leftmost and rightmost edges of the bins along each dimension (if not specified ex-
plicitly in the bins parameters): [[xmin, xmax], [ymin, ymax]]. All values
outside of this range will be considered outliers and not tallied in the histogram.
normed : bool, optional
If False, returns the number of samples in each bin. If True, returns the bin density, i.e.
the bin count divided by the bin area.
weights : array_like, shape(N,), optional
An array of values w_i weighing each sample (x_i, y_i). Weights are normalized
to 1 if normed is True. If normed is False, the values of the returned histogram are equal
to the sum of the weights belonging to the samples falling into each bin.
Returns
H : ndarray, shape(nx, ny)
The bi-dimensional histogram of samples x and y. Values in x are histogrammed along
the first dimension and values in y are histogrammed along the second dimension.
xedges : ndarray, shape(nx,)
The bin edges along the first dimension.
yedges : ndarray, shape(ny,)
The bin edges along the second dimension.
See Also:

histogram
1D histogram
histogramdd
Multidimensional histogram

Notes
When normed is True, then the returned histogram is the sample density, defined such that:

X ny1
nx1 X
Hi,j xi yj = 1
i=0 j=0

where H is the histogram array and xi yi the area of bin {i,j}.

Please note that the histogram does not follow the Cartesian convention where x values are on the abcissa and
y values on the ordinate axis. Rather, x is histogrammed along the first dimension of the array (vertical), and y
along the second dimension of the array (horizontal). This ensures compatibility with histogramdd.

708 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x, y = np.random.randn(2, 100)

>>> H, xedges, yedges = np.histogram2d(x, y, bins=(5, 8))
>>> H.shape, xedges.shape, yedges.shape
((5, 8), (6,), (9,))

We can now use the Matplotlib to visualize this 2-dimensional histogram:

>>> extent = [xedges[0], xedges[-1], yedges[0], yedges[-1]]

>>> import matplotlib.pyplot as plt
>>> plt.imshow(H, extent=extent)
<matplotlib.image.AxesImage object at ...>
>>> plt.show()

2
2 1 0 1 2

histogramdd(sample, bins=10, range=None, normed=False, weights=None)

Compute the multidimensional histogram of some data.
Parameters
sample : array_like
The data to be histogrammed. It must be an (N,D) array or data that can be converted to
such. The rows of the resulting array are the coordinates of points in a D dimensional
polytope.
bins : sequence or int, optional
The bin specification:
A sequence of arrays describing the bin edges along each dimension.
The number of bins for each dimension (nx, ny, ... =bins)
The number of bins for all dimensions (nx=ny=...=bins).
range : sequence, optional
A sequence of lower and upper bin edges to be used if the edges are not given explicitely
in bins. Defaults to the minimum and maximum values along each dimension.
normed : boolean, optional

3.12. Statistics 709

NumPy Reference, Release 2.0.0.dev8464

If False, returns the number of samples in each bin. If True, returns the bin density, ie,
the bin count divided by the bin hypervolume.
weights : array_like (N,), optional
An array of values w_i weighing each sample (x_i, y_i, z_i, ...). Weights are normalized
to 1 if normed is True. If normed is False, the values of the returned histogram are equal
to the sum of the weights belonging to the samples falling into each bin.
Returns
H : ndarray
The multidimensional histogram of sample x. See normed and weights for the different
possible semantics.
edges : list
A list of D arrays describing the bin edges for each dimension.
See Also:

histogram
1D histogram
histogram2d
2D histogram

Examples

>>> r = np.random.randn(100,3)
>>> H, edges = np.histogramdd(r, bins = (5, 8, 4))
>>> H.shape, edges[0].size, edges[1].size, edges[2].size
((5, 8, 4), 6, 9, 5)

bincount(x, weights=None)
Count number of occurrences of each value in array of non-negative ints.
The number of bins (of size 1) is one larger than the largest value in x. Each bin gives the number of occurrences
of its index value in x. If weights is specified the input array is weighted by it, i.e. if a value n is found at position
i, out[n] += weight[i] instead of out[n] += 1.
Parameters
x : array_like, 1 dimension, nonnegative ints
Input array.
weights : array_like, optional
Weights, array of the same shape as x.
Returns
out : ndarray of ints
The result of binning the input array. The length of out is equal to np.amax(x)+1.
Raises
ValueError :
If the input is not 1-dimensional, or contains elements with negative values.
TypeError :
If the type of the input is float or complex.

710 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:
histogram, digitize, unique

Examples

>>> np.bincount(np.arange(5))
array([1, 1, 1, 1, 1])
>>> np.bincount(np.array([0, 1, 1, 3, 2, 1, 7]))
array([1, 3, 1, 1, 0, 0, 0, 1])

>>> x = np.array([0, 1, 1, 3, 2, 1, 7, 23])

>>> np.bincount(x).size == np.amax(x)+1
True

>>> np.bincount(np.arange(5, dtype=np.float))

Traceback (most recent call last):
File "<stdin>", line 1, in <module>
TypeError: array cannot be safely cast to required type

A possible use of bincount is to perform sums over variable-size chunks of an array, using the weights
keyword.

>>> w = np.array([0.3, 0.5, 0.2, 0.7, 1., -0.6]) # weights

>>> x = np.array([0, 1, 1, 2, 2, 2])
>>> np.bincount(x, weights=w)
array([ 0.3, 0.7, 1.1])

digitize(x, bins)
Return the indices of the bins to which each value in input array belongs.
Each index i returned is such that bins[i-1] <= x < bins[i] if bins is monotonically increasing, or
bins[i-1] > x >= bins[i] if bins is monotonically decreasing. If values in x are beyond the bounds of
bins, 0 or len(bins) is returned as appropriate.
Parameters
x : array_like
Input array to be binned. It has to be 1-dimensional.
bins : array_like
Array of bins. It has to be 1-dimensional and monotonic.
Returns
out : ndarray of ints
Output array of indices, of same shape as x.
Raises
ValueError :
If the input is not 1-dimensional, or if bins is not monotonic.
TypeError :
If the type of the input is complex.
See Also:
bincount, histogram, unique

3.12. Statistics 711

NumPy Reference, Release 2.0.0.dev8464

Notes
If values in x are such that they fall outside the bin range, attempting to index bins with the indices that digitize
returns will result in an IndexError.

Examples

>>> x = np.array([0.2, 6.4, 3.0, 1.6])

>>> bins = np.array([0.0, 1.0, 2.5, 4.0, 10.0])
>>> inds = np.digitize(x, bins)
>>> inds
array([1, 4, 3, 2])
>>> for n in range(x.size):
... print bins[inds[n]-1], "<=", x[n], "<", bins[inds[n]]
...
0.0 <= 0.2 < 1.0
4.0 <= 6.4 < 10.0
2.5 <= 3.0 < 4.0
1.0 <= 1.6 < 2.5

3.13 Mathematical functions

3.13.1 Trigonometric functions

sin() Trigonometric sine, element-wise.

cos() Cosine elementwise.
tan() Compute tangent element-wise.
arcsin() Inverse sine elementwise.
arccos() Trigonometric inverse cosine, element-wise.
arctan() Trigonometric inverse tangent, element-wise.
hypot(x1) Given the legs of a right triangle, return its hypotenuse.
arctan2(x1) Elementwise arc tangent of x1/x2 choosing the quadrant correctly.
degrees() Convert angles from radians to degrees.
radians() Convert angles from degrees to radians.
unwrap(p[, discont, axis]) Unwrap by changing deltas between values to 2*pi complement.
deg2rad() Convert angles from degrees to radians.
rad2deg() Convert angles from radians to degrees.
sin
Trigonometric sine, element-wise.
Parameters
x : array_like
Angle, in radians (2 rad equals 360 degrees).
Returns
y : array_like
The sine of each element of x.
See Also:
arcsin, sinh, cos

712 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notes
The sine is one of the fundamental functions of trigonometry (the mathematical study of triangles). Consider
a circle of radius 1 centered on the origin. A ray comes in from the +x axis, makes an angle at the origin
(measured counter-clockwise from that axis), and departs from the origin. The y coordinate of the outgoing
rays intersection with the unit circle is the sine of that angle. It ranges from -1 for x = 3/2 to +1 for /2.
The function has zeroes where the angle is a multiple of . Sines of angles between and 2 are negative. The
numerous properties of the sine and related functions are included in any standard trigonometry text.

Examples
Print sine of one angle:

>>> np.sin(np.pi/2.)
1.0

Print sines of an array of angles given in degrees:

>>> np.sin(np.array((0., 30., 45., 60., 90.)) * np.pi / 180. )

array([ 0. , 0.5 , 0.70710678, 0.8660254 , 1. ])

Plot the sine function:

>>> import matplotlib.pylab as plt

>>> x = np.linspace(-np.pi, np.pi, 201)
>>> plt.plot(x, np.sin(x))
>>> plt.xlabel(Angle [rad])
>>> plt.ylabel(sin(x))
>>> plt.axis(tight)
>>> plt.show()

1.0

0.5
sin(x)

0.0

0.5

1.0
3 2 1 0 1 2 3
Angle [rad]

cos
Cosine elementwise.
Parameters
x : array_like
Input array in radians.

3.13. Mathematical functions 713

NumPy Reference, Release 2.0.0.dev8464

out : ndarray, optional

Output array of same shape as x.
Returns
y : ndarray
The corresponding cosine values.
Raises
ValueError: invalid return array shape :
if out is provided and out.shape != x.shape (See Examples)

Notes
If out is provided, the function writes the result into it, and returns a reference to out. (See Examples)

References
M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions. New York, NY: Dover, 1972.

Examples

>>> np.cos(np.array([0, np.pi/2, np.pi]))

array([ 1.00000000e+00, 6.12303177e-17, -1.00000000e+00])
>>>
>>> # Example of providing the optional output parameter
>>> out2 = np.cos([0.1], out1)
>>> out2 is out1
True
>>>
>>> # Example of ValueError due to provision of shape mis-matched out
>>> np.cos(np.zeros((3,3)),np.zeros((2,2)))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: invalid return array shape

tan
Compute tangent element-wise.
Equivalent to np.sin(x)/np.cos(x) element-wise.
Parameters
x : array_like
Input array.
out : ndarray, optional
Output array of same shape as x.
Returns
y : ndarray
The corresponding tangent values.
Raises
ValueError: invalid return array shape :
if out is provided and out.shape != x.shape (See Examples)

714 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notes
If out is provided, the function writes the result into it, and returns a reference to out. (See Examples)

References
M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions. New York, NY: Dover, 1972.

Examples

>>> from math import pi

>>> np.tan(np.array([-pi,pi/2,pi]))
array([ 1.22460635e-16, 1.63317787e+16, -1.22460635e-16])
>>>
>>> # Example of providing the optional output parameter illustrating
>>> # that what is returned is a reference to said parameter
>>> out2 = np.cos([0.1], out1)
>>> out2 is out1
True
>>>
>>> # Example of ValueError due to provision of shape mis-matched out
>>> np.cos(np.zeros((3,3)),np.zeros((2,2)))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: invalid return array shape

arcsin
Inverse sine elementwise.
Parameters
x : array_like
y-coordinate on the unit circle.
out : ndarray, optional
Array of the same shape as x, to store results in. See doc.ufuncs (Section Output
arguments) for more details.
Returns
angle : ndarray
The angle of the ray intersecting the unit circle at the given y-coordinate in radians
[-pi, pi]. If x is a scalar then a scalar is returned, otherwise an array is returned.
See Also:
sin, cos, arccos, tan, arctan, arctan2, emath.arcsin

Notes
arcsin is a multivalued function: for each x there are infinitely many numbers z such that sin(z) = x. The
convention is to return the angle z whose real part lies in [-pi/2, pi/2].
For real-valued input data types, arcsin always returns real output. For each value that cannot be expressed as a
real number or infinity, it yields nan and sets the invalid floating point error flag.
For complex-valued input, arcsin is a complex analytical function that has branch cuts [-inf, -1] and [1, inf] and
is continuous from above on the former and from below on the latter.
The inverse sine is also known as asin or sin^-1.

3.13. Mathematical functions 715

NumPy Reference, Release 2.0.0.dev8464

References
[R6], [R7]

Examples

>>> np.arcsin(1) # pi/2

1.5707963267948966
>>> np.arcsin(-1) # -pi/2
-1.5707963267948966
>>> np.arcsin(0)
0.0

arccos
Trigonometric inverse cosine, element-wise.
The inverse of cos so that, if y = cos(x), then x = arccos(y).
Parameters
x : array_like
x-coordinate on the unit circle. For real arguments, the domain is [-1, 1].
out : ndarray, optional
Array of the same shape as a, to store results in. See doc.ufuncs (Section Output
arguments) for more details.
Returns
angle : ndarray
The angle of the ray intersecting the unit circle at the given x-coordinate in radians [0,
pi]. If x is a scalar then a scalar is returned, otherwise an array of the same shape as x is
returned.
See Also:
cos, arctan, arcsin, emath.arccos

Notes
arccos is a multivalued function: for each x there are infinitely many numbers z such that cos(z) = x. The
convention is to return the angle z whose real part lies in [0, pi].
For real-valued input data types, arccos always returns real output. For each value that cannot be expressed as a
real number or infinity, it yields nan and sets the invalid floating point error flag.
For complex-valued input, arccos is a complex analytical function that has branch cuts [-inf, -1] and [1, inf] and
is continuous from above on the former and from below on the latter.
The inverse cos is also known as acos or cos^-1.

References
[R2], [R3]

Examples
We expect the arccos of 1 to be 0, and of -1 to be pi:

>>> np.arccos([1, -1])

array([ 0. , 3.14159265])

716 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Plot arccos:

>>> import matplotlib.pyplot as plt

>>> x = np.linspace(-1, 1, num=100)
>>> plt.plot(x, np.arccos(x))
>>> plt.axis(tight)
>>> plt.show()

3.0
2.5
2.0
1.5
1.0
0.5
0.0
1.0 0.5 0.0 0.5 1.0

arctan
Trigonometric inverse tangent, element-wise.
The inverse of tan, so that if y = tan(x) then x = arctan(y).
Parameters
x : array_like
Input values. arctan is applied to each element of x.
Returns
out : ndarray
Out has the same shape as x. Its real part is in [-pi/2, pi/2]. It is a scalar if x is a
scalar.
See Also:

arctan2
Calculate the arctan of y/x.

Notes
arctan is a multivalued function: for each x there are infinitely many numbers z such that tan(z) = x. The
convention is to return the angle z whose real part lies in [-pi/2, pi/2].
For real-valued input data types, arctan always returns real output. For each value that cannot be expressed as a
real number or infinity, it yields nan and sets the invalid floating point error flag.
For complex-valued input, arctan is a complex analytical function that has branch cuts [1j, infj] and [-1j, -infj]
and is continuous from the left on the former and from the right on the latter.
The inverse tangent is also known as atan or tan^-1.

3.13. Mathematical functions 717

NumPy Reference, Release 2.0.0.dev8464

References
[R10], [R11]

Examples
We expect the arctan of 0 to be 0, and of 1 to be /4:

>>> np.arctan([0, 1])

array([ 0. , 0.78539816])

>>> np.pi/4
0.78539816339744828

Plot arctan:

>>> import matplotlib.pyplot as plt

>>> x = np.linspace(-10, 10)
>>> plt.plot(x, np.arctan(x))
>>> plt.axis(tight)
>>> plt.show()

1.0
0.5
0.0
0.5
1.0

10 5 0 5 10

hypot
Given the legs of a right triangle, return its hypotenuse.
Equivalent to sqrt(x1**2 + x2**2), element-wise. If x1 or x2 is scalar_like (i.e., unambiguously cast-
able to a scalar type), it is broadcast for use with each element of the other argument. (See Examples)
Parameters
x1, x2 : array_like
Leg of the triangle(s).
out : ndarray, optional
Array into which the output is placed. Its type is preserved and it must be of the right
shape to hold the output. See doc.ufuncs.
Returns
z : ndarray

718 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

The hypotenuse of the triangle(s).

Examples

>>> np.hypot(3*np.ones((3, 3)), 4*np.ones((3, 3)))

array([[ 5., 5., 5.],
[ 5., 5., 5.],
[ 5., 5., 5.]])

Example showing broadcast of scalar_like argument:

>>> np.hypot(3*np.ones((3, 3)), [4])

array([[ 5., 5., 5.],
[ 5., 5., 5.],
[ 5., 5., 5.]])

arctan2
Elementwise arc tangent of x1/x2 choosing the quadrant correctly.
The quadrant (ie. branch) is chosen so that arctan2(x1, x2) is the signed angle in radians between the line
segments (0,0) - (1,0) and (0,0) - (x2,x1). This function is defined also for x2 = 0.
arctan2 is not defined for complex-valued arguments.
Parameters
x1 : array_like, real-valued
y-coordinates.
x2 : array_like, real-valued
x-coordinates. x2 must be broadcastable to match the shape of x1, or vice versa.
Returns
angle : ndarray
Array of angles in radians, in the range [-pi, pi].
See Also:
arctan, tan

Notes
arctan2 is identical to the atan2 function of the underlying C library. The following special values are defined
in the C standard [2]:
x1 x2 arctan2(x1,x2)
+/- 0 +0 +/- 0
+/- 0 -0 +/- pi
>0 +/-inf +0 / +pi
<0 +/-inf -0 / -pi
+/-inf +inf +/- (pi/4)
+/-inf -inf +/- (3*pi/4)
Note that +0 and -0 are distinct floating point numbers.

References
[R12], [R13]

3.13. Mathematical functions 719

NumPy Reference, Release 2.0.0.dev8464

Examples
Consider four points in different quadrants:

>>> x = np.array([-1, +1, +1, -1])

>>> y = np.array([-1, -1, +1, +1])
>>> np.arctan2(y, x) * 180 / np.pi
array([-135., -45., 45., 135.])

Note the order of the parameters. arctan2 is defined also when x2 = 0 and at several other special points,
obtaining values in the range [-pi, pi]:

>>> np.arctan2([1., -1.], [0., 0.])

array([ 1.57079633, -1.57079633])
>>> np.arctan2([0., 0., np.inf], [+0., -0., np.inf])
array([ 0. , 3.14159265, 0.78539816])

degrees
Convert angles from radians to degrees.
Parameters
x : array_like
Input array in radians.
out : ndarray, optional
Output array of same shape as x.
Returns
y : ndarray of floats
The corresponding degree values.
See Also:

rad2deg
equivalent function

Examples
Convert a radian array to degrees

>>> rad = np.arange(12.)*np.pi/6

>>> np.degrees(rad)
array([ 0., 30., 60., 90., 120., 150., 180., 210., 240.,
270., 300., 330.])

>>> out = np.zeros((rad.shape))

>>> r = degrees(rad, out)
>>> np.all(r == out)
True

radians
Convert angles from degrees to radians.
Parameters
x : array_like
Input array in degrees.

720 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

out : ndarray, optional

Output array of same shape as x.
Returns
y : ndarray
The corresponding radian values.
See Also:

deg2rad
equivalent function

Examples
Convert a degree array to radians

>>> deg = np.arange(12.) * 30.

>>> np.radians(deg)
array([ 0. , 0.52359878, 1.04719755, 1.57079633, 2.0943951 ,
2.61799388, 3.14159265, 3.66519143, 4.1887902 , 4.71238898,
5.23598776, 5.75958653])

>>> out = np.zeros((deg.shape))

>>> ret = np.radians(deg, out)
>>> ret is out
True

unwrap(p, discont=3.1415926535897931, axis=-1)

Unwrap by changing deltas between values to 2*pi complement.
Unwrap radian phase p by changing absolute jumps greater than discont to their 2*pi complement along the
given axis.
Parameters
p : array_like
Input array.
discont : float, optional
Maximum discontinuity between values, default is pi.
axis : int, optional
Axis along which unwrap will operate, default is the last axis.
Returns
out : ndarray
Output array.
See Also:
rad2deg, deg2rad

Notes
If the discontinuity in p is smaller than pi, but larger than discont, no unwrapping is done because taking the
2*pi complement would only make the discontinuity larger.

3.13. Mathematical functions 721

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> phase = np.linspace(0, np.pi, num=5)

>>> phase[3:] += np.pi
>>> phase
array([ 0. , 0.78539816, 1.57079633, 5.49778714, 6.28318531])
>>> np.unwrap(phase)
array([ 0. , 0.78539816, 1.57079633, -0.78539816, 0. ])

deg2rad
Convert angles from degrees to radians.
Parameters
x : array_like
Angles in degrees.
Returns
y : ndarray
The corresponding angle in radians.
See Also:

rad2deg
Convert angles from radians to degrees.
unwrap
Remove large jumps in angle by wrapping.

Notes
New in version 1.3.0. deg2rad(x) is x * pi / 180.

Examples

>>> np.deg2rad(180)
3.1415926535897931

rad2deg
Convert angles from radians to degrees.
Parameters
x : array_like
Angle in radians.
out : ndarray, optional
Array into which the output is placed. Its type is preserved and it must be of the right
shape to hold the output. See doc.ufuncs.
Returns
y : ndarray
The corresponding angle in degrees.
See Also:

deg2rad
Convert angles from degrees to radians.

722 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

unwrap
Remove large jumps in angle by wrapping.

Notes
New in version 1.3.0. rad2deg(x) is 180 * x / pi.

Examples

>>> np.rad2deg(np.pi/2)
90.0

3.13.2 Hyperbolic functions

sinh() Hyperbolic sine, element-wise.

cosh() Hyperbolic cosine, element-wise.
tanh() Compute hyperbolic tangent element-wise.
arcsinh() Inverse hyperbolic sine elementwise.
arccosh() Inverse hyperbolic cosine, elementwise.
arctanh() Inverse hyperbolic tangent elementwise.
sinh
Hyperbolic sine, element-wise.
Equivalent to 1/2 * (np.exp(x) - np.exp(-x)) or -1j * np.sin(1j*x).
Parameters
x : array_like
Input array.
out : ndarray, optional
Output array of same shape as x.
Returns
y : ndarray
The corresponding hyperbolic sine values.
Raises
ValueError: invalid return array shape :
if out is provided and out.shape != x.shape (See Examples)

Notes
If out is provided, the function writes the result into it, and returns a reference to out. (See Examples)

References
M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions. New York, NY: Dover, 1972, pg. 83.

Examples

>>> np.sinh(0)
0.0
>>> np.sinh(np.pi*1j/2)
1j
>>> np.sinh(np.pi*1j) # (exact value is 0)

3.13. Mathematical functions 723

NumPy Reference, Release 2.0.0.dev8464

1.2246063538223773e-016j
>>> # Discrepancy due to vagaries of floating point arithmetic.

>>> # Example of providing the optional output parameter

>>> out2 = np.sinh([0.1], out1)
>>> out2 is out1
True

>>> # Example of ValueError due to provision of shape mis-matched out

>>> np.sinh(np.zeros((3,3)),np.zeros((2,2)))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: invalid return array shape

cosh
Hyperbolic cosine, element-wise.
Equivalent to 1/2 * (np.exp(x) + np.exp(-x)) and np.cos(1j*x).
Parameters
x : array_like
Input array.
Returns
out : ndarray
Output array of same shape as x.

Examples

>>> np.cosh(0)
1.0

The hyperbolic cosine describes the shape of a hanging cable:

>>> import matplotlib.pyplot as plt

>>> x = np.linspace(-4, 4, 1000)
>>> plt.plot(x, np.cosh(x))
>>> plt.show()

724 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

30
25
20
15
10
5
0
4 3 2 1 0 1 2 3 4

tanh
Compute hyperbolic tangent element-wise.
Equivalent to np.sinh(x)/np.cosh(x) or -1j * np.tan(1j*x).
Parameters
x : array_like
Input array.
out : ndarray, optional
Output array of same shape as x.
Returns
y : ndarray
The corresponding hyperbolic tangent values.
Raises
ValueError: invalid return array shape :
if out is provided and out.shape != x.shape (See Examples)

Notes
If out is provided, the function writes the result into it, and returns a reference to out. (See Examples)

References
[R272], [R273]

Examples

>>> np.tanh((0, np.pi*1j, np.pi*1j/2))

array([ 0. +0.00000000e+00j, 0. -1.22460635e-16j, 0. +1.63317787e+16j])

>>> # Example of providing the optional output parameter illustrating

>>> # that what is returned is a reference to said parameter
>>> out2 = np.tanh([0.1], out1)

3.13. Mathematical functions 725

NumPy Reference, Release 2.0.0.dev8464

>>> out2 is out1

True

>>> # Example of ValueError due to provision of shape mis-matched out

>>> np.tanh(np.zeros((3,3)),np.zeros((2,2)))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: invalid return array shape

arcsinh
Inverse hyperbolic sine elementwise.
Parameters
x : array_like
Input array.
out : ndarray, optional
Array into which the output is placed. Its type is preserved and it must be of the right
shape to hold the output. See doc.ufuncs.
Returns
out : ndarray
Array of of the same shape as x.

Notes
arcsinh is a multivalued function: for each x there are infinitely many numbers z such that sinh(z) = x. The
convention is to return the z whose imaginary part lies in [-pi/2, pi/2].
For real-valued input data types, arcsinh always returns real output. For each value that cannot be expressed as
a real number or infinity, it returns nan and sets the invalid floating point error flag.
For complex-valued input, arccos is a complex analytical function that has branch cuts [1j, infj] and [-1j, -infj]
and is continuous from the right on the former and from the left on the latter.
The inverse hyperbolic sine is also known as asinh or sinh^-1.

References
[R8], [R9]

Examples

>>> np.arcsinh(np.array([np.e, 10.0]))

array([ 1.72538256, 2.99822295])

arccosh
Inverse hyperbolic cosine, elementwise.
Parameters
x : array_like
Input array.
out : ndarray, optional
Array of the same shape as x, to store results in. See doc.ufuncs (Section Output
arguments) for details.

726 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Returns
y : ndarray
Array of the same shape as x.
See Also:
cosh, arcsinh, sinh, arctanh, tanh

Notes
arccosh is a multivalued function: for each x there are infinitely many numbers z such that cosh(z) = x. The
convention is to return the z whose imaginary part lies in [-pi, pi] and the real part in [0, inf].
For real-valued input data types, arccosh always returns real output. For each value that cannot be expressed as
a real number or infinity, it yields nan and sets the invalid floating point error flag.
For complex-valued input, arccosh is a complex analytical function that has a branch cut [-inf, 1] and is contin-
uous from above on it.

References
[R4], [R5]

Examples

>>> np.arccosh([np.e, 10.0])

array([ 1.65745445, 2.99322285])
>>> np.arccosh(1)
0.0

arctanh
Inverse hyperbolic tangent elementwise.
Parameters
x : array_like
Input array.
Returns
out : ndarray
Array of the same shape as x.
See Also:
emath.arctanh

Notes
arctanh is a multivalued function: for each x there are infinitely many numbers z such that tanh(z) = x. The
convention is to return the z whose imaginary part lies in [-pi/2, pi/2].
For real-valued input data types, arctanh always returns real output. For each value that cannot be expressed as
a real number or infinity, it yields nan and sets the invalid floating point error flag.
For complex-valued input, arctanh is a complex analytical function that has branch cuts [-1, -inf] and [1, inf]
and is continuous from above on the former and from below on the latter.
The inverse hyperbolic tangent is also known as atanh or tanh^-1.

References
[R14], [R15]

3.13. Mathematical functions 727

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.arctanh([0, -0.5])

array([ 0. , -0.54930614])

3.13.3 Rounding

around(a[, decimals, out]) Evenly round to the given number of decimals.

round_(a[, decimals, out]) Round an array to the given number of decimals.
rint() Round elements of the array to the nearest integer.
fix(x[, y]) Round to nearest integer towards zero.
floor() Return the floor of the input, element-wise.
ceil() Return the ceiling of the input, element-wise.
trunc() Return the truncated value of the input, element-wise.
around(a, decimals=0, out=None)
Evenly round to the given number of decimals.
Parameters
a : array_like
Input data.
decimals : int, optional
Number of decimal places to round to (default: 0). If decimals is negative, it specifies
the number of positions to the left of the decimal point.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as
the expected output, but the type of the output values will be cast if necessary. See
doc.ufuncs (Section Output arguments) for details.
Returns
rounded_array : ndarray
An array of the same type as a, containing the rounded values. Unless out was specified,
a new array is created. A reference to the result is returned.
The real and imaginary parts of complex numbers are rounded separately. The result of
rounding a float is a float.
See Also:

ndarray.round
equivalent method

ceil, fix, floor, rint, trunc

Notes
For values exactly halfway between rounded decimal values, Numpy rounds to the nearest even value. Thus
1.5 and 2.5 round to 2.0, -0.5 and 0.5 round to 0.0, etc. Results may also be surprising due to the inexact
representation of decimal fractions in the IEEE floating point standard [R16] and errors introduced when scaling
by powers of ten.

728 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

References
[R16], [R17]

Examples

>>> np.around([0.37, 1.64])

array([ 0., 2.])
>>> np.around([0.37, 1.64], decimals=1)
array([ 0.4, 1.6])
>>> np.around([.5, 1.5, 2.5, 3.5, 4.5]) # rounds to nearest even value
array([ 0., 2., 2., 4., 4.])
>>> np.around([1,2,3,11], decimals=1) # ndarray of ints is returned
array([ 1, 2, 3, 11])
>>> np.around([1,2,3,11], decimals=-1)
array([ 0, 0, 0, 10])

round_(a, decimals=0, out=None)

Round an array to the given number of decimals.
Refer to around for full documentation.
See Also:

around
equivalent function

rint
Round elements of the array to the nearest integer.
Parameters
x : array_like
Input array.
Returns
out : {ndarray, scalar}
Output array is same shape and type as x.
See Also:
ceil, floor, trunc

Examples

>>> a = np.array([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0])

>>> np.rint(a)
array([-2., -2., -0., 0., 2., 2., 2.])

fix(x, y=None)
Round to nearest integer towards zero.
Round an array of floats element-wise to nearest integer towards zero. The rounded values are returned as floats.
Parameters
x : array_like
An array of floats to be rounded
y : ndarray, optional

3.13. Mathematical functions 729

NumPy Reference, Release 2.0.0.dev8464

Output array
Returns
out : ndarray of floats
The array of rounded numbers
See Also:
trunc, floor, ceil
around
Round to given number of decimals

Examples

>>> np.fix(3.14)
3.0
>>> np.fix(3)
3.0
>>> np.fix([2.1, 2.9, -2.1, -2.9])
array([ 2., 2., -2., -2.])

floor
Return the floor of the input, element-wise.
The floor of the scalar x is the largest integer i, such that i . It is often denoted as bxc.
Parameters
x : array_like
Input data.
Returns
y : {ndarray, scalar}
The floor of each element in x.
See Also:
ceil, trunc, rint

Notes
Some spreadsheet programs calculate the floor-towards-zero, in other words floor(-2.5) == -2.
NumPy, however, uses the a definition of floor such that floor(-2.5) == -3.

Examples

>>> a = np.array([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0])

>>> np.floor(a)
array([-2., -2., -1., 0., 1., 1., 2.])

ceil
Return the ceiling of the input, element-wise.
The ceil of the scalar x is the smallest integer i, such that i >= x. It is often denoted as dxe.
Parameters
x : array_like
Input data.

730 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Returns
y : {ndarray, scalar}
The ceiling of each element in x, with float dtype.
See Also:
floor, trunc, rint

Examples

>>> a = np.array([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0])

>>> np.ceil(a)
array([-1., -1., -0., 1., 2., 2., 2.])

trunc
Return the truncated value of the input, element-wise.
The truncated value of the scalar x is the nearest integer i which is closer to zero than x is. In short, the fractional
part of the signed number x is discarded.
Parameters
x : array_like
Input data.
Returns
y : {ndarray, scalar}
The truncated value of each element in x.
See Also:
ceil, floor, rint

Notes
New in version 1.3.0.

Examples

>>> a = np.array([-1.7, -1.5, -0.2, 0.2, 1.5, 1.7, 2.0])

>>> np.trunc(a)
array([-1., -1., -0., 0., 1., 1., 2.])

3.13.4 Sums, products, differences

prod(a[, axis, dtype, out]) Return the product of array elements over a given axis.
sum(a[, axis, dtype, out]) Sum of array elements over a given axis.
nansum(a[, axis]) Return the sum of array elements over a given axis treating
cumprod(a[, axis, dtype, out]) Return the cumulative product of elements along a given axis.
cumsum(a[, axis, dtype, out]) Return the cumulative sum of the elements along a given axis.
diff(a[, n, axis]) Calculate the n-th order discrete difference along given axis.
ediff1d(ary[, to_end, to_begin]) The differences between consecutive elements of an array.
gradient(f, *varargs) Return the gradient of an N-dimensional array.
cross(a, b[, axisa, axisb, axisc, axis]) Return the cross product of two (arrays of) vectors.
trapz(y[, x, dx, axis]) Integrate along the given axis using the composite trapezoidal rule.

3.13. Mathematical functions 731

NumPy Reference, Release 2.0.0.dev8464

prod(a, axis=None, dtype=None, out=None)

Return the product of array elements over a given axis.
Parameters
a : array_like
Input data.
axis : int, optional
Axis over which the product is taken. By default, the product of all elements is calcu-
lated.
dtype : data-type, optional
The data-type of the returned array, as well as of the accumulator in which the elements
are multiplied. By default, if a is of integer type, dtype is the default platform integer.
(Note: if the type of a is unsigned, then so is dtype.) Otherwise, the dtype is the same
as that of a.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output, but the type of the output values will be cast if necessary.
Returns
product_along_axis : ndarray, see dtype parameter above.
An array shaped as a but with the specified axis removed. Returns a reference to out if
specified.
See Also:

ndarray.prod
equivalent method
numpy.doc.ufuncs
Section Output arguments

Notes
Arithmetic is modular when using integer types, and no error is raised on overflow. That means that, on a 32-bit
platform:

>>> x = np.array([536870910, 536870910, 536870910, 536870910])

>>> np.prod(x) #random
16

Examples
By default, calculate the product of all elements:

>>> np.prod([1.,2.])
2.0

Even when the input array is two-dimensional:

>>> np.prod([[1.,2.],[3.,4.]])
24.0

But we can also specify the axis over which to multiply:

732 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.prod([[1.,2.],[3.,4.]], axis=1)

array([ 2., 12.])

If the type of x is unsigned, then the output type is the unsigned platform integer:

>>> x = np.array([1, 2, 3], dtype=np.uint8)

>>> np.prod(x).dtype == np.uint
True

If x is of a signed integer type, then the output type is the default platform integer:

>>> x = np.array([1, 2, 3], dtype=np.int8)

>>> np.prod(x).dtype == np.int
True

sum(a, axis=None, dtype=None, out=None)

Sum of array elements over a given axis.
Parameters
a : array_like
Elements to sum.
axis : integer, optional
Axis over which the sum is taken. By default axis is None, and all elements are summed.
dtype : dtype, optional
The type of the returned array and of the accumulator in which the elements are
summed. By default, the dtype of a is used. An exception is when a has an integer
type with less precision than the default platform integer. In that case, the default plat-
form integer is used instead.
out : ndarray, optional
Array into which the output is placed. By default, a new array is created. If out is
given, it must be of the appropriate shape (the shape of a with axis removed, i.e.,
numpy.delete(a.shape, axis)). Its type is preserved. See doc.ufuncs (Sec-
tion Output arguments) for more details.
Returns
sum_along_axis : ndarray
An array with the same shape as a, with the specified axis removed. If a is a 0-d array,
or if axis is None, a scalar is returned. If an output array is specified, a reference to out
is returned.
See Also:

ndarray.sum
Equivalent method.
cumsum
Cumulative sum of array elements.
trapz
Integration of array values using the composite trapezoidal rule.

mean, average

3.13. Mathematical functions 733

NumPy Reference, Release 2.0.0.dev8464

Notes
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> np.sum([0.5, 1.5])

2.0
>>> np.sum([0.5, 0.7, 0.2, 1.5], dtype=np.int32)
1
>>> np.sum([[0, 1], [0, 5]])
6
>>> np.sum([[0, 1], [0, 5]], axis=0)
array([0, 6])
>>> np.sum([[0, 1], [0, 5]], axis=1)
array([1, 5])

If the accumulator is too small, overflow occurs:

>>> np.ones(128, dtype=np.int8).sum(dtype=np.int8)

-128

nansum(a, axis=None)
Return the sum of array elements over a given axis treating Not a Numbers (NaNs) as zero.
Parameters
a : array_like
Array containing numbers whose sum is desired. If a is not an array, a conversion is
attempted.
axis : int, optional
Axis along which the sum is computed. The default is to compute the sum of the
flattened array.
Returns
y : ndarray
An array with the same shape as a, with the specified axis removed. If a is a 0-d array,
or if axis is None, a scalar is returned with the same dtype as a.
See Also:

numpy.sum
Sum across array including Not a Numbers.
isnan
Shows which elements are Not a Number (NaN).
isfinite
Shows which elements are not: Not a Number, positive and negative infinity

Notes
Numpy uses the IEEE Standard for Binary Floating-Point for Arithmetic (IEEE 754). This means that Not a
Number is not equivalent to infinity. If positive or negative infinity are present the result is positive or negative
infinity. But if both positive and negative infinity are present, the result is Not A Number (NaN).
Arithmetic is modular when using integer types (all elements of a must be finite i.e. no elements that are NaNs,
positive infinity and negative infinity because NaNs are floating point types), and no error is raised on overflow.

734 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.nansum(1)
1
>>> np.nansum([1])
1
>>> np.nansum([1, np.nan])
1.0
>>> a = np.array([[1, 1], [1, np.nan]])
>>> np.nansum(a)
3.0
>>> np.nansum(a, axis=0)
array([ 2., 1.])

When positive infinity and negative infinity are present

>>> np.nansum([1, np.nan, np.inf])

inf
>>> np.nansum([1, np.nan, np.NINF])
-inf
>>> np.nansum([1, np.nan, np.inf, np.NINF])
nan

cumprod(a, axis=None, dtype=None, out=None)

Return the cumulative product of elements along a given axis.
Parameters
a : array_like
Input array.
axis : int, optional
Axis along which the cumulative product is computed. By default the input is flattened.
dtype : dtype, optional
Type of the returned array, as well as of the accumulator in which the elements are
multiplied. If dtype is not specified, it defaults to the dtype of a, unless a has an integer
dtype with a precision less than that of the default platform integer. In that case, the
default platform integer is used instead.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type of the resulting values will be cast if
necessary.
Returns
cumprod : ndarray
A new array holding the result is returned unless out is specified, in which case a refer-
ence to out is returned.
See Also:

numpy.doc.ufuncs
Section Output arguments

3.13. Mathematical functions 735

NumPy Reference, Release 2.0.0.dev8464

Notes
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> a = np.array([1,2,3])
>>> np.cumprod(a) # intermediate results 1, 1*2
... # total product 1*2*3 = 6
array([1, 2, 6])
>>> a = np.array([[1, 2, 3], [4, 5, 6]])
>>> np.cumprod(a, dtype=float) # specify type of output
array([ 1., 2., 6., 24., 120., 720.])

The cumulative product for each column (i.e., over the rows of) a:

>>> np.cumprod(a, axis=0)

array([[ 1, 2, 3],
[ 4, 10, 18]])

The cumulative product for each row (i.e. over the columns of) a:

>>> np.cumprod(a,axis=1)
array([[ 1, 2, 6],
[ 4, 20, 120]])

cumsum(a, axis=None, dtype=None, out=None)

Return the cumulative sum of the elements along a given axis.
Parameters
a : array_like
Input array.
axis : int, optional
Axis along which the cumulative sum is computed. The default (None) is to compute
the cumsum over the flattened array.
dtype : dtype, optional
Type of the returned array and of the accumulator in which the elements are summed. If
dtype is not specified, it defaults to the dtype of a, unless a has an integer dtype with a
precision less than that of the default platform integer. In that case, the default platform
integer is used.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary. See doc.ufuncs
(Section Output arguments) for more details.
Returns
cumsum_along_axis : ndarray.
A new array holding the result is returned unless out is specified, in which case a ref-
erence to out is returned. The result has the same size as a, and the same shape as a if
axis is not None or a is a 1-d array.
See Also:

736 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

sum
Sum array elements.
trapz
Integration of array values using the composite trapezoidal rule.

Notes
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> a = np.array([[1,2,3], [4,5,6]])

>>> a
array([[1, 2, 3],
[4, 5, 6]])
>>> np.cumsum(a)
array([ 1, 3, 6, 10, 15, 21])
>>> np.cumsum(a, dtype=float) # specifies type of output value(s)
array([ 1., 3., 6., 10., 15., 21.])

>>> np.cumsum(a,axis=0) # sum over rows for each of the 3 columns

array([[1, 2, 3],
[5, 7, 9]])
>>> np.cumsum(a,axis=1) # sum over columns for each of the 2 rows
array([[ 1, 3, 6],
[ 4, 9, 15]])

diff(a, n=1, axis=-1)

Calculate the n-th order discrete difference along given axis.
The first order difference is given by out[n] = a[n+1] - a[n] along the given axis, higher order differ-
ences are calculated by using diff recursively.
Parameters
a : array_like
Input array
n : int, optional
The number of times values are differenced.
axis : int, optional
The axis along which the difference is taken, default is the last axis.
Returns
out : ndarray
The n order differences. The shape of the output is the same as a except along axis
where the dimension is smaller by n.
See Also:
gradient, ediff1d

Examples

>>> x = np.array([1, 2, 4, 7, 0])

>>> np.diff(x)

3.13. Mathematical functions 737

NumPy Reference, Release 2.0.0.dev8464

array([ 1, 2, 3, -7])
>>> np.diff(x, n=2)
array([ 1, 1, -10])

>>> x = np.array([[1, 3, 6, 10], [0, 5, 6, 8]])

>>> np.diff(x)
array([[2, 3, 4],
[5, 1, 2]])
>>> np.diff(x, axis=0)
array([[-1, 2, 0, -2]])

ediff1d(ary, to_end=None, to_begin=None)

The differences between consecutive elements of an array.
Parameters
ary : array_like
If necessary, will be flattened before the differences are taken.
to_end : array_like, optional
Number(s) to append at the end of the returned differences.
to_begin : array_like, optional
Number(s) to prepend at the beginning of the returned differences.
Returns
ed : ndarray
The differences. Loosely, this is ary.flat[1:] - ary.flat[:-1].
See Also:
diff, gradient

Notes
When applied to masked arrays, this function drops the mask information if the to_begin and/or to_end param-
eters are used.

Examples

>>> x = np.array([1, 2, 4, 7, 0])

>>> np.ediff1d(x)
array([ 1, 2, 3, -7])

>>> np.ediff1d(x, to_begin=-99, to_end=np.array([88, 99]))

array([-99, 1, 2, 3, -7, 88, 99])

The returned array is always 1D.

>>> y = [[1, 2, 4], [1, 6, 24]]

>>> np.ediff1d(y)
array([ 1, 2, -3, 5, 18])

gradient(f, *varargs)
Return the gradient of an N-dimensional array.

738 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

The gradient is computed using central differences in the interior and first differences at the boundaries. The
returned gradient hence has the same shape as the input array.
Parameters
f : array_like
An N-dimensional array containing samples of a scalar function.
*varargs : scalars
0, 1, or N scalars specifying the sample distances in each direction, that is: dx, dy, dz,
... The default distance is 1.
Returns
g : ndarray
N arrays of the same shape as f giving the derivative of f with respect to each dimension.

Examples

>>> x = np.array([1, 2, 4, 7, 11, 16], dtype=np.float)

>>> np.gradient(x)
array([ 1. , 1.5, 2.5, 3.5, 4.5, 5. ])
>>> np.gradient(x, 2)
array([ 0.5 , 0.75, 1.25, 1.75, 2.25, 2.5 ])

>>> np.gradient(np.array([[1, 2, 6], [3, 4, 5]], dtype=np.float))

[array([[ 2., 2., -1.],
[ 2., 2., -1.]]),
array([[ 1. , 2.5, 4. ],
[ 1. , 1. , 1. ]])]

cross(a, b, axisa=-1, axisb=-1, axisc=-1, axis=None)

Return the cross product of two (arrays of) vectors.
The cross product of a and b in R3 is a vector perpendicular to both a and b. If a and b are arrays of vectors,
the vectors are defined by the last axis of a and b by default, and these axes can have dimensions 2 or 3. Where
the dimension of either a or b is 2, the third component of the input vector is assumed to be zero and the cross
product calculated accordingly. In cases where both input vectors have dimension 2, the z-component of the
cross product is returned.
Parameters
a : array_like
Components of the first vector(s).
b : array_like
Components of the second vector(s).
axisa : int, optional
Axis of a that defines the vector(s). By default, the last axis.
axisb : int, optional
Axis of b that defines the vector(s). By default, the last axis.
axisc : int, optional
Axis of c containing the cross product vector(s). By default, the last axis.
axis : int, optional

3.13. Mathematical functions 739

NumPy Reference, Release 2.0.0.dev8464

If defined, the axis of a, b and c that defines the vector(s) and cross product(s). Overrides
axisa, axisb and axisc.
Returns
c : ndarray
Vector cross product(s).
Raises
ValueError :
When the dimension of the vector(s) in a and/or b does not equal 2 or 3.
See Also:

inner
Inner product
outer
Outer product.
ix_
Construct index arrays.

Examples
Vector cross-product.

>>> x = [1, 2, 3]
>>> y = [4, 5, 6]
>>> np.cross(x, y)
array([-3, 6, -3])

One vector with dimension 2.

>>> x = [1, 2]
>>> y = [4, 5, 6]
>>> np.cross(x, y)
array([12, -6, -3])

Equivalently:

>>> x = [1, 2, 0]
>>> y = [4, 5, 6]
>>> np.cross(x, y)
array([12, -6, -3])

Both vectors with dimension 2.

>>> x = [1,2]
>>> y = [4,5]
>>> np.cross(x, y)
-3

Multiple vector cross-products. Note that the direction of the cross product vector is defined by the right-hand
rule.

740 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> x = np.array([[1,2,3], [4,5,6]])

>>> y = np.array([[4,5,6], [1,2,3]])
>>> np.cross(x, y)
array([[-3, 6, -3],
[ 3, -6, 3]])

The orientation of c can be changed using the axisc keyword.

>>> np.cross(x, y, axisc=0)

array([[-3, 3],
[ 6, -6],
[-3, 3]])

Change the vector definition of x and y using axisa and axisb.

>>> x = np.array([[1,2,3], [4,5,6], [7, 8, 9]])

>>> y = np.array([[7, 8, 9], [4,5,6], [1,2,3]])
>>> np.cross(x, y)
array([[ -6, 12, -6],
[ 0, 0, 0],
[ 6, -12, 6]])
>>> np.cross(x, y, axisa=0, axisb=0)
array([[-24, 48, -24],
[-30, 60, -30],
[-36, 72, -36]])

trapz(y, x=None, dx=1.0, axis=-1)

Integrate along the given axis using the composite trapezoidal rule.
Integrate y (x) along given axis.
Parameters
y : array_like
Input array to integrate.
x : array_like, optional
If x is None, then spacing between all y elements is dx.
dx : scalar, optional
If x is None, spacing given by dx is assumed. Default is 1.
axis : int, optional
Specify the axis.
Returns
out : float
Definite integral as approximated by trapezoidal rule.
See Also:
sum, cumsum

Notes
Image [R275] illustrates trapezoidal rule y-axis locations of points will be taken from y array, by default x-axis
distances between points will be 1.0, alternatively they can be provided with x array or with dx scalar. Return
value will be equal to combined area under the red lines.

3.13. Mathematical functions 741

NumPy Reference, Release 2.0.0.dev8464

References
[R274], [R275]

Examples

>>> np.trapz([1,2,3])
4.0
>>> np.trapz([1,2,3], x=[4,6,8])
8.0
>>> np.trapz([1,2,3], dx=2)
8.0
>>> a = np.arange(6).reshape(2, 3)
>>> a
array([[0, 1, 2],
[3, 4, 5]])
>>> np.trapz(a, axis=0)
array([ 1.5, 2.5, 3.5])
>>> np.trapz(a, axis=1)
array([ 2., 8.])

3.13.5 Exponents and logarithms

exp() Calculate the exponential of all elements in the input array.

expm1() Calculate exp(x) - 1 for all elements in the array.
exp2() Calculate 2**p for all p in the input array.
log() Natural logarithm, element-wise.
log10() Return the base 10 logarithm of the input array, element-wise.
log2() Base-2 logarithm of x.
log1p() Return the natural logarithm of one plus the input array, element-wise.
logaddexp(x1) Logarithm of the sum of exponentiations of the inputs.
logaddexp2(x1) Logarithm of the sum of exponentiations of the inputs in base-2.
exp
Calculate the exponential of all elements in the input array.
Parameters
x : array_like
Input values.
Returns
out : ndarray
Output array, element-wise exponential of x.
See Also:

expm1
Calculate exp(x) - 1 for all elements in the array.
exp2
Calculate 2**x for all elements in the array.

Notes
The irrational number e is also known as Eulers number. It is approximately 2.718281, and is the base of the
natural logarithm, ln (this means that, if x = ln y = loge y, then ex = y. For real input, exp(x) is always

742 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

positive.
For complex arguments, x = a + ib, we can write ex = ea eib . The first term, ea , is already known (it is
the real argument, described above). The second term, eib , is cos b + i sin b, a function with magnitude 1 and a
periodic phase.

References
[R28], [R29]

Examples
Plot the magnitude and phase of exp(x) in the complex plane:

>>> import matplotlib.pyplot as plt

>>> x = np.linspace(-2*np.pi, 2*np.pi, 100)

>>> xx = x + 1j * x[:, np.newaxis] # a + ib over complex plane
>>> out = np.exp(xx)

>>> plt.subplot(121)
>>> plt.imshow(np.abs(out),
... extent=[-2*np.pi, 2*np.pi, -2*np.pi, 2*np.pi])
>>> plt.title(Magnitude of exp(x))

>>> plt.subplot(122)
>>> plt.imshow(np.angle(out),
... extent=[-2*np.pi, 2*np.pi, -2*np.pi, 2*np.pi])
>>> plt.title(Phase (angle) of exp(x))
>>> plt.show()

Magnitude of exp(x) Phase (angle) of exp(x)

6 6
4 4
2 2
0 0
2 2
4 4
6 6
6 4 2 0 2 4 6 6 4 2 0 2 4 6

expm1
Calculate exp(x) - 1 for all elements in the array.
Parameters
x : array_like
Input values.

3.13. Mathematical functions 743

NumPy Reference, Release 2.0.0.dev8464

Returns
out : ndarray
Element-wise exponential minus one: out = exp(x) - 1.
See Also:

log1p
log(1 + x), the inverse of expm1.

Notes
This function provides greater precision than the formula exp(x) - 1 for small values of x.

Examples
The true value of exp(1e-10) - 1 is 1.00000000005e-10 to about 32 significant digits. This example
shows the superiority of expm1 in this case.

>>> np.expm1(1e-10)
1.00000000005e-10
>>> np.exp(1e-10) - 1
1.000000082740371e-10

exp2
Calculate 2**p for all p in the input array.
Parameters
x : array_like
Input values.
out : ndarray, optional
Array to insert results into.
Returns
out : ndarray
Element-wise 2 to the power x.
See Also:

exp
calculate x**p.

Notes
New in version 1.3.0.

Examples

>>> np.exp2([2, 3])

array([ 4., 8.])

log
Natural logarithm, element-wise.
The natural logarithm log is the inverse of the exponential function, so that log(exp(x)) = x. The natural logarithm
is logarithm in base e.

744 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
x : array_like
Input value.
Returns
y : ndarray
The natural logarithm of x, element-wise.
See Also:
log10, log2, log1p, emath.log

Notes
Logarithm is a multivalued function: for each x there is an infinite number of z such that exp(z) = x. The
convention is to return the z whose imaginary part lies in [-pi, pi].
For real-valued input data types, log always returns real output. For each value that cannot be expressed as a
real number or infinity, it yields nan and sets the invalid floating point error flag.
For complex-valued input, log is a complex analytical function that has a branch cut [-inf, 0] and is continuous
from above on it. log handles the floating-point negative zero as an infinitesimal negative number, conforming
to the C99 standard.

References
[R51], [R52]

Examples

>>> np.log([1, np.e, np.e**2, 0])

array([ 0., 1., 2., -Inf])

log10
Return the base 10 logarithm of the input array, element-wise.
Parameters
x : array_like
Input values.
Returns
y : ndarray
The logarithm to the base 10 of x, element-wise. NaNs are returned where x is negative.
See Also:
emath.log10

Notes
Logarithm is a multivalued function: for each x there is an infinite number of z such that 10**z = x. The
convention is to return the z whose imaginary part lies in [-pi, pi].
For real-valued input data types, log10 always returns real output. For each value that cannot be expressed as a
real number or infinity, it yields nan and sets the invalid floating point error flag.
For complex-valued input, log10 is a complex analytical function that has a branch cut [-inf, 0] and is continuous
from above on it. log10 handles the floating-point negative zero as an infinitesimal negative number, conforming
to the C99 standard.

3.13. Mathematical functions 745

NumPy Reference, Release 2.0.0.dev8464

References
[R53], [R54]

Examples

>>> np.log10([1e-15, -3.])

array([-15., NaN])

log2
Base-2 logarithm of x.
Parameters
x : array_like
Input values.
Returns
y : ndarray
Base-2 logarithm of x.
See Also:
log, log10, log1p, emath.log2

Notes
New in version 1.3.0. Logarithm is a multivalued function: for each x there is an infinite number of z such that
2**z = x. The convention is to return the z whose imaginary part lies in [-pi, pi].
For real-valued input data types, log2 always returns real output. For each value that cannot be expressed as a
real number or infinity, it yields nan and sets the invalid floating point error flag.
For complex-valued input, log2 is a complex analytical function that has a branch cut [-inf, 0] and is continuous
from above on it. log2 handles the floating-point negative zero as an infinitesimal negative number, conforming
to the C99 standard.

Examples

>>> x = np.array([0, 1, 2, 2**4])

>>> np.log2(x)
array([-Inf, 0., 1., 4.])

>>> xi = np.array([0+1.j, 1, 2+0.j, 4.j])

>>> np.log2(xi)
array([ 0.+2.26618007j, 0.+0.j , 1.+0.j , 2.+2.26618007j])

log1p
Return the natural logarithm of one plus the input array, element-wise.
Calculates log(1 + x).
Parameters
x : array_like
Input values.
Returns
y : ndarray
Natural logarithm of 1 + x, element-wise.

746 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:

expm1
exp(x) - 1, the inverse of log1p.

Notes
For real-valued input, log1p is accurate also for x so small that 1 + x == 1 in floating-point accuracy.
Logarithm is a multivalued function: for each x there is an infinite number of z such that exp(z) = 1 + x. The
convention is to return the z whose imaginary part lies in [-pi, pi].
For real-valued input data types, log1p always returns real output. For each value that cannot be expressed as a
real number or infinity, it yields nan and sets the invalid floating point error flag.
For complex-valued input, log1p is a complex analytical function that has a branch cut [-inf, -1] and is con-
tinuous from above on it. log1p handles the floating-point negative zero as an infinitesimal negative number,
conforming to the C99 standard.

References
[R55], [R56]

Examples

>>> np.log1p(1e-99)
1e-99
>>> np.log(1 + 1e-99)
0.0

logaddexp
Logarithm of the sum of exponentiations of the inputs.
Calculates log(exp(x1) + exp(x2)). This function is useful in statistics where the calculated proba-
bilities of events may be so small as to exceed the range of normal floating point numbers. In such cases the
logarithm of the calculated probability is stored. This function allows adding probabilities stored in such a
fashion.
Parameters
x1, x2 : array_like
Input values.
Returns
result : ndarray
Logarithm of exp(x1) + exp(x2).
See Also:

logaddexp2
Logarithm of the sum of exponentiations of inputs in base-2.

Notes
New in version 1.3.0.

Examples

3.13. Mathematical functions 747

NumPy Reference, Release 2.0.0.dev8464

>>> prob1 = np.log(1e-50)

>>> prob2 = np.log(2.5e-50)
>>> prob12 = np.logaddexp(prob1, prob2)
>>> prob12
-113.87649168120691
>>> np.exp(prob12)
3.5000000000000057e-50

logaddexp2
Logarithm of the sum of exponentiations of the inputs in base-2.
Calculates log2(2**x1 + 2**x2). This function is useful in machine learning when the calculated prob-
abilities of events may be so small as to exceed the range of normal floating point numbers. In such cases the
base-2 logarithm of the calculated probability can be used instead. This function allows adding probabilities
stored in such a fashion.
Parameters
x1, x2 : array_like
Input values.
out : ndarray, optional
Array to store results in.
Returns
result : ndarray
Base-2 logarithm of 2**x1 + 2**x2.
See Also:

logaddexp
Logarithm of the sum of exponentiations of the inputs.

Notes
New in version 1.3.0.

Examples

>>> prob1 = np.log2(1e-50)

>>> prob2 = np.log2(2.5e-50)
>>> prob12 = np.logaddexp2(prob1, prob2)
>>> prob1, prob2, prob12
(-166.09640474436813, -164.77447664948076, -164.28904982231052)
>>> 2**prob12
3.4999999999999914e-50

3.13.6 Other special functions

i0(x) Modified Bessel function of the first kind, order 0.

sinc(x) Return the sinc function.
i0(x)
Modified Bessel function of the first kind, order 0.
Usually denoted I0 . This function does broadcast, but will not up-cast int dtype arguments unless accompa-
nied by at least one float or complex dtype argument (see Raises below).

748 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
x : array_like, dtype float or complex
Argument of the Bessel function.
Returns
out : ndarray, shape = x.shape, dtype = x.dtype
The modified Bessel function evaluated at each of the elements of x.
Raises
TypeError: array cannot be safely cast to required type :
If argument consists exclusively of int dtypes.
See Also:
scipy.special.iv, scipy.special.ive

Notes
We use the algorithm published by Clenshaw [R38] and referenced by Abramowitz and Stegun [R39], for which
the function domain is partitioned into the two intervals [0,8] and (8,inf), and Chebyshev polynomial expansions
are employed in each interval. Relative error on the domain [0,30] using IEEE arithmetic is documented [R40]
as having a peak of 5.8e-16 with an rms of 1.4e-16 (n = 30000).

References
[R38], [R39], [R40]

Examples

>>> np.i0([0.])
array(1.0)
>>> np.i0([0., 1. + 2j])
array([ 1.00000000+0.j , 0.18785373+0.64616944j])

sinc(x)
Return the sinc function.
The sinc function is sin(x)/(x).
Parameters
x : ndarray
Array (possibly multi-dimensional) of values for which to to calculate sinc(x).
Returns
out : ndarray
sinc(x), which has the same shape as the input.

Notes
sinc(0) is the limit value 1.
The name sinc is short for sine cardinal or sinus cardinalis.
The sinc function is used in various signal processing applications, including in anti-aliasing, in the construction
of a Lanczos resampling filter, and in interpolation.
For bandlimited interpolation of discrete-time signals, the ideal interpolation kernel is proportional to the sinc
function.

3.13. Mathematical functions 749

NumPy Reference, Release 2.0.0.dev8464

References
[R270], [R271]

Examples

>>> x = np.arange(-20., 21.)/5.

>>> np.sinc(x)
array([ -3.89804309e-17, -4.92362781e-02, -8.40918587e-02,
-8.90384387e-02, -5.84680802e-02, 3.89804309e-17,
6.68206631e-02, 1.16434881e-01, 1.26137788e-01,
8.50444803e-02, -3.89804309e-17, -1.03943254e-01,
-1.89206682e-01, -2.16236208e-01, -1.55914881e-01,
3.89804309e-17, 2.33872321e-01, 5.04551152e-01,
7.56826729e-01, 9.35489284e-01, 1.00000000e+00,
9.35489284e-01, 7.56826729e-01, 5.04551152e-01,
2.33872321e-01, 3.89804309e-17, -1.55914881e-01,
-2.16236208e-01, -1.89206682e-01, -1.03943254e-01,
-3.89804309e-17, 8.50444803e-02, 1.26137788e-01,
1.16434881e-01, 6.68206631e-02, 3.89804309e-17,
-5.84680802e-02, -8.90384387e-02, -8.40918587e-02,
-4.92362781e-02, -3.89804309e-17])

>>> import matplotlib.pyplot as plt

>>> plt.plot(x, np.sinc(x))
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.title("Sinc Function")
<matplotlib.text.Text object at 0x...>
>>> plt.ylabel("Amplitude")
<matplotlib.text.Text object at 0x...>
>>> plt.xlabel("X")
<matplotlib.text.Text object at 0x...>
>>> plt.show()

It works in 2-D as well:

>>> x = np.arange(-200., 201.)/50.

>>> xx = np.outer(x, x)
>>> plt.imshow(np.sinc(xx))
<matplotlib.image.AxesImage object at 0x...>

750 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Sinc Function
0

100

Amplitude
200

300

400
0 100 200 300 400
X

3.13.7 Floating point routines

signbit() Returns element-wise True where signbit is set (less than zero).
copysign(x1) Change the sign of x1 to that of x2, element-wise.
frexp(out1) Split the number, x, into a normalized fraction (y1) and exponent (y2)
ldexp(x1) Compute y = x1 * 2**x2.
signbit
Returns element-wise True where signbit is set (less than zero).
Parameters
x: array_like :
The input value(s).
out : ndarray, optional
Array into which the output is placed. Its type is preserved and it must be of the right
shape to hold the output. See doc.ufuncs.
Returns
out : array_like, bool
Output array.

Examples

>>> np.signbit(-1.2)
True
>>> np.signbit(np.array([1, -2.3, 2.1]))
array([False, True, False], dtype=bool)

copysign
Change the sign of x1 to that of x2, element-wise.
If both arguments are arrays or sequences, they have to be of the same length. If x2 is a scalar, its sign will be
copied to all elements of x1.

3.13. Mathematical functions 751

NumPy Reference, Release 2.0.0.dev8464

Parameters
x1: array_like :
Values to change the sign of.
x2: array_like :
The sign of x2 is copied to x1.
out : ndarray, optional
Array into which the output is placed. Its type is preserved and it must be of the right
shape to hold the output. See doc.ufuncs.
Returns
out : array_like
The values of x1 with the sign of x2.

Examples

>>> np.copysign(1.3, -1)

-1.3
>>> 1/np.copysign(0, 1)
inf
>>> 1/np.copysign(0, -1)
-inf

>>> np.copysign([-1, 0, 1], -1.1)

array([-1., -0., -1.])
>>> np.copysign([-1, 0, 1], np.arange(3)-1)
array([-1., 0., 1.])

frexp
Split the number, x, into a normalized fraction (y1) and exponent (y2)
ldexp
Compute y = x1 * 2**x2.

3.13.8 Arithmetic operations

add(x1) Add arguments element-wise.

reciprocal() Return the reciprocal of the argument, element-wise.
negative() Returns an array with the negative of each element of the original array.
multiply(x1) Multiply arguments element-wise.
divide(x1) Divide arguments element-wise.
power(x1) Returns element-wise base array raised to power from second array.
subtract(x1) Subtract arguments, element-wise.
true_divide(x1) Returns a true division of the inputs, element-wise.
floor_divide(x1) Return the largest integer smaller or equal to the division of the inputs.
fmod(x1) Return the element-wise remainder of division.
mod(x1) Return element-wise remainder of division.
modf(out1) Return the fractional and integral parts of an array, element-wise.
remainder(x1) Return element-wise remainder of division.
add
Add arguments element-wise.

752 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
x1, x2 : array_like
The arrays to be added.
Returns
y : {ndarray, scalar}
The sum of x1 and x2, element-wise. Returns scalar if both x1 and x2 are scalars.

Notes
Equivalent to x1 + x2 in terms of array broadcasting.

Examples

>>> np.add(1.0, 4.0)

5.0
>>> x1 = np.arange(9.0).reshape((3, 3))
>>> x2 = np.arange(3.0)
>>> np.add(x1, x2)
array([[ 0., 2., 4.],
[ 3., 5., 7.],
[ 6., 8., 10.]])

reciprocal
Return the reciprocal of the argument, element-wise.
Calculates 1/x.
Parameters
x : array_like
Input array.
Returns
y : ndarray
Return array.

Notes
Note: This function is not designed to work with integers.
For integer arguments with absolute value larger than 1 the result is always zero because of the way Python
handles integer division. For integer zero the result is an overflow.

Examples

>>> np.reciprocal(2.)
0.5
>>> np.reciprocal([1, 2., 3.33])
array([ 1. , 0.5 , 0.3003003])

negative
Returns an array with the negative of each element of the original array.
Parameters
x : {array_like, scalar}
Input array.

3.13. Mathematical functions 753

NumPy Reference, Release 2.0.0.dev8464

Returns
y : {ndarray, scalar}
Returned array or scalar y=-x.

Examples

>>> np.negative([1.,-1.])
array([-1., 1.])

multiply
Multiply arguments element-wise.
Parameters
x1, x2 : array_like
Input arrays to be multiplied.
Returns
y : ndarray
The product of x1 and x2, element-wise. Returns a scalar if both x1 and x2 are scalars.

Notes
Equivalent to x1 * x2 in terms of array broadcasting.

Examples

>>> np.multiply(2.0, 4.0)

8.0

>>> x1 = np.arange(9.0).reshape((3, 3))

>>> x2 = np.arange(3.0)
>>> np.multiply(x1, x2)
array([[ 0., 1., 4.],
[ 0., 4., 10.],
[ 0., 7., 16.]])

divide
Divide arguments element-wise.
Parameters
x1 : array_like
Dividend array.
x2 : array_like
Divisor array.
out : ndarray, optional
Array into which the output is placed. Its type is preserved and it must be of the right
shape to hold the output. See doc.ufuncs.
Returns
y : {ndarray, scalar}
The quotient x1/x2, element-wise. Returns a scalar if both x1 and x2 are scalars.
See Also:

754 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

seterr
Set whether to raise or warn on overflow, underflow and division by zero.

Notes
Equivalent to x1 / x2 in terms of array-broadcasting.
Behavior on division by zero can be changed using seterr.
When both x1 and x2 are of an integer type, divide will return integers and throw away the fractional part.
Moreover, division by zero always yields zero in integer arithmetic.

Examples

>>> np.divide(2.0, 4.0)

0.5
>>> x1 = np.arange(9.0).reshape((3, 3))
>>> x2 = np.arange(3.0)
>>> np.divide(x1, x2)
array([[ NaN, 1. , 1. ],
[ Inf, 4. , 2.5],
[ Inf, 7. , 4. ]])

Note the behavior with integer types:

>>> np.divide(2, 4)
0
>>> np.divide(2, 4.)
0.5

Division by zero always yields zero in integer arithmetic, and does not raise an exception or a warning:

>>> np.divide(np.array([0, 1], dtype=int), np.array([0, 0], dtype=int))

array([0, 0])

Division by zero can, however, be caught using seterr:

>>> old_err_state = np.seterr(divide=raise)

>>> np.divide(1, 0)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
FloatingPointError: divide by zero encountered in divide

>>> ignored_states = np.seterr(**old_err_state)

>>> np.divide(1, 0)
0

power
Returns element-wise base array raised to power from second array.
Raise each base in x1 to the power of the exponents in x2. This requires that x1 and x2 must be broadcastable to
the same shape.
Parameters
x1 : array_like
The bases.
x2 : array_like

3.13. Mathematical functions 755

NumPy Reference, Release 2.0.0.dev8464

The exponents.
Returns
y : ndarray
The bases in x1 raised to the exponents in x2.

Examples
Cube each element in a list.

>>> x1 = range(6)
>>> x1
[0, 1, 2, 3, 4, 5]
>>> np.power(x1, 3)
array([ 0, 1, 8, 27, 64, 125])

Raise the bases to different exponents.

>>> x2 = [1.0, 2.0, 3.0, 3.0, 2.0, 1.0]

>>> np.power(x1, x2)
array([ 0., 1., 8., 27., 16., 5.])

The effect of broadcasting.

>>> x2 = np.array([[1, 2, 3, 3, 2, 1], [1, 2, 3, 3, 2, 1]])

>>> x2
array([[1, 2, 3, 3, 2, 1],
[1, 2, 3, 3, 2, 1]])
>>> np.power(x1, x2)
array([[ 0, 1, 8, 27, 16, 5],
[ 0, 1, 8, 27, 16, 5]])

subtract
Subtract arguments, element-wise.
Parameters
x1, x2 : array_like
The arrays to be subtracted from each other.
Returns
y : ndarray
The difference of x1 and x2, element-wise. Returns a scalar if both x1 and x2 are scalars.

Notes
Equivalent to x1 - x2 in terms of array broadcasting.

Examples

>>> np.subtract(1.0, 4.0)

-3.0

>>> x1 = np.arange(9.0).reshape((3, 3))

>>> x2 = np.arange(3.0)
>>> np.subtract(x1, x2)
array([[ 0., 0., 0.],

756 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

[ 3., 3., 3.],

[ 6., 6., 6.]])

true_divide
Returns a true division of the inputs, element-wise.
Instead of the Python traditional floor division, this returns a true division. True division adjusts the output
type to present the best answer, regardless of input types.
Parameters
x1 : array_like
Dividend array.
x2 : array_like
Divisor array.
Returns
out : ndarray
Result is scalar if both inputs are scalar, ndarray otherwise.

Notes
The floor division operator // was added in Python 2.2 making // and / equivalent operators. The default floor
division operation of / can be replaced by true division with from __future__ import division.
In Python 3.0, // is the floor division operator and / the true division operator. The true_divide(x1,
x2) function is equivalent to true division in Python.

Examples

>>> x = np.arange(5)
>>> np.true_divide(x, 4)
array([ 0. , 0.25, 0.5 , 0.75, 1. ])

>>> x/4
array([0, 0, 0, 0, 1])
>>> x//4
array([0, 0, 0, 0, 1])

>>> from __future__ import division

>>> x/4
array([ 0. , 0.25, 0.5 , 0.75, 1. ])
>>> x//4
array([0, 0, 0, 0, 1])

floor_divide
Return the largest integer smaller or equal to the division of the inputs.
Parameters
x1 : array_like
Numerator.
x2 : array_like
Denominator.

3.13. Mathematical functions 757

NumPy Reference, Release 2.0.0.dev8464

Returns
y : ndarray
y = floor(x1/x2)
See Also:

divide
Standard division.
floor
Round a number to the nearest integer toward minus infinity.
ceil
Round a number to the nearest integer toward infinity.

Examples

>>> np.floor_divide(7,3)
2
>>> np.floor_divide([1., 2., 3., 4.], 2.5)
array([ 0., 0., 1., 1.])

fmod
Return the element-wise remainder of division.
This is the NumPy implementation of the Python modulo operator %.
Parameters
x1 : array_like
Dividend.
x2 : array_like
Divisor.
Returns
y : array_like
The remainder of the division of x1 by x2.
See Also:

remainder
Modulo operation where the quotient is floor(x1/x2).

divide

Notes
The result of the modulo operation for negative dividend and divisors is bound by conventions. In fmod, the sign
of the remainder is the sign of the dividend. In remainder, the sign of the divisor does not affect the sign of the
result.

Examples

>>> np.fmod([-3, -2, -1, 1, 2, 3], 2)

array([-1, 0, -1, 1, 0, 1])
>>> np.remainder([-3, -2, -1, 1, 2, 3], 2)
array([1, 0, 1, 1, 0, 1])

758 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.fmod([5, 3], [2, 2.])

array([ 1., 1.])
>>> a = np.arange(-3, 3).reshape(3, 2)
>>> a
array([[-3, -2],
[-1, 0],
[ 1, 2]])
>>> np.fmod(a, [2,2])
array([[-1, 0],
[-1, 0],
[ 1, 0]])

mod
Return element-wise remainder of division.
Computes x1 - floor(x1 / x2) * x2.
Parameters
x1 : array_like
Dividend array.
x2 : array_like
Divisor array.
out : ndarray, optional
Array into which the output is placed. Its type is preserved and it must be of the right
shape to hold the output. See doc.ufuncs.
Returns
y : ndarray
The remainder of the quotient x1/x2, element-wise. Returns a scalar if both x1 and x2
are scalars.
See Also:
divide, floor

Notes
Returns 0 when x2 is 0 and both x1 and x2 are (arrays of) integers.

Examples

>>> np.remainder([4, 7], [2, 3])

array([0, 1])
>>> np.remainder(np.arange(7), 5)
array([0, 1, 2, 3, 4, 0, 1])

modf
Return the fractional and integral parts of an array, element-wise.
The fractional and integral parts are negative if the given number is negative.
Parameters
x : array_like
Input array.

3.13. Mathematical functions 759

NumPy Reference, Release 2.0.0.dev8464

Returns
y1 : ndarray
Fractional part of x.
y2 : ndarray
Integral part of x.

Notes
For integer input the return values are floats.

Examples

>>> np.modf([0, 3.5])

(array([ 0. , 0.5]), array([ 0., 3.]))
>>> np.modf(-0.5)
(-0.5, -0)

remainder
Return element-wise remainder of division.
Computes x1 - floor(x1 / x2) * x2.
Parameters
x1 : array_like
Dividend array.
x2 : array_like
Divisor array.
out : ndarray, optional
Array into which the output is placed. Its type is preserved and it must be of the right
shape to hold the output. See doc.ufuncs.
Returns
y : ndarray
The remainder of the quotient x1/x2, element-wise. Returns a scalar if both x1 and x2
are scalars.
See Also:
divide, floor

Notes
Returns 0 when x2 is 0 and both x1 and x2 are (arrays of) integers.

Examples

>>> np.remainder([4, 7], [2, 3])

array([0, 1])
>>> np.remainder(np.arange(7), 5)
array([0, 1, 2, 3, 4, 0, 1])

760 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

3.13.9 Handling complex numbers

angle(z[, deg]) Return the angle of the complex argument.

real(val) Return the real part of the elements of the array.
imag(val) Return the imaginary part of the elements of the array.
conj() Return the complex conjugate, element-wise.
angle(z, deg=0)
Return the angle of the complex argument.
Parameters
z : array_like
A complex number or sequence of complex numbers.
deg : bool, optional
Return angle in degrees if True, radians if False (default).
Returns
angle : {ndarray, scalar}
The counterclockwise angle from the positive real axis on the complex plane, with dtype
as numpy.float64.
See Also:
arctan2, absolute

Examples

>>> np.angle([1.0, 1.0j, 1+1j]) # in radians

array([ 0. , 1.57079633, 0.78539816])
>>> np.angle(1+1j, deg=True) # in degrees
45.0

real(val)
Return the real part of the elements of the array.
Parameters
val : array_like
Input array.
Returns
out : ndarray
Output array. If val is real, the type of val is used for the output. If val has complex
elements, the returned type is float.
See Also:
real_if_close, imag, angle

Examples

>>> a = np.array([1+2j, 3+4j, 5+6j])

>>> a.real
array([ 1., 3., 5.])
>>> a.real = 9
>>> a
array([ 9.+2.j, 9.+4.j, 9.+6.j])

3.13. Mathematical functions 761

NumPy Reference, Release 2.0.0.dev8464

>>> a.real = np.array([9, 8, 7])

>>> a
array([ 9.+2.j, 8.+4.j, 7.+6.j])

imag(val)
Return the imaginary part of the elements of the array.
Parameters
val : array_like
Input array.
Returns
out : ndarray
Output array. If val is real, the type of val is used for the output. If val has complex
elements, the returned type is float.
See Also:
real, angle, real_if_close

Examples

>>> a = np.array([1+2j, 3+4j, 5+6j])

>>> a.imag
array([ 2., 4., 6.])
>>> a.imag = np.array([8, 10, 12])
>>> a
array([ 1. +8.j, 3.+10.j, 5.+12.j])

conj
Return the complex conjugate, element-wise.
The complex conjugate of a complex number is obtained by changing the sign of its imaginary part.
Parameters
x : array_like
Input value.
Returns
y : ndarray
The complex conjugate of x, with same dtype as y.

Examples

>>> np.conjugate(1+2j)
(1-2j)

>>> x = np.eye(2) + 1j * np.eye(2)

>>> np.conjugate(x)
array([[ 1.-1.j, 0.-0.j],
[ 0.-0.j, 1.-1.j]])

762 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

3.13.10 Miscellaneous

convolve(a, v[, mode]) Returns the discrete, linear convolution of two one-dimensional sequences.
clip(a, a_min, a_max[, out]) Clip (limit) the values in an array.
sqrt() Return the positive square-root of an array, element-wise.
square() Return the element-wise square of the input.
absolute() Calculate the absolute value element-wise.
fabs() Compute the absolute values elementwise.
sign() Returns an element-wise indication of the sign of a number.
maximum(x1) Element-wise maximum of array elements.
minimum(x1) Element-wise minimum of array elements.
nan_to_num(x) Replace nan with zero and inf with finite numbers.
real_if_close(a[, tol]) If complex input returns a real array if complex parts are close to zero.
interp(x, xp, fp[, left, right]) One-dimensional linear interpolation.
convolve(a, v, mode=full)
Returns the discrete, linear convolution of two one-dimensional sequences.
The convolution operator is often seen in signal processing, where it models the effect of a linear time-invariant
system on a signal [R27]. In probability theory, the sum of two independent random variables is distributed
according to the convolution of their individual distributions.
Parameters
a : (N,) array_like
First one-dimensional input array.
v : (M,) array_like
Second one-dimensional input array.
mode : {full, valid, same}, optional
full:
By default, mode is full. This returns the convolution at each point of overlap, with
an output shape of (N+M-1,). At the end-points of the convolution, the signals do not
overlap completely, and boundary effects may be seen.
same:
Mode same returns output of length max(M, N). Boundary effects are still visible.
valid:
Mode valid returns output of length max(M, N) - min(M, N) + 1. The con-
volution product is only given for points where the signals overlap completely. Values
outside the signal boundary have no effect.
Returns
out : ndarray
Discrete, linear convolution of a and v.
See Also:

scipy.signal.fftconvolve
Convolve two arrays using the Fast Fourier Transform.
scipy.linalg.toeplitz
Used to construct the convolution operator.

3.13. Mathematical functions 763

NumPy Reference, Release 2.0.0.dev8464

Notes
The discrete convolution operation is defined as

X
(f g)[n] = f [m]g[n m]
m=

It can be shown that a convolution x(t) y(t) in time/space is equivalent to the multiplication X(f )Y (f ) in
the Fourier domain, after appropriate padding (padding is necessary to prevent circular convolution). Since
multiplication is more efficient (faster) than convolution, the function scipy.signal.fftconvolve exploits the FFT
to calculate the convolution of large data-sets.

References
[R27]

Examples
Note how the convolution operator flips the second array before sliding the two across one another:

>>> np.convolve([1, 2, 3], [0, 1, 0.5])

array([ 0. , 1. , 2.5, 4. , 1.5])

Only return the middle values of the convolution. Contains boundary effects, where zeros are taken into account:

>>> np.convolve([1,2,3],[0,1,0.5], same)

array([ 1. , 2.5, 4. ])

The two arrays are of the same length, so there is only one position where they completely overlap:

>>> np.convolve([1,2,3],[0,1,0.5], valid)

array([ 2.5])

clip(a, a_min, a_max, out=None)

Clip (limit) the values in an array.
Given an interval, values outside the interval are clipped to the interval edges. For example, if an interval of
[0, 1] is specified, values smaller than 0 become 0, and values larger than 1 become 1.
Parameters
a : array_like
Array containing elements to clip.
a_min : scalar or array_like
Minimum value.
a_max : scalar or array_like
Maximum value. If a_min or a_max are array_like, then they will be broadcasted to the
shape of a.
out : ndarray, optional
The results will be placed in this array. It may be the input array for in-place clipping.
out must be of the right shape to hold the output. Its type is preserved.
Returns
clipped_array : ndarray

764 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

An array with the elements of a, but where values < a_min are replaced with a_min, and
those > a_max with a_max.
See Also:

numpy.doc.ufuncs
Section Output arguments

Examples

>>> a = np.arange(10)
>>> np.clip(a, 1, 8)
array([1, 1, 2, 3, 4, 5, 6, 7, 8, 8])
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> np.clip(a, 3, 6, out=a)
array([3, 3, 3, 3, 4, 5, 6, 6, 6, 6])
>>> a = np.arange(10)
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> np.clip(a, [3,4,1,1,1,4,4,4,4,4], 8)
array([3, 4, 2, 3, 4, 5, 6, 7, 8, 8])

sqrt
Return the positive square-root of an array, element-wise.
Parameters
x : array_like
The square root of each element in this array is calculated.
Returns
y : ndarray
An array of the same shape as x, containing the square-root of each element in x. If any
element in x is complex, a complex array is returned. If all of the elements of x are real,
negative elements return numpy.nan elements.
See Also:

numpy.lib.scimath.sqrt
A version which returns complex numbers when given negative reals.

Notes
sqrt has a branch cut [-inf, 0) and is continuous from above on it.

Examples

>>> np.sqrt([1,4,9])
array([ 1., 2., 3.])

>>> np.sqrt([4, -1, -3+4J])

array([ 2.+0.j, 0.+1.j, 1.+2.j])

>>> np.sqrt([4, -1, numpy.inf])

array([ 2., NaN, Inf])

3.13. Mathematical functions 765

NumPy Reference, Release 2.0.0.dev8464

square
Return the element-wise square of the input.
Parameters
x : array_like
Input data.
Returns
out : ndarray
Element-wise x*x, of the same shape and dtype as x. Returns scalar if x is a scalar.
See Also:
numpy.linalg.matrix_power, sqrt, power

Examples

>>> np.square([-1j, 1])

array([-1.-0.j, 1.+0.j])

absolute
Calculate the absolute value element-wise.
Parameters
x : array_like
Input array.
Returns
res : ndarray
An ndarray containing theabsolute value of each element in x. For complex input, a +
ib, the absolute value is a2 + b2 .

Examples

>>> x = np.array([-1.2, 1.2])

>>> np.absolute(x)
array([ 1.2, 1.2])
>>> np.absolute(1.2 + 1j)
1.5620499351813308

Plot the function over [-10, 10]:

>>> import matplotlib.pyplot as plt

>>> x = np.linspace(-10, 10, 101)

>>> plt.plot(x, np.absolute(x))
>>> plt.show()

Plot the function over the complex plane:

>>> xx = x + 1j * x[:, np.newaxis]

>>> plt.imshow(np.abs(xx), extent=[-10, 10, -10, 10])
>>> plt.show()

766 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

10

10
10 5 0 5 10

fabs
Compute the absolute values elementwise.
This function returns the absolute values (positive magnitude) of the data in x. Complex values are not handled,
use absolute to find the absolute values of complex data.
Parameters
x : array_like
The array of numbers for which the absolute values are required. If x is a scalar, the
result y will also be a scalar.
out : ndarray, optional
Array into which the output is placed. Its type is preserved and it must be of the right
shape to hold the output. See doc.ufuncs.
Returns
y : {ndarray, scalar}
The absolute values of x, the returned values are always floats.
See Also:

absolute
Absolute values including complex types.

Examples

>>> np.fabs(-1)
1.0
>>> np.fabs([-1.2, 1.2])
array([ 1.2, 1.2])

sign
Returns an element-wise indication of the sign of a number.
The sign function returns -1 if x < 0, 0 if x==0, 1 if x > 0.

3.13. Mathematical functions 767

NumPy Reference, Release 2.0.0.dev8464

Parameters
x : array_like
Input values.
Returns
y : ndarray
The sign of x.

Examples

>>> np.sign([-5., 4.5])

array([-1., 1.])
>>> np.sign(0)
0

maximum
Element-wise maximum of array elements.
Compare two arrays and returns a new array containing the element-wise maxima. If one of the elements being
compared is a nan, then that element is returned. If both elements are nans then the first is returned. The latter
distinction is important for complex nans, which are defined as at least one of the real or imaginary parts being
a nan. The net effect is that nans are propagated.
Parameters
x1, x2 : array_like
The arrays holding the elements to be compared. They must have the same shape, or
shapes that can be broadcast to a single shape.
Returns
y : {ndarray, scalar}
The maximum of x1 and x2, element-wise. Returns scalar if both x1 and x2 are scalars.
See Also:

minimum
element-wise minimum
fmax
element-wise maximum that ignores nans unless both inputs are nans.
fmin
element-wise minimum that ignores nans unless both inputs are nans.

Notes
Equivalent to np.where(x1 > x2, x1, x2) but faster and does proper broadcasting.

Examples

>>> np.maximum([2, 3, 4], [1, 5, 2])

array([2, 5, 4])

>>> np.maximum(np.eye(2), [0.5, 2])

array([[ 1. , 2. ],
[ 0.5, 2. ]])

768 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.maximum([np.nan, 0, np.nan], [0, np.nan, np.nan])

array([ NaN, NaN, NaN])
>>> np.maximum(np.Inf, 1)
inf

minimum
Element-wise minimum of array elements.
Compare two arrays and returns a new array containing the element-wise minima. If one of the elements being
compared is a nan, then that element is returned. If both elements are nans then the first is returned. The latter
distinction is important for complex nans, which are defined as at least one of the real or imaginary parts being
a nan. The net effect is that nans are propagated.
Parameters
x1, x2 : array_like
The arrays holding the elements to be compared. They must have the same shape, or
shapes that can be broadcast to a single shape.
Returns
y : {ndarray, scalar}
The minimum of x1 and x2, element-wise. Returns scalar if both x1 and x2 are scalars.
See Also:

maximum
element-wise minimum that propagates nans.
fmax
element-wise maximum that ignores nans unless both inputs are nans.
fmin
element-wise minimum that ignores nans unless both inputs are nans.

Notes
The minimum is equivalent to np.where(x1 <= x2, x1, x2) when neither x1 nor x2 are nans, but it is
faster and does proper broadcasting.

Examples

>>> np.minimum([2, 3, 4], [1, 5, 2])

array([1, 3, 2])

>>> np.minimum(np.eye(2), [0.5, 2]) # broadcasting

array([[ 0.5, 0. ],
[ 0. , 1. ]])

>>> np.minimum([np.nan, 0, np.nan],[0, np.nan, np.nan])

array([ NaN, NaN, NaN])

nan_to_num(x)
Replace nan with zero and inf with finite numbers.
Returns an array or scalar replacing Not a Number (NaN) with zero, (positive) infinity with a very large number
and negative infinity with a very small (or negative) number.

3.13. Mathematical functions 769

NumPy Reference, Release 2.0.0.dev8464

Parameters
x : array_like
Input data.
Returns
out : ndarray, float
Array with the same shape as x and dtype of the element in x with the greatest precision.
NaN is replaced by zero, and infinity (-infinity) is replaced by the largest (smallest or
most negative) floating point value that fits in the output dtype. All finite numbers are
upcast to the output dtype (default float64).
See Also:

isinf
Shows which elements are negative or negative infinity.
isneginf
Shows which elements are negative infinity.
isposinf
Shows which elements are positive infinity.
isnan
Shows which elements are Not a Number (NaN).
isfinite
Shows which elements are finite (not NaN, not infinity)

Notes
Numpy uses the IEEE Standard for Binary Floating-Point for Arithmetic (IEEE 754). This means that Not a
Number is not equivalent to infinity.

Examples

>>> np.set_printoptions(precision=8)
>>> x = np.array([np.inf, -np.inf, np.nan, -128, 128])
>>> np.nan_to_num(x)
array([ 1.79769313e+308, -1.79769313e+308, 0.00000000e+000,
-1.28000000e+002, 1.28000000e+002])

real_if_close(a, tol=100)
If complex input returns a real array if complex parts are close to zero.
Close to zero is defined as tol * (machine epsilon of the type for a).
Parameters
a : array_like
Input array.
tol : float
Tolerance in machine epsilons for the complex part of the elements in the array.
Returns
out : ndarray
If a is real, the type of a is used for the output. If a has complex elements, the returned
type is float.

770 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:
real, imag, angle

Notes
Machine epsilon varies from machine to machine and between data types but Python floats on most platforms
have a machine epsilon equal to 2.2204460492503131e-16. You can use np.finfo(np.float).eps to print out the
machine epsilon for floats.

Examples

>>> np.finfo(np.float).eps
2.2204460492503131e-16

>>> np.real_if_close([2.1 + 4e-14j], tol=1000)

array([ 2.1])
>>> np.real_if_close([2.1 + 4e-13j], tol=1000)
array([ 2.1 +4.00000000e-13j])

interp(x, xp, fp, left=None, right=None)

One-dimensional linear interpolation.
Returns the one-dimensional piecewise linear interpolant to a function with given values at discrete data-points.
Parameters
x : array_like
The x-coordinates of the interpolated values.
xp : 1-D sequence of floats
The x-coordinates of the data points, must be increasing.
fp : 1-D sequence of floats
The y-coordinates of the data points, same length as xp.
left : float, optional
Value to return for x , default is fp[0].
right : float, optional
Value to return for x > xp[-1], defaults is fp[-1].
Returns
y : {float, ndarray}
The interpolated values, same shape as x.
Raises
ValueError :
If xp and fp have different length

Notes
Does not check that the x-coordinate sequence xp is increasing. If xp is not increasing, the results are nonsense.
A simple check for increasingness is:

np.all(np.diff(xp) > 0)

3.13. Mathematical functions 771

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> xp = [1, 2, 3]
>>> fp = [3, 2, 0]
>>> np.interp(2.5, xp, fp)
1.0
>>> np.interp([0, 1, 1.5, 2.72, 3.14], xp, fp)
array([ 3. , 3. , 2.5 , 0.56, 0. ])
>>> UNDEF = -99.0
>>> np.interp(3.14, xp, fp, right=UNDEF)
-99.0

Plot an interpolant to the sine function:

>>> x = np.linspace(0, 2*np.pi, 10)

>>> y = np.sin(x)
>>> xvals = np.linspace(0, 2*np.pi, 50)
>>> yinterp = np.interp(xvals, x, y)
>>> import matplotlib.pyplot as plt
>>> plt.plot(x, y, o)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.plot(xvals, yinterp, -x)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.show()

1.0

0.5

0.0

0.5

1.0
0 1 2 3 4 5 6 7

3.14 Functional programming

apply_along_axis(func1d, axis, arr, *args) Apply a function to 1-D slices along the given axis.
apply_over_axes(func, a, axes) Apply a function repeatedly over multiple axes.
vectorize(pyfunc[, otypes, doc]) Generalized function class.
frompyfunc(func, nin, nout) Takes an arbitrary Python function and returns a Numpy ufunc.
piecewise(x, condlist, funclist, *args, **kw) Evaluate a piecewise-defined function.
apply_along_axis(func1d, axis, arr, *args)
Apply a function to 1-D slices along the given axis.

772 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Execute func1d(a, *args) where func1d operates on 1-D arrays and a is a 1-D slice of arr along axis.
Parameters
func1d : function
This function should accept 1-D arrays. It is applied to 1-D slices of arr along the
specified axis.
axis : integer
Axis along which arr is sliced.
arr : ndarray
Input array.
args : any
Additional arguments to func1d.
Returns
outarr : ndarray
The output array. The shape of outarr is identical to the shape of arr, except along the
axis dimension, where the length of outarr is equal to the size of the return value of
func1d. If func1d returns a scalar outarr will have one fewer dimensions than arr.
See Also:

apply_over_axes
Apply a function repeatedly over multiple axes.

Examples

>>> def my_func(a):

... """Average first and last element of a 1-D array"""
... return (a[0] + a[-1]) * 0.5
>>> b = np.array([[1,2,3], [4,5,6], [7,8,9]])
>>> np.apply_along_axis(my_func, 0, b)
array([ 4., 5., 6.])
>>> np.apply_along_axis(my_func, 1, b)
array([ 2., 5., 8.])

For a function that doesnt return a scalar, the number of dimensions in outarr is the same as arr.

>>> def new_func(a):

... """Divide elements of a by 2."""
... return a * 0.5
>>> b = np.array([[1,2,3], [4,5,6], [7,8,9]])
>>> np.apply_along_axis(new_func, 0, b)
array([[ 0.5, 1. , 1.5],
[ 2. , 2.5, 3. ],
[ 3.5, 4. , 4.5]])

apply_over_axes(func, a, axes)
Apply a function repeatedly over multiple axes.
func is called as res = func(a, axis), where axis is the first element of axes. The result res of the function call must
have either the same dimensions as a or one less dimension. If res has one less dimension than a, a dimension
is inserted before axis. The call to func is then repeated for each axis in axes, with res as the first argument.

3.14. Functional programming 773

NumPy Reference, Release 2.0.0.dev8464

Parameters
func : function
This function must take two arguments, func(a, axis).
a : ndarray
Input array.
axes : array_like
Axes over which func is applied, the elements must be integers.
Returns
val : ndarray
The output array. The number of dimensions is the same as a, but the shape can be
different. This depends on whether func changes the shape of its output with respect to
its input.
See Also:

apply_along_axis
Apply a function to 1-D slices of an array along the given axis.

Examples

>>> a = np.arange(24).reshape(2,3,4)
>>> a
array([[[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]],
[[12, 13, 14, 15],
[16, 17, 18, 19],
[20, 21, 22, 23]]])

Sum over axes 0 and 2. The result has same number of dimensions as the original array:

>>> np.apply_over_axes(np.sum, a, [0,2])

array([[[ 60],
[ 92],
[124]]])

class vectorize(pyfunc, otypes=, doc=None)

Generalized function class.
Define a vectorized function which takes a nested sequence of objects or numpy arrays as inputs and returns a
numpy array as output. The vectorized function evaluates pyfunc over successive tuples of the input arrays like
the python map function, except it uses the broadcasting rules of numpy.
The data type of the output of vectorized is determined by calling the function with the first element of the input.
This can be avoided by specifying the otypes argument.
Parameters
pyfunc : callable
A python function or method.
otypes : str or list of dtypes, optional
The output data type. It must be specified as either a string of typecode characters or a
list of data type specifiers. There should be one data type specifier for each output.

774 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

doc : str, optional

The docstring for the function. If None, the docstring will be the pyfunc one.

Examples

>>> def myfunc(a, b):

... """Return a-b if a>b, otherwise return a+b"""
... if a > b:
... return a - b
... else:
... return a + b

>>> vfunc = np.vectorize(myfunc)

>>> vfunc([1, 2, 3, 4], 2)
array([3, 4, 1, 2])

The docstring is taken from the input function to vectorize unless it is specified

>>> vfunc.__doc__
Return a-b if a>b, otherwise return a+b
>>> vfunc = np.vectorize(myfunc, doc=Vectorized myfunc)
>>> vfunc.__doc__
Vectorized myfunc

The output type is determined by evaluating the first element of the input, unless it is specified

>>> out = vfunc([1, 2, 3, 4], 2)

>>> type(out[0])
<type numpy.int32>
>>> vfunc = np.vectorize(myfunc, otypes=[np.float])
>>> out = vfunc([1, 2, 3, 4], 2)
>>> type(out[0])
<type numpy.float64>

frompyfunc(func, nin, nout)

Takes an arbitrary Python function and returns a Numpy ufunc.
Can be used, for example, to add broadcasting to a built-in Python function (see Examples section).
Parameters
func : Python function object
An arbitrary Python function.
nin : int
The number of input arguments.
nout : int
The number of objects returned by func.
Returns
out : ufunc
Returns a Numpy universal function (ufunc) object.

Notes
The returned ufunc always returns PyObject arrays.

3.14. Functional programming 775

NumPy Reference, Release 2.0.0.dev8464

Examples
Use frompyfunc to add broadcasting to the Python function oct:

>>> oct_array = np.frompyfunc(oct, 1, 1)

>>> oct_array(np.array((10, 30, 100)))
array([012, 036, 0144], dtype=object)
>>> np.array((oct(10), oct(30), oct(100))) # for comparison
array([012, 036, 0144],
dtype=|S4)

piecewise(x, condlist, funclist, *args, **kw)

Evaluate a piecewise-defined function.
Given a set of conditions and corresponding functions, evaluate each function on the input data wherever its
condition is true.
Parameters
x : ndarray
The input domain.
condlist : list of bool arrays
Each boolean array corresponds to a function in funclist. Wherever condlist[i] is True,
funclist[i](x) is used as the output value.
Each boolean array in condlist selects a piece of x, and should therefore be of the same
shape as x.
The length of condlist must correspond to that of funclist. If one extra function is given,
i.e. if len(funclist) - len(condlist) == 1, then that extra function is the
default value, used wherever all conditions are false.
funclist : list of callables, f(x,*args,**kw), or scalars
Each function is evaluated over x wherever its corresponding condition is True. It should
take an array as input and give an array or a scalar value as output. If, instead of
a callable, a scalar is provided then a constant function (lambda x: scalar) is
assumed.
args : tuple, optional
Any further arguments given to piecewise are passed to the functions upon execution,
i.e., if called piecewise(..., ..., 1, a), then each function is called as
f(x, 1, a).
kw : dict, optional
Keyword arguments used in calling piecewise are passed to the functions upon exe-
cution, i.e., if called piecewise(..., ..., lambda=1), then each function is
called as f(x, lambda=1).
Returns
out : ndarray
The output is the same shape and type as x and is found by calling the functions in
funclist on the appropriate portions of x, as defined by the boolean arrays in condlist.
Portions not covered by any condition have undefined values.
See Also:
choose, select, where

776 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notes
This is similar to choose or select, except that functions are evaluated on elements of x that satisfy the corre-
sponding condition from condlist.
The result is:

|--
|funclist[0](x[condlist[0]])
out = |funclist[1](x[condlist[1]])
|...
|funclist[n2](x[condlist[n2]])
|--

Examples
Define the sigma function, which is -1 for x < 0 and +1 for x >= 0.

>>> x = np.arange(6) - 2.5

>>> np.piecewise(x, [x < 0, x >= 0], [-1, 1])
array([-1., -1., -1., 1., 1., 1.])

Define the absolute value, which is -x for x <0 and x for x >= 0.

>>> np.piecewise(x, [x < 0, x >= 0], [lambda x: -x, lambda x: x])

array([ 2.5, 1.5, 0.5, 0.5, 1.5, 2.5])

3.15 Polynomials

3.15.1 Basics

poly1d(c_or_r[, r, variable]) A one-dimensional polynomial class.

polyval(p, x) Evaluate a polynomial at specific values.
poly(seq_of_zeros) Find the coefficients of a polynomial with the given sequence of roots.
roots(p) Return the roots of a polynomial with coefficients given in p.
class poly1d(c_or_r, r=0, variable=None)
A one-dimensional polynomial class.
A convenience class, used to encapsulate natural operations on polynomials so that said operations may take
on their customary form in code (see Examples).
Parameters
c_or_r : array_like
The polynomials coefficients, in decreasing powers, or if the value of the second param-
eter is True, the polynomials roots (values where the polynomial evaluates to 0). For
example, poly1d([1, 2, 3]) returns an object that represents x2 +2x+3, whereas
poly1d([1, 2, 3], True) returns one that represents (x 1)(x 2)(x 3) =
x3 6x2 + 11x 6.
r : bool, optional
If True, c_or_r specifies the polynomials roots; the default is False.
variable : str, optional

3.15. Polynomials 777

NumPy Reference, Release 2.0.0.dev8464

Changes the variable used when printing p from x to variable (see Examples).

Examples
Construct the polynomial x2 + 2x + 3:

>>> p = np.poly1d([1, 2, 3])

>>> print np.poly1d(p)
2
1 x + 2 x + 3

Evaluate the polynomial at x = 0.5:

>>> p(0.5)
4.25

Find the roots:

>>> p.r
array([-1.+1.41421356j, -1.-1.41421356j])
>>> p(p.r)
array([ -4.44089210e-16+0.j, -4.44089210e-16+0.j])

These numbers in the previous line represent (0, 0) to machine precision

Show the coefficients:

>>> p.c
array([1, 2, 3])

Display the order (the leading zero-coefficients are removed):

>>> p.order
2

Show the coefficient of the k-th power in the polynomial (which is equivalent to p.c[-(i+1)]):

>>> p[1]
2

Polynomials can be added, subtracted, multiplied, and divided (returns quotient and remainder):

>>> p * p
poly1d([ 1, 4, 10, 12, 9])

>>> (p**3 + 4) / p
(poly1d([ 1., 4., 10., 12., 9.]), poly1d([ 4.]))

asarray(p) gives the coefficient array, so polynomials can be used in all functions that accept arrays:

>>> p**2 # square of polynomial

poly1d([ 1, 4, 10, 12, 9])

>>> np.square(p) # square of individual coefficients

array([1, 4, 9])

The variable used in the string representation of p can be modified, using the variable parameter:

778 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> p = np.poly1d([1,2,3], variable=z)

>>> print p
2
1 z + 2 z + 3

Construct a polynomial from its roots:

>>> np.poly1d([1, 2], True)

poly1d([ 1, -3, 2])

This is the same polynomial as obtained by:

>>> np.poly1d([1, -1]) * np.poly1d([1, -2])

poly1d([ 1, -3, 2])

Attributes
coeffs
order
variable

Methods
deriv([m]) Return a derivative of this polynomial.
integ([m, k]) Return an antiderivative (indefinite integral) of this polynomial.
deriv(m=1)
Return a derivative of this polynomial.
Refer to polyder for full documentation.
See Also:

polyder
equivalent function

integ(m=1, k=0)
Return an antiderivative (indefinite integral) of this polynomial.
Refer to polyint for full documentation.
See Also:

polyint
equivalent function

polyval(p, x)
Evaluate a polynomial at specific values.
If p is of length N, this function returns the value:
p[0]*x**(N-1) + p[1]*x**(N-2) + ... + p[N-2]*x + p[N-1]
If x is a sequence, then p(x) is returned for each element of x. If x is another polynomial then the composite
polynomial p(x(t)) is returned.
Parameters
p : array_like or poly1d object

3.15. Polynomials 779

NumPy Reference, Release 2.0.0.dev8464

1D array of polynomial coefficients (including coefficients equal to zero) from highest

degree to the constant term, or an instance of poly1d.
x : array_like or poly1d object
A number, a 1D array of numbers, or an instance of poly1d, at which to evaluate p.
Returns
values : ndarray or poly1d
If x is a poly1d instance, the result is the composition of the two polynomials, i.e., x
is substituted in p and the simplified result is returned. In addition, the type of x -
array_like or poly1d - governs the type of the output: x array_like => values array_like,
x a poly1d object => values is also.
See Also:

poly1d
A polynomial class.

Notes
Horners scheme [R68] is used to evaluate the polynomial. Even so, for polynomials of high degree the values
may be inaccurate due to rounding errors. Use carefully.

References
[R68]

Examples

>>> np.polyval([3,0,1], 5) # 3 * 5**2 + 0 * 5**1 + 1

76
>>> np.polyval([3,0,1], np.poly1d(5))
poly1d([ 76.])
>>> np.polyval(np.poly1d([3,0,1]), 5)
76
>>> np.polyval(np.poly1d([3,0,1]), np.poly1d(5))
poly1d([ 76.])

poly(seq_of_zeros)
Find the coefficients of a polynomial with the given sequence of roots.
Returns the coefficients of the polynomial whose leading coefficient is one for the given sequence of zeros
(multiple roots must be included in the sequence as many times as their multiplicity; see Examples). A square
matrix (or array, which will be treated as a matrix) can also be given, in which case the coefficients of the
characteristic polynomial of the matrix are returned.
Parameters
seq_of_zeros : array_like, shape (N,) or (N, N)
A sequence of polynomial roots, or a square array or matrix object.
Returns
c : ndarray
1D array of polynomial coefficients from highest to lowest degree:
c[0] * x**(N) + c[1] * x**(N-1) + ... + c[N-1] * x + c[N]
where c[0] always equals 1.

780 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Raises
ValueError :
If input is the wrong shape (the input must be a 1-D or square 2-D array).
See Also:

polyval
Evaluate a polynomial at a point.
roots
Return the roots of a polynomial.
polyfit
Least squares polynomial fit.
poly1d
A one-dimensional polynomial class.

Notes
Specifying the roots of a polynomial still leaves one degree of freedom, typically represented by an undetermined
leading coefficient. [R64] In the case of this function, that coefficient - the first one in the returned array - is
always taken as one. (If for some reason you have one other point, the only automatic way presently to leverage
that information is to use polyfit.)
The characteristic polynomial, pa (t), of an n-by-n matrix A is given by
pa (t) = det(t I A),
where I is the n-by-n identity matrix. [R65]

References
[R64], [R65]

Examples
Given a sequence of a polynomials zeros:

>>> np.poly((0, 0, 0)) # Multiple root example

array([1, 0, 0, 0])

The line above represents z**3 + 0*z**2 + 0*z + 0.

>>> np.poly((-1./2, 0, 1./2))

array([ 1. , 0. , -0.25, 0. ])

The line above represents z**3 - z/4

>>> np.poly((np.random.random(1.)[0], 0, np.random.random(1.)[0]))

array([ 1. , -0.77086955, 0.08618131, 0. ]) #random

Given a square array object:

>>> P = np.array([[0, 1./3], [-1./2, 0]])

>>> np.poly(P)
array([ 1. , 0. , 0.16666667])

Or a square matrix object:

3.15. Polynomials 781

NumPy Reference, Release 2.0.0.dev8464

>>> np.poly(np.matrix(P))
array([ 1. , 0. , 0.16666667])

Note how in all cases the leading coefficient is always 1.

roots(p)
Return the roots of a polynomial with coefficients given in p.
The values in the rank-1 array p are coefficients of a polynomial. If the length of p is n+1 then the polynomial
is described by p[0] * x**n + p[1] * x**(n-1) + ... + p[n-1]*x + p[n]
Parameters
p : array_like of shape(M,)
Rank-1 array of polynomial co-efficients.
Returns
out : ndarray
An array containing the complex roots of the polynomial.
Raises
ValueError: :
When p cannot be converted to a rank-1 array.
See Also:

poly
Find the coefficients of a polynomial with a given sequence of roots.
polyval
Evaluate a polynomial at a point.
polyfit
Least squares polynomial fit.
poly1d
A one-dimensional polynomial class.

Notes
The algorithm relies on computing the eigenvalues of the companion matrix [R267].

References
[R267]

Examples

>>> coeff = [3.2, 2, 1]

>>> np.roots(coeff)
array([-0.3125+0.46351241j, -0.3125-0.46351241j])

3.15.2 Fitting

polyfit(x, y, deg[, rcond, full]) Least squares polynomial fit.

782 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

polyfit(x, y, deg, rcond=None, full=False)

Least squares polynomial fit.
Fit a polynomial p(x) = p[0] * x**deg + ... + p[deg] of degree deg to points (x, y). Returns a
vector of coefficients p that minimises the squared error.
Parameters
x : array_like, shape (M,)
x-coordinates of the M sample points (x[i], y[i]).
y : array_like, shape (M,) or (M, K)
y-coordinates of the sample points. Several data sets of sample points sharing the same
x-coordinates can be fitted at once by passing in a 2D-array that contains one dataset
per column.
deg : int
Degree of the fitting polynomial
rcond : float, optional
Relative condition number of the fit. Singular values smaller than this relative to the
largest singular value will be ignored. The default value is len(x)*eps, where eps is the
relative precision of the float type, about 2e-16 in most cases.
full : bool, optional
Switch determining nature of return value. When it is False (the default) just the coef-
ficients are returned, when True diagnostic information from the singular value decom-
position is also returned.
Returns
p : ndarray, shape (M,) or (M, K)
Polynomial coefficients, highest power first. If y was 2-D, the coefficients for k-th data
set are in p[:,k].
residuals, rank, singular_values, rcond : present only if full = True
Residuals of the least-squares fit, the effective rank of the scaled Vandermonde coeffi-
cient matrix, its singular values, and the specified value of rcond. For more details, see
linalg.lstsq.
See Also:

polyval
Computes polynomial values.
linalg.lstsq
Computes a least-squares fit.
scipy.interpolate.UnivariateSpline
Computes spline fits.

Notes
The solution minimizes the squared error
k
X
E= |p(xj ) yj |2
j=0

3.15. Polynomials 783

NumPy Reference, Release 2.0.0.dev8464

in the equations:

x[0]**n * p[n] + ... + x[0] * p[1] + p[0] = y[0]

x[1]**n * p[n] + ... + x[1] * p[1] + p[0] = y[1]
...
x[k]**n * p[n] + ... + x[k] * p[1] + p[0] = y[k]

The coefficient matrix of the coefficients p is a Vandermonde matrix.

polyfit issues a RankWarning when the least-squares fit is badly conditioned. This implies that the best fit is
not well-defined due to numerical error. The results may be improved by lowering the polynomial degree or
by replacing x by x - x.mean(). The rcond parameter can also be set to a value smaller than its default, but the
resulting fit may be spurious: including contributions from the small singular values can add numerical noise to
the result.
Note that fitting polynomial coefficients is inherently badly conditioned when the degree of the polynomial is
large or the interval of sample points is badly centered. The quality of the fit should always be checked in these
cases. When polynomial fits are not satisfactory, splines may be a good alternative.

References
[R66], [R67]

Examples

>>> x = np.array([0.0, 1.0, 2.0, 3.0, 4.0, 5.0])

>>> y = np.array([0.0, 0.8, 0.9, 0.1, -0.8, -1.0])
>>> z = np.polyfit(x, y, 3)
>>> z
array([ 0.08703704, -0.81349206, 1.69312169, -0.03968254])

It is convenient to use poly1d objects for dealing with polynomials:

>>> p = np.poly1d(z)
>>> p(0.5)
0.6143849206349179
>>> p(3.5)
-0.34732142857143039
>>> p(10)
22.579365079365115

High-order polynomials may oscillate wildly:

>>> p30 = np.poly1d(np.polyfit(x, y, 30))

/... RankWarning: Polyfit may be poorly conditioned...
>>> p30(4)
-0.80000000000000204
>>> p30(5)
-0.99999999999999445
>>> p30(4.5)
-0.10547061179440398

Illustration:

>>> import matplotlib.pyplot as plt

>>> xp = np.linspace(-2, 6, 100)
>>> plt.plot(x, y, ., xp, p(xp), -, xp, p30(xp), --)
[<matplotlib.lines.Line2D object at 0x...>, <matplotlib.lines.Line2D object at 0x...>, <matplotl

784 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> plt.ylim(-2,2)
(-2, 2)
>>> plt.show()

2.0
1.5
1.0
0.5
0.0
0.5
1.0
1.5
2.0
2 1 0 1 2 3 4 5 6

3.15.3 Calculus

polyder(p[, m]) Return the derivative of the specified order of a polynomial.

polyint(p[, m, k]) Return an antiderivative (indefinite integral) of a polynomial.
polyder(p, m=1)
Return the derivative of the specified order of a polynomial.
Parameters
p : poly1d or sequence
Polynomial to differentiate. A sequence is interpreted as polynomial coefficients, see
poly1d.
m : int, optional
Order of differentiation (default: 1)
Returns
der : poly1d
A new polynomial representing the derivative.
See Also:

polyint
Anti-derivative of a polynomial.
poly1d
Class for one-dimensional polynomials.

Examples
The derivative of the polynomial x3 + x2 + x1 + 1 is:

3.15. Polynomials 785

NumPy Reference, Release 2.0.0.dev8464

>>> p = np.poly1d([1,1,1,1])
>>> p2 = np.polyder(p)
>>> p2
poly1d([3, 2, 1])

which evaluates to:

>>> p2(2.)
17.0

We can verify this, approximating the derivative with (f(x + h) - f(x))/h:

>>> (p(2. + 0.001) - p(2.)) / 0.001

17.007000999997857

The fourth-order derivative of a 3rd-order polynomial is zero:

>>> np.polyder(p, 2)
poly1d([6, 2])
>>> np.polyder(p, 3)
poly1d([6])
>>> np.polyder(p, 4)
poly1d([ 0.])

polyint(p, m=1, k=None)

Return an antiderivative (indefinite integral) of a polynomial.
m
d
The returned order m antiderivative P of polynomial p satisfies dx m P (x) = p(x) and is defined up to m - 1

integration constants k. The constants determine the low-order polynomial part

km1 0 k0
x + ... + xm1
0! (m 1)!

of P so that P (j) (0) = kmj1 .

Parameters
p : {array_like, poly1d}
Polynomial to differentiate. A sequence is interpreted as polynomial coefficients, see
poly1d.
m : int, optional
Order of the antiderivative. (Default: 1)
k : {None, list of m scalars, scalar}, optional
Integration constants. They are given in the order of integration: those corresponding
to highest-order terms come first.
If None (default), all constants are assumed to be zero. If m = 1, a single scalar can be
given instead of a list.
See Also:

polyder
derivative of a polynomial

786 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

poly1d.integ
equivalent method

Examples
The defining property of the antiderivative:

>>> p = np.poly1d([1,1,1])
>>> P = np.polyint(p)
>>> P
poly1d([ 0.33333333, 0.5 , 1. , 0. ])
>>> np.polyder(P) == p
True

The integration constants default to zero, but can be specified:

>>> P = np.polyint(p, 3)
>>> P(0)
0.0
>>> np.polyder(P)(0)
0.0
>>> np.polyder(P, 2)(0)
0.0
>>> P = np.polyint(p, 3, k=[6,5,3])
>>> P
poly1d([ 0.01666667, 0.04166667, 0.16666667, 3. , 5. , 3. ])

Note that 3 = 6 / 2!, and that the constants are given in the order of integrations. Constant of the highest-order
polynomial term comes first:

>>> np.polyder(P, 2)(0)

6.0
>>> np.polyder(P, 1)(0)
5.0
>>> P(0)
3.0

3.15.4 Arithmetic

polyadd(a1, a2) Find the sum of two polynomials.

polydiv(u, v) Returns the quotient and remainder of polynomial division.
polymul(a1, a2) Find the product of two polynomials.
polysub(a1, a2) Difference (subtraction) of two polynomials.
polyadd(a1, a2)
Find the sum of two polynomials.
Returns the polynomial resulting from the sum of two input polynomials. Each input must be either a poly1d
object or a 1D sequence of polynomial coefficients, from highest to lowest degree.
Parameters
a1, a2 : array_like or poly1d object
Input polynomials.
Returns
out : ndarray or poly1d object

3.15. Polynomials 787

NumPy Reference, Release 2.0.0.dev8464

The sum of the inputs. If either input is a poly1d object, then the output is also a poly1d
object. Otherwise, it is a 1D array of polynomial coefficients from highest to lowest
degree.
See Also:

poly1d
A one-dimensional polynomial class.

poly, polyadd, polyder, polydiv, polyfit, polyint, polysub, polyval

Examples

>>> np.polyadd([1, 2], [9, 5, 4])

array([9, 6, 6])

Using poly1d objects:

>>> p1 = np.poly1d([1, 2])

>>> p2 = np.poly1d([9, 5, 4])
>>> print p1
1 x + 2
>>> print p2
2
9 x + 5 x + 4
>>> print np.polyadd(p1, p2)
2
9 x + 6 x + 6

polydiv(u, v)
Returns the quotient and remainder of polynomial division.
The input arrays are the coefficients (including any coefficients equal to zero) of the numerator (dividend) and
denominator (divisor) polynomials, respectively.
Parameters
u : array_like or poly1d
Dividend polynomials coefficients.
v : array_like or poly1d
Divisor polynomials coefficients.
Returns
q : ndarray
Coefficients, including those equal to zero, of the quotient.
r : ndarray
Coefficients, including those equal to zero, of the remainder.
See Also:
poly, polyadd, polyder, polydiv, polyfit, polyint, polymul, polysub, polyval

Notes
Both u and v must be 0-d or 1-d (ndim = 0 or 1), but u.ndim need not equal v.ndim. In other words, all four
possible combinations - u.ndim = v.ndim = 0, u.ndim = v.ndim = 1, u.ndim = 1, v.ndim
= 0, and u.ndim = 0, v.ndim = 1 - work.

788 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

3x2 + 5x + 2
= 1.5x + 1.75, remainder0.25
2x + 1

>>> x = np.array([3.0, 5.0, 2.0])

>>> y = np.array([2.0, 1.0])
>>> np.polydiv(x, y)
(array([ 1.5 , 1.75]), array([ 0.25]))

polymul(a1, a2)
Find the product of two polynomials.
Finds the polynomial resulting from the multiplication of the two input polynomials. Each input must be either
a poly1d object or a 1D sequence of polynomial coefficients, from highest to lowest degree.
Parameters
a1, a2 : array_like or poly1d object
Input polynomials.
Returns
out : ndarray or poly1d object
The polynomial resulting from the multiplication of the inputs. If either inputs is a
poly1d object, then the output is also a poly1d object. Otherwise, it is a 1D array of
polynomial coefficients from highest to lowest degree.
See Also:

poly1d
A one-dimensional polynomial class.

poly, polyadd, polyder, polydiv, polyfit, polyint, polysub, polyval

Examples

>>> np.polymul([1, 2, 3], [9, 5, 1])

array([ 9, 23, 38, 17, 3])

Using poly1d objects:

>>> p1 = np.poly1d([1, 2, 3])

>>> p2 = np.poly1d([9, 5, 1])
>>> print p1
2
1 x + 2 x + 3
>>> print p2
2
9 x + 5 x + 1
>>> print np.polymul(p1, p2)
4 3 2
9 x + 23 x + 38 x + 17 x + 3

polysub(a1, a2)
Difference (subtraction) of two polynomials.

3.15. Polynomials 789

NumPy Reference, Release 2.0.0.dev8464

Given two polynomials a1 and a2, returns a1 - a2. a1 and a2 can be either array_like sequences of the
polynomials coefficients (including coefficients equal to zero), or poly1d objects.
Parameters
a1, a2 : array_like or poly1d
Minuend and subtrahend polynomials, respectively.
Returns
out : ndarray or poly1d
Array or poly1d object of the difference polynomials coefficients.
See Also:
polyval, polydiv, polymul, polyadd

Examples

(2x2 + 10x 2) (3x2 + 10x 4) = (x2 + 2)

>>> np.polysub([2, 10, -2], [3, 10, -4])

array([-1, 0, 2])

3.15.5 Warnings

RankWarning Issued by polyfit when the Vandermonde matrix is rank deficient.

exception RankWarning
Issued by polyfit when the Vandermonde matrix is rank deficient.
For more information, a way to suppress the warning, and an example of RankWarning being issued, see polyfit.

3.16 Financial functions

3.16.1 Simple financial functions

fv(rate, nper, pmt, pv[, when]) Compute the future value.

pv(rate, nper, pmt[, fv, when]) Compute the present value.
npv(rate, values) Returns the NPV (Net Present Value) of a cash flow series.
pmt(rate, nper, pv[, fv, when]) Compute the payment against loan principal plus interest.
ppmt(rate, per, nper, pv[, fv, when]) Not implemented.
ipmt(rate, per, nper, pv[, fv, when]) Not implemented.
irr(values) Return the Internal Rate of Return (IRR).
mirr(values, finance_rate, reinvest_rate) Modified internal rate of return.
nper(rate, pmt, pv[, fv, when]) Compute the number of periodic payments.
rate(nper, pmt, pv, fv[, when, guess, tol, ...]) Compute the rate of interest per period.
fv(rate, nper, pmt, pv, when=end)
Compute the future value.
Given:

790 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

a present value, pv
an interest rate compounded once per period, of which there are
nper total
a (fixed) payment, pmt, paid either
at the beginning (when = {begin, 1}) or the end (when = {end, 0}) of each period
Return:
the value at the end of the nper periods

Parameters
rate : scalar or array_like of shape(M, )
Rate of interest as decimal (not per cent) per period
nper : scalar or array_like of shape(M, )
Number of compounding periods
pmt : scalar or array_like of shape(M, )
Payment
pv : scalar or array_like of shape(M, )
Present value
when : {{begin, 1}, {end, 0}}, {string, int}, optional
When payments are due (begin (1) or end (0)). Defaults to {end, 0}.
Returns
out : ndarray
Future values. If all input is scalar, returns a scalar float. If any input is array_like,
returns future values for each input element. If multiple inputs are array_like, they all
must have the same shape.

Notes
The future value is computed by solving the equation:

fv +
pv*(1+rate)**nper +
pmt*(1 + rate*when)/rate*((1 + rate)**nper - 1) == 0

or, when rate == 0:

fv + pv + pmt * nper == 0

References
[WRW]

Examples
What is the future value after 10 years of saving $100 now, with an additional monthly savings of $100. Assume
the interest rate is 5% (annually) compounded monthly?

3.16. Financial functions 791

NumPy Reference, Release 2.0.0.dev8464

>>> np.fv(0.05/12, 10*12, -100, -100)

15692.928894335748

By convention, the negative sign represents cash flow out (i.e. money not available today). Thus, saving $100 a
month at 5% annual interest leads to $15,692.93 available to spend in 10 years.
If any input is array_like, returns an array of equal shape. Lets compare different interest rates from the example
above.

>>> a = np.array((0.05, 0.06, 0.07))/12

>>> np.fv(a, 10*12, -100, -100)
array([ 15692.92889434, 16569.87435405, 17509.44688102])

pv(rate, nper, pmt, fv=0.0, when=end)

Compute the present value.
Given:

a future value, fv
an interest rate compounded once per period, of which there are
nper total
a (fixed) payment, pmt, paid either
at the beginning (when = {begin, 1}) or the end (when = {end, 0}) of each period
Return:
the value now

Parameters
rate : array_like
Rate of interest (per period)
nper : array_like
Number of compounding periods
pmt : array_like
Payment
fv : array_like, optional
Future value
when : {{begin, 1}, {end, 0}}, {string, int}, optional
When payments are due (begin (1) or end (0))
Returns
out : ndarray, float
Present value of a series of payments or investments.

Notes
The present value is computed by solving the equation:

792 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

fv +
pv*(1 + rate)**nper +
pmt*(1 + rate*when)/rate*((1 + rate)**nper - 1) = 0

or, when rate = 0:

fv + pv + pmt * nper = 0

for pv, which is then returned.

References
[WRW]

Examples
What is the present value (e.g., the initial investment) of an investment that needs to total $15692.93 after 10
years of saving $100 every month? Assume the interest rate is 5% (annually) compounded monthly.

>>> np.pv(0.05/12, 10*12, -100, 15692.93)

-100.00067131625819

By convention, the negative sign represents cash flow out (i.e., money not available today). Thus, to end up with
$15,692.93 in 10 years saving $100 a month at 5% annual interest, ones initial deposit should also be $100.
If any input is array_like, pv returns an array of equal shape. Lets compare different interest rates in the
example above:

>>> a = np.array((0.05, 0.04, 0.03))/12

>>> np.pv(a, 10*12, -100, 15692.93)
array([ -100.00067132, -649.26771385, -1273.78633713])

So, to end up with the same $15692.93 under the same $100 per month savings plan, for annual interest rates
of 4% and 3%, one would need initial investments of $649.27 and $1273.79, respectively.
npv(rate, values)
Returns the NPV (Net Present Value) of a cash flow series.
Parameters
rate : scalar
The discount rate.
values : array_like, shape(M, )
The values of the time series of cash flows. The (fixed) time interval between cash flow
events must be the same as that for which rate is given (i.e., if rate is per year, then
precisely a year is understood to elapse between each cash flow event). By convention,
investments or deposits are negative, income or withdrawals are positive; values
must begin with the initial investment, thus values[0] will typically be negative.
Returns
out : float
The NPV of the input cash flow series values at the discount rate.

3.16. Financial functions 793

NumPy Reference, Release 2.0.0.dev8464

Notes
Returns the result of: [G61]
M
X valuest
t=0
(1 + rate)t

References
[G61]

Examples

>>> np.npv(0.281,[-100, 39, 59, 55, 20])

-0.0066187288356340801

(Compare with the Example given for numpy.lib.financial.irr)

pmt(rate, nper, pv, fv=0, when=end)
Compute the payment against loan principal plus interest.
Given:

a present value, pv (e.g., an amount borrowed)

a future value, fv (e.g., 0)
an interest rate compounded once per period, of which there are
nper total
and (optional) specification of whether payment is made at the beginning (when = {begin, 1}) or the
end (when = {end, 0}) of each period
Return:
the (fixed) periodic payment.

Parameters
rate : array_like
Rate of interest (per period)
nper : array_like
Number of compounding periods
pv : array_like
Present value
fv : array_like (optional)
Future value (default = 0)
when : {{begin, 1}, {end, 0}}, {string, int}
When payments are due (begin (1) or end (0))
Returns
out : ndarray

794 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Payment against loan plus interest. If all input is scalar, returns a scalar float. If any
input is array_like, returns payment for each input element. If multiple inputs are ar-
ray_like, they all must have the same shape.

Notes
The payment is computed by solving the equation:

fv +
pv*(1 + rate)**nper +
pmt*(1 + rate*when)/rate*((1 + rate)**nper - 1) == 0

or, when rate == 0:

fv + pv + pmt * nper == 0

for pmt.
Note that computing a monthly mortgage payment is only one use for this function. For example, pmt returns the
periodic deposit one must make to achieve a specified future balance given an initial deposit, a fixed, periodically
compounded interest rate, and the total number of periods.

References
[WRW]

Examples
What is the monthly payment needed to pay off a $200,000 loan in 15 years at an annual interest rate of 7.5%?

>>> np.pmt(0.075/12, 12*15, 200000)

-1854.0247200054619

In order to pay-off (i.e., have a future-value of 0) the $200,000 obtained today, a monthly payment of $1,854.02
would be required. Note that this example illustrates usage of fv having a default value of 0.
ppmt(rate, per, nper, pv, fv=0.0, when=end)
Not implemented. Compute the payment against loan principal.
Parameters
rate : array_like
Rate of interest (per period)
per : array_like, int
Amount paid against the loan changes. The per is the period of interest.
nper : array_like
Number of compounding periods
pv : array_like
Present value
fv : array_like, optional
Future value
when : {{begin, 1}, {end, 0}}, {string, int}
When payments are due (begin (1) or end (0))

3.16. Financial functions 795

NumPy Reference, Release 2.0.0.dev8464

See Also:
pmt, pv, ipmt
ipmt(rate, per, nper, pv, fv=0.0, when=end)
Not implemented. Compute the payment portion for loan interest.
Parameters
rate : scalar or array_like of shape(M, )
Rate of interest as decimal (not per cent) per period
per : scalar or array_like of shape(M, )
Interest paid against the loan changes during the life or the loan. The per is the payment
period to calculate the interest amount.
nper : scalar or array_like of shape(M, )
Number of compounding periods
pv : scalar or array_like of shape(M, )
Present value
fv : scalar or array_like of shape(M, ), optional
Future value
when : {{begin, 1}, {end, 0}}, {string, int}, optional
When payments are due (begin (1) or end (0)). Defaults to {end, 0}.
Returns
out : ndarray
Interest portion of payment. If all input is scalar, returns a scalar float. If any input
is array_like, returns interest payment for each input element. If multiple inputs are
array_like, they all must have the same shape.
See Also:
ppmt, pmt, pv

Notes
The total payment is made up of payment against principal plus interest.
pmt = ppmt + ipmt
irr(values)
Return the Internal Rate of Return (IRR).
This is the average periodically compounded rate of return that gives a net present value of 0.0; for a more
complete explanation, see Notes below.
Parameters
values : array_like, shape(N,)
Input cash flows per time period. By convention, net deposits are negative and net
withdrawals are positive. Thus, for example, at least the first element of values, which
represents the initial investment, will typically be negative.
Returns
out : float
Internal Rate of Return for periodic input values.

796 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notes
The IRR is perhaps best understood through an example (illustrated using np.irr in the Examples section below).
Suppose one invests 100 units and then makes the following withdrawals at regular (fixed) intervals: 39, 59, 55,
20. Assuming the ending value is 0, ones 100 unit investment yields 173 units; however, due to the combi-
nation of compounding and the periodic withdrawals, the average rate of return is neither simply 0.73/4 nor
(1.73)^0.25-1. Rather, it is the solution (for r) of the equation:
39 59 55 20
100 + + 2
+ 3
+ =0
1 + r (1 + r) (1 + r) (1 + r)4

In general, for values = [v0 , v1 , ...vM ], irr is the solution of the equation: [G41]
M
X vt
=0
t=0
(1 + irr)t

References
[G41]

Examples

>>> np.irr([-100, 39, 59, 55, 20])

0.2809484211599611

(Compare with the Example given for numpy.lib.financial.npv)

mirr(values, finance_rate, reinvest_rate)
Modified internal rate of return.
Parameters
values : array_like
Cash flows (must contain at least one positive and one negative value) or nan is returned.
The first value is considered a sunk cost at time zero.
finance_rate : scalar
Interest rate paid on the cash flows
reinvest_rate : scalar
Interest rate received on the cash flows upon reinvestment
Returns
out : float
Modified internal rate of return
nper(rate, pmt, pv, fv=0, when=end)
Compute the number of periodic payments.
Parameters
rate : array_like
Rate of interest (per period)
pmt : array_like
Payment

3.16. Financial functions 797

NumPy Reference, Release 2.0.0.dev8464

pv : array_like
Present value
fv : array_like, optional
Future value
when : {{begin, 1}, {end, 0}}, {string, int}, optional
When payments are due (begin (1) or end (0))

Notes
The number of periods nper is computed by solving the equation:

fv + pv*(1+rate)**nper + pmt*(1+rate*when)/rate*((1+rate)**nper-1) = 0

but if rate = 0 then:

fv + pv + pmt*nper = 0

Examples
If you only had $150/month to pay towards the loan, how long would it take to pay-off a loan of $8,000 at 7%
annual interest?

>>> np.nper(0.07/12, -150, 8000)

64.073348770661852

So, over 64 months would be required to pay off the loan.

The same analysis could be done with several different interest rates and/or payments and/or total amounts to
produce an entire table.

>>> np.nper(*(np.ogrid[0.07/12: 0.08/12: 0.01/12,

... -150 : -99 : 50 ,
... 8000 : 9001 : 1000]))
array([[[ 64.07334877, 74.06368256],
[ 108.07548412, 127.99022654]],
[[ 66.12443902, 76.87897353],
[ 114.70165583, 137.90124779]]])

rate(nper, pmt, pv, fv, when=end, guess=0.10000000000000001, tol=9.9999999999999995e-07, maxiter=100)

Compute the rate of interest per period.
Parameters
nper : array_like
Number of compounding periods
pmt : array_like
Payment
pv : array_like
Present value
fv : array_like
Future value

798 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

when : {{begin, 1}, {end, 0}}, {string, int}, optional

When payments are due (begin (1) or end (0))
guess : float, optional
Starting guess for solving the rate of interest
tol : float, optional
Required tolerance for the solution
maxiter : int, optional
Maximum iterations in finding the solution

Notes
The rate of interest is computed by iteratively solving the (non-linear) equation:

fv + pv*(1+rate)**nper + pmt*(1+rate*when)/rate * ((1+rate)**nper - 1) = 0

for rate.

References
Wheeler, D. A., E. Rathke, and R. Weir (Eds.) (2009, May). Open Document Format for Office Applica-
tions (OpenDocument)v1.2, Part 2: Recalculated Formula (OpenFormula) Format - Annotated Version, Pre-
Draft 12. Organization for the Advancement of Structured Information Standards (OASIS). Billerica, MA,
USA. [ODT Document]. Available: http://www.oasis-open.org/committees/documents.php?wg_abbrev=office-
formula OpenDocument-formula-20090508.odt

3.17 Set routines

3.17.1 Making proper sets

unique(ar[, return_index, return_inverse]) Find the unique elements of an array.

unique(ar, return_index=False, return_inverse=False)
Find the unique elements of an array.
Returns the sorted unique elements of an array. There are two optional outputs in addition to the unique elements:
the indices of the input array that give the unique values, and the indices of the unique array that reconstruct the
input array.
Parameters
ar : array_like
Input array. This will be flattened if it is not already 1-D.
return_index : bool, optional
If True, also return the indices of ar that result in the unique array.
return_inverse : bool, optional
If True, also return the indices of the unique array that can be used to reconstruct ar.
Returns
unique : ndarray
The sorted unique values.

3.17. Set routines 799

NumPy Reference, Release 2.0.0.dev8464

unique_indices : ndarray, optional

The indices of the unique values in the (flattened) original array. Only provided if
return_index is True.
unique_inverse : ndarray, optional
The indices to reconstruct the (flattened) original array from the unique array. Only
provided if return_inverse is True.
See Also:

numpy.lib.arraysetops
Module with a number of other functions for performing set operations on arrays.

Examples

>>> np.unique([1, 1, 2, 2, 3, 3])

array([1, 2, 3])
>>> a = np.array([[1, 1], [2, 3]])
>>> np.unique(a)
array([1, 2, 3])

Return the indices of the original array that give the unique values:

>>> a = np.array([a, b, b, c, a])

>>> u, indices = np.unique(a, return_index=True)
>>> u
array([a, b, c],
dtype=|S1)
>>> indices
array([0, 1, 3])
>>> a[indices]
array([a, b, c],
dtype=|S1)

Reconstruct the input array from the unique values:

>>> a = np.array([1, 2, 6, 4, 2, 3, 2])

>>> u, indices = np.unique(a, return_inverse=True)
>>> u
array([1, 2, 3, 4, 6])
>>> indices
array([0, 1, 4, 3, 1, 2, 1])
>>> u[indices]
array([1, 2, 6, 4, 2, 3, 2])

3.17.2 Boolean operations

in1d(ar1, ar2[, assume_unique]) Test whether each element of a 1D array is also present in a second
array.
intersect1d(ar1, ar2[, Find the intersection of two arrays.
assume_unique])
setdiff1d(ar1, ar2[, assume_unique]) Find the set difference of two arrays.
setxor1d(ar1, ar2[, assume_unique]) Find the set exclusive-or of two arrays.
union1d(ar1, ar2) Find the union of two arrays.

800 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

in1d(ar1, ar2, assume_unique=False)

Test whether each element of a 1D array is also present in a second array.
Returns a boolean array the same length as ar1 that is True where an element of ar1 is in ar2 and False otherwise.
Parameters
ar1 : array_like, shape (M,)
Input array.
ar2 : array_like
The values against which to test each value of ar1.
assume_unique : bool, optional
If True, the input arrays are both assumed to be unique, which can speed up the calcu-
lation. Default is False.
Returns
mask : ndarray of bools, shape(M,)
The values ar1[mask] are in ar2.
See Also:

numpy.lib.arraysetops
Module with a number of other functions for performing set operations on arrays.

Notes
in1d can be considered as an element-wise function version of the python keyword in, for 1D sequences.
in1d(a, b) is roughly equivalent to np.array([item in b for item in a]). New in version
1.4.0.

Examples

>>> test = np.array([0, 1, 2, 5, 0])

>>> states = [0, 2]
>>> mask = np.in1d(test, states)
>>> mask
array([ True, False, True, False, True], dtype=bool)
>>> test[mask]
array([0, 2, 0])

intersect1d(ar1, ar2, assume_unique=False)

Find the intersection of two arrays.
Return the sorted, unique values that are in both of the input arrays.
Parameters
ar1, ar2 : array_like
Input arrays.
assume_unique : bool
If True, the input arrays are both assumed to be unique, which can speed up the calcu-
lation. Default is False.
Returns
out : ndarray
Sorted 1D array of common and unique elements.

3.17. Set routines 801

NumPy Reference, Release 2.0.0.dev8464

See Also:

numpy.lib.arraysetops
Module with a number of other functions for performing set operations on arrays.

Examples

>>> np.intersect1d([1, 3, 4, 3], [3, 1, 2, 1])

array([1, 3])

setdiff1d(ar1, ar2, assume_unique=False)

Find the set difference of two arrays.
Return the sorted, unique values in ar1 that are not in ar2.
Parameters
ar1 : array_like
Input array.
ar2 : array_like
Input comparison array.
assume_unique : bool
If True, the input arrays are both assumed to be unique, which can speed up the calcu-
lation. Default is False.
Returns
difference : ndarray
Sorted 1D array of values in ar1 that are not in ar2.
See Also:

numpy.lib.arraysetops
Module with a number of other functions for performing set operations on arrays.

Examples

>>> a = np.array([1, 2, 3, 2, 4, 1])

>>> b = np.array([3, 4, 5, 6])
>>> np.setdiff1d(a, b)
array([1, 2])

setxor1d(ar1, ar2, assume_unique=False)

Find the set exclusive-or of two arrays.
Return the sorted, unique values that are in only one (not both) of the input arrays.
Parameters
ar1, ar2 : array_like
Input arrays.
assume_unique : bool
If True, the input arrays are both assumed to be unique, which can speed up the calcu-
lation. Default is False.

802 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Returns
xor : ndarray
Sorted 1D array of unique values that are in only one of the input arrays.

Examples

>>> a = np.array([1, 2, 3, 2, 4])

>>> b = np.array([2, 3, 5, 7, 5])
>>> np.setxor1d(a,b)
array([1, 4, 5, 7])

union1d(ar1, ar2)
Find the union of two arrays.
Return the unique, sorted array of values that are in either of the two input arrays.
Parameters
ar1, ar2 : array_like
Input arrays. They are flattened if they are not already 1D.
Returns
union : ndarray
Unique, sorted union of the input arrays.
See Also:

numpy.lib.arraysetops
Module with a number of other functions for performing set operations on arrays.

Examples

>>> np.union1d([-1, 0, 1], [-2, 0, 2])

array([-2, -1, 0, 1, 2])

3.18 Window functions

3.18.1 Various windows

bartlett(M) Return the Bartlett window.

blackman(M) Return the Blackman window.
hamming(M) Return the Hamming window.
hanning(M) Return the Hanning window.
kaiser(M, beta) Return the Kaiser window.
bartlett(M)
Return the Bartlett window.
The Bartlett window is very similar to a triangular window, except that the end points are at zero. It is often
used in signal processing for tapering a signal, without generating too much ripple in the frequency domain.
Parameters
M : int
Number of points in the output window. If zero or less, an empty array is returned.

3.18. Window functions 803

NumPy Reference, Release 2.0.0.dev8464

Returns
out : array
The triangular window, normalized to one (the value one appears only if the number of
samples is odd), with the first and last samples equal to zero.
See Also:
blackman, hamming, hanning, kaiser

Notes
The Bartlett window is defined as
  
2 M 1  M 1 
w(n) = n
M 1 2 2 

Most references to the Bartlett window come from the signal processing literature, where it is used as one of
many windowing functions for smoothing values. Note that convolution with this window produces linear in-
terpolation. It is also known as an apodization (which meansremoving the foot, i.e. smoothing discontinuities
at the beginning and end of the sampled signal) or tapering function. The fourier transform of the Bartlett is the
product of two sinc functions. Note the excellent discussion in Kanasewich.

References
[R18], [R19], [R20], [R21], [R22]

Examples

>>> np.bartlett(12)
array([ 0. , 0.18181818, 0.36363636, 0.54545455, 0.72727273,
0.90909091, 0.90909091, 0.72727273, 0.54545455, 0.36363636,
0.18181818, 0. ])

Plot the window and its frequency response (requires SciPy and matplotlib):

>>> from numpy import clip, log10, array, bartlett, linspace

>>> from numpy.fft import fft, fftshift
>>> import matplotlib.pyplot as plt

>>> window = bartlett(51)

>>> plt.plot(window)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.title("Bartlett window")
<matplotlib.text.Text object at 0x...>
>>> plt.ylabel("Amplitude")
<matplotlib.text.Text object at 0x...>
>>> plt.xlabel("Sample")
<matplotlib.text.Text object at 0x...>
>>> plt.show()

>>> plt.figure()
<matplotlib.figure.Figure object at 0x...>
>>> A = fft(window, 2048) / 25.5
>>> mag = abs(fftshift(A))
>>> freq = linspace(-0.5,0.5,len(A))
>>> response = 20*log10(mag)

804 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> response = clip(response,-100,100)

>>> plt.plot(freq, response)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.title("Frequency response of Bartlett window")
<matplotlib.text.Text object at 0x...>
>>> plt.ylabel("Magnitude [dB]")
<matplotlib.text.Text object at 0x...>
>>> plt.xlabel("Normalized frequency [cycles per sample]")
<matplotlib.text.Text object at 0x...>
>>> plt.axis(tight)
(-0.5, 0.5, -100.0, ...)
>>> plt.show()

Bartlett window
1.0

0.8

0.6
Amplitude

0.4

0.2

0.0
0 10 20 30 40 50
Sample

Frequency response of Bartlett window

20
Magnitude [dB]

40

60

80

100
0.4 0.2 0.0 0.2 0.4
Normalized frequency [cycles per sample]

blackman(M)
Return the Blackman window.
The Blackman window is a taper formed by using the the first three terms of a summation of cosines. It was
designed to have close to the minimal leakage possible. It is close to optimal, only slightly worse than a Kaiser
window.

3.18. Window functions 805

NumPy Reference, Release 2.0.0.dev8464

Parameters
M : int
Number of points in the output window. If zero or less, an empty array is returned.
Returns
out : array
The window, normalized to one (the value one appears only if the number of samples is
odd).
See Also:
bartlett, hamming, hanning, kaiser

Notes
The Blackman window is defined as

w(n) = 0.42 0.5 cos(2n/M ) + 0.08 cos(4n/M )

Most references to the Blackman window come from the signal processing literature, where it is used as one of
many windowing functions for smoothing values. It is also known as an apodization (which means removing
the foot, i.e. smoothing discontinuities at the beginning and end of the sampled signal) or tapering function. It
is known as a near optimal tapering function, almost as good (by some measures) as the kaiser window.

References
[R24], [R25], [R26]

Examples

>>> from numpy import blackman

>>> blackman(12)
array([ -1.38777878e-17, 3.26064346e-02, 1.59903635e-01,
4.14397981e-01, 7.36045180e-01, 9.67046769e-01,
9.67046769e-01, 7.36045180e-01, 4.14397981e-01,
1.59903635e-01, 3.26064346e-02, -1.38777878e-17])

Plot the window and the frequency response:

>>> from numpy import clip, log10, array, blackman, linspace

>>> from numpy.fft import fft, fftshift
>>> import matplotlib.pyplot as plt

>>> window = blackman(51)

>>> plt.plot(window)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.title("Blackman window")
<matplotlib.text.Text object at 0x...>
>>> plt.ylabel("Amplitude")
<matplotlib.text.Text object at 0x...>
>>> plt.xlabel("Sample")
<matplotlib.text.Text object at 0x...>
>>> plt.show()

806 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> plt.figure()
<matplotlib.figure.Figure object at 0x...>
>>> A = fft(window, 2048) / 25.5
>>> mag = abs(fftshift(A))
>>> freq = linspace(-0.5,0.5,len(A))
>>> response = 20*log10(mag)
>>> response = clip(response,-100,100)
>>> plt.plot(freq, response)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.title("Frequency response of Blackman window")
<matplotlib.text.Text object at 0x...>
>>> plt.ylabel("Magnitude [dB]")
<matplotlib.text.Text object at 0x...>
>>> plt.xlabel("Normalized frequency [cycles per sample]")
<matplotlib.text.Text object at 0x...>
>>> plt.axis(tight)
(-0.5, 0.5, -100.0, ...)
>>> plt.show()

Blackman window
1.0
0.8
0.6
Amplitude

0.4
0.2
0.0
0.2
0 10 20 30 40 50
Sample

Frequency response of Blackman window

20
Magnitude [dB]

40

60

80

100
0.4 0.2 0.0 0.2 0.4
Normalized frequency [cycles per sample]

3.18. Window functions 807

NumPy Reference, Release 2.0.0.dev8464

hamming(M)
Return the Hamming window.
The Hamming window is a taper formed by using a weighted cosine.
Parameters
M : int
Number of points in the output window. If zero or less, an empty array is returned.
Returns
out : ndarray
The window, normalized to one (the value one appears only if the number of samples is
odd).
See Also:
bartlett, blackman, hanning, kaiser

Notes
The Hamming window is defined as
 
2n
w(n) = 0.54 + 0.46cos 0nM 1
M 1

The Hamming was named for R. W. Hamming, an associate of J. W. Tukey and is described in Blackman and
Tukey. It was recommended for smoothing the truncated autocovariance function in the time domain. Most
references to the Hamming window come from the signal processing literature, where it is used as one of many
windowing functions for smoothing values. It is also known as an apodization (which means removing the
foot, i.e. smoothing discontinuities at the beginning and end of the sampled signal) or tapering function.

References
[R30], [R31], [R32], [R33]

Examples

>>> np.hamming(12)
array([ 0.08 , 0.15302337, 0.34890909, 0.60546483, 0.84123594,
0.98136677, 0.98136677, 0.84123594, 0.60546483, 0.34890909,
0.15302337, 0.08 ])

Plot the window and the frequency response:

>>> from numpy.fft import fft, fftshift

>>> import matplotlib.pyplot as plt

>>> window = np.hamming(51)

>>> plt.plot(window)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.title("Hamming window")
<matplotlib.text.Text object at 0x...>
>>> plt.ylabel("Amplitude")
<matplotlib.text.Text object at 0x...>
>>> plt.xlabel("Sample")
<matplotlib.text.Text object at 0x...>
>>> plt.show()

808 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> plt.figure()
<matplotlib.figure.Figure object at 0x...>
>>> A = fft(window, 2048) / 25.5
>>> mag = np.abs(fftshift(A))
>>> freq = np.linspace(-0.5, 0.5, len(A))
>>> response = 20 * np.log10(mag)
>>> response = np.clip(response, -100, 100)
>>> plt.plot(freq, response)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.title("Frequency response of Hamming window")
<matplotlib.text.Text object at 0x...>
>>> plt.ylabel("Magnitude [dB]")
<matplotlib.text.Text object at 0x...>
>>> plt.xlabel("Normalized frequency [cycles per sample]")
<matplotlib.text.Text object at 0x...>
>>> plt.axis(tight)
(-0.5, 0.5, -100.0, ...)
>>> plt.show()

Hamming window
1.0

0.8

0.6
Amplitude

0.4

0.2

0.0
0 10 20 30 40 50
Sample

Frequency response of Hamming window

0

20
Magnitude [dB]

40

60

80

100
0.4 0.2 0.0 0.2 0.4
Normalized frequency [cycles per sample]

3.18. Window functions 809

NumPy Reference, Release 2.0.0.dev8464

hanning(M)
Return the Hanning window.
The Hanning window is a taper formed by using a weighted cosine.
Parameters
M : int
Number of points in the output window. If zero or less, an empty array is returned.
Returns
out : ndarray, shape(M,)
The window, normalized to one (the value one appears only if M is odd).
See Also:
bartlett, blackman, hamming, kaiser

Notes
The Hanning window is defined as
 
2n
w(n) = 0.5 0.5cos 0nM 1
M 1

The Hanning was named for Julius van Hann, an Austrian meterologist. It is also known as the Cosine Bell.
Some authors prefer that it be called a Hann window, to help avoid confusion with the very similar Hamming
window.
Most references to the Hanning window come from the signal processing literature, where it is used as one of
many windowing functions for smoothing values. It is also known as an apodization (which means removing
the foot, i.e. smoothing discontinuities at the beginning and end of the sampled signal) or tapering function.

References
[R34], [R35], [R36], [R37]

Examples

>>> from numpy import hanning

>>> hanning(12)
array([ 0. , 0.07937323, 0.29229249, 0.57115742, 0.82743037,
0.97974649, 0.97974649, 0.82743037, 0.57115742, 0.29229249,
0.07937323, 0. ])

Plot the window and its frequency response:

>>> from numpy.fft import fft, fftshift

>>> import matplotlib.pyplot as plt

>>> window = np.hanning(51)

>>> plt.plot(window)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.title("Hann window")
<matplotlib.text.Text object at 0x...>
>>> plt.ylabel("Amplitude")
<matplotlib.text.Text object at 0x...>
>>> plt.xlabel("Sample")

810 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

<matplotlib.text.Text object at 0x...>

>>> plt.show()

>>> plt.figure()
<matplotlib.figure.Figure object at 0x...>
>>> A = fft(window, 2048) / 25.5
>>> mag = abs(fftshift(A))
>>> freq = np.linspace(-0.5,0.5,len(A))
>>> response = 20*np.log10(mag)
>>> response = np.clip(response,-100,100)
>>> plt.plot(freq, response)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.title("Frequency response of the Hann window")
<matplotlib.text.Text object at 0x...>
>>> plt.ylabel("Magnitude [dB]")
<matplotlib.text.Text object at 0x...>
>>> plt.xlabel("Normalized frequency [cycles per sample]")
<matplotlib.text.Text object at 0x...>
>>> plt.axis(tight)
(-0.5, 0.5, -100.0, ...)
>>> plt.show()

Hann window
1.0

0.8

0.6
Amplitude

0.4

0.2

0.0
0 10 20 30 40 50
Sample

3.18. Window functions 811

NumPy Reference, Release 2.0.0.dev8464

Frequency response of the Hann window

20
Magnitude [dB]

40

60

80

100
0.4 0.2 0.0 0.2 0.4
Normalized frequency [cycles per sample]

kaiser(M, beta)
Return the Kaiser window.
The Kaiser window is a taper formed by using a Bessel function.
Parameters
M : int
Number of points in the output window. If zero or less, an empty array is returned.
beta : float
Shape parameter for window.
Returns
out : array
The window, normalized to one (the value one appears only if the number of samples is
odd).
See Also:
bartlett, blackman, hamming, hanning

Notes
The Kaiser window is defined as
s !
4n2
w(n) = I0 1 /I0 ()
(M 1)2

with
M 1 M 1
n ,
2 2

where I0 is the modified zeroth-order Bessel function.

The Kaiser was named for Jim Kaiser, who discovered a simple approximation to the DPSS window based on
Bessel functions. The Kaiser window is a very good approximation to the Digital Prolate Spheroidal Sequence,

812 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

or Slepian window, which is the transform which maximizes the energy in the main lobe of the window relative
to total energy.
The Kaiser can approximate many other windows by varying the beta parameter.
beta Window shape
0 Rectangular
5 Similar to a Hamming
6 Similar to a Hanning
8.6 Similar to a Blackman
A beta value of 14 is probably a good starting point. Note that as beta gets large, the window narrows, and so
the number of samples needs to be large enough to sample the increasingly narrow spike, otherwise nans will
get returned.
Most references to the Kaiser window come from the signal processing literature, where it is used as one of
many windowing functions for smoothing values. It is also known as an apodization (which means removing
the foot, i.e. smoothing discontinuities at the beginning and end of the sampled signal) or tapering function.

References
[R43], [R44], [R45]

Examples

>>> from numpy import kaiser

>>> kaiser(12, 14)
array([ 7.72686684e-06, 3.46009194e-03, 4.65200189e-02,
2.29737120e-01, 5.99885316e-01, 9.45674898e-01,
9.45674898e-01, 5.99885316e-01, 2.29737120e-01,
4.65200189e-02, 3.46009194e-03, 7.72686684e-06])

Plot the window and the frequency response:

>>> from numpy import clip, log10, array, kaiser, linspace

>>> from numpy.fft import fft, fftshift
>>> import matplotlib.pyplot as plt

>>> window = kaiser(51, 14)

>>> plt.plot(window)
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.title("Kaiser window")
<matplotlib.text.Text object at 0x...>
>>> plt.ylabel("Amplitude")
<matplotlib.text.Text object at 0x...>
>>> plt.xlabel("Sample")
<matplotlib.text.Text object at 0x...>
>>> plt.show()

>>> plt.figure()
<matplotlib.figure.Figure object at 0x...>
>>> A = fft(window, 2048) / 25.5
>>> mag = abs(fftshift(A))
>>> freq = linspace(-0.5,0.5,len(A))
>>> response = 20*log10(mag)
>>> response = clip(response,-100,100)
>>> plt.plot(freq, response)
[<matplotlib.lines.Line2D object at 0x...>]

3.18. Window functions 813

NumPy Reference, Release 2.0.0.dev8464

>>> plt.title("Frequency response of Kaiser window")

<matplotlib.text.Text object at 0x...>
>>> plt.ylabel("Magnitude [dB]")
<matplotlib.text.Text object at 0x...>
>>> plt.xlabel("Normalized frequency [cycles per sample]")
<matplotlib.text.Text object at 0x...>
>>> plt.axis(tight)
(-0.5, 0.5, -100.0, ...)
>>> plt.show()

Kaiser window
1.0

0.8

0.6
Amplitude

0.4

0.2

0.0
0 10 20 30 40 50
Sample

Frequency response of Kaiser window

20
Magnitude [dB]

40

60

80

100
0.4 0.2 0.0 0.2 0.4
Normalized frequency [cycles per sample]

814 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

3.19 Floating point error handling

3.19.1 Setting and getting error handling

seterr([all, divide, over, under, invalid]) Set how floating-point errors are handled.
geterr() Get the current way of handling floating-point errors.
seterrcall(func) Set the floating-point error callback function or log object.
geterrcall() Return the current callback function used on floating-point errors.
errstate(**kwargs) Context manager for floating-point error handling.
seterr(all=None, divide=None, over=None, under=None, invalid=None)
Set how floating-point errors are handled.
Note that operations on integer scalar types (such as int16) are handled like floating point, and are affected by
these settings.
Parameters
all : {ignore, warn, raise, call, print, log}, optional
Set treatment for all types of floating-point errors at once:
ignore: Take no action when the exception occurs.
warn: Print a RuntimeWarning (via the Python warnings module).
raise: Raise a FloatingPointError.
call: Call a function specified using the seterrcall function.
print: Print a warning directly to stdout.
log: Record error in a Log object specified by seterrcall.
The default is not to change the current behavior.
divide : {ignore, warn, raise, call, print, log}, optional
Treatment for division by zero.
over : {ignore, warn, raise, call, print, log}, optional
Treatment for floating-point overflow.
under : {ignore, warn, raise, call, print, log}, optional
Treatment for floating-point underflow.
invalid : {ignore, warn, raise, call, print, log}, optional
Treatment for invalid floating-point operation.
Returns
old_settings : dict
Dictionary containing the old settings.
See Also:

seterrcall
Set a callback function for the call mode.

geterr, geterrcall

3.19. Floating point error handling 815

NumPy Reference, Release 2.0.0.dev8464

Notes
The floating-point exceptions are defined in the IEEE 754 standard [1]:
Division by zero: infinite result obtained from finite numbers.
Overflow: result too large to be expressed.
Underflow: result so close to zero that some precision was lost.
Invalid operation: result is not an expressible number, typically indicates that a NaN was produced.

Examples

>>> old_settings = np.seterr(all=ignore) #seterr to known value

>>> np.seterr(over=raise)
{over: ignore, divide: ignore, invalid: ignore,
under: ignore}
>>> np.seterr(all=ignore) # reset to default
{over: raise, divide: ignore, invalid: ignore, under: ignore}

>>> np.int16(32000) * np.int16(3)

30464
>>> old_settings = np.seterr(all=warn, over=raise)
>>> np.int16(32000) * np.int16(3)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
FloatingPointError: overflow encountered in short_scalars

>>> old_settings = np.seterr(all=print)

>>> np.geterr()
{over: print, divide: print, invalid: print, under: print}
>>> np.int16(32000) * np.int16(3)
Warning: overflow encountered in short_scalars
30464

geterr()
Get the current way of handling floating-point errors.
Returns
res : dict
A dictionary with keys divide, over, under, and invalid, whose values are from
the strings ignore, print, log, warn, raise, and call. The keys represent pos-
sible floating-point exceptions, and the values define how these exceptions are handled.
See Also:
geterrcall, seterr, seterrcall

Notes
For complete documentation of the types of floating-point exceptions and treatment options, see seterr.

Examples

>>> np.geterr() # default is all set to ignore

{over: ignore, divide: ignore, invalid: ignore,
under: ignore}

816 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.arange(3.) / np.arange(3.)

array([ NaN, 1., 1.])

>>> oldsettings = np.seterr(all=warn, over=raise)

>>> np.geterr()
{over: raise, divide: warn, invalid: warn, under: warn}
>>> np.arange(3.) / np.arange(3.)
__main__:1: RuntimeWarning: invalid value encountered in divide
array([ NaN, 1., 1.])

seterrcall(func)
Set the floating-point error callback function or log object.
There are two ways to capture floating-point error messages. The first is to set the error-handler to call, using
seterr. Then, set the function to call using this function.
The second is to set the error-handler to log, using seterr. Floating-point errors then trigger a call to the write
method of the provided object.
Parameters
func : callable f(err, flag) or object with write method
Function to call upon floating-point errors (call-mode) or object whose write method
is used to log such message (log-mode).
The call function takes two arguments. The first is the type of error (one of divide,
over, under, or invalid), and the second is the status flag. The flag is a byte, whose
least-significant bits indicate the status:

[0 0 0 0 invalid over under invalid]

In other words, flags = divide + 2*over + 4*under + 8*invalid.

If an object is provided, its write method should take one argument, a string.
Returns
h : callable, log instance or None
The old error handler.
See Also:
seterr, geterr, geterrcall

Examples
Callback upon error:

>>> def err_handler(type, flag):

... print "Floating point error (%s), with flag %s" % (type, flag)
...

>>> saved_handler = np.seterrcall(err_handler)

>>> save_err = np.seterr(all=call)

>>> np.array([1, 2, 3]) / 0.0

Floating point error (divide by zero), with flag 1
array([ Inf, Inf, Inf])

3.19. Floating point error handling 817

NumPy Reference, Release 2.0.0.dev8464

>>> np.seterrcall(saved_handler)
<function err_handler at 0x...>
>>> np.seterr(**save_err)
{over: call, divide: call, invalid: call, under: call}

Log error message:

>>> class Log(object):

... def write(self, msg):
... print "LOG: %s" % msg
...

>>> log = Log()

>>> saved_handler = np.seterrcall(log)
>>> save_err = np.seterr(all=log)

>>> np.array([1, 2, 3]) / 0.0

LOG: Warning: divide by zero encountered in divide
<BLANKLINE>
array([ Inf, Inf, Inf])

>>> np.seterrcall(saved_handler)
<__main__.Log object at 0x...>
>>> np.seterr(**save_err)
{over: log, divide: log, invalid: log, under: log}

geterrcall()
Return the current callback function used on floating-point errors.
When the error handling for a floating-point error (one of divide, over, under, or invalid) is set to call
or log, the function that is called or the log instance that is written to is returned by geterrcall. This function
or log instance has been set with seterrcall.
Returns
errobj : callable, log instance or None
The current error handler. If no handler was set through seterrcall, None is returned.
See Also:
seterrcall, seterr, geterr

Notes
For complete documentation of the types of floating-point exceptions and treatment options, see seterr.

Examples

>>> np.geterrcall() # we did not yet set a handler, returns None

>>> oldsettings = np.seterr(all=call)

>>> def err_handler(type, flag):
... print "Floating point error (%s), with flag %s" % (type, flag)
>>> oldhandler = np.seterrcall(err_handler)
>>> np.array([1, 2, 3]) / 0.0

818 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Floating point error (divide by zero), with flag 1

array([ Inf, Inf, Inf])

>>> cur_handler = np.geterrcall()

>>> cur_handler is err_handler
True

class errstate(**kwargs)
Context manager for floating-point error handling.
Using an instance of errstate as a context manager allows statements in that context to execute with a known
error handling behavior. Upon entering the context the error handling is set with seterr and seterrcall, and upon
exiting it is reset to what it was before.
Parameters
kwargs : {divide, over, under, invalid}
Keyword arguments. The valid keywords are the possible floating-point exceptions.
Each keyword should have a string value that defines the treatment for the particular
error. Possible values are {ignore, warn, raise, call, print, log}.
See Also:
seterr, geterr, seterrcall, geterrcall

Notes
The with statement was introduced in Python 2.5, and can only be used there by importing it: from
__future__ import with_statement. In earlier Python versions the with statement is not avail-
able.
For complete documentation of the types of floating-point exceptions and treatment options, see seterr.

Examples

>>> from __future__ import with_statement # use with in Python 2.5

>>> olderr = np.seterr(all=ignore) # Set error handling to known state.

>>> np.arange(3) / 0.
array([ NaN, Inf, Inf])
>>> with np.errstate(divide=warn):
... np.arange(3) / 0.
...
__main__:2: RuntimeWarning: divide by zero encountered in divide
array([ NaN, Inf, Inf])

>>> np.sqrt(-1)
nan
>>> with np.errstate(invalid=raise):
... np.sqrt(-1)
Traceback (most recent call last):
File "<stdin>", line 2, in <module>
FloatingPointError: invalid value encountered in sqrt

Outside the context the error handling behavior has not changed:

3.19. Floating point error handling 819

NumPy Reference, Release 2.0.0.dev8464

>>> np.geterr()
{over: ignore, divide: ignore, invalid: ignore,
under: ignore}

3.19.2 Internal functions

seterrobj(errobj) Set the object that defines floating-point error handling.

geterrobj() Return the current object that defines floating-point error handling.
seterrobj(errobj)
Set the object that defines floating-point error handling.
The error object contains all information that defines the error handling behavior in Numpy. seterrobj is used
internally by the other functions that set error handling behavior (seterr, seterrcall).
Parameters
errobj : list
The error object, a list containing three elements: [internal numpy buffer size, error
mask, error callback function].
The error mask is a single integer that holds the treatment information on all four float-
ing point errors. The information for each error type is contained in three bits of the
integer. If we print it in base 8, we can see what treatment is set for invalid, under,
over, and divide (in that order). The printed string can be interpreted with
0 : ignore
1 : warn
2 : raise
3 : call
4 : print
5 : log
See Also:
geterrobj, seterr, geterr, seterrcall, geterrcall, getbufsize, setbufsize

Notes
For complete documentation of the types of floating-point exceptions and treatment options, see seterr.

Examples

>>> old_errobj = np.geterrobj() # first get the defaults

>>> old_errobj
[10000, 0, None]

>>> def err_handler(type, flag):

... print "Floating point error (%s), with flag %s" % (type, flag)
...
>>> new_errobj = [20000, 12, err_handler]
>>> np.seterrobj(new_errobj)
>>> np.base_repr(12, 8) # int for divide=4 (print) and over=1 (warn)
14

820 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.geterr()
{over: warn, divide: print, invalid: ignore, under: ignore}
>>> np.geterrcall() is err_handler
True

geterrobj()
Return the current object that defines floating-point error handling.
The error object contains all information that defines the error handling behavior in Numpy. geterrobj is used
internally by the other functions that get and set error handling behavior (geterr, seterr, geterrcall, seterrcall).
Returns
errobj : list
The error object, a list containing three elements: [internal numpy buffer size, error
mask, error callback function].
The error mask is a single integer that holds the treatment information on all four float-
ing point errors. The information for each error type is contained in three bits of the
integer. If we print it in base 8, we can see what treatment is set for invalid, under,
over, and divide (in that order). The printed string can be interpreted with
0 : ignore
1 : warn
2 : raise
3 : call
4 : print
5 : log
See Also:
seterrobj, seterr, geterr, seterrcall, geterrcall, getbufsize, setbufsize

Notes
For complete documentation of the types of floating-point exceptions and treatment options, see seterr.

Examples

>>> np.geterrobj() # first get the defaults

[10000, 0, None]

>>> def err_handler(type, flag):

... print "Floating point error (%s), with flag %s" % (type, flag)
...
>>> old_bufsize = np.setbufsize(20000)
>>> old_err = np.seterr(divide=raise)
>>> old_handler = np.seterrcall(err_handler)
>>> np.geterrobj()
[20000, 2, <function err_handler at 0x91dcaac>]

>>> old_err = np.seterr(all=ignore)

>>> np.base_repr(np.geterrobj()[1], 8)
0
>>> old_err = np.seterr(divide=warn, over=log, under=call,

3.19. Floating point error handling 821

NumPy Reference, Release 2.0.0.dev8464

invalid=print)
>>> np.base_repr(np.geterrobj()[1], 8)
4351

3.20 Masked array operations

3.20.1 Constants

ma.MaskType Numpys Boolean type. Character code: ?. Alias: bool8

MaskType
alias of bool_

3.20.2 Creation

From existing data

ma.masked_array An array class with possibly masked values.

ma.array(data[, dtype, copy, order, mask, ...]) An array class with possibly masked values.
ma.copy copy
ma.frombuffer(buffer[, dtype, count, offset]) Interpret a buffer as a 1-dimensional array.
ma.fromfunction(function, shape, **kwargs) Construct an array by executing a function over each coordinate.
ma.MaskedArray.copy([order]) Return a copy of the array.
masked_array
alias of MaskedArray
array(data, dtype=None, copy=False, order=False, mask=False, fill_value=None, keep_mask=True,
hard_mask=False, shrink=True, subok=True, ndmin=0)
An array class with possibly masked values.
Masked values of True exclude the corresponding element from any computation.
Construction:

x = MaskedArray(data, mask=nomask, dtype=None, copy=True,

fill_value=None, keep_mask=True, hard_mask=False,
shrink=True)

Parameters
data : array_like
Input data.
mask : sequence, optional
Mask. Must be convertible to an array of booleans with the same shape as data. True
indicates a masked (i.e. invalid) data.
dtype : dtype, optional
Data type of the output. If dtype is None, the type of the data argument (data.dtype)
is used. If dtype is not None and different from data.dtype, a copy is performed.
copy : bool, optional

822 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Whether to copy the input data (True), or to use a reference instead. Default is False.
subok : bool, optional
Whether to return a subclass of MaskedArray if possible (True) or a plain MaskedArray.
Default is True.
ndmin : int, optional
Minimum number of dimensions. Default is 0.
fill_value : scalar, optional
Value used to fill in the masked values when necessary. If None, a default based on the
data-type is used.
keep_mask : bool, optional
Whether to combine mask with the mask of the input data, if any (True), or to use only
mask for the output (False). Default is True.
hard_mask : bool, optional
Whether to use a hard mask or not. With a hard mask, masked values cannot be un-
masked. Default is False.
shrink : bool, optional
Whether to force compression of an empty mask. Default is True.

copy
copy a.copy(order=C)
Return a copy of the array.

Parameters
order : {C, F, A}, optional
By default, the result is stored in C-contiguous (row-major) order in memory. If order is
F, the result has Fortran (column-major) order. If order is A (Any), then the result
has the same order as the input.

Examples

>>> x = np.array([[1,2,3],[4,5,6]], order=F)

>>> y = x.copy()

>>> x.fill(0)

>>> x
array([[0, 0, 0],
[0, 0, 0]])

>>> y
array([[1, 2, 3],
[4, 5, 6]])

3.20. Masked array operations 823

NumPy Reference, Release 2.0.0.dev8464

>>> y.flags[C_CONTIGUOUS]
True

frombuffer
Interpret a buffer as a 1-dimensional array.
Parameters
buffer :
An object that exposes the buffer interface.
dtype : data-type, optional
Data type of the returned array.
count : int, optional
Number of items to read. -1 means all data in the buffer.
offset : int, optional
Start reading the buffer from this offset.

Notes
If the buffer has data that is not in machine byte-order, this should be specified as part of the data-type, e.g.:

>>> dt = np.dtype(int)
>>> dt = dt.newbyteorder(>)
>>> np.frombuffer(buf, dtype=dt)

The data of the resulting array will not be byteswapped, but will be interpreted correctly.

Examples

>>> s = hello world

>>> np.frombuffer(s, dtype=S1, count=5, offset=6)
array([w, o, r, l, d],
dtype=|S1)

fromfunction
Construct an array by executing a function over each coordinate.
The resulting array therefore has a value fn(x, y, z) at coordinate (x, y, z).
Parameters
function : callable
The function is called with N parameters, each of which represents the coordinates of
the array varying along a specific axis. For example, if shape were (2, 2), then the
parameters would be two arrays, [[0, 0], [1, 1]] and [[0, 1], [0, 1]].
function must be capable of operating on arrays, and should return a scalar value.
shape : (N,) tuple of ints
Shape of the output array, which also determines the shape of the coordinate arrays
passed to function.
dtype : data-type, optional
Data-type of the coordinate arrays passed to function. By default, dtype is float.

824 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Returns
out : any
The result of the call to function is passed back directly. Therefore the type and shape
of out is completely determined by function.
See Also:
indices, meshgrid

Notes
Keywords other than shape and dtype are passed to function.

Examples

>>> np.fromfunction(lambda i, j: i == j, (3, 3), dtype=int)

array([[ True, False, False],
[False, True, False],
[False, False, True]], dtype=bool)

>>> np.fromfunction(lambda i, j: i + j, (3, 3), dtype=int)

array([[0, 1, 2],
[1, 2, 3],
[2, 3, 4]])

copy(order=C)
Return a copy of the array.
Parameters
order : {C, F, A}, optional
By default, the result is stored in C-contiguous (row-major) order in memory. If order is
F, the result has Fortran (column-major) order. If order is A (Any), then the result
has the same order as the input.

Examples

>>> x = np.array([[1,2,3],[4,5,6]], order=F)

>>> y = x.copy()

>>> x.fill(0)

>>> x
array([[0, 0, 0],
[0, 0, 0]])

>>> y
array([[1, 2, 3],
[4, 5, 6]])

>>> y.flags[C_CONTIGUOUS]
True

3.20. Masked array operations 825

NumPy Reference, Release 2.0.0.dev8464

Ones and zeros

ma.empty(shape[, dtype, order]) Return a new array of given shape and type, without initializing entries.
ma.empty_like(a) Return a new array with the same shape and type as a given array.
ma.masked_all(shape[, dtype]) Empty masked array with all elements masked.
ma.masked_all_like(arr) Empty masked array with the properties of an existing array.
ma.ones(shape[, dtype, order]) Return a new array of given shape and type, filled with ones.
ma.zeros(shape[, dtype, order]) Return a new array of given shape and type, filled with zeros.
empty
Return a new array of given shape and type, without initializing entries.
Parameters
shape : int or tuple of int
Shape of the empty array
dtype : data-type, optional
Desired output data-type.
order : {C, F}, optional
Whether to store multi-dimensional data in C (row-major) or Fortran (column-major)
order in memory.
See Also:
empty_like, zeros, ones

Notes
empty, unlike zeros, does not set the array values to zero, and may therefore be marginally faster. On the other
hand, it requires the user to manually set all the values in the array, and should be used with caution.

Examples

>>> np.empty([2, 2])

array([[ -9.74499359e+001, 6.69583040e-309],
[ 2.13182611e-314, 3.06959433e-309]]) #random

>>> np.empty([2, 2], dtype=int)

array([[-1073741821, -1067949133],
[ 496041986, 19249760]]) #random

empty_like
Return a new array with the same shape and type as a given array.
Parameters
a : array_like
The shape and data-type of a define the parameters of the returned array.
Returns
out : ndarray
Array of random data with the same shape and type as a.
See Also:

ones_like
Return an array of ones with shape and type of input.

826 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

zeros_like
Return an array of zeros with shape and type of input.
empty
Return a new uninitialized array.
ones
Return a new array setting values to one.
zeros
Return a new array setting values to zero.

Notes
This function does not initialize the returned array; to do that use zeros_like or ones_like instead. It may be
marginally faster than the functions that do set the array values.

Examples

>>> a = ([1,2,3], [4,5,6]) # a is array-like

>>> np.empty_like(a)
array([[-1073741821, -1073741821, 3], #random
[ 0, 0, -1073741821]])
>>> a = np.array([[1., 2., 3.],[4.,5.,6.]])
>>> np.empty_like(a)
array([[ -2.00000715e+000, 1.48219694e-323, -2.00000572e+000], #random
[ 4.38791518e-305, -2.00000715e+000, 4.17269252e-309]])

masked_all(shape, dtype=<type float>)

Empty masked array with all elements masked.
Return an empty masked array of the given shape and dtype, where all the data are masked.
Parameters
shape : tuple
Shape of the required MaskedArray.
dtype : dtype, optional
Data type of the output.
Returns
a : MaskedArray
A masked array with all data masked.
See Also:

masked_all_like
Empty masked array modelled on an existing array.

Examples

>>> import numpy.ma as ma

>>> ma.masked_all((3, 3))
masked_array(data =
[[-- -- --]
[-- -- --]
[-- -- --]],
mask =

3.20. Masked array operations 827

NumPy Reference, Release 2.0.0.dev8464

[[ True True True]

[ True True True]
[ True True True]],
fill_value=1e+20)

The dtype parameter defines the underlying data type.

>>> a = ma.masked_all((3, 3))

>>> a.dtype
dtype(float64)
>>> a = ma.masked_all((3, 3), dtype=np.int32)
>>> a.dtype
dtype(int32)

masked_all_like(arr)
Empty masked array with the properties of an existing array.
Return an empty masked array of the same shape and dtype as the array arr, where all the data are masked.
Parameters
arr : ndarray
An array describing the shape and dtype of the required MaskedArray.
Returns
a : MaskedArray
A masked array with all data masked.
Raises
AttributeError :
If arr doesnt have a shape attribute (i.e. not an ndarray)
See Also:

masked_all
Empty masked array with all elements masked.

Examples

>>> import numpy.ma as ma

>>> arr = np.zeros((2, 3), dtype=np.float32)
>>> arr
array([[ 0., 0., 0.],
[ 0., 0., 0.]], dtype=float32)
>>> ma.masked_all_like(arr)
masked_array(data =
[[-- -- --]
[-- -- --]],
mask =
[[ True True True]
[ True True True]],
fill_value=1e+20)

The dtype of the masked array matches the dtype of arr.

>>> arr.dtype
dtype(float32)

828 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> ma.masked_all_like(arr).dtype
dtype(float32)

ones
Return a new array of given shape and type, filled with ones.
Please refer to the documentation for zeros.
See Also:
zeros

Examples

>>> np.ones(5)
array([ 1., 1., 1., 1., 1.])

>>> np.ones((5,), dtype=np.int)

array([1, 1, 1, 1, 1])

>>> np.ones((2, 1))

array([[ 1.],
[ 1.]])

>>> s = (2,2)
>>> np.ones(s)
array([[ 1., 1.],
[ 1., 1.]])

zeros
Return a new array of given shape and type, filled with zeros.
Parameters
shape : int or sequence of ints
Shape of the new array, e.g., (2, 3) or 2.
dtype : data-type, optional
The desired data-type for the array, e.g., numpy.int8. Default is numpy.float64.
order : {C, F}, optional
Whether to store multidimensional data in C- or Fortran-contiguous (row- or column-
wise) order in memory.
Returns
out : ndarray
Array of zeros with the given shape, dtype, and order.
See Also:

zeros_like
Return an array of zeros with shape and type of input.
ones_like
Return an array of ones with shape and type of input.

3.20. Masked array operations 829

NumPy Reference, Release 2.0.0.dev8464

empty_like
Return an empty array with shape and type of input.
ones
Return a new array setting values to one.
empty
Return a new uninitialized array.

Examples

>>> np.zeros(5)
array([ 0., 0., 0., 0., 0.])

>>> np.zeros((5,), dtype=numpy.int)

array([0, 0, 0, 0, 0])

>>> np.zeros((2, 1))

array([[ 0.],
[ 0.]])

>>> s = (2,2)
>>> np.zeros(s)
array([[ 0., 0.],
[ 0., 0.]])

>>> np.zeros((2,), dtype=[(x, i4), (y, i4)]) # custom dtype

array([(0, 0), (0, 0)],
dtype=[(x, <i4), (y, <i4)])

3.20.3 Inspecting the array

ma.all(self[, axis, out]) Check if all of the elements of a are true.

ma.any(self[, axis, out]) Check if any of the elements of a are true.
ma.count(a[, axis]) Count the non-masked elements of the array along the given axis.
ma.count_masked(arr[, axis]) Count the number of masked elements along the given axis.
ma.getmask(a) Return the mask of a masked array, or nomask.
ma.getmaskarray(arr) Return the mask of a masked array, or full boolean array of False.
ma.getdata(a[, subok]) Return the data of a masked array as an ndarray.
ma.nonzero(self) Return the indices of unmasked elements that are not zero.
ma.shape(obj) Return the shape of an array.
ma.size(obj[, axis]) Return the number of elements along a given axis.
ma.MaskedArray.data
ma.MaskedArray.mask Mask
ma.MaskedArray.recordmask
ma.MaskedArray.all([axis, out]) Check if all of the elements of a are true.
ma.MaskedArray.any([axis, out]) Check if any of the elements of a are true.
ma.MaskedArray.count([axis]) Count the non-masked elements of the array along the given axis.
ma.MaskedArray.nonzero() Return the indices of unmasked elements that are not zero.
ma.shape(obj) Return the shape of an array.
ma.size(obj[, axis]) Return the number of elements along a given axis.

830 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

all
Check if all of the elements of a are true.
Performs a logical_and over the given axis and returns the result. Masked values are considered as True
during computation. For convenience, the output array is masked where ALL the values along the current axis
are masked: if the output would have been a scalar and that all the values are masked, then the output is masked.
Parameters
axis : {None, integer}
Axis to perform the operation over. If None, perform over flattened array.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
See Also:

all
equivalent function

Examples

>>> np.ma.array([1,2,3]).all()
True
>>> a = np.ma.array([1,2,3], mask=True)
>>> (a.all() is np.ma.masked)
True

any
Check if any of the elements of a are true.
Performs a logical_or over the given axis and returns the result. Masked values are considered as False during
computation.
Parameters
axis : {None, integer}
Axis to perform the operation over. If None, perform over flattened array and return a
scalar.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
See Also:

any
equivalent function

count(a, axis=None)
Count the non-masked elements of the array along the given axis.
Parameters
axis : int, optional
Axis along which to count the non-masked elements. If axis is None, all non-masked
elements are counted.

3.20. Masked array operations 831

NumPy Reference, Release 2.0.0.dev8464

Returns
result : int or ndarray
If axis is None, an integer count is returned. When axis is not None, an array with shape
determined by the lengths of the remaining axes, is returned.
See Also:

count_masked
Count masked elements in array or along a given axis.

Examples

>>> import numpy.ma as ma

>>> a = ma.arange(6).reshape((2, 3))
>>> a[1, :] = ma.masked
>>> a
masked_array(data =
[[0 1 2]
[-- -- --]],
mask =
[[False False False]
[ True True True]],
fill_value = 999999)
>>> a.count()
3

When the axis keyword is specified an array of appropriate size is returned.

>>> a.count(axis=0)
array([1, 1, 1])
>>> a.count(axis=1)
array([3, 0])

count_masked(arr, axis=None)
Count the number of masked elements along the given axis.
Parameters
arr : array_like
An array with (possibly) masked elements.
axis : int, optional
Axis along which to count. If None (default), a flattened version of the array is used.
Returns
count : int, ndarray
The total number of masked elements (axis=None) or the number of masked elements
along each slice of the given axis.
See Also:

MaskedArray.count
Count non-masked elements.

832 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> import numpy.ma as ma

>>> a = np.arange(9).reshape((3,3))
>>> a = ma.array(a)
>>> a[1, 0] = ma.masked
>>> a[1, 2] = ma.masked
>>> a[2, 1] = ma.masked
>>> a
masked_array(data =
[[0 1 2]
[-- 4 --]
[6 -- 8]],
mask =
[[False False False]
[ True False True]
[False True False]],
fill_value=999999)
>>> ma.count_masked(a)
3

When the axis keyword is used an array is returned.

>>> ma.count_masked(a, axis=0)

array([1, 1, 1])
>>> ma.count_masked(a, axis=1)
array([0, 2, 1])

getmask(a)
Return the mask of a masked array, or nomask.
Return the mask of a as an ndarray if a is a MaskedArray and the mask is not nomask, else return nomask. To
guarantee a full array of booleans of the same shape as a, use getmaskarray.
Parameters
a : array_like
Input MaskedArray for which the mask is required.
See Also:

getdata
Return the data of a masked array as an ndarray.
getmaskarray
Return the mask of a masked array, or full array of False.

Examples

>>> import numpy.ma as ma

>>> a = ma.masked_equal([[1,2],[3,4]], 2)
>>> a
masked_array(data =
[[1 --]
[3 4]],
mask =
[[False True]
[False False]],
fill_value=999999)

3.20. Masked array operations 833

NumPy Reference, Release 2.0.0.dev8464

>>> ma.getmask(a)
array([[False, True],
[False, False]], dtype=bool)

Equivalently use the MaskedArray mask attribute.

>>> a.mask
array([[False, True],
[False, False]], dtype=bool)

Result when mask == nomask

>>> b = ma.masked_array([[1,2],[3,4]])
>>> b
masked_array(data =
[[1 2]
[3 4]],
mask =
False,
fill_value=999999)
>>> ma.nomask
False
>>> ma.getmask(b) == ma.nomask
True
>>> b.mask == ma.nomask
True

getmaskarray(arr)
Return the mask of a masked array, or full boolean array of False.
Return the mask of arr as an ndarray if arr is a MaskedArray and the mask is not nomask, else return a full
boolean array of False of the same shape as arr.
Parameters
arr : array_like
Input MaskedArray for which the mask is required.
See Also:

getmask
Return the mask of a masked array, or nomask.
getdata
Return the data of a masked array as an ndarray.

Examples

>>> import numpy.ma as ma

>>> a = ma.masked_equal([[1,2],[3,4]], 2)
>>> a
masked_array(data =
[[1 --]
[3 4]],
mask =
[[False True]
[False False]],
fill_value=999999)

834 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> ma.getmaskarray(a)
array([[False, True],
[False, False]], dtype=bool)

Result when mask == nomask

>>> b = ma.masked_array([[1,2],[3,4]])
>>> b
masked_array(data =
[[1 2]
[3 4]],
mask =
False,
fill_value=999999)
>>> >ma.getmaskarray(b)
array([[False, False],
[False, False]], dtype=bool)

getdata(a, subok=True)
Return the data of a masked array as an ndarray.
Return the data of a (if any) as an ndarray if a is a MaskedArray, else return a as a ndarray or subclass
(depending on subok) if not.
Parameters
a : array_like
Input MaskedArray, alternatively a ndarray or a subclass thereof.
subok : bool
Whether to force the output to be a pure ndarray (False) or to return a subclass of ndarray
if appropriate (True, default).
See Also:

getmask
Return the mask of a masked array, or nomask.
getmaskarray
Return the mask of a masked array, or full array of False.

Examples

>>> import numpy.ma as ma

>>> a = ma.masked_equal([[1,2],[3,4]], 2)
>>> a
masked_array(data =
[[1 --]
[3 4]],
mask =
[[False True]
[False False]],
fill_value=999999)
>>> ma.getdata(a)
array([[1, 2],
[3, 4]])

Equivalently use the MaskedArray data attribute.

3.20. Masked array operations 835

NumPy Reference, Release 2.0.0.dev8464

>>> a.data
array([[1, 2],
[3, 4]])

nonzero
Return the indices of unmasked elements that are not zero.
Returns a tuple of arrays, one for each dimension, containing the indices of the non-zero elements in that
dimension. The corresponding non-zero values can be obtained with:

a[a.nonzero()]

To group the indices by element, rather than dimension, use instead:

np.transpose(a.nonzero())

The result of this is always a 2d array, with a row for each non-zero element.
Parameters
None :
Returns
tuple_of_arrays : tuple
Indices of elements that are non-zero.
See Also:

numpy.nonzero
Function operating on ndarrays.
flatnonzero
Return indices that are non-zero in the flattened version of the input array.
ndarray.nonzero
Equivalent ndarray method.

Examples

>>> import numpy.ma as ma

>>> x = ma.array(np.eye(3))
>>> x
masked_array(data =
[[ 1. 0. 0.]
[ 0. 1. 0.]
[ 0. 0. 1.]],
mask =
False,
fill_value=1e+20)
>>> x.nonzero()
(array([0, 1, 2]), array([0, 1, 2]))

Masked elements are ignored.

>>> x[1, 1] = ma.masked

>>> x
masked_array(data =
[[1.0 0.0 0.0]

836 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

[0.0 -- 0.0]
[0.0 0.0 1.0]],
mask =
[[False False False]
[False True False]
[False False False]],
fill_value=1e+20)
>>> x.nonzero()
(array([0, 2]), array([0, 2]))

Indices can also be grouped by element.

>>> np.transpose(x.nonzero())
array([[0, 0],
[2, 2]])

A common use for nonzero is to find the indices of an array, where a condition is True. Given an array a, the
condition a > 3 is a boolean array and since False is interpreted as 0, ma.nonzero(a > 3) yields the indices of the
a where the condition is true.

>>> a = ma.array([[1,2,3],[4,5,6],[7,8,9]])
>>> a > 3
masked_array(data =
[[False False False]
[ True True True]
[ True True True]],
mask =
False,
fill_value=999999)
>>> ma.nonzero(a > 3)
(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

The nonzero method of the condition array can also be called.

>>> (a > 3).nonzero()

(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

shape(obj)
Return the shape of an array.
Parameters
a : array_like
Input array.
Returns
shape : tuple of ints
The elements of the shape tuple give the lengths of the corresponding array dimensions.
See Also:
alen
ndarray.shape
Equivalent array method.

3.20. Masked array operations 837

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.shape(np.eye(3))
(3, 3)
>>> np.shape([[1, 2]])
(1, 2)
>>> np.shape([0])
(1,)
>>> np.shape(0)
()

>>> a = np.array([(1, 2), (3, 4)], dtype=[(x, i4), (y, i4)])

>>> np.shape(a)
(2,)
>>> a.shape
(2,)

size(obj, axis=None)
Return the number of elements along a given axis.
Parameters
a : array_like
Input data.
axis : int, optional
Axis along which the elements are counted. By default, give the total number of ele-
ments.
Returns
element_count : int
Number of elements along the specified axis.
See Also:

shape
dimensions of array
ndarray.shape
dimensions of array
ndarray.size
number of elements in array

Examples

>>> a = np.array([[1,2,3],[4,5,6]])
>>> np.size(a)
6
>>> np.size(a,1)
3
>>> np.size(a,0)
2

data

838 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

mask
Mask
recordmask

all(axis=None, out=None)
Check if all of the elements of a are true.
Performs a logical_and over the given axis and returns the result. Masked values are considered as True
during computation. For convenience, the output array is masked where ALL the values along the current axis
are masked: if the output would have been a scalar and that all the values are masked, then the output is masked.
Parameters
axis : {None, integer}
Axis to perform the operation over. If None, perform over flattened array.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
See Also:

all
equivalent function

Examples

>>> np.ma.array([1,2,3]).all()
True
>>> a = np.ma.array([1,2,3], mask=True)
>>> (a.all() is np.ma.masked)
True

any(axis=None, out=None)
Check if any of the elements of a are true.
Performs a logical_or over the given axis and returns the result. Masked values are considered as False during
computation.
Parameters
axis : {None, integer}
Axis to perform the operation over. If None, perform over flattened array and return a
scalar.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
See Also:

any
equivalent function

count(axis=None)
Count the non-masked elements of the array along the given axis.

3.20. Masked array operations 839

NumPy Reference, Release 2.0.0.dev8464

Parameters
axis : int, optional
Axis along which to count the non-masked elements. If axis is None, all non-masked
elements are counted.
Returns
result : int or ndarray
If axis is None, an integer count is returned. When axis is not None, an array with shape
determined by the lengths of the remaining axes, is returned.
See Also:

count_masked
Count masked elements in array or along a given axis.

Examples

>>> import numpy.ma as ma

>>> a = ma.arange(6).reshape((2, 3))
>>> a[1, :] = ma.masked
>>> a
masked_array(data =
[[0 1 2]
[-- -- --]],
mask =
[[False False False]
[ True True True]],
fill_value = 999999)
>>> a.count()
3

When the axis keyword is specified an array of appropriate size is returned.

>>> a.count(axis=0)
array([1, 1, 1])
>>> a.count(axis=1)
array([3, 0])

nonzero()
Return the indices of unmasked elements that are not zero.
Returns a tuple of arrays, one for each dimension, containing the indices of the non-zero elements in that
dimension. The corresponding non-zero values can be obtained with:

a[a.nonzero()]

To group the indices by element, rather than dimension, use instead:

np.transpose(a.nonzero())

The result of this is always a 2d array, with a row for each non-zero element.
Parameters
None :
Returns
tuple_of_arrays : tuple

840 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Indices of elements that are non-zero.

See Also:

numpy.nonzero
Function operating on ndarrays.
flatnonzero
Return indices that are non-zero in the flattened version of the input array.
ndarray.nonzero
Equivalent ndarray method.

Examples

>>> import numpy.ma as ma

>>> x = ma.array(np.eye(3))
>>> x
masked_array(data =
[[ 1. 0. 0.]
[ 0. 1. 0.]
[ 0. 0. 1.]],
mask =
False,
fill_value=1e+20)
>>> x.nonzero()
(array([0, 1, 2]), array([0, 1, 2]))

Masked elements are ignored.

>>> x[1, 1] = ma.masked

>>> x
masked_array(data =
[[1.0 0.0 0.0]
[0.0 -- 0.0]
[0.0 0.0 1.0]],
mask =
[[False False False]
[False True False]
[False False False]],
fill_value=1e+20)
>>> x.nonzero()
(array([0, 2]), array([0, 2]))

Indices can also be grouped by element.

>>> np.transpose(x.nonzero())
array([[0, 0],
[2, 2]])

A common use for nonzero is to find the indices of an array, where a condition is True. Given an array a, the
condition a > 3 is a boolean array and since False is interpreted as 0, ma.nonzero(a > 3) yields the indices of the
a where the condition is true.

>>> a = ma.array([[1,2,3],[4,5,6],[7,8,9]])
>>> a > 3
masked_array(data =
[[False False False]

3.20. Masked array operations 841

NumPy Reference, Release 2.0.0.dev8464

[ True True True]

[ True True True]],
mask =
False,
fill_value=999999)
>>> ma.nonzero(a > 3)
(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

The nonzero method of the condition array can also be called.

>>> (a > 3).nonzero()

(array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))

shape(obj)
Return the shape of an array.
Parameters
a : array_like
Input array.
Returns
shape : tuple of ints
The elements of the shape tuple give the lengths of the corresponding array dimensions.
See Also:
alen
ndarray.shape
Equivalent array method.

Examples

>>> np.shape(np.eye(3))
(3, 3)
>>> np.shape([[1, 2]])
(1, 2)
>>> np.shape([0])
(1,)
>>> np.shape(0)
()

>>> a = np.array([(1, 2), (3, 4)], dtype=[(x, i4), (y, i4)])

>>> np.shape(a)
(2,)
>>> a.shape
(2,)

size(obj, axis=None)
Return the number of elements along a given axis.
Parameters
a : array_like
Input data.
axis : int, optional

842 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Axis along which the elements are counted. By default, give the total number of ele-
ments.
Returns
element_count : int
Number of elements along the specified axis.
See Also:

shape
dimensions of array
ndarray.shape
dimensions of array
ndarray.size
number of elements in array

Examples

>>> a = np.array([[1,2,3],[4,5,6]])
>>> np.size(a)
6
>>> np.size(a,1)
3
>>> np.size(a,0)
2

3.20.4 Manipulating a MaskedArray

Changing the shape

ma.ravel(self) Returns a 1D version of self, as a view.

ma.reshape(a, new_shape[, order]) Returns an array containing the same data with a new
shape.
ma.resize(x, new_shape) Return a new masked array with the specified size and
shape.
ma.MaskedArray.flatten([order]) Return a copy of the array collapsed into one dimension.
ma.MaskedArray.ravel() Returns a 1D version of self, as a view.
ma.MaskedArray.reshape(*s, **kwargs) Give a new shape to the array without changing its data.
ma.MaskedArray.resize(newshape[, refcheck,
...])
ravel
Returns a 1D version of self, as a view.
Returns
MaskedArray :
Output view is of shape (self.size,) (or
(np.ma.product(self.shape),)).

3.20. Masked array operations 843

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.ravel()
[1 -- 3 -- 5 -- 7 -- 9]

reshape(a, new_shape, order=C)

Returns an array containing the same data with a new shape.
Refer to MaskedArray.reshape for full documentation.
See Also:

MaskedArray.reshape
equivalent function

resize(x, new_shape)
Return a new masked array with the specified size and shape.
This is the masked equivalent of the numpy.resize function. The new array is filled with repeated copies of
x (in the order that the data are stored in memory). If x is masked, the new array will be masked, and the new
mask will be a repetition of the old one.
See Also:

numpy.resize
Equivalent function in the top level NumPy module.

Examples

>>> import numpy.ma as ma

>>> a = ma.array([[1, 2] ,[3, 4]])
>>> a[0, 1] = ma.masked
>>> a
masked_array(data =
[[1 --]
[3 4]],
mask =
[[False True]
[False False]],
fill_value = 999999)
>>> np.resize(a, (3, 3))
array([[1, 2, 3],
[4, 1, 2],
[3, 4, 1]])
>>> ma.resize(a, (3, 3))
masked_array(data =
[[1 -- 3]
[4 1 --]
[3 4 1]],
mask =
[[False True False]
[False False True]

844 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

[False False False]],

fill_value = 999999)

A MaskedArray is always returned, regardless of the input type.

>>> a = np.array([[1, 2] ,[3, 4]])

>>> ma.resize(a, (3, 3))
masked_array(data =
[[1 2 3]
[4 1 2]
[3 4 1]],
mask =
False,
fill_value = 999999)

flatten(order=C)
Return a copy of the array collapsed into one dimension.
Parameters
order : {C, F}, optional
Whether to flatten in C (row-major) or Fortran (column-major) order. The default is
C.
Returns
y : ndarray
A copy of the input array, flattened to one dimension.
See Also:

ravel
Return a flattened array.
flat
A 1-D flat iterator over the array.

Examples

>>> a = np.array([[1,2], [3,4]])

>>> a.flatten()
array([1, 2, 3, 4])
>>> a.flatten(F)
array([1, 3, 2, 4])

ravel()
Returns a 1D version of self, as a view.
Returns
MaskedArray :
Output view is of shape (self.size,) (or
(np.ma.product(self.shape),)).

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]

3.20. Masked array operations 845

NumPy Reference, Release 2.0.0.dev8464

[-- 5 --]
[7 -- 9]]
>>> print x.ravel()
[1 -- 3 -- 5 -- 7 -- 9]

reshape(*s, **kwargs)
Give a new shape to the array without changing its data.
Returns a masked array containing the same data, but with a new shape. The result is a view on the original
array; if this is not possible, a ValueError is raised.
Parameters
shape : int or tuple of ints
The new shape should be compatible with the original shape. If an integer is supplied,
then the result will be a 1-D array of that length.
order : {C, F}, optional
Determines whether the array data should be viewed as in C (row-major) or FORTRAN
(column-major) order.
Returns
reshaped_array : array
A new view on the array.
See Also:

reshape
Equivalent function in the masked array module.
numpy.ndarray.reshape
Equivalent method on ndarray object.
numpy.reshape
Equivalent function in the NumPy module.

Notes
The reshaping operation cannot guarantee that a copy will not be made, to modify the shape in place, use
a.shape = s

Examples

>>> x = np.ma.array([[1,2],[3,4]], mask=[1,0,0,1])

>>> print x
[[-- 2]
[3 --]]
>>> x = x.reshape((4,1))
>>> print x
[[--]
[2]
[3]
[--]]

resize(newshape, refcheck=True, order=False)

Warning: This method does nothing, except raise a ValueError exception. A masked array does not own
its data and therefore cannot safely be resized in place. Use the numpy.ma.resize function instead.

846 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

This method is difficult to implement safely and may be deprecated in future releases of NumPy.

Modifying axes

ma.swapaxes swapaxes
ma.transpose(a[, axes]) Permute the dimensions of an array.
ma.MaskedArray.swapaxes(axis1, axis2) Return a view of the array with axis1 and axis2 interchanged.
ma.MaskedArray.transpose(*axes) Returns a view of the array with axes transposed.
swapaxes
swapaxes a.swapaxes(axis1, axis2)
Return a view of the array with axis1 and axis2 interchanged.
Refer to numpy.swapaxes for full documentation.
See Also:

numpy.swapaxes
equivalent function

transpose(a, axes=None)
Permute the dimensions of an array.
This function is exactly equivalent to numpy.transpose.
See Also:

numpy.transpose
Equivalent function in top-level NumPy module.

Examples

>>> import numpy.ma as ma

>>> x = ma.arange(4).reshape((2,2))
>>> x[1, 1] = ma.masked
>>>> x
masked_array(data =
[[0 1]
[2 --]],
mask =
[[False False]
[False True]],
fill_value = 999999)
>>> ma.transpose(x)
masked_array(data =
[[0 2]
[1 --]],
mask =
[[False False]
[False True]],
fill_value = 999999)

swapaxes(axis1, axis2)
Return a view of the array with axis1 and axis2 interchanged.
Refer to numpy.swapaxes for full documentation.
See Also:

3.20. Masked array operations 847

NumPy Reference, Release 2.0.0.dev8464

numpy.swapaxes
equivalent function

transpose(*axes)
Returns a view of the array with axes transposed.
For a 1-D array, this has no effect. (To change between column and row vectors, first cast the 1-D array into a ma-
trix object.) For a 2-D array, this is the usual matrix transpose. For an n-D array, if axes are given, their order in-
dicates how the axes are permuted (see Examples). If axes are not provided and a.shape = (i[0], i[1],
... i[n-2], i[n-1]), then a.transpose().shape = (i[n-1], i[n-2], ... i[1],
i[0]).
Parameters
axes : None, tuple of ints, or n ints
None or no argument: reverses the order of the axes.
tuple of ints: i in the j-th place in the tuple means as i-th axis becomes a.transpose()s j-th
axis.
n ints: same as an n-tuple of the same ints (this form is intended simply as a convenience
alternative to the tuple form)
Returns
out : ndarray
View of a, with axes suitably permuted.
See Also:

ndarray.T
Array property returning the array transposed.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> a
array([[1, 2],
[3, 4]])
>>> a.transpose()
array([[1, 3],
[2, 4]])
>>> a.transpose((1, 0))
array([[1, 3],
[2, 4]])
>>> a.transpose(1, 0)
array([[1, 3],
[2, 4]])

848 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Changing the number of dimensions

ma.atleast_1d(*arys) Convert inputs to arrays with at least one dimension.

ma.atleast_2d(*arys) View inputs as arrays with at least two dimensions.
ma.atleast_3d(*arys) View inputs as arrays with at least three dimensions.
ma.expand_dims(x, axis) Expand the shape of an array.
ma.squeeze(a) Remove single-dimensional entries from the shape of an array.
ma.MaskedArray.squeeze() Remove single-dimensional entries from the shape of a.
ma.column_stack(tup) Stack 1-D arrays as columns into a 2-D array.
ma.concatenate(arrays[, axis]) Concatenate a sequence of arrays along the given axis.
ma.dstack(tup) Stack arrays in sequence depth wise (along third axis).
ma.hstack(tup) Stack arrays in sequence horizontally (column wise).
ma.hsplit(ary, indices_or_sections) Split an array into multiple sub-arrays horizontally (column-wise).
ma.mr_ Translate slice objects to concatenation along the first axis.
ma.row_stack(tup) Stack arrays in sequence vertically (row wise).
ma.vstack(tup) Stack arrays in sequence vertically (row wise).
atleast_1d(*arys)
Convert inputs to arrays with at least one dimension.
Scalar inputs are converted to 1-dimensional arrays, whilst higher-dimensional inputs are preserved.
Parameters
array1, array2, ... : array_like
One or more input arrays.
Returns
ret : ndarray
An array, or sequence of arrays, each with a.ndim >= 1. Copies are made only if
necessary.
See Also:
atleast_2d, atleast_3d

Examples

>>> np.atleast_1d(1.0)
array([ 1.])

>>> x = np.arange(9.0).reshape(3,3)
>>> np.atleast_1d(x)
array([[ 0., 1., 2.],
[ 3., 4., 5.],
[ 6., 7., 8.]])
>>> np.atleast_1d(x) is x
True

>>> np.atleast_1d(1, [3, 4])

[array([1]), array([3, 4])]

atleast_2d(*arys)
View inputs as arrays with at least two dimensions.
Parameters
array1, array2, ... : array_like

3.20. Masked array operations 849

NumPy Reference, Release 2.0.0.dev8464

One or more array-like sequences. Non-array inputs are converted to arrays. Arrays
that already have two or more dimensions are preserved.
Returns
res, res2, ... : ndarray
An array, or tuple of arrays, each with a.ndim >= 2. Copies are avoided where
possible, and views with two or more dimensions are returned.
See Also:
atleast_1d, atleast_3d

Examples

>>> np.atleast_2d(3.0)
array([[ 3.]])

>>> x = np.arange(3.0)
>>> np.atleast_2d(x)
array([[ 0., 1., 2.]])
>>> np.atleast_2d(x).base is x
True

>>> np.atleast_2d(1, [1, 2], [[1, 2]])

[array([[1]]), array([[1, 2]]), array([[1, 2]])]

atleast_3d(*arys)
View inputs as arrays with at least three dimensions.
Parameters
array1, array2, ... : array_like
One or more array-like sequences. Non-array inputs are converted to arrays. Arrays
that already have three or more dimensions are preserved.
Returns
res1, res2, ... : ndarray
An array, or tuple of arrays, each with a.ndim >= 3. Copies are avoided where
possible, and views with three or more dimensions are returned. For example, a 1-D
array of shape N becomes a view of shape (1, N, 1). A 2-D array of shape (M, N)
becomes a view of shape (M, N, 1).
See Also:
atleast_1d, atleast_2d

Examples

>>> np.atleast_3d(3.0)
array([[[ 3.]]])

>>> x = np.arange(3.0)
>>> np.atleast_3d(x).shape
(1, 3, 1)

850 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> x = np.arange(12.0).reshape(4,3)
>>> np.atleast_3d(x).shape
(4, 3, 1)
>>> np.atleast_3d(x).base is x
True

>>> for arr in np.atleast_3d([1, 2], [[1, 2]], [[[1, 2]]]):

... print arr, arr.shape
...
[[[1]
[2]]] (1, 2, 1)
[[[1]
[2]]] (1, 2, 1)
[[[1 2]]] (1, 1, 2)

expand_dims(x, axis)
Expand the shape of an array.
Expands the shape of the array by including a new axis before the one specified by the axis parameter. This
function behaves the same as numpy.expand_dims but preserves masked elements.
See Also:

numpy.expand_dims
Equivalent function in top-level NumPy module.

Examples

>>> import numpy.ma as ma

>>> x = ma.array([1, 2, 4])
>>> x[1] = ma.masked
>>> x
masked_array(data = [1 -- 4],
mask = [False True False],
fill_value = 999999)
>>> np.expand_dims(x, axis=0)
array([[1, 2, 4]])
>>> ma.expand_dims(x, axis=0)
masked_array(data =
[[1 -- 4]],
mask =
[[False True False]],
fill_value = 999999)

The same result can be achieved using slicing syntax with np.newaxis.

>>> x[np.newaxis, :]
masked_array(data =
[[1 -- 4]],
mask =
[[False True False]],
fill_value = 999999)

squeeze(a)
Remove single-dimensional entries from the shape of an array.

3.20. Masked array operations 851

NumPy Reference, Release 2.0.0.dev8464

Parameters
a : array_like
Input data.
Returns
squeezed : ndarray
The input array, but with with all dimensions of length 1 removed. Whenever possible,
a view on a is returned.

Examples

>>> x = np.array([[[0], [1], [2]]])

>>> x.shape
(1, 3, 1)
>>> np.squeeze(x).shape
(3,)

squeeze()
Remove single-dimensional entries from the shape of a.
Refer to numpy.squeeze for full documentation.
See Also:

numpy.squeeze
equivalent function

column_stack

Stack 1-D arrays as columns into a 2-D array.

Take a sequence of 1-D arrays and stack them as columns to make a single 2-D array. 2-D arrays are
stacked as-is, just like with hstack. 1-D arrays are turned into 2-D columns first.

Parameters
tup : sequence of 1-D or 2-D arrays.
Arrays to stack. All of them must have the same first dimension.
Returns
stacked : 2-D array
The array formed by stacking the given arrays.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.column_stack((a,b))
array([[1, 2],
[2, 3],
[3, 4]])

852 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

concatenate(arrays, axis=0)
Concatenate a sequence of arrays along the given axis.
Parameters
arrays : sequence of array_like
The arrays must have the same shape, except in the dimension corresponding to axis
(the first, by default).
axis : int, optional
The axis along which the arrays will be joined. Default is 0.
Returns
result : MaskedArray
The concatenated array with any masked entries preserved.
See Also:

numpy.concatenate
Equivalent function in the top-level NumPy module.

Examples

>>> import numpy.ma as ma

>>> a = ma.arange(3)
>>> a[1] = ma.masked
>>> b = ma.arange(2, 5)
>>> a
masked_array(data = [0 -- 2],
mask = [False True False],
fill_value = 999999)
>>> b
masked_array(data = [2 3 4],
mask = False,
fill_value = 999999)
>>> ma.concatenate([a, b])
masked_array(data = [0 -- 2 2 3 4],
mask = [False True False False False False],
fill_value = 999999)

dstack

Stack arrays in sequence depth wise (along third axis).

Takes a sequence of arrays and stack them along the third axis to make a single array. Rebuilds
arrays divided by dsplit. This is a simple way to stack 2D arrays (images) into a single 3D array for
processing.

Parameters
tup : sequence of arrays
Arrays to stack. All of them must have the same shape along all but the third axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

3.20. Masked array operations 853

NumPy Reference, Release 2.0.0.dev8464

See Also:

vstack
Stack along first axis.
hstack
Stack along second axis.
concatenate
Join arrays.
dsplit
Split array along third axis.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.dstack((a,b))
array([[[1, 2],
[2, 3],
[3, 4]]])

>>> a = np.array([[1],[2],[3]])
>>> b = np.array([[2],[3],[4]])
>>> np.dstack((a,b))
array([[[1, 2]],
[[2, 3]],
[[3, 4]]])

hstack

Stack arrays in sequence horizontally (column wise).

Take a sequence of arrays and stack them horizontally to make a single array. Rebuild arrays divided
by hsplit.

Parameters
tup : sequence of ndarrays
All arrays must have the same shape along all but the second axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

See Also:

vstack
Stack arrays in sequence vertically (row wise).
dstack
Stack arrays in sequence depth wise (along third axis).

854 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

concatenate
Join a sequence of arrays together.
hsplit
Split array along second axis.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.hstack((a,b))
array([1, 2, 3, 2, 3, 4])
>>> a = np.array([[1],[2],[3]])
>>> b = np.array([[2],[3],[4]])
>>> np.hstack((a,b))
array([[1, 2],
[2, 3],
[3, 4]])

hsplit

Split an array into multiple sub-arrays horizontally (column-wise).

Please refer to the split documentation. hsplit is equivalent to split with axis=1, the array is always
split along the second axis regardless of the array dimension.
See Also:

split
Split an array into multiple sub-arrays of equal size.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> x = np.arange(16.0).reshape(4, 4)
>>> x
array([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.],
[ 12., 13., 14., 15.]])
>>> np.hsplit(x, 2)
[array([[ 0., 1.],
[ 4., 5.],
[ 8., 9.],
[ 12., 13.]]),
array([[ 2., 3.],
[ 6., 7.],
[ 10., 11.],
[ 14., 15.]])]
>>> np.hsplit(x, np.array([3, 6]))
[array([[ 0., 1., 2.],

3.20. Masked array operations 855

NumPy Reference, Release 2.0.0.dev8464

[ 4., 5., 6.],

[ 8., 9., 10.],
[ 12., 13., 14.]]),
array([[ 3.],
[ 7.],
[ 11.],
[ 15.]]),
array([], dtype=float64)]

With a higher dimensional array the split is still along the second axis.

>>> x = np.arange(8.0).reshape(2, 2, 2)
>>> x
array([[[ 0., 1.],
[ 2., 3.]],
[[ 4., 5.],
[ 6., 7.]]])
>>> np.hsplit(x, 2)
[array([[[ 0., 1.]],
[[ 4., 5.]]]),
array([[[ 2., 3.]],
[[ 6., 7.]]])]

mr_
Translate slice objects to concatenation along the first axis.
This is the masked array version of lib.index_tricks.RClass.
See Also:
lib.index_tricks.RClass

Examples

>>> np.ma.mr_[np.ma.array([1,2,3]), 0, 0, np.ma.array([4,5,6])]

array([1, 2, 3, 0, 0, 4, 5, 6])

row_stack

Stack arrays in sequence vertically (row wise).

Take a sequence of arrays and stack them vertically to make a single array. Rebuild arrays divided
by vsplit.

Parameters
tup : sequence of ndarrays
Tuple containing arrays to be stacked. The arrays must have the same shape along all
but the first axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

See Also:

856 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

hstack
Stack arrays in sequence horizontally (column wise).
dstack
Stack arrays in sequence depth wise (along third dimension).
concatenate
Join a sequence of arrays together.
vsplit
Split array into a list of multiple sub-arrays vertically.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array([1, 2, 3])

>>> b = np.array([2, 3, 4])
>>> np.vstack((a,b))
array([[1, 2, 3],
[2, 3, 4]])

>>> a = np.array([[1], [2], [3]])

>>> b = np.array([[2], [3], [4]])
>>> np.vstack((a,b))
array([[1],
[2],
[3],
[2],
[3],
[4]])

vstack

Stack arrays in sequence vertically (row wise).

Take a sequence of arrays and stack them vertically to make a single array. Rebuild arrays divided
by vsplit.

Parameters
tup : sequence of ndarrays
Tuple containing arrays to be stacked. The arrays must have the same shape along all
but the first axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

See Also:

hstack
Stack arrays in sequence horizontally (column wise).
dstack
Stack arrays in sequence depth wise (along third dimension).

3.20. Masked array operations 857

NumPy Reference, Release 2.0.0.dev8464

concatenate
Join a sequence of arrays together.
vsplit
Split array into a list of multiple sub-arrays vertically.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array([1, 2, 3])

>>> b = np.array([2, 3, 4])
>>> np.vstack((a,b))
array([[1, 2, 3],
[2, 3, 4]])

>>> a = np.array([[1], [2], [3]])

>>> b = np.array([[2], [3], [4]])
>>> np.vstack((a,b))
array([[1],
[2],
[3],
[2],
[3],
[4]])

Joining arrays

ma.column_stack(tup) Stack 1-D arrays as columns into a 2-D array.

ma.concatenate(arrays[, axis]) Concatenate a sequence of arrays along the given axis.
ma.dstack(tup) Stack arrays in sequence depth wise (along third axis).
ma.hstack(tup) Stack arrays in sequence horizontally (column wise).
ma.vstack(tup) Stack arrays in sequence vertically (row wise).
column_stack

Stack 1-D arrays as columns into a 2-D array.

Take a sequence of 1-D arrays and stack them as columns to make a single 2-D array. 2-D arrays are
stacked as-is, just like with hstack. 1-D arrays are turned into 2-D columns first.

Parameters
tup : sequence of 1-D or 2-D arrays.
Arrays to stack. All of them must have the same first dimension.
Returns
stacked : 2-D array
The array formed by stacking the given arrays.

Notes
The function is applied to both the _data and the _mask, if any.

858 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.column_stack((a,b))
array([[1, 2],
[2, 3],
[3, 4]])

concatenate(arrays, axis=0)
Concatenate a sequence of arrays along the given axis.
Parameters
arrays : sequence of array_like
The arrays must have the same shape, except in the dimension corresponding to axis
(the first, by default).
axis : int, optional
The axis along which the arrays will be joined. Default is 0.
Returns
result : MaskedArray
The concatenated array with any masked entries preserved.
See Also:

numpy.concatenate
Equivalent function in the top-level NumPy module.

Examples

>>> import numpy.ma as ma

>>> a = ma.arange(3)
>>> a[1] = ma.masked
>>> b = ma.arange(2, 5)
>>> a
masked_array(data = [0 -- 2],
mask = [False True False],
fill_value = 999999)
>>> b
masked_array(data = [2 3 4],
mask = False,
fill_value = 999999)
>>> ma.concatenate([a, b])
masked_array(data = [0 -- 2 2 3 4],
mask = [False True False False False False],
fill_value = 999999)

dstack

Stack arrays in sequence depth wise (along third axis).

Takes a sequence of arrays and stack them along the third axis to make a single array. Rebuilds
arrays divided by dsplit. This is a simple way to stack 2D arrays (images) into a single 3D array for
processing.

3.20. Masked array operations 859

NumPy Reference, Release 2.0.0.dev8464

Parameters
tup : sequence of arrays
Arrays to stack. All of them must have the same shape along all but the third axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

See Also:

vstack
Stack along first axis.
hstack
Stack along second axis.
concatenate
Join arrays.
dsplit
Split array along third axis.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.dstack((a,b))
array([[[1, 2],
[2, 3],
[3, 4]]])

>>> a = np.array([[1],[2],[3]])
>>> b = np.array([[2],[3],[4]])
>>> np.dstack((a,b))
array([[[1, 2]],
[[2, 3]],
[[3, 4]]])

hstack

Stack arrays in sequence horizontally (column wise).

Take a sequence of arrays and stack them horizontally to make a single array. Rebuild arrays divided
by hsplit.

Parameters
tup : sequence of ndarrays
All arrays must have the same shape along all but the second axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

860 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:

vstack
Stack arrays in sequence vertically (row wise).
dstack
Stack arrays in sequence depth wise (along third axis).
concatenate
Join a sequence of arrays together.
hsplit
Split array along second axis.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array((1,2,3))
>>> b = np.array((2,3,4))
>>> np.hstack((a,b))
array([1, 2, 3, 2, 3, 4])
>>> a = np.array([[1],[2],[3]])
>>> b = np.array([[2],[3],[4]])
>>> np.hstack((a,b))
array([[1, 2],
[2, 3],
[3, 4]])

vstack

Stack arrays in sequence vertically (row wise).

Take a sequence of arrays and stack them vertically to make a single array. Rebuild arrays divided
by vsplit.

Parameters
tup : sequence of ndarrays
Tuple containing arrays to be stacked. The arrays must have the same shape along all
but the first axis.
Returns
stacked : ndarray
The array formed by stacking the given arrays.

See Also:

hstack
Stack arrays in sequence horizontally (column wise).
dstack
Stack arrays in sequence depth wise (along third dimension).
concatenate
Join a sequence of arrays together.

3.20. Masked array operations 861

NumPy Reference, Release 2.0.0.dev8464

vsplit
Split array into a list of multiple sub-arrays vertically.

Notes
The function is applied to both the _data and the _mask, if any.

Examples

>>> a = np.array([1, 2, 3])

>>> b = np.array([2, 3, 4])
>>> np.vstack((a,b))
array([[1, 2, 3],
[2, 3, 4]])

>>> a = np.array([[1], [2], [3]])

>>> b = np.array([[2], [3], [4]])
>>> np.vstack((a,b))
array([[1],
[2],
[3],
[2],
[3],
[4]])

3.20.5 Operations on masks

Creating a mask

ma.make_mask(m[, copy, shrink, flag, dtype]) Create a boolean mask from an array.
ma.make_mask_none(newshape[, dtype]) Return a boolean mask of the given shape, filled with False.
ma.mask_or(m1, m2[, copy, shrink]) Combine two masks with the logical_or operator.
ma.make_mask_descr(ndtype) Construct a dtype description list from a given dtype.
make_mask(m, copy=False, shrink=True, flag=None, dtype=<type numpy.bool_>)
Create a boolean mask from an array.
Return m as a boolean mask, creating a copy if necessary or requested. The function can accept any sequence
that is convertible to integers, or nomask. Does not require that contents must be 0s and 1s, values of 0 are
interepreted as False, everything else as True.
Parameters
m : array_like
Potential mask.
copy : bool, optional
Whether to return a copy of m (True) or m itself (False).
shrink : bool, optional
Whether to shrink m to nomask if all its values are False.
flag : bool, optional
Deprecated equivalent of shrink.

862 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

dtype : dtype, optional

Data-type of the output mask. By default, the output mask has a dtype of MaskType
(bool). If the dtype is flexible, each field has a boolean dtype.
Returns
result : ndarray
A boolean mask derived from m.

Examples

>>> import numpy.ma as ma

>>> m = [True, False, True, True]
>>> ma.make_mask(m)
array([ True, False, True, True], dtype=bool)
>>> m = [1, 0, 1, 1]
>>> ma.make_mask(m)
array([ True, False, True, True], dtype=bool)
>>> m = [1, 0, 2, -3]
>>> ma.make_mask(m)
array([ True, False, True, True], dtype=bool)

Effect of the shrink parameter.

>>> m = np.zeros(4)
>>> m
array([ 0., 0., 0., 0.])
>>> ma.make_mask(m)
False
>>> ma.make_mask(m, shrink=False)
array([False, False, False, False], dtype=bool)

Using a flexible dtype.

>>> m = [1, 0, 1, 1]
>>> n = [0, 1, 0, 0]
>>> arr = []
>>> for man, mouse in zip(m, n):
... arr.append((man, mouse))
>>> arr
[(1, 0), (0, 1), (1, 0), (1, 0)]
>>> dtype = np.dtype({names:[man, mouse],
formats:[np.int, np.int]})
>>> arr = np.array(arr, dtype=dtype)
>>> arr
array([(1, 0), (0, 1), (1, 0), (1, 0)],
dtype=[(man, <i4), (mouse, <i4)])
>>> ma.make_mask(arr, dtype=dtype)
array([(True, False), (False, True), (True, False), (True, False)],
dtype=[(man, |b1), (mouse, |b1)])

make_mask_none(newshape, dtype=None)
Return a boolean mask of the given shape, filled with False.
This function returns a boolean ndarray with all entries False, that can be used in common mask manipulations.
If a complex dtype is specified, the type of each field is converted to a boolean type.

3.20. Masked array operations 863

NumPy Reference, Release 2.0.0.dev8464

Parameters
newshape : tuple
A tuple indicating the shape of the mask.
dtype: {None, dtype}, optional :
If None, use a MaskType instance. Otherwise, use a new datatype with the same fields
as dtype, converted to boolean types.
Returns
result : ndarray
An ndarray of appropriate shape and dtype, filled with False.
See Also:

make_mask
Create a boolean mask from an array.
make_mask_descr
Construct a dtype description list from a given dtype.

Examples

>>> import numpy.ma as ma

>>> ma.make_mask_none((3,))
array([False, False, False], dtype=bool)

Defining a more complex dtype.

>>> dtype = np.dtype({names:[foo, bar],

formats:[np.float32, np.int]})
>>> dtype
dtype([(foo, <f4), (bar, <i4)])
>>> ma.make_mask_none((3,), dtype=dtype)
array([(False, False), (False, False), (False, False)],
dtype=[(foo, |b1), (bar, |b1)])

mask_or(m1, m2, copy=False, shrink=True)

Combine two masks with the logical_or operator.
The result may be a view on m1 or m2 if the other is nomask (i.e. False).
Parameters
m1, m2 : array_like
Input masks.
copy : bool, optional
If copy is False and one of the inputs is nomask, return a view of the other input mask.
Defaults to False.
shrink : bool, optional
Whether to shrink the output to nomask if all its values are False. Defaults to True.
Returns
mask : output mask
The result masks values that are masked in either m1 or m2.

864 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Raises
ValueError :
If m1 and m2 have different flexible dtypes.

Examples

>>> m1 = np.ma.make_mask([0, 1, 1, 0])

>>> m2 = np.ma.make_mask([1, 0, 0, 0])
>>> np.ma.mask_or(m1, m2)
array([ True, True, True, False], dtype=bool)

make_mask_descr(ndtype)
Construct a dtype description list from a given dtype.
Returns a new dtype object, with the type of all fields in ndtype to a boolean type. Field names are not altered.
Parameters
ndtype : dtype
The dtype to convert.
Returns
result : dtype
A dtype that looks like ndtype, the type of all fields is boolean.

Examples

>>> import numpy.ma as ma

>>> dtype = np.dtype({names:[foo, bar],
formats:[np.float32, np.int]})
>>> dtype
dtype([(foo, <f4), (bar, <i4)])
>>> ma.make_mask_descr(dtype)
dtype([(foo, |b1), (bar, |b1)])
>>> ma.make_mask_descr(np.float32)
<type numpy.bool_>

Accessing a mask

ma.getmask(a) Return the mask of a masked array, or nomask.

ma.getmaskarray(arr) Return the mask of a masked array, or full boolean array of False.
ma.masked_array.mask Mask
getmask(a)
Return the mask of a masked array, or nomask.
Return the mask of a as an ndarray if a is a MaskedArray and the mask is not nomask, else return nomask. To
guarantee a full array of booleans of the same shape as a, use getmaskarray.
Parameters
a : array_like
Input MaskedArray for which the mask is required.
See Also:

3.20. Masked array operations 865

NumPy Reference, Release 2.0.0.dev8464

getdata
Return the data of a masked array as an ndarray.
getmaskarray
Return the mask of a masked array, or full array of False.

Examples

>>> import numpy.ma as ma

>>> a = ma.masked_equal([[1,2],[3,4]], 2)
>>> a
masked_array(data =
[[1 --]
[3 4]],
mask =
[[False True]
[False False]],
fill_value=999999)
>>> ma.getmask(a)
array([[False, True],
[False, False]], dtype=bool)

Equivalently use the MaskedArray mask attribute.

>>> a.mask
array([[False, True],
[False, False]], dtype=bool)

Result when mask == nomask

>>> b = ma.masked_array([[1,2],[3,4]])
>>> b
masked_array(data =
[[1 2]
[3 4]],
mask =
False,
fill_value=999999)
>>> ma.nomask
False
>>> ma.getmask(b) == ma.nomask
True
>>> b.mask == ma.nomask
True

getmaskarray(arr)
Return the mask of a masked array, or full boolean array of False.
Return the mask of arr as an ndarray if arr is a MaskedArray and the mask is not nomask, else return a full
boolean array of False of the same shape as arr.
Parameters
arr : array_like
Input MaskedArray for which the mask is required.
See Also:

866 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

getmask
Return the mask of a masked array, or nomask.
getdata
Return the data of a masked array as an ndarray.

Examples

>>> import numpy.ma as ma

>>> a = ma.masked_equal([[1,2],[3,4]], 2)
>>> a
masked_array(data =
[[1 --]
[3 4]],
mask =
[[False True]
[False False]],
fill_value=999999)
>>> ma.getmaskarray(a)
array([[False, True],
[False, False]], dtype=bool)

Result when mask == nomask

>>> b = ma.masked_array([[1,2],[3,4]])
>>> b
masked_array(data =
[[1 2]
[3 4]],
mask =
False,
fill_value=999999)
>>> >ma.getmaskarray(b)
array([[False, False],
[False, False]], dtype=bool)

mask
Mask

Finding masked data

ma.flatnotmasked_contiguous(a) Find contiguous unmasked data in a masked array along the given axis.
ma.flatnotmasked_edges(a) Find the indices of the first and last unmasked values.
ma.notmasked_contiguous(a[, axis]) Find contiguous unmasked data in a masked array along the given axis.
ma.notmasked_edges(a[, axis]) Find the indices of the first and last unmasked values along an axis.
flatnotmasked_contiguous(a)
Find contiguous unmasked data in a masked array along the given axis.
Parameters
a : narray
The input array.
Returns
slice_list : list
A sorted sequence of slices (start index, end index).

3.20. Masked array operations 867

NumPy Reference, Release 2.0.0.dev8464

See Also:
flatnotmasked_edges, notmasked_contiguous, notmasked_edges

Notes
Only accepts 2-D arrays at most.

Examples

>>> a = np.arange(10)
>>> mask = (a < 3) | (a > 8) | (a == 5)
>>> ma = np.ma.array(a, mask=mask)
>>> np.array(ma[~ma.mask])
array([3, 4, 6, 7, 8])

>>> np.ma.extras.flatnotmasked_contiguous(ma)
[slice(3, 4, None), slice(6, 8, None)]
>>> ma = np.ma.array(a, mask=np.ones_like(a))
>>> print np.ma.extras.flatnotmasked_edges(ma)
None

flatnotmasked_edges(a)
Find the indices of the first and last unmasked values.
Expects a 1-D MaskedArray, returns None if all values are masked.
Parameters
arr : array_like
Input 1-D MaskedArray
Returns
edges : ndarray or None
The indices of first and last non-masked value in the array. Returns None if all values
are masked.
See Also:
flatnotmasked_contiguous, notmasked_contiguous, notmasked_edges

Notes
Only accepts 1-D arrays.

Examples

>>> a = np.arange(10)
>>> mask = (a < 3) | (a > 8) | (a == 5)

>>> ma = np.ma.array(a, mask=m)

>>> np.array(ma[~ma.mask])
array([3, 4, 6, 7, 8])

>>> flatnotmasked_edges(ma)
array([3, 8])

868 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> ma = np.ma.array(a, mask=np.ones_like(a))

>>> print flatnotmasked_edges(ma)
None

notmasked_contiguous(a, axis=None)
Find contiguous unmasked data in a masked array along the given axis.
Parameters
a : array_like
The input array.
axis : int, optional
Axis along which to perform the operation. If None (default), applies to a flattened
version of the array.
Returns
endpoints : list
A list of slices (start and end indexes) of unmasked indexes in the array.
See Also:
flatnotmasked_edges, flatnotmasked_contiguous, notmasked_edges

Notes
Only accepts 2-D arrays at most.

Examples

>>> a = np.arange(9).reshape((3, 3))

>>> mask = np.zeros_like(a)
>>> mask[1:, 1:] = 1

>>> ma = np.ma.array(a, mask=mask)

>>> np.array(ma[~ma.mask])
array([0, 1, 2, 3, 6])

>>> np.ma.extras.notmasked_contiguous(ma)
[slice(0, 3, None), slice(6, 6, None)]

notmasked_edges(a, axis=None)
Find the indices of the first and last unmasked values along an axis.
If all values are masked, return None. Otherwise, return a list of two tuples, corresponding to the indices of the
first and last unmasked values respectively.
Parameters
a : array_like
The input array.
axis : int, optional
Axis along which to perform the operation. If None (default), applies to a flattened
version of the array.
Returns
edges : ndarray or list

3.20. Masked array operations 869

NumPy Reference, Release 2.0.0.dev8464

An array of start and end indexes if there are any masked data in the array. If there are
no masked data in the array, edges is a list of the first and last index.
See Also:
flatnotmasked_contiguous, flatnotmasked_edges, notmasked_contiguous

Examples

>>> a = np.arange(9).reshape((3, 3))

>>> m = np.zeros_like(a)
>>> m[1:, 1:] = 1

>>> ma = np.ma.array(a, mask=m)

>>> np.array(ma[~ma.mask])
array([0, 1, 2, 3, 6])

>>> np.ma.extras.notmasked_edges(ma)
array([0, 6])

Modifying a mask

ma.mask_cols(a[, axis]) Mask columns of a 2D array that contain masked values.

ma.mask_or(m1, m2[, copy, shrink]) Combine two masks with the logical_or operator.
ma.mask_rowcols(a[, axis]) Mask rows and/or columns of a 2D array that contain masked values.
ma.mask_rows(a[, axis]) Mask rows of a 2D array that contain masked values.
ma.harden_mask(self) Force the mask to hard.
ma.soften_mask(self) Force the mask to soft.
ma.MaskedArray.harden_mask() Force the mask to hard.
ma.MaskedArray.soften_mask() Force the mask to soft.
ma.MaskedArray.shrink_mask() Reduce a mask to nomask when possible.
ma.MaskedArray.unshare_mask() Copy the mask and set the sharedmask flag to False.
mask_cols(a, axis=None)
Mask columns of a 2D array that contain masked values.
This function is a shortcut to mask_rowcols with axis equal to 1.
See Also:

mask_rowcols
Mask rows and/or columns of a 2D array.
masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.zeros((3, 3), dtype=np.int)
>>> a[1, 1] = 1
>>> a
array([[0, 0, 0],
[0, 1, 0],
[0, 0, 0]])
>>> a = ma.masked_equal(a, 1)

870 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> a
masked_array(data =
[[0 0 0]
[0 -- 0]
[0 0 0]],
mask =
[[False False False]
[False True False]
[False False False]],
fill_value=999999)
>>> ma.mask_cols(a)
masked_array(data =
[[0 -- 0]
[0 -- 0]
[0 -- 0]],
mask =
[[False True False]
[False True False]
[False True False]],
fill_value=999999)

mask_or(m1, m2, copy=False, shrink=True)

Combine two masks with the logical_or operator.
The result may be a view on m1 or m2 if the other is nomask (i.e. False).
Parameters
m1, m2 : array_like
Input masks.
copy : bool, optional
If copy is False and one of the inputs is nomask, return a view of the other input mask.
Defaults to False.
shrink : bool, optional
Whether to shrink the output to nomask if all its values are False. Defaults to True.
Returns
mask : output mask
The result masks values that are masked in either m1 or m2.
Raises
ValueError :
If m1 and m2 have different flexible dtypes.

Examples

>>> m1 = np.ma.make_mask([0, 1, 1, 0])

>>> m2 = np.ma.make_mask([1, 0, 0, 0])
>>> np.ma.mask_or(m1, m2)
array([ True, True, True, False], dtype=bool)

mask_rowcols(a, axis=None)
Mask rows and/or columns of a 2D array that contain masked values.
Mask whole rows and/or columns of a 2D array that contain masked values. The masking behavior is selected
using the axis parameter.

3.20. Masked array operations 871

NumPy Reference, Release 2.0.0.dev8464

If axis is None, rows and columns are masked.

If axis is 0, only rows are masked.
If axis is 1 or -1, only columns are masked.

Parameters
a : array_like, MaskedArray
The array to mask. If not a MaskedArray instance (or if no array elements are masked).
The result is a MaskedArray with mask set to nomask (False). Must be a 2D array.
axis : int, optional
Axis along which to perform the operation. If None, applies to a flattened version of the
array.
Returns
a : MaskedArray
A modified version of the input array, masked depending on the value of the axis pa-
rameter.
Raises
NotImplementedError :
If input array a is not 2D.

See Also:

mask_rows
Mask rows of a 2D array that contain masked values.
mask_cols
Mask cols of a 2D array that contain masked values.
masked_where
Mask where a condition is met.

Notes
The input arrays mask is modified by this function.

Examples

>>> import numpy.ma as ma

>>> a = np.zeros((3, 3), dtype=np.int)
>>> a[1, 1] = 1
>>> a
array([[0, 0, 0],
[0, 1, 0],
[0, 0, 0]])
>>> a = ma.masked_equal(a, 1)
>>> a
masked_array(data =
[[0 0 0]
[0 -- 0]
[0 0 0]],
mask =
[[False False False]
[False True False]

872 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

[False False False]],

fill_value=999999)
>>> ma.mask_rowcols(a)
masked_array(data =
[[0 -- 0]
[-- -- --]
[0 -- 0]],
mask =
[[False True False]
[ True True True]
[False True False]],
fill_value=999999)

mask_rows(a, axis=None)
Mask rows of a 2D array that contain masked values.
This function is a shortcut to mask_rowcols with axis equal to 0.
See Also:

mask_rowcols
Mask rows and/or columns of a 2D array.
masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.zeros((3, 3), dtype=np.int)
>>> a[1, 1] = 1
>>> a
array([[0, 0, 0],
[0, 1, 0],
[0, 0, 0]])
>>> a = ma.masked_equal(a, 1)
>>> a
masked_array(data =
[[0 0 0]
[0 -- 0]
[0 0 0]],
mask =
[[False False False]
[False True False]
[False False False]],
fill_value=999999)
>>> ma.mask_rows(a)
masked_array(data =
[[0 0 0]
[-- -- --]
[0 0 0]],
mask =
[[False False False]
[ True True True]
[False False False]],
fill_value=999999)

harden_mask

3.20. Masked array operations 873

NumPy Reference, Release 2.0.0.dev8464

Force the mask to hard.

Whether the mask of a masked array is hard or soft is determined by its hardmask property. harden_mask sets
hardmask to True.
See Also:
hardmask
soften_mask
Force the mask to soft.
Whether the mask of a masked array is hard or soft is determined by its hardmask property. soften_mask sets
hardmask to False.
See Also:
hardmask
harden_mask()
Force the mask to hard.
Whether the mask of a masked array is hard or soft is determined by its hardmask property. harden_mask sets
hardmask to True.
See Also:
hardmask
soften_mask()
Force the mask to soft.
Whether the mask of a masked array is hard or soft is determined by its hardmask property. soften_mask sets
hardmask to False.
See Also:
hardmask
shrink_mask()
Reduce a mask to nomask when possible.
Parameters
None :
Returns
None :

Examples

>>> x = np.ma.array([[1,2 ], [3, 4]], mask=[0]*4)

>>> x.mask
array([[False, False],
[False, False]], dtype=bool)
>>> x.shrink_mask()
>>> x.mask
False

unshare_mask()
Copy the mask and set the sharedmask flag to False.
Whether the mask is shared between masked arrays can be seen from the sharedmask property. unshare_mask
ensures the mask is not shared. A copy of the mask is only made if it was shared.
See Also:

874 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

sharedmask

3.20.6 Conversion operations

> to a masked array

ma.asarray(a[, dtype, order]) Convert the input to a masked array of the given data-type.
ma.asanyarray(a[, dtype]) Convert the input to a masked array, conserving subclasses.
ma.fix_invalid(a[, mask, copy, fill_value]) Return input with invalid data masked and replaced by a fill
value.
ma.masked_equal(x, value[, copy]) Mask an array where equal to a given value.
ma.masked_greater(x, value[, copy]) Mask an array where greater than a given value.
ma.masked_greater_equal(x, value[, Mask an array where greater than or equal to a given value.
copy])
ma.masked_inside(x, v1, v2[, copy]) Mask an array inside a given interval.
ma.masked_invalid(a[, copy]) Mask an array where invalid values occur (NaNs or infs).
ma.masked_less(x, value[, copy]) Mask an array where less than a given value.
ma.masked_less_equal(x, value[, copy]) Mask an array where less than or equal to a given value.
ma.masked_not_equal(x, value[, copy]) Mask an array where not equal to a given value.
ma.masked_object(x, value[, copy, shrink]) Mask the array x where the data are exactly equal to value.
ma.masked_outside(x, v1, v2[, copy]) Mask an array outside a given interval.
ma.masked_values(x, value[, rtol, atol, ...]) Mask using floating point equality.
ma.masked_where(condition, a[, copy]) Mask an array where a condition is met.
asarray(a, dtype=None, order=None)
Convert the input to a masked array of the given data-type.
No copy is performed if the input is already an ndarray. If a is a subclass of MaskedArray, a base class
MaskedArray is returned.
Parameters
a : array_like
Input data, in any form that can be converted to a masked array. This includes lists, lists
of tuples, tuples, tuples of tuples, tuples of lists, ndarrays and masked arrays.
dtype : dtype, optional
By default, the data-type is inferred from the input data.
order : {C, F}, optional
Whether to use row-major (C) or column-major (FORTRAN) memory representa-
tion. Default is C.
Returns
out : MaskedArray
Masked array interpretation of a.
See Also:

asanyarray
Similar to asarray, but conserves subclasses.

3.20. Masked array operations 875

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.arange(10.).reshape(2, 5)
>>> x
array([[ 0., 1., 2., 3., 4.],
[ 5., 6., 7., 8., 9.]])
>>> np.ma.asarray(x)
masked_array(data =
[[ 0. 1. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]],
mask =
False,
fill_value = 1e+20)
>>> type(np.ma.asarray(x))
<class numpy.ma.core.MaskedArray>

asanyarray(a, dtype=None)
Convert the input to a masked array, conserving subclasses.
If a is a subclass of MaskedArray, its class is conserved. No copy is performed if the input is already an ndarray.
Parameters
a : array_like
Input data, in any form that can be converted to an array.
dtype : dtype, optional
By default, the data-type is inferred from the input data.
order : {C, F}, optional
Whether to use row-major (C) or column-major (FORTRAN) memory representa-
tion. Default is C.
Returns
out : MaskedArray
MaskedArray interpretation of a.
See Also:

asarray
Similar to asanyarray, but does not conserve subclass.

Examples

>>> x = np.arange(10.).reshape(2, 5)
>>> x
array([[ 0., 1., 2., 3., 4.],
[ 5., 6., 7., 8., 9.]])
>>> np.ma.asanyarray(x)
masked_array(data =
[[ 0. 1. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]],
mask =
False,
fill_value = 1e+20)
>>> type(np.ma.asanyarray(x))
<class numpy.ma.core.MaskedArray>

876 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

fix_invalid(a, mask=False, copy=True, fill_value=None)

Return input with invalid data masked and replaced by a fill value.
Invalid data means values of nan, inf, etc.
Parameters
a : array_like
Input array, a (subclass of) ndarray.
copy : bool, optional
Whether to use a copy of a (True) or to fix a in place (False). Default is True.
fill_value : scalar, optional
Value used for fixing invalid data. Default is None, in which case the a.fill_value
is used.
Returns
b : MaskedArray
The input array with invalid entries fixed.

Notes
A copy is performed by default.

Examples

>>> x = np.ma.array([1., -1, np.nan, np.inf], mask=[1] + [0]*3)

>>> x
masked_array(data = [-- -1.0 nan inf],
mask = [ True False False False],
fill_value = 1e+20)
>>> np.ma.fix_invalid(x)
masked_array(data = [-- -1.0 -- --],
mask = [ True False True True],
fill_value = 1e+20)

>>> fixed = np.ma.fix_invalid(x)

>>> fixed.data
array([ 1.00000000e+00, -1.00000000e+00, 1.00000000e+20,
1.00000000e+20])
>>> x.data
array([ 1., -1., NaN, Inf])

masked_equal(x, value, copy=True)

Mask an array where equal to a given value.
This function is a shortcut to masked_where, with condition = (x == value). For floating point arrays, consider
using masked_values(x, value).
See Also:

masked_where
Mask where a condition is met.
masked_values
Mask using floating point equality.

3.20. Masked array operations 877

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_equal(a, 2)
masked_array(data = [0 1 -- 3],
mask = [False False True False],
fill_value=999999)

masked_greater(x, value, copy=True)

Mask an array where greater than a given value.
This function is a shortcut to masked_where, with condition = (x > value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_greater(a, 2)
masked_array(data = [0 1 2 --],
mask = [False False False True],
fill_value=999999)

masked_greater_equal(x, value, copy=True)

Mask an array where greater than or equal to a given value.
This function is a shortcut to masked_where, with condition = (x >= value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_greater_equal(a, 2)
masked_array(data = [0 1 -- --],
mask = [False False True True],
fill_value=999999)

masked_inside(x, v1, v2, copy=True)

Mask an array inside a given interval.
Shortcut to masked_where, where condition is True for x inside the interval [v1,v2] (v1 <= x <= v2). The
boundaries v1 and v2 can be given in either order.

878 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:

masked_where
Mask where a condition is met.

Notes
The array x is prefilled with its filling value.

Examples

>>> import numpy.ma as ma

>>> x = [0.31, 1.2, 0.01, 0.2, -0.4, -1.1]
>>> ma.masked_inside(x, -0.3, 0.3)
masked_array(data = [0.31 1.2 -- -- -0.4 -1.1],
mask = [False False True True False False],
fill_value=1e+20)

The order of v1 and v2 doesnt matter.

>>> ma.masked_inside(x, 0.3, -0.3)

masked_array(data = [0.31 1.2 -- -- -0.4 -1.1],
mask = [False False True True False False],
fill_value=1e+20)

masked_invalid(a, copy=True)
Mask an array where invalid values occur (NaNs or infs).
This function is a shortcut to masked_where, with condition = ~(np.isfinite(a)). Any pre-existing mask is
conserved. Only applies to arrays with a dtype where NaNs or infs make sense (i.e. floating point types), but
accepts any array_like object.
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(5, dtype=np.float)
>>> a[2] = np.NaN
>>> a[3] = np.PINF
>>> a
array([ 0., 1., NaN, Inf, 4.])
>>> ma.masked_invalid(a)
masked_array(data = [0.0 1.0 -- -- 4.0],
mask = [False False True True False],
fill_value=1e+20)

masked_less(x, value, copy=True)

Mask an array where less than a given value.
This function is a shortcut to masked_where, with condition = (x < value).
See Also:

3.20. Masked array operations 879

NumPy Reference, Release 2.0.0.dev8464

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_less(a, 2)
masked_array(data = [-- -- 2 3],
mask = [ True True False False],
fill_value=999999)

masked_less_equal(x, value, copy=True)

Mask an array where less than or equal to a given value.
This function is a shortcut to masked_where, with condition = (x <= value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_less_equal(a, 2)
masked_array(data = [-- -- -- 3],
mask = [ True True True False],
fill_value=999999)

masked_not_equal(x, value, copy=True)

Mask an array where not equal to a given value.
This function is a shortcut to masked_where, with condition = (x != value).
See Also:

masked_where
Mask where a condition is met.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_not_equal(a, 2)
masked_array(data = [-- -- 2 --],
mask = [ True True False True],
fill_value=999999)

masked_object(x, value, copy=True, shrink=True)

Mask the array x where the data are exactly equal to value.

880 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

This function is similar to masked_values, but only suitable for object arrays: for floating point, use
masked_values instead.
Parameters
x : array_like
Array to mask
value : object
Comparison value
copy : {True, False}, optional
Whether to return a copy of x.
shrink : {True, False}, optional
Whether to collapse a mask full of False to nomask
Returns
result : MaskedArray
The result of masking x where equal to value.
See Also:

masked_where
Mask where a condition is met.
masked_equal
Mask where equal to a given value (integers).
masked_values
Mask using floating point equality.

Examples

>>> import numpy.ma as ma

>>> food = np.array([green_eggs, ham], dtype=object)
>>> # dont eat spoiled food
>>> eat = ma.masked_object(food, green_eggs)
>>> print eat
[-- ham]
>>> # plain ol ham is boring
>>> fresh_food = np.array([cheese, ham, pineapple], dtype=object)
>>> eat = ma.masked_object(fresh_food, green_eggs)
>>> print eat
[cheese ham pineapple]

Note that mask is set to nomask if possible.

>>> eat
masked_array(data = [cheese ham pineapple],
mask = False,
fill_value=?)

masked_outside(x, v1, v2, copy=True)

Mask an array outside a given interval.
Shortcut to masked_where, where condition is True for x outside the interval [v1,v2] (x < v1)|(x > v2). The
boundaries v1 and v2 can be given in either order.

3.20. Masked array operations 881

NumPy Reference, Release 2.0.0.dev8464

See Also:

masked_where
Mask where a condition is met.

Notes
The array x is prefilled with its filling value.

Examples

>>> import numpy.ma as ma

>>> x = [0.31, 1.2, 0.01, 0.2, -0.4, -1.1]
>>> ma.masked_outside(x, -0.3, 0.3)
masked_array(data = [-- -- 0.01 0.2 -- --],
mask = [ True True False False True True],
fill_value=1e+20)

The order of v1 and v2 doesnt matter.

>>> ma.masked_outside(x, 0.3, -0.3)

masked_array(data = [-- -- 0.01 0.2 -- --],
mask = [ True True False False True True],
fill_value=1e+20)

masked_values(x, value, rtol=1.0000000000000001e-05, atol=1e-08, copy=True, shrink=True)

Mask using floating point equality.
Return a MaskedArray, masked where the data in array x are approximately equal to value, i.e. where the
following condition is True
(abs(x - value) <= atol+rtol*abs(value))
The fill_value is set to value and the mask is set to nomask if possible. For integers, consider using
masked_equal.
Parameters
x : array_like
Array to mask.
value : float
Masking value.
rtol : float, optional
Tolerance parameter.
atol : float, optional
Tolerance parameter (1e-8).
copy : bool, optional
Whether to return a copy of x.
shrink : bool, optional
Whether to collapse a mask full of False to nomask.
Returns
result : MaskedArray

882 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

The result of masking x where approximately equal to value.

See Also:

masked_where
Mask where a condition is met.
masked_equal
Mask where equal to a given value (integers).

Examples

>>> import numpy.ma as ma

>>> x = np.array([1, 1.1, 2, 1.1, 3])
>>> ma.masked_values(x, 1.1)
masked_array(data = [1.0 -- 2.0 -- 3.0],
mask = [False True False True False],
fill_value=1.1)

Note that mask is set to nomask if possible.

>>> ma.masked_values(x, 1.5)

masked_array(data = [ 1. 1.1 2. 1.1 3. ],
mask = False,
fill_value=1.5)

For integers, the fill value will be different in general to the result of masked_equal.

>>> x = np.arange(5)
>>> x
array([0, 1, 2, 3, 4])
>>> ma.masked_values(x, 2)
masked_array(data = [0 1 -- 3 4],
mask = [False False True False False],
fill_value=2)
>>> ma.masked_equal(x, 2)
masked_array(data = [0 1 -- 3 4],
mask = [False False True False False],
fill_value=999999)

masked_where(condition, a, copy=True)
Mask an array where a condition is met.
Return a as an array masked where condition is True. Any masked values of a or condition are also masked in
the output.
Parameters
condition : array_like
Masking condition. When condition tests floating point values for equality, consider
using masked_values instead.
a : array_like
Array to mask.
copy : bool
If True (default) make a copy of a in the result. If False modify a in place and return a
view.

3.20. Masked array operations 883

NumPy Reference, Release 2.0.0.dev8464

Returns
result : MaskedArray
The result of masking a where condition is True.
See Also:

masked_values
Mask using floating point equality.
masked_equal
Mask where equal to a given value.
masked_not_equal
Mask where not equal to a given value.
masked_less_equal
Mask where less than or equal to a given value.
masked_greater_equal
Mask where greater than or equal to a given value.
masked_less
Mask where less than a given value.
masked_greater
Mask where greater than a given value.
masked_inside
Mask inside a given interval.
masked_outside
Mask outside a given interval.
masked_invalid
Mask invalid values (NaNs or infs).

Examples

>>> import numpy.ma as ma

>>> a = np.arange(4)
>>> a
array([0, 1, 2, 3])
>>> ma.masked_where(a <= 2, a)
masked_array(data = [-- -- -- 3],
mask = [ True True True False],
fill_value=999999)

Mask array b conditional on a.

>>> b = [a, b, c, d]
>>> ma.masked_where(a == 2, b)
masked_array(data = [a b -- d],
mask = [False False True False],
fill_value=N/A)

Effect of the copy argument.

884 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> c = ma.masked_where(a <= 2, a)

>>> c
masked_array(data = [-- -- -- 3],
mask = [ True True True False],
fill_value=999999)
>>> c[0] = 99
>>> c
masked_array(data = [99 -- -- 3],
mask = [False True True False],
fill_value=999999)
>>> a
array([0, 1, 2, 3])
>>> c = ma.masked_where(a <= 2, a, copy=False)
>>> c[0] = 99
>>> c
masked_array(data = [99 -- -- 3],
mask = [False True True False],
fill_value=999999)
>>> a
array([99, 1, 2, 3])

When condition or a contain masked values.

>>> a = np.arange(4)
>>> a = ma.masked_where(a == 2, a)
>>> a
masked_array(data = [0 1 -- 3],
mask = [False False True False],
fill_value=999999)
>>> b = np.arange(4)
>>> b = ma.masked_where(b == 0, b)
>>> b
masked_array(data = [-- 1 2 3],
mask = [ True False False False],
fill_value=999999)
>>> ma.masked_where(a == 3, b)
masked_array(data = [-- 1 -- --],
mask = [ True False True True],
fill_value=999999)

> to a ndarray

ma.compress_cols(a) Suppress whole columns of a 2-D array that contain masked values.
ma.compress_rowcols(x[, axis]) Suppress the rows and/or columns of a 2-D array that contain
ma.compress_rows(a) Suppress whole rows of a 2-D array that contain masked values.
ma.compressed(x) Return all the non-masked data as a 1-D array.
ma.filled(a[, fill_value]) Return input as an array with masked data replaced by a fill value.
ma.MaskedArray.compressed() Return all the non-masked data as a 1-D array.
ma.MaskedArray.filled([fill_value]) Return a copy of self, with masked values filled with a given value.
compress_cols(a)
Suppress whole columns of a 2-D array that contain masked values.
This is equivalent to np.ma.extras.compress_rowcols(a, 1), see extras.compress_rowcols for de-
tails.
See Also:

3.20. Masked array operations 885

NumPy Reference, Release 2.0.0.dev8464

extras.compress_rowcols
compress_rowcols(x, axis=None)
Suppress the rows and/or columns of a 2-D array that contain masked values.
The suppression behavior is selected with the axis parameter.
If axis is None, both rows and columns are suppressed.
If axis is 0, only rows are suppressed.
If axis is 1 or -1, only columns are suppressed.

Parameters
axis : int, optional
Axis along which to perform the operation. Default is None.
Returns
compressed_array : ndarray
The compressed array.

Examples

>>> x = np.ma.array(np.arange(9).reshape(3, 3), mask=[[1, 0, 0],

... [1, 0, 0],
... [0, 0, 0]])
>>> x
masked_array(data =
[[-- 1 2]
[-- 4 5]
[6 7 8]],
mask =
[[ True False False]
[ True False False]
[False False False]],
fill_value = 999999)

>>> np.ma.extras.compress_rowcols(x)
array([[7, 8]])
>>> np.ma.extras.compress_rowcols(x, 0)
array([[6, 7, 8]])
>>> np.ma.extras.compress_rowcols(x, 1)
array([[1, 2],
[4, 5],
[7, 8]])

compress_rows(a)
Suppress whole rows of a 2-D array that contain masked values.
This is equivalent to np.ma.extras.compress_rowcols(a, 0), see extras.compress_rowcols for de-
tails.
See Also:
extras.compress_rowcols
compressed(x)
Return all the non-masked data as a 1-D array.

886 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

This function is equivalent to calling the compressed method of a MaskedArray, see MaskedArray.compressed
for details.
See Also:

MaskedArray.compressed
Equivalent method.

filled(a, fill_value=None)
Return input as an array with masked data replaced by a fill value.
If a is not a MaskedArray, a itself is returned. If a is a MaskedArray and fill_value is None, fill_value is set to
a.fill_value.
Parameters
a : MaskedArray or array_like
An input object.
fill_value : scalar, optional
Filling value. Default is None.
Returns
a : ndarray
The filled array.
See Also:
compressed

Examples

>>> x = np.ma.array(np.arange(9).reshape(3, 3), mask=[[1, 0, 0],

... [1, 0, 0],
... [0, 0, 0]])
>>> x.filled()
array([[999999, 1, 2],
[999999, 4, 5],
[ 6, 7, 8]])

compressed()
Return all the non-masked data as a 1-D array.
Returns
data : ndarray
A new ndarray holding the non-masked data is returned.

Notes
The result is not a MaskedArray!

Examples

>>> x = np.ma.array(np.arange(5), mask=[0]*2 + [1]*3)

>>> x.compressed()
array([0, 1])
>>> type(x.compressed())
<type numpy.ndarray>

3.20. Masked array operations 887

NumPy Reference, Release 2.0.0.dev8464

filled(fill_value=None)
Return a copy of self, with masked values filled with a given value.
Parameters
fill_value : scalar, optional
The value to use for invalid entries (None by default). If None, the fill_value attribute
of the array is used instead.

Notes
The result is not a MaskedArray!

Examples

>>> x = np.ma.array([1,2,3,4,5], mask=[0,0,1,0,1], fill_value=-999)

>>> x.filled()
array([1, 2, -999, 4, -999])
>>> type(x.filled())
<type numpy.ndarray>

Subclassing is preserved. This means that if the data part of the masked array is a matrix, filled returns a matrix:

>>> x = np.ma.array(np.matrix([[1, 2], [3, 4]]), mask=[[0, 1], [1, 0]])

>>> x.filled()
matrix([[ 1, 999999],
[999999, 4]])

> to another object

ma.MaskedArray.tofile(fid[, sep, Save a masked array to a file in binary format.

format])
ma.MaskedArray.tolist([fill_value]) Return the data portion of the masked array as a hierarchical
Python list.
ma.MaskedArray.torecords() Transforms a masked array into a flexible-type array.
ma.MaskedArray.tostring([fill_value, Return the array data as a string containing the raw bytes in the
order]) array.
tofile(fid, sep=, format=%s)
Save a masked array to a file in binary format.

Warning: This function is not implemented yet.

Raises
NotImplementedError :
When tofile is called.
tolist(fill_value=None)
Return the data portion of the masked array as a hierarchical Python list.
Data items are converted to the nearest compatible Python type. Masked values are converted to fill_value. If
fill_value is None, the corresponding entries in the output list will be None.
Parameters
fill_value : scalar, optional
The value to use for invalid entries. Default is None.

888 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Returns
result : list
The Python list representation of the masked array.

Examples

>>> x = np.ma.array([[1,2,3], [4,5,6], [7,8,9]], mask=[0] + [1,0]*4)

>>> x.tolist()
[[1, None, 3], [None, 5, None], [7, None, 9]]
>>> x.tolist(-999)
[[1, -999, 3], [-999, 5, -999], [7, -999, 9]]

torecords()
Transforms a masked array into a flexible-type array.
The flexible type array that is returned will have two fields:
the _data field stores the _data part of the array.
the _mask field stores the _mask part of the array.

Parameters
None :
Returns
record : ndarray
A new flexible-type ndarray with two fields: the first element containing a value, the
second element containing the corresponding mask boolean. The returned record shape
matches self.shape.

Notes
A side-effect of transforming a masked array into a flexible ndarray is that meta information (fill_value,
...) will be lost.

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.toflex()
[[(1, False) (2, True) (3, False)]
[(4, True) (5, False) (6, True)]
[(7, False) (8, True) (9, False)]]

tostring(fill_value=None, order=C)
Return the array data as a string containing the raw bytes in the array.
The array is filled with a fill value before the string conversion.
Parameters
fill_value : scalar, optional
Value used to fill in the masked values. Deafult is None, in which case MaskedAr-
ray.fill_value is used.
order : {C,F,A}, optional

3.20. Masked array operations 889

NumPy Reference, Release 2.0.0.dev8464

Order of the data item in the copy. Default is C.

C C order (row major).
F Fortran order (column major).
A Any, current order of array.
None Same as A.
See Also:
ndarray.tostring, tolist, tofile

Notes
As for ndarray.tostring, information about the shape, dtype, etc., but also about fill_value, will be lost.

Examples

>>> x = np.ma.array(np.array([[1, 2], [3, 4]]), mask=[[0, 1], [1, 0]])

>>> x.tostring()
\x01\x00\x00\x00?B\x0f\x00?B\x0f\x00\x04\x00\x00\x00

Pickling and unpickling

ma.dump(a, F) Pickle a masked array to a file.

ma.dumps(a) Return a string corresponding to the pickling of a masked array.
ma.load(F) Wrapper around cPickle.load which accepts either a file-like object
ma.loads(strg) Load a pickle from the current string.
dump(a, F)
Pickle a masked array to a file.
This is a wrapper around cPickle.dump.
Parameters
a : MaskedArray
The array to be pickled.
F : str or file-like object
The file to pickle a to. If a string, the full path to the file.
dumps(a)
Return a string corresponding to the pickling of a masked array.
This is a wrapper around cPickle.dumps.
Parameters
a : MaskedArray
The array for which the string representation of the pickle is returned.
load(F)
Wrapper around cPickle.load which accepts either a file-like object or a filename.
Parameters
F : str or file
The file or file name to load.
See Also:

890 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

dump
Pickle an array

Notes
This is different from numpy.load, which does not use cPickle but loads the NumPy binary .npy format.
loads(strg)
Load a pickle from the current string.
The result of cPickle.loads(strg) is returned.
Parameters
strg : str
The string to load.
See Also:

dumps
Return a string corresponding to the pickling of a masked array.

Filling a masked array

ma.common_fill_value(a, b) Return the common filling value of two masked arrays, if any.
ma.default_fill_value(obj) Return the default fill value for the argument object.
ma.maximum_fill_value(obj) Return the minimum value that can be represented by the dtype of
an object.
ma.maximum_fill_value(obj) Return the minimum value that can be represented by the dtype of
an object.
ma.set_fill_value(a, fill_value) Set the filling value of a, if a is a masked array.
ma.MaskedArray.get_fill_value() Return the filling value of the masked array.
Set the filling value of the masked array.
ma.MaskedArray.set_fill_value([value])
ma.MaskedArray.fill_value Filling value.
common_fill_value(a, b)
Return the common filling value of two masked arrays, if any.
If a.fill_value == b.fill_value, return the fill value, otherwise return None.
Parameters
a, b : MaskedArray
The masked arrays for which to compare fill values.
Returns
fill_value : scalar or None
The common fill value, or None.

Examples

>>> x = np.ma.array([0, 1.], fill_value=3)

>>> y = np.ma.array([0, 1.], fill_value=3)
>>> np.ma.common_fill_value(x, y)
3.0

default_fill_value(obj)
Return the default fill value for the argument object.

3.20. Masked array operations 891

NumPy Reference, Release 2.0.0.dev8464

The default filling value depends on the datatype of the input array or the type of the input scalar:
datatype default
bool True
int 999999
float 1.e20
complex 1.e20+0j
object ?
string N/A

Parameters
obj : ndarray, dtype or scalar
The array data-type or scalar for which the default fill value is returned.
Returns
fill_value : scalar
The default fill value.

Examples

>>> np.ma.default_fill_value(1)
999999
>>> np.ma.default_fill_value(np.array([1.1, 2., np.pi]))
1e+20
>>> np.ma.default_fill_value(np.dtype(complex))
(1e+20+0j)

maximum_fill_value(obj)
Return the minimum value that can be represented by the dtype of an object.
This function is useful for calculating a fill value suitable for taking the maximum of an array with a given dtype.
Parameters
obj : {ndarray, dtype}
An object that can be queried for its numeric type.
Returns
val : scalar
The minimum representable value.
Raises
TypeError :
If obj isnt a suitable numeric type.
See Also:

minimum_fill_value
The inverse function.
set_fill_value
Set the filling value of a masked array.
MaskedArray.fill_value
Return current fill value.

892 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> import numpy.ma as ma

>>> a = np.int8()
>>> ma.maximum_fill_value(a)
-128
>>> a = np.int32()
>>> ma.maximum_fill_value(a)
-2147483648

An array of numeric data can also be passed.

>>> a = np.array([1, 2, 3], dtype=np.int8)

>>> ma.maximum_fill_value(a)
-128
>>> a = np.array([1, 2, 3], dtype=np.float32)
>>> ma.maximum_fill_value(a)
-inf

maximum_fill_value(obj)
Return the minimum value that can be represented by the dtype of an object.
This function is useful for calculating a fill value suitable for taking the maximum of an array with a given dtype.
Parameters
obj : {ndarray, dtype}
An object that can be queried for its numeric type.
Returns
val : scalar
The minimum representable value.
Raises
TypeError :
If obj isnt a suitable numeric type.
See Also:

minimum_fill_value
The inverse function.
set_fill_value
Set the filling value of a masked array.
MaskedArray.fill_value
Return current fill value.

Examples

>>> import numpy.ma as ma

>>> a = np.int8()
>>> ma.maximum_fill_value(a)
-128
>>> a = np.int32()
>>> ma.maximum_fill_value(a)
-2147483648

An array of numeric data can also be passed.

3.20. Masked array operations 893

NumPy Reference, Release 2.0.0.dev8464

>>> a = np.array([1, 2, 3], dtype=np.int8)

>>> ma.maximum_fill_value(a)
-128
>>> a = np.array([1, 2, 3], dtype=np.float32)
>>> ma.maximum_fill_value(a)
-inf

set_fill_value(a, fill_value)
Set the filling value of a, if a is a masked array.
This function changes the fill value of the masked array a in place. If a is not a masked array, the function
returns silently, without doing anything.
Parameters
a : array_like
Input array.
fill_value : dtype
Filling value. A consistency test is performed to make sure the value is compatible with
the dtype of a.
Returns
None :
Nothing returned by this function.
See Also:

maximum_fill_value
Return the default fill value for a dtype.
MaskedArray.fill_value
Return current fill value.
MaskedArray.set_fill_value
Equivalent method.

Examples

>>> import numpy.ma as ma

>>> a = np.arange(5)
>>> a
array([0, 1, 2, 3, 4])
>>> a = ma.masked_where(a < 3, a)
>>> a
masked_array(data = [-- -- -- 3 4],
mask = [ True True True False False],
fill_value=999999)
>>> ma.set_fill_value(a, -999)
>>> a
masked_array(data = [-- -- -- 3 4],
mask = [ True True True False False],
fill_value=-999)

Nothing happens if a is not a masked array.

894 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> a = range(5)
>>> a
[0, 1, 2, 3, 4]
>>> ma.set_fill_value(a, 100)
>>> a
[0, 1, 2, 3, 4]
>>> a = np.arange(5)
>>> a
array([0, 1, 2, 3, 4])
>>> ma.set_fill_value(a, 100)
>>> a
array([0, 1, 2, 3, 4])

get_fill_value()
Return the filling value of the masked array.
Returns
fill_value : scalar
The filling value.

Examples

>>> for dt in [np.int32, np.int64, np.float64, np.complex128]:

... np.ma.array([0, 1], dtype=dt).get_fill_value()
...
999999
999999
1e+20
(1e+20+0j)

>>> x = np.ma.array([0, 1.], fill_value=-np.inf)

>>> x.get_fill_value()
-inf

set_fill_value(value=None)
Set the filling value of the masked array.
Parameters
value : scalar, optional
The new filling value. Default is None, in which case a default based on the data type
is used.
See Also:

ma.set_fill_value
Equivalent function.

Examples

>>> x = np.ma.array([0, 1.], fill_value=-np.inf)

>>> x.fill_value
-inf
>>> x.set_fill_value(np.pi)
>>> x.fill_value
3.1415926535897931

3.20. Masked array operations 895

NumPy Reference, Release 2.0.0.dev8464

Reset to default:

>>> x.set_fill_value()
>>> x.fill_value
1e+20

fill_value
Filling value.

3.20.7 Masked arrays arithmetics

Arithmetics

ma.anom(self[, axis, dtype]) Compute the anomalies (deviations from the arithmetic mean) along
the given axis.
ma.anomalies(self[, axis, dtype]) Compute the anomalies (deviations from the arithmetic mean) along
the given axis.
ma.average(a[, axis, weights, Return the weighted average of array over the given axis.
returned])
ma.conjugate() Return the complex conjugate, element-wise.
ma.corrcoef(x[, y, rowvar, bias, Return correlation coefficients of the input array.
allow_masked])
ma.cov(x[, y, rowvar, bias, Estimate the covariance matrix.
allow_masked])
ma.cumsum(self[, axis, dtype, out]) Return the cumulative sum of the elements along the given axis.
ma.cumprod(self[, axis, dtype, out]) Return the cumulative product of the elements along the given axis.
ma.mean(self[, axis, dtype, out]) Returns the average of the array elements.
ma.median(a[, axis, out, Compute the median along the specified axis.
overwrite_input])
ma.power(a, b[, third]) Returns element-wise base array raised to power from second array.
ma.prod(self[, axis, dtype, out]) Return the product of the array elements over the given axis.
ma.std(self[, axis, dtype, out, ddof]) Compute the standard deviation along the specified axis.
ma.sum(self[, axis, dtype, out]) Return the sum of the array elements over the given axis.
ma.var(self[, axis, dtype, out, ddof]) Compute the variance along the specified axis.
ma.MaskedArray.anom([axis, dtype]) Compute the anomalies (deviations from the arithmetic mean) along
the given axis.
ma.MaskedArray.cumprod([axis, Return the cumulative product of the elements along the given axis.
dtype, out])
ma.MaskedArray.cumsum([axis, Return the cumulative sum of the elements along the given axis.
dtype, out])
ma.MaskedArray.mean([axis, dtype, Returns the average of the array elements.
out])
ma.MaskedArray.prod([axis, dtype, Return the product of the array elements over the given axis.
out])
ma.MaskedArray.std([axis, dtype, Compute the standard deviation along the specified axis.
out, ddof])
ma.MaskedArray.sum([axis, dtype, Return the sum of the array elements over the given axis.
out])
ma.MaskedArray.var([axis, dtype, Compute the variance along the specified axis.
out, ddof])

896 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

anom
Compute the anomalies (deviations from the arithmetic mean) along the given axis.
Returns an array of anomalies, with the same shape as the input and where the arithmetic mean is computed
along the given axis.
Parameters
axis : int, optional
Axis over which the anomalies are taken. The default is to use the mean of the flattened
array as reference.
dtype : dtype, optional
Type to use in computing the variance. For arrays of integer type
the default is float32; for arrays of float types it is the same as the array type.
See Also:

mean
Compute the mean of the array.

Examples

>>> a = np.ma.array([1,2,3])
>>> a.anom()
masked_array(data = [-1. 0. 1.],
mask = False,
fill_value = 1e+20)

anomalies
Compute the anomalies (deviations from the arithmetic mean) along the given axis.
Returns an array of anomalies, with the same shape as the input and where the arithmetic mean is computed
along the given axis.
Parameters
axis : int, optional
Axis over which the anomalies are taken. The default is to use the mean of the flattened
array as reference.
dtype : dtype, optional
Type to use in computing the variance. For arrays of integer type
the default is float32; for arrays of float types it is the same as the array type.
See Also:

mean
Compute the mean of the array.

Examples

>>> a = np.ma.array([1,2,3])
>>> a.anom()
masked_array(data = [-1. 0. 1.],
mask = False,
fill_value = 1e+20)

3.20. Masked array operations 897

NumPy Reference, Release 2.0.0.dev8464

average(a, axis=None, weights=None, returned=False)

Return the weighted average of array over the given axis.
Parameters
a : array_like
Data to be averaged. Masked entries are not taken into account in the computation.
axis : int, optional
Axis along which the variance is computed. The default is to compute the variance of
the flattened array.
weights : array_like, optional
The importance that each element has in the computation of the average. The weights
array can either be 1-D (in which case its length must be the size of a along the given
axis) or of the same shape as a. If weights=None, then all data in a are assumed to
have a weight equal to one.
returned : bool, optional
Flag indicating whether a tuple (result, sum of weights) should be returned
as output (True), or just the result (False). Default is False.
Returns
average, [sum_of_weights] : (tuple of) scalar or MaskedArray
The average along the specified axis. When returned is True, return a tuple with the
average as the first element and the sum of the weights as the second element. The
return type is np.float64 if a is of integer type, otherwise it is of the same type as a. If
returned, sum_of_weights is of the same type as average.

Examples

>>> a = np.ma.array([1., 2., 3., 4.], mask=[False, False, True, True])

>>> np.ma.average(a, weights=[3, 1, 0, 0])
1.25

>>> x = np.ma.arange(6.).reshape(3, 2)
>>> print x
[[ 0. 1.]
[ 2. 3.]
[ 4. 5.]]
>>> avg, sumweights = np.ma.average(x, axis=0, weights=[1, 2, 3],
... returned=True)
>>> print avg
[2.66666666667 3.66666666667]

conjugate
Return the complex conjugate, element-wise.
The complex conjugate of a complex number is obtained by changing the sign of its imaginary part.
Parameters
x : array_like
Input value.
Returns
y : ndarray

898 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

The complex conjugate of x, with same dtype as y.

Examples

>>> np.conjugate(1+2j)
(1-2j)

>>> x = np.eye(2) + 1j * np.eye(2)

>>> np.conjugate(x)
array([[ 1.-1.j, 0.-0.j],
[ 0.-0.j, 1.-1.j]])

corrcoef(x, y=None, rowvar=True, bias=False, allow_masked=True)

Return correlation coefficients of the input array.
Except for the handling of missing data this function does the same as numpy.corrcoef. For more details
and examples, see numpy.corrcoef.
Parameters
x : array_like
A 1-D or 2-D array containing multiple variables and observations. Each row of x
represents a variable, and each column a single observation of all those variables. Also
see rowvar below.
y : array_like, optional
An additional set of variables and observations. y has the same shape as x.
rowvar : bool, optional
If rowvar is True (default), then each row represents a variable, with observations in the
columns. Otherwise, the relationship is transposed: each column represents a variable,
while the rows contain observations.
bias : bool, optional
Default normalization (False) is by (N-1), where N is the number of observations given
(unbiased estimate). If bias is 1, then normalization is by N.
allow_masked : bool, optional
If True, masked values are propagated pair-wise: if a value is masked in x, the corre-
sponding value is masked in y. If False, raises an exception.
See Also:

numpy.corrcoef
Equivalent function in top-level NumPy module.
cov
Estimate the covariance matrix.

cov(x, y=None, rowvar=True, bias=False, allow_masked=True)

Estimate the covariance matrix.
Except for the handling of missing data this function does the same as numpy.cov. For more details and
examples, see numpy.cov.
By default, masked values are recognized as such. If x and y have the same shape, a common mask is allocated:
if x[i,j] is masked, then y[i,j] will also be masked. Setting allow_masked to False will raise an exception
if values are missing in either of the input arrays.

3.20. Masked array operations 899

NumPy Reference, Release 2.0.0.dev8464

Parameters
x : array_like
A 1-D or 2-D array containing multiple variables and observations. Each row of x
represents a variable, and each column a single observation of all those variables. Also
see rowvar below.
y : array_like, optional
An additional set of variables and observations. y has the same form as x.
rowvar : bool, optional
If rowvar is True (default), then each row represents a variable, with observations in the
columns. Otherwise, the relationship is transposed: each column represents a variable,
while the rows contain observations.
bias : bool, optional
Default normalization (False) is by (N-1), where N is the number of observations given
(unbiased estimate). If bias is True, then normalization is by N.
allow_masked : bool, optional
If True, masked values are propagated pair-wise: if a value is masked in x, the cor-
responding value is masked in y. If False, raises a ValueError exception when some
values are missing.
Raises
ValueError: :
Raised if some values are missing and allow_masked is False.
See Also:
numpy.cov
cumsum
Return the cumulative sum of the elements along the given axis. The cumulative sum is calculated over the
flattened array by default, otherwise over the specified axis.
Masked values are set to 0 internally during the computation. However, their position is saved, and the result
will be masked at the same locations.
Parameters
axis : {None, -1, int}, optional
Axis along which the sum is computed. The default (axis = None) is to compute over
the flattened array. axis may be negative, in which case it counts from the last to the
first axis.
dtype : {None, dtype}, optional
Type of the returned array and of the accumulator in which the elements are summed. If
dtype is not specified, it defaults to the dtype of a, unless a has an integer dtype with a
precision less than that of the default platform integer. In that case, the default platform
integer is used.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
cumsum : ndarray.

900 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

A new array holding the result is returned unless out is specified, in which case a
reference to out is returned.

Notes
The mask is lost if out is not a valid MaskedArray !
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> marr = np.ma.array(np.arange(10), mask=[0,0,0,1,1,1,0,0,0,0])

>>> print marr.cumsum()
[0 1 3 -- -- -- 9 16 24 33]

cumprod
Return the cumulative product of the elements along the given axis. The cumulative product is taken over the
flattened array by default, otherwise over the specified axis.
Masked values are set to 1 internally during the computation. However, their position is saved, and the result
will be masked at the same locations.
Parameters
axis : {None, -1, int}, optional
Axis along which the product is computed. The default (axis = None) is to compute
over the flattened array.
dtype : {None, dtype}, optional
Determines the type of the returned array and of the accumulator where the elements
are multiplied. If dtype has the value None and the type of a is an integer type
of precision less than the default platform integer, then the default platform integer
precision is used. Otherwise, the dtype is the same as that of a.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
cumprod : ndarray
A new array holding the result is returned unless out is specified, in which case a refer-
ence to out is returned.

Notes
The mask is lost if out is not a valid MaskedArray !
Arithmetic is modular when using integer types, and no error is raised on overflow.
mean
Returns the average of the array elements.
Masked entries are ignored. The average is taken over the flattened array by default, otherwise over the specified
axis. Refer to numpy.mean for the full documentation.
Parameters
a : array_like
Array containing numbers whose mean is desired. If a is not an array, a conversion is
attempted.

3.20. Masked array operations 901

NumPy Reference, Release 2.0.0.dev8464

axis : int, optional

Axis along which the means are computed. The default is to compute the mean of the
flattened array.
dtype : dtype, optional
Type to use in computing the mean. For integer inputs, the default is float64; for floating
point, inputs it is the same as the input dtype.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary.
Returns
mean : ndarray, see dtype parameter above
If out=None, returns a new array containing the mean values, otherwise a reference to
the output array is returned.
See Also:

numpy.ma.mean
Equivalent function.
numpy.mean
Equivalent function on non-masked arrays.
numpy.ma.average
Weighted average.

Examples

>>> a = np.ma.array([1,2,3], mask=[False, False, True])

>>> a
masked_array(data = [1 2 --],
mask = [False False True],
fill_value = 999999)
>>> a.mean()
1.5

median(a, axis=None, out=None, overwrite_input=False)

Compute the median along the specified axis.
Returns the median of the array elements.
Parameters
a : array_like
Input array or object that can be converted to an array.
axis : int, optional
Axis along which the medians are computed. The default (None) is to compute the
median along a flattened version of the array.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
overwrite_input : bool, optional

902 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

If True, then allow use of memory of input array (a) for calculations. The input array
will be modified by the call to median. This will save memory when you do not need to
preserve the contents of the input array. Treat the input as undefined, but it will probably
be fully or partially sorted. Default is False. Note that, if overwrite_input is True, and
the input is not already an ndarray, an error will be raised.
Returns
median : ndarray
A new array holding the result is returned unless out is specified, in which case a refer-
ence to out is returned. Return data-type is float64 for integers and floats smaller than
float64, or the input data-type, otherwise.
See Also:
mean

Notes
Given a vector V with N non masked values, the median of V is the middle value of a sorted copy of V (Vs) - i.e.
Vs[(N-1)/2], when N is odd, or {Vs[N/2 - 1] + Vs[N/2]}/2 when N is even.

Examples

>>> x = np.ma.array(np.arange(8), mask=[0]*4 + [1]*4)

>>> np.ma.extras.median(x)
1.5

>>> x = np.ma.array(np.arange(10).reshape(2, 5), mask=[0]*6 + [1]*4)

>>> np.ma.extras.median(x)
2.5
>>> np.ma.extras.median(x, axis=-1, overwrite_input=True)
masked_array(data = [ 2. 5.],
mask = False,
fill_value = 1e+20)

power(a, b, third=None)
Returns element-wise base array raised to power from second array.
This is the masked array version of numpy.power. For details see numpy.power.
See Also:
numpy.power

Notes
The out argument to numpy.power is not supported, third has to be None.
prod
Return the product of the array elements over the given axis. Masked elements are set to 1 internally for
computation.
Parameters
axis : {None, int}, optional
Axis over which the product is taken. If None is used, then the product is over all the
array elements.
dtype : {None, dtype}, optional

3.20. Masked array operations 903

NumPy Reference, Release 2.0.0.dev8464

Determines the type of the returned array and of the accumulator where the elements are
multiplied. If dtype has the value None and the type of a is an integer type of precision
less than the default platform integer, then the default platform integer precision is used.
Otherwise, the dtype is the same as that of a.
out : {None, array}, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary.
Returns
product_along_axis : {array, scalar}, see dtype parameter above.
Returns an array whose shape is the same as a with the specified axis removed. Returns
a 0d array when a is 1d or axis=None. Returns a reference to the specified output array
if specified.
See Also:

prod
equivalent function

Notes
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> np.prod([1.,2.])
2.0
>>> np.prod([1.,2.], dtype=np.int32)
2
>>> np.prod([[1.,2.],[3.,4.]])
24.0
>>> np.prod([[1.,2.],[3.,4.]], axis=1)
array([ 2., 12.])

std
Compute the standard deviation along the specified axis.
Returns the standard deviation, a measure of the spread of a distribution, of the array elements. The standard
deviation is computed for the flattened array by default, otherwise over the specified axis.
Parameters
a : array_like
Calculate the standard deviation of these values.
axis : int, optional
Axis along which the standard deviation is computed. The default is to compute the
standard deviation of the flattened array.
dtype : dtype, optional
Type to use in computing the standard deviation. For arrays of integer type the default
is float64, for arrays of float types it is the same as the array type.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type (of the calculated values) will be cast if necessary.

904 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

ddof : int, optional

Means Delta Degrees of Freedom. The divisor used in calculations is N - ddof,
where N represents the number of elements. By default ddof is zero.
Returns
standard_deviation : ndarray, see dtype parameter above.
If out is None, return a new array containing the standard deviation, otherwise return a
reference to the output array.
See Also:
var, mean
numpy.doc.ufuncs
Section Output arguments

Notes
The standard deviation is the square root of the average of the squared deviations from the mean, i.e., std =
sqrt(mean(abs(x - x.mean())**2)).
The average squared deviation is normally calculated as x.sum() / N, where N = len(x). If, however,
ddof is specified, the divisor N - ddof is used instead. In standard statistical practice, ddof=1 provides an
unbiased estimator of the variance of the infinite population. ddof=0 provides a maximum likelihood estimate
of the variance for normally distributed variables. The standard deviation computed in this function is the
square root of the estimated variance, so even with ddof=1, it will not be an unbiased estimate of the standard
deviation per se.
Note that, for complex numbers, std takes the absolute value before squaring, so that the result is always real
and nonnegative.
For floating-point input, the std is computed using the same precision the input has. Depending on the input
data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a higher-
accuracy accumulator using the dtype keyword can alleviate this issue.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> np.std(a)
1.1180339887498949
>>> np.std(a, axis=0)
array([ 1., 1.])
>>> np.std(a, axis=1)
array([ 0.5, 0.5])

In single precision, std() can be inaccurate:

>>> a = np.zeros((2,512*512), dtype=np.float32)

>>> a[0,:] = 1.0
>>> a[1,:] = 0.1
>>> np.std(a)
0.45172946707416706

Computing the standard deviation in float64 is more accurate:

>>> np.std(a, dtype=np.float64)

0.44999999925552653

3.20. Masked array operations 905

NumPy Reference, Release 2.0.0.dev8464

sum
Return the sum of the array elements over the given axis. Masked elements are set to 0 internally.
Parameters
axis : {None, -1, int}, optional
Axis along which the sum is computed. The default (axis = None) is to compute over
the flattened array.
dtype : {None, dtype}, optional
Determines the type of the returned array and of the accumulator where the elements
are summed. If dtype has the value None and the type of a is an integer type of precision
less than the default platform integer, then the default platform integer precision is used.
Otherwise, the dtype is the same as that of a.
out : {None, ndarray}, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
sum_along_axis : MaskedArray or scalar
An array with the same shape as self, with the specified axis removed. If self is a 0-d
array, or if axis is None, a scalar is returned. If an output array is specified, a reference
to out is returned.

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.sum()
25
>>> print x.sum(axis=1)
[4 5 16]
>>> print x.sum(axis=0)
[8 5 12]
>>> print type(x.sum(axis=0, dtype=np.int64)[0])
<type numpy.int64>

var
Compute the variance along the specified axis.
Returns the variance of the array elements, a measure of the spread of a distribution. The variance is computed
for the flattened array by default, otherwise over the specified axis.
Parameters
a : array_like
Array containing numbers whose variance is desired. If a is not an array, a conversion
is attempted.
axis : int, optional
Axis along which the variance is computed. The default is to compute the variance of
the flattened array.
dtype : dtype, optional

906 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Type to use in computing the variance. For arrays of integer type the default is float32;
for arrays of float types it is the same as the array type.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type is cast if necessary.
ddof : int, optional
Delta Degrees of Freedom: the divisor used in calculation is N - ddof, where N
represents the number of elements. By default ddof is zero.
Returns
variance : ndarray, see dtype parameter above
If out=None, returns a new array containing the variance; otherwise a reference to the
output array is returned.
See Also:

std
Standard deviation
mean
Average
numpy.doc.ufuncs
Section Output arguments

Notes
The variance is the average of the squared deviations from the mean, i.e., var = mean(abs(x -
x.mean())**2).
The mean is normally calculated as x.sum() / N, where N = len(x). If, however, ddof is specified, the
divisor N - ddof is used instead. In standard statistical practice, ddof=1 provides an unbiased estimator of
the variance of the infinite population. ddof=0 provides a maximum likelihood estimate of the variance for
normally distributed variables.
Note that for complex numbers, the absolute value is taken before squaring, so that the result is always real and
nonnegative.
For floating-point input, the variance is computed using the same precision the input has. Depending on the
input data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a
higher-accuracy accumulator using the dtype keyword can alleviate this issue.

Examples

>>> a = np.array([[1,2],[3,4]])
>>> np.var(a)
1.25
>>> np.var(a,0)
array([ 1., 1.])
>>> np.var(a,1)
array([ 0.25, 0.25])

In single precision, var() can be inaccurate:

3.20. Masked array operations 907

NumPy Reference, Release 2.0.0.dev8464

>>> a = np.zeros((2,512*512), dtype=np.float32)

>>> a[0,:] = 1.0
>>> a[1,:] = 0.1
>>> np.var(a)
0.20405951142311096

Computing the standard deviation in float64 is more accurate:

>>> np.var(a, dtype=np.float64)

0.20249999932997387
>>> ((1-0.55)**2 + (0.1-0.55)**2)/2
0.20250000000000001

anom(axis=None, dtype=None)
Compute the anomalies (deviations from the arithmetic mean) along the given axis.
Returns an array of anomalies, with the same shape as the input and where the arithmetic mean is computed
along the given axis.
Parameters
axis : int, optional
Axis over which the anomalies are taken. The default is to use the mean of the flattened
array as reference.
dtype : dtype, optional
Type to use in computing the variance. For arrays of integer type
the default is float32; for arrays of float types it is the same as the array type.
See Also:

mean
Compute the mean of the array.

Examples

>>> a = np.ma.array([1,2,3])
>>> a.anom()
masked_array(data = [-1. 0. 1.],
mask = False,
fill_value = 1e+20)

cumprod(axis=None, dtype=None, out=None)

Return the cumulative product of the elements along the given axis. The cumulative product is taken over the
flattened array by default, otherwise over the specified axis.
Masked values are set to 1 internally during the computation. However, their position is saved, and the result
will be masked at the same locations.
Parameters
axis : {None, -1, int}, optional
Axis along which the product is computed. The default (axis = None) is to compute
over the flattened array.
dtype : {None, dtype}, optional
Determines the type of the returned array and of the accumulator where the elements
are multiplied. If dtype has the value None and the type of a is an integer type

908 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

of precision less than the default platform integer, then the default platform integer
precision is used. Otherwise, the dtype is the same as that of a.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
cumprod : ndarray
A new array holding the result is returned unless out is specified, in which case a refer-
ence to out is returned.

Notes
The mask is lost if out is not a valid MaskedArray !
Arithmetic is modular when using integer types, and no error is raised on overflow.
cumsum(axis=None, dtype=None, out=None)
Return the cumulative sum of the elements along the given axis. The cumulative sum is calculated over the
flattened array by default, otherwise over the specified axis.
Masked values are set to 0 internally during the computation. However, their position is saved, and the result
will be masked at the same locations.
Parameters
axis : {None, -1, int}, optional
Axis along which the sum is computed. The default (axis = None) is to compute over
the flattened array. axis may be negative, in which case it counts from the last to the
first axis.
dtype : {None, dtype}, optional
Type of the returned array and of the accumulator in which the elements are summed. If
dtype is not specified, it defaults to the dtype of a, unless a has an integer dtype with a
precision less than that of the default platform integer. In that case, the default platform
integer is used.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
cumsum : ndarray.
A new array holding the result is returned unless out is specified, in which case a
reference to out is returned.

Notes
The mask is lost if out is not a valid MaskedArray !
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> marr = np.ma.array(np.arange(10), mask=[0,0,0,1,1,1,0,0,0,0])

>>> print marr.cumsum()
[0 1 3 -- -- -- 9 16 24 33]

3.20. Masked array operations 909

NumPy Reference, Release 2.0.0.dev8464

mean(axis=None, dtype=None, out=None)

Returns the average of the array elements.
Masked entries are ignored. The average is taken over the flattened array by default, otherwise over the specified
axis. Refer to numpy.mean for the full documentation.
Parameters
a : array_like
Array containing numbers whose mean is desired. If a is not an array, a conversion is
attempted.
axis : int, optional
Axis along which the means are computed. The default is to compute the mean of the
flattened array.
dtype : dtype, optional
Type to use in computing the mean. For integer inputs, the default is float64; for floating
point, inputs it is the same as the input dtype.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary.
Returns
mean : ndarray, see dtype parameter above
If out=None, returns a new array containing the mean values, otherwise a reference to
the output array is returned.
See Also:

numpy.ma.mean
Equivalent function.
numpy.mean
Equivalent function on non-masked arrays.
numpy.ma.average
Weighted average.

Examples

>>> a = np.ma.array([1,2,3], mask=[False, False, True])

>>> a
masked_array(data = [1 2 --],
mask = [False False True],
fill_value = 999999)
>>> a.mean()
1.5

prod(axis=None, dtype=None, out=None)

Return the product of the array elements over the given axis. Masked elements are set to 1 internally for
computation.
Parameters
axis : {None, int}, optional
Axis over which the product is taken. If None is used, then the product is over all the
array elements.

910 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

dtype : {None, dtype}, optional

Determines the type of the returned array and of the accumulator where the elements are
multiplied. If dtype has the value None and the type of a is an integer type of precision
less than the default platform integer, then the default platform integer precision is used.
Otherwise, the dtype is the same as that of a.
out : {None, array}, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type will be cast if necessary.
Returns
product_along_axis : {array, scalar}, see dtype parameter above.
Returns an array whose shape is the same as a with the specified axis removed. Returns
a 0d array when a is 1d or axis=None. Returns a reference to the specified output array
if specified.
See Also:

prod
equivalent function

Notes
Arithmetic is modular when using integer types, and no error is raised on overflow.

Examples

>>> np.prod([1.,2.])
2.0
>>> np.prod([1.,2.], dtype=np.int32)
2
>>> np.prod([[1.,2.],[3.,4.]])
24.0
>>> np.prod([[1.,2.],[3.,4.]], axis=1)
array([ 2., 12.])

std(axis=None, dtype=None, out=None, ddof=0)

Compute the standard deviation along the specified axis.
Returns the standard deviation, a measure of the spread of a distribution, of the array elements. The standard
deviation is computed for the flattened array by default, otherwise over the specified axis.
Parameters
a : array_like
Calculate the standard deviation of these values.
axis : int, optional
Axis along which the standard deviation is computed. The default is to compute the
standard deviation of the flattened array.
dtype : dtype, optional
Type to use in computing the standard deviation. For arrays of integer type the default
is float64, for arrays of float types it is the same as the array type.
out : ndarray, optional

3.20. Masked array operations 911

NumPy Reference, Release 2.0.0.dev8464

Alternative output array in which to place the result. It must have the same shape as the
expected output but the type (of the calculated values) will be cast if necessary.
ddof : int, optional
Means Delta Degrees of Freedom. The divisor used in calculations is N - ddof,
where N represents the number of elements. By default ddof is zero.
Returns
standard_deviation : ndarray, see dtype parameter above.
If out is None, return a new array containing the standard deviation, otherwise return a
reference to the output array.
See Also:
var, mean
numpy.doc.ufuncs
Section Output arguments

Notes
The standard deviation is the square root of the average of the squared deviations from the mean, i.e., std =
sqrt(mean(abs(x - x.mean())**2)).
The average squared deviation is normally calculated as x.sum() / N, where N = len(x). If, however,
ddof is specified, the divisor N - ddof is used instead. In standard statistical practice, ddof=1 provides an
unbiased estimator of the variance of the infinite population. ddof=0 provides a maximum likelihood estimate
of the variance for normally distributed variables. The standard deviation computed in this function is the
square root of the estimated variance, so even with ddof=1, it will not be an unbiased estimate of the standard
deviation per se.
Note that, for complex numbers, std takes the absolute value before squaring, so that the result is always real
and nonnegative.
For floating-point input, the std is computed using the same precision the input has. Depending on the input
data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a higher-
accuracy accumulator using the dtype keyword can alleviate this issue.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> np.std(a)
1.1180339887498949
>>> np.std(a, axis=0)
array([ 1., 1.])
>>> np.std(a, axis=1)
array([ 0.5, 0.5])

In single precision, std() can be inaccurate:

>>> a = np.zeros((2,512*512), dtype=np.float32)

>>> a[0,:] = 1.0
>>> a[1,:] = 0.1
>>> np.std(a)
0.45172946707416706

Computing the standard deviation in float64 is more accurate:

912 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.std(a, dtype=np.float64)

0.44999999925552653

sum(axis=None, dtype=None, out=None)

Return the sum of the array elements over the given axis. Masked elements are set to 0 internally.
Parameters
axis : {None, -1, int}, optional
Axis along which the sum is computed. The default (axis = None) is to compute over
the flattened array.
dtype : {None, dtype}, optional
Determines the type of the returned array and of the accumulator where the elements
are summed. If dtype has the value None and the type of a is an integer type of precision
less than the default platform integer, then the default platform integer precision is used.
Otherwise, the dtype is the same as that of a.
out : {None, ndarray}, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
Returns
sum_along_axis : MaskedArray or scalar
An array with the same shape as self, with the specified axis removed. If self is a 0-d
array, or if axis is None, a scalar is returned. If an output array is specified, a reference
to out is returned.

Examples

>>> x = np.ma.array([[1,2,3],[4,5,6],[7,8,9]], mask=[0] + [1,0]*4)

>>> print x
[[1 -- 3]
[-- 5 --]
[7 -- 9]]
>>> print x.sum()
25
>>> print x.sum(axis=1)
[4 5 16]
>>> print x.sum(axis=0)
[8 5 12]
>>> print type(x.sum(axis=0, dtype=np.int64)[0])
<type numpy.int64>

var(axis=None, dtype=None, out=None, ddof=0)

Compute the variance along the specified axis.
Returns the variance of the array elements, a measure of the spread of a distribution. The variance is computed
for the flattened array by default, otherwise over the specified axis.
Parameters
a : array_like
Array containing numbers whose variance is desired. If a is not an array, a conversion
is attempted.
axis : int, optional

3.20. Masked array operations 913

NumPy Reference, Release 2.0.0.dev8464

Axis along which the variance is computed. The default is to compute the variance of
the flattened array.
dtype : dtype, optional
Type to use in computing the variance. For arrays of integer type the default is float32;
for arrays of float types it is the same as the array type.
out : ndarray, optional
Alternative output array in which to place the result. It must have the same shape as the
expected output but the type is cast if necessary.
ddof : int, optional
Delta Degrees of Freedom: the divisor used in calculation is N - ddof, where N
represents the number of elements. By default ddof is zero.
Returns
variance : ndarray, see dtype parameter above
If out=None, returns a new array containing the variance; otherwise a reference to the
output array is returned.
See Also:

std
Standard deviation
mean
Average
numpy.doc.ufuncs
Section Output arguments

Notes
The variance is the average of the squared deviations from the mean, i.e., var = mean(abs(x -
x.mean())**2).
The mean is normally calculated as x.sum() / N, where N = len(x). If, however, ddof is specified, the
divisor N - ddof is used instead. In standard statistical practice, ddof=1 provides an unbiased estimator of
the variance of the infinite population. ddof=0 provides a maximum likelihood estimate of the variance for
normally distributed variables.
Note that for complex numbers, the absolute value is taken before squaring, so that the result is always real and
nonnegative.
For floating-point input, the variance is computed using the same precision the input has. Depending on the
input data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a
higher-accuracy accumulator using the dtype keyword can alleviate this issue.

Examples

>>> a = np.array([[1,2],[3,4]])
>>> np.var(a)
1.25
>>> np.var(a,0)
array([ 1., 1.])
>>> np.var(a,1)
array([ 0.25, 0.25])

914 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

In single precision, var() can be inaccurate:

>>> a = np.zeros((2,512*512), dtype=np.float32)

>>> a[0,:] = 1.0
>>> a[1,:] = 0.1
>>> np.var(a)
0.20405951142311096

Computing the standard deviation in float64 is more accurate:

>>> np.var(a, dtype=np.float64)

0.20249999932997387
>>> ((1-0.55)**2 + (0.1-0.55)**2)/2
0.20250000000000001

Minimum/maximum

ma.argmax(a[, axis, fill_value]) Function version of the eponymous method.

ma.argmin(a[, axis, fill_value]) Returns array of indices of the maximum values along the
given axis.
ma.max(obj[, axis, out, fill_value]) Return the maximum along a given axis.
ma.min(obj[, axis, out, fill_value]) Return the minimum along a given axis.
ma.ptp(obj[, axis, out, fill_value]) Return (maximum - minimum) along the the given dimension
(i.e.
ma.MaskedArray.argmax([axis, fill_value, Returns array of indices of the maximum values along the
out]) given axis.
ma.MaskedArray.argmin([axis, fill_value, Return array of indices to the minimum values along the given
out]) axis.
ma.MaskedArray.max([axis, out, Return the maximum along a given axis.
fill_value])
ma.MaskedArray.min([axis, out, Return the minimum along a given axis.
fill_value])
ma.MaskedArray.ptp([axis, out, Return (maximum - minimum) along the the given dimension
fill_value]) (i.e.
argmax(a, axis=None, fill_value=None)
Function version of the eponymous method.
argmin(a, axis=None, fill_value=None)
Returns array of indices of the maximum values along the given axis. Masked values are treated as if they had
the value fill_value.
Parameters
axis : {None, integer}
If None, the index is into the flattened array, otherwise along the specified axis
fill_value : {var}, optional
Value used to fill in the masked values. If None, the output of maxi-
mum_fill_value(self._data) is used instead.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.

3.20. Masked array operations 915

NumPy Reference, Release 2.0.0.dev8464

Returns
index_array : {integer_array}

Examples

>>> a = np.arange(6).reshape(2,3)
>>> a.argmax()
5
>>> a.argmax(0)
array([1, 1, 1])
>>> a.argmax(1)
array([2, 2])

max(obj, axis=None, out=None, fill_value=None)

Return the maximum along a given axis.
Parameters
axis : {None, int}, optional
Axis along which to operate. By default, axis is None and the flattened input is used.
out : array_like, optional
Alternative output array in which to place the result. Must be of the same shape and
buffer length as the expected output.
fill_value : {var}, optional
Value used to fill in the masked values. If None, use the output of maxi-
mum_fill_value().
Returns
amax : array_like
New array holding the result. If out was specified, out is returned.
See Also:

maximum_fill_value
Returns the maximum filling value for a given datatype.

min(obj, axis=None, out=None, fill_value=None)

Return the minimum along a given axis.
Parameters
axis : {None, int}, optional
Axis along which to operate. By default, axis is None and the flattened input is used.
out : array_like, optional
Alternative output array in which to place the result. Must be of the same shape and
buffer length as the expected output.
fill_value : {var}, optional
Value used to fill in the masked values. If None, use the output of minimum_fill_value.
Returns
amin : array_like
New array holding the result. If out was specified, out is returned.
See Also:

916 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

minimum_fill_value
Returns the minimum filling value for a given datatype.

ptp(obj, axis=None, out=None, fill_value=None)

Return (maximum - minimum) along the the given dimension (i.e. peak-to-peak value).
Parameters
axis : {None, int}, optional
Axis along which to find the peaks. If None (default) the flattened array is used.
out : {None, array_like}, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
fill_value : {var}, optional
Value used to fill in the masked values.
Returns
ptp : ndarray.
A new array holding the result, unless out was specified, in which case a reference to
out is returned.
argmax(axis=None, fill_value=None, out=None)
Returns array of indices of the maximum values along the given axis. Masked values are treated as if they had
the value fill_value.
Parameters
axis : {None, integer}
If None, the index is into the flattened array, otherwise along the specified axis
fill_value : {var}, optional
Value used to fill in the masked values. If None, the output of maxi-
mum_fill_value(self._data) is used instead.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
Returns
index_array : {integer_array}

Examples

>>> a = np.arange(6).reshape(2,3)
>>> a.argmax()
5
>>> a.argmax(0)
array([1, 1, 1])
>>> a.argmax(1)
array([2, 2])

argmin(axis=None, fill_value=None, out=None)

Return array of indices to the minimum values along the given axis.
Parameters
axis : {None, integer}

3.20. Masked array operations 917

NumPy Reference, Release 2.0.0.dev8464

If None, the index is into the flattened array, otherwise along the specified axis
fill_value : {var}, optional
Value used to fill in the masked values. If None, the output of mini-
mum_fill_value(self._data) is used instead.
out : {None, array}, optional
Array into which the result can be placed. Its type is preserved and it must be of the
right shape to hold the output.
Returns
{ndarray, scalar} :
If multi-dimension input, returns a new ndarray of indices to the minimum values along
the given axis. Otherwise, returns a scalar of index to the minimum values along the
given axis.

Examples

>>> x = np.ma.array(arange(4), mask=[1,1,0,0])

>>> x.shape = (2,2)
>>> print x
[[-- --]
[2 3]]
>>> print x.argmin(axis=0, fill_value=-1)
[0 0]
>>> print x.argmin(axis=0, fill_value=9)
[1 1]

max(axis=None, out=None, fill_value=None)

Return the maximum along a given axis.
Parameters
axis : {None, int}, optional
Axis along which to operate. By default, axis is None and the flattened input is used.
out : array_like, optional
Alternative output array in which to place the result. Must be of the same shape and
buffer length as the expected output.
fill_value : {var}, optional
Value used to fill in the masked values. If None, use the output of maxi-
mum_fill_value().
Returns
amax : array_like
New array holding the result. If out was specified, out is returned.
See Also:

maximum_fill_value
Returns the maximum filling value for a given datatype.

min(axis=None, out=None, fill_value=None)

Return the minimum along a given axis.

918 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
axis : {None, int}, optional
Axis along which to operate. By default, axis is None and the flattened input is used.
out : array_like, optional
Alternative output array in which to place the result. Must be of the same shape and
buffer length as the expected output.
fill_value : {var}, optional
Value used to fill in the masked values. If None, use the output of minimum_fill_value.
Returns
amin : array_like
New array holding the result. If out was specified, out is returned.
See Also:

minimum_fill_value
Returns the minimum filling value for a given datatype.

ptp(axis=None, out=None, fill_value=None)

Return (maximum - minimum) along the the given dimension (i.e. peak-to-peak value).
Parameters
axis : {None, int}, optional
Axis along which to find the peaks. If None (default) the flattened array is used.
out : {None, array_like}, optional
Alternative output array in which to place the result. It must have the same shape and
buffer length as the expected output but the type will be cast if necessary.
fill_value : {var}, optional
Value used to fill in the masked values.
Returns
ptp : ndarray.
A new array holding the result, unless out was specified, in which case a reference to
out is returned.

Sorting

ma.argsort(a[, axis, kind, order, fill_value]) Return an ndarray of indices that sort the array along the
specified axis.
ma.sort(a[, axis, kind, order, endwith, ...]) Sort the array, in-place
ma.MaskedArray.argsort([axis, kind, Return an ndarray of indices that sort the array along the
order, ...]) specified axis.
ma.MaskedArray.sort([axis, kind, order, Sort the array, in-place
...])
argsort(a, axis=None, kind=quicksort, order=None, fill_value=None)
Return an ndarray of indices that sort the array along the specified axis. Masked values are filled beforehand to
fill_value.

3.20. Masked array operations 919

NumPy Reference, Release 2.0.0.dev8464

Parameters
axis : int, optional
Axis along which to sort. The default is -1 (last axis). If None, the flattened array is
used.
fill_value : var, optional
Value used to fill the array before sorting. The default is the fill_value attribute of the
input array.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm.
order : list, optional
When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
Returns
index_array : ndarray, int
Array of indices that sort a along the specified axis. In other words,
a[index_array] yields a sorted a.
See Also:

sort
Describes sorting algorithms used.
lexsort
Indirect stable sort with multiple keys.
ndarray.sort
Inplace sort.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = np.ma.array([3,2,1], mask=[False, False, True])

>>> a
masked_array(data = [3 2 --],
mask = [False False True],
fill_value = 999999)
>>> a.argsort()
array([1, 0, 2])

sort(a, axis=-1, kind=quicksort, order=None, endwith=True, fill_value=None)

Sort the array, in-place
Parameters
a : array_like
Array to be sorted.
axis : int, optional
Axis along which to sort. If None, the array is flattened before sorting. The default is
-1, which sorts along the last axis.

920 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

kind : {quicksort, mergesort, heapsort}, optional

Sorting algorithm. Default is quicksort.
order : list, optional
When a is a structured array, this argument specifies which fields to compare first, sec-
ond, and so on. This list does not need to include all of the fields.
endwith : {True, False}, optional
Whether missing values (if any) should be forced in the upper indices (at the end of the
array) (True) or lower indices (at the beginning).
fill_value : {var}, optional
Value used internally for the masked values. If fill_value is not None, it supersedes
endwith.
Returns
sorted_array : ndarray
Array of the same type and shape as a.
See Also:

ndarray.sort
Method to sort an array in-place.
argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in a sorted array.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # Default
>>> a.sort()
>>> print a
[1 3 5 -- --]

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # Put missing values in the front
>>> a.sort(endwith=False)
>>> print a
[-- -- 1 3 5]

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # fill_value takes over endwith
>>> a.sort(endwith=False, fill_value=3)
>>> print a
[1 -- -- 3 5]

3.20. Masked array operations 921

NumPy Reference, Release 2.0.0.dev8464

argsort(axis=None, kind=quicksort, order=None, fill_value=None)

Return an ndarray of indices that sort the array along the specified axis. Masked values are filled beforehand to
fill_value.
Parameters
axis : int, optional
Axis along which to sort. The default is -1 (last axis). If None, the flattened array is
used.
fill_value : var, optional
Value used to fill the array before sorting. The default is the fill_value attribute of the
input array.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm.
order : list, optional
When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
Returns
index_array : ndarray, int
Array of indices that sort a along the specified axis. In other words,
a[index_array] yields a sorted a.
See Also:

sort
Describes sorting algorithms used.
lexsort
Indirect stable sort with multiple keys.
ndarray.sort
Inplace sort.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = np.ma.array([3,2,1], mask=[False, False, True])

>>> a
masked_array(data = [3 2 --],
mask = [False False True],
fill_value = 999999)
>>> a.argsort()
array([1, 0, 2])

sort(axis=-1, kind=quicksort, order=None, endwith=True, fill_value=None)

Sort the array, in-place
Parameters
a : array_like
Array to be sorted.

922 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

axis : int, optional

Axis along which to sort. If None, the array is flattened before sorting. The default is
-1, which sorts along the last axis.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm. Default is quicksort.
order : list, optional
When a is a structured array, this argument specifies which fields to compare first, sec-
ond, and so on. This list does not need to include all of the fields.
endwith : {True, False}, optional
Whether missing values (if any) should be forced in the upper indices (at the end of the
array) (True) or lower indices (at the beginning).
fill_value : {var}, optional
Value used internally for the masked values. If fill_value is not None, it supersedes
endwith.
Returns
sorted_array : ndarray
Array of the same type and shape as a.
See Also:

ndarray.sort
Method to sort an array in-place.
argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in a sorted array.

Notes
See sort for notes on the different sorting algorithms.

Examples

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # Default
>>> a.sort()
>>> print a
[1 3 5 -- --]

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # Put missing values in the front
>>> a.sort(endwith=False)
>>> print a
[-- -- 1 3 5]

3.20. Masked array operations 923

NumPy Reference, Release 2.0.0.dev8464

>>> a = ma.array([1, 2, 5, 4, 3],mask=[0, 1, 0, 1, 0])

>>> # fill_value takes over endwith
>>> a.sort(endwith=False, fill_value=3)
>>> print a
[1 -- -- 3 5]

Algebra

ma.diag(v[, k]) Extract a diagonal or construct a diagonal array.

ma.dot(a, b[, strict]) Return the dot product of two arrays.
ma.identity(n[, dtype]) Return the identity array.
ma.inner(a, b) Inner product of two arrays.
ma.innerproduct(a, b) Inner product of two arrays.
ma.outer(a, b) Compute the outer product of two vectors.
ma.outerproduct(a, b) Compute the outer product of two vectors.
ma.trace(self[, offset, axis1, axis2, ...]) Return the sum along diagonals of the array.
ma.transpose(a[, axes]) Permute the dimensions of an array.
ma.MaskedArray.trace([offset, axis1, axis2, ...]) Return the sum along diagonals of the array.
ma.MaskedArray.transpose(*axes) Returns a view of the array with axes transposed.
diag(v, k=0)
Extract a diagonal or construct a diagonal array.
This function is the equivalent of numpy.diag that takes masked values into account, see numpy.diag for
details.
See Also:

numpy.diag
Equivalent function for ndarrays.

dot(a, b, strict=False)
Return the dot product of two arrays.
Note: Works only with 2-D arrays at the moment.
This function is the equivalent of numpy.dot that takes masked values into account, see numpy.dot for
details.
Parameters
a, b : ndarray
Inputs arrays.
strict : bool, optional
Whether masked data are propagated (True) or set to 0 (False) for the computation.
Default is False. Propagating the mask means that if a masked value appears in a row
or column, the whole row or column is considered masked.
See Also:

numpy.dot
Equivalent function for ndarrays.

924 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = ma.array([[1, 2, 3], [4, 5, 6]], mask=[[1, 0, 0], [0, 0, 0]])

>>> b = ma.array([[1, 2], [3, 4], [5, 6]], mask=[[1, 0], [0, 0], [0, 0]])
>>> np.ma.dot(a, b)
masked_array(data =
[[21 26]
[45 64]],
mask =
[[False False]
[False False]],
fill_value = 999999)
>>> np.ma.dot(a, b, strict=True)
masked_array(data =
[[-- --]
[-- 64]],
mask =
[[ True True]
[ True False]],
fill_value = 999999)

identity
Return the identity array.
The identity array is a square array with ones on the main diagonal.
Parameters
n : int
Number of rows (and columns) in n x n output.
dtype : data-type, optional
Data-type of the output. Defaults to float.
Returns
out : ndarray
n x n array with its main diagonal set to one, and all other elements 0.

Examples

>>> np.identity(3)
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])

inner(a, b)
Inner product of two arrays.
Ordinary inner product of vectors for 1-D arrays (without complex conjugation), in higher dimensions a sum
product over the last axes.
Parameters
a, b : array_like
If a and b are nonscalar, their last dimensions of must match.
Returns
out : ndarray
out.shape = a.shape[:-1] + b.shape[:-1]

3.20. Masked array operations 925

NumPy Reference, Release 2.0.0.dev8464

Raises
ValueError :
If the last dimension of a and b has different size.
See Also:

tensordot
Sum products over arbitrary axes.
dot
Generalised matrix product, using second last dimension of b.

Notes
Masked values are replaced by 0.

Examples
Ordinary inner product for vectors:

>>> a = np.array([1,2,3])
>>> b = np.array([0,1,0])
>>> np.inner(a, b)
2

A multidimensional example:

>>> a = np.arange(24).reshape((2,3,4))
>>> b = np.arange(4)
>>> np.inner(a, b)
array([[ 14, 38, 62],
[ 86, 110, 134]])

An example where b is a scalar:

>>> np.inner(np.eye(2), 7)
array([[ 7., 0.],
[ 0., 7.]])

innerproduct(a, b)
Inner product of two arrays.
Ordinary inner product of vectors for 1-D arrays (without complex conjugation), in higher dimensions a sum
product over the last axes.
Parameters
a, b : array_like
If a and b are nonscalar, their last dimensions of must match.
Returns
out : ndarray
out.shape = a.shape[:-1] + b.shape[:-1]
Raises
ValueError :
If the last dimension of a and b has different size.
See Also:

926 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

tensordot
Sum products over arbitrary axes.
dot
Generalised matrix product, using second last dimension of b.

Notes
Masked values are replaced by 0.

Examples
Ordinary inner product for vectors:

>>> a = np.array([1,2,3])
>>> b = np.array([0,1,0])
>>> np.inner(a, b)
2

A multidimensional example:

>>> a = np.arange(24).reshape((2,3,4))
>>> b = np.arange(4)
>>> np.inner(a, b)
array([[ 14, 38, 62],
[ 86, 110, 134]])

An example where b is a scalar:

>>> np.inner(np.eye(2), 7)
array([[ 7., 0.],
[ 0., 7.]])

outer(a, b)
Compute the outer product of two vectors.
Given two vectors, a = [a0, a1, ..., aM] and b = [b0, b1, ..., bN], the outer product [R57]
is:

[[a0*b0 a0*b1 ... a0*bN ]

[a1*b0 .
[ ... .
[aM*b0 aM*bN ]]

Parameters
a, b : array_like, shape (M,), (N,)
First and second input vectors. Inputs are flattened if they are not already 1-dimensional.
Returns
out : ndarray, shape (M, N)
out[i, j] = a[i] * b[j]

Notes
Masked values are replaced by 0.

3.20. Masked array operations 927

NumPy Reference, Release 2.0.0.dev8464

References
[R57]

Examples
Make a (very coarse) grid for computing a Mandelbrot set:

>>> rl = np.outer(np.ones((5,)), np.linspace(-2, 2, 5))

>>> rl
array([[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.]])
>>> im = np.outer(1j*np.linspace(2, -2, 5), np.ones((5,)))
>>> im
array([[ 0.+2.j, 0.+2.j, 0.+2.j, 0.+2.j, 0.+2.j],
[ 0.+1.j, 0.+1.j, 0.+1.j, 0.+1.j, 0.+1.j],
[ 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[ 0.-1.j, 0.-1.j, 0.-1.j, 0.-1.j, 0.-1.j],
[ 0.-2.j, 0.-2.j, 0.-2.j, 0.-2.j, 0.-2.j]])
>>> grid = rl + im
>>> grid
array([[-2.+2.j, -1.+2.j, 0.+2.j, 1.+2.j, 2.+2.j],
[-2.+1.j, -1.+1.j, 0.+1.j, 1.+1.j, 2.+1.j],
[-2.+0.j, -1.+0.j, 0.+0.j, 1.+0.j, 2.+0.j],
[-2.-1.j, -1.-1.j, 0.-1.j, 1.-1.j, 2.-1.j],
[-2.-2.j, -1.-2.j, 0.-2.j, 1.-2.j, 2.-2.j]])

An example using a vector of letters:

>>> x = np.array([a, b, c], dtype=object)

>>> np.outer(x, [1, 2, 3])
array([[a, aa, aaa],
[b, bb, bbb],
[c, cc, ccc]], dtype=object)

outerproduct(a, b)
Compute the outer product of two vectors.
Given two vectors, a = [a0, a1, ..., aM] and b = [b0, b1, ..., bN], the outer product [R58]
is:

[[a0*b0 a0*b1 ... a0*bN ]

[a1*b0 .
[ ... .
[aM*b0 aM*bN ]]

Parameters
a, b : array_like, shape (M,), (N,)
First and second input vectors. Inputs are flattened if they are not already 1-dimensional.
Returns
out : ndarray, shape (M, N)
out[i, j] = a[i] * b[j]

928 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Notes
Masked values are replaced by 0.

References
[R58]

Examples
Make a (very coarse) grid for computing a Mandelbrot set:

>>> rl = np.outer(np.ones((5,)), np.linspace(-2, 2, 5))

>>> rl
array([[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.],
[-2., -1., 0., 1., 2.]])
>>> im = np.outer(1j*np.linspace(2, -2, 5), np.ones((5,)))
>>> im
array([[ 0.+2.j, 0.+2.j, 0.+2.j, 0.+2.j, 0.+2.j],
[ 0.+1.j, 0.+1.j, 0.+1.j, 0.+1.j, 0.+1.j],
[ 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[ 0.-1.j, 0.-1.j, 0.-1.j, 0.-1.j, 0.-1.j],
[ 0.-2.j, 0.-2.j, 0.-2.j, 0.-2.j, 0.-2.j]])
>>> grid = rl + im
>>> grid
array([[-2.+2.j, -1.+2.j, 0.+2.j, 1.+2.j, 2.+2.j],
[-2.+1.j, -1.+1.j, 0.+1.j, 1.+1.j, 2.+1.j],
[-2.+0.j, -1.+0.j, 0.+0.j, 1.+0.j, 2.+0.j],
[-2.-1.j, -1.-1.j, 0.-1.j, 1.-1.j, 2.-1.j],
[-2.-2.j, -1.-2.j, 0.-2.j, 1.-2.j, 2.-2.j]])

An example using a vector of letters:

>>> x = np.array([a, b, c], dtype=object)

>>> np.outer(x, [1, 2, 3])
array([[a, aa, aaa],
[b, bb, bbb],
[c, cc, ccc]], dtype=object)

trace

Return the sum along diagonals of the array.

Refer to numpy.trace for full documentation.
See Also:

numpy.trace
equivalent function

transpose(a, axes=None)
Permute the dimensions of an array.
This function is exactly equivalent to numpy.transpose.
See Also:

3.20. Masked array operations 929

NumPy Reference, Release 2.0.0.dev8464

numpy.transpose
Equivalent function in top-level NumPy module.

Examples

>>> import numpy.ma as ma

>>> x = ma.arange(4).reshape((2,2))
>>> x[1, 1] = ma.masked
>>>> x
masked_array(data =
[[0 1]
[2 --]],
mask =
[[False False]
[False True]],
fill_value = 999999)
>>> ma.transpose(x)
masked_array(data =
[[0 2]
[1 --]],
mask =
[[False False]
[False True]],
fill_value = 999999)

trace(offset=0, axis1=0, axis2=1, dtype=None, out=None)

Return the sum along diagonals of the array.
Refer to numpy.trace for full documentation.
See Also:

numpy.trace
equivalent function

transpose(*axes)
Returns a view of the array with axes transposed.
For a 1-D array, this has no effect. (To change between column and row vectors, first cast the 1-D array into a ma-
trix object.) For a 2-D array, this is the usual matrix transpose. For an n-D array, if axes are given, their order in-
dicates how the axes are permuted (see Examples). If axes are not provided and a.shape = (i[0], i[1],
... i[n-2], i[n-1]), then a.transpose().shape = (i[n-1], i[n-2], ... i[1],
i[0]).
Parameters
axes : None, tuple of ints, or n ints
None or no argument: reverses the order of the axes.
tuple of ints: i in the j-th place in the tuple means as i-th axis becomes a.transpose()s j-th
axis.
n ints: same as an n-tuple of the same ints (this form is intended simply as a convenience
alternative to the tuple form)
Returns
out : ndarray
View of a, with axes suitably permuted.

930 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:

ndarray.T
Array property returning the array transposed.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> a
array([[1, 2],
[3, 4]])
>>> a.transpose()
array([[1, 3],
[2, 4]])
>>> a.transpose((1, 0))
array([[1, 3],
[2, 4]])
>>> a.transpose(1, 0)
array([[1, 3],
[2, 4]])

Polynomial fit

ma.vander(x[, n]) Generate a Van der Monde matrix.

ma.polyfit(x, y, deg[, rcond, full]) Least squares polynomial fit.
vander(x, n=None)
Generate a Van der Monde matrix.
The columns of the output matrix are decreasing powers of the input vector. Specifically, the i-th output column
is the input vector to the power of N - i - 1. Such a matrix with a geometric progression in each row is
named Van Der Monde, or Vandermonde matrix, from Alexandre-Theophile Vandermonde.
Parameters
x : array_like
1-D input array.
N : int, optional
Order of (number of columns in) the output. If N is not specified, a square array is
returned (N = len(x)).
Returns
out : ndarray
Van der Monde matrix of order N. The first column is x^(N-1), the second x^(N-2)
and so forth.

Notes
Masked values in the input array result in rows of zeros.

References
[R61]

3.20. Masked array operations 931

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.array([1, 2, 3, 5])

>>> N = 3
>>> np.vander(x, N)
array([[ 1, 1, 1],
[ 4, 2, 1],
[ 9, 3, 1],
[25, 5, 1]])

>>> np.column_stack([x**(N-1-i) for i in range(N)])

array([[ 1, 1, 1],
[ 4, 2, 1],
[ 9, 3, 1],
[25, 5, 1]])

>>> x = np.array([1, 2, 3, 5])

>>> np.vander(x)
array([[ 1, 1, 1, 1],
[ 8, 4, 2, 1],
[ 27, 9, 3, 1],
[125, 25, 5, 1]])

The determinant of a square Vandermonde matrix is the product of the differences between the values of the
input vector:

>>> np.linalg.det(np.vander(x))
48.000000000000043
>>> (5-3)*(5-2)*(5-1)*(3-2)*(3-1)*(2-1)
48

polyfit(x, y, deg, rcond=None, full=False)

Least squares polynomial fit.
Fit a polynomial p(x) = p[0] * x**deg + ... + p[deg] of degree deg to points (x, y). Returns a
vector of coefficients p that minimises the squared error.
Parameters
x : array_like, shape (M,)
x-coordinates of the M sample points (x[i], y[i]).
y : array_like, shape (M,) or (M, K)
y-coordinates of the sample points. Several data sets of sample points sharing the same
x-coordinates can be fitted at once by passing in a 2D-array that contains one dataset
per column.
deg : int
Degree of the fitting polynomial
rcond : float, optional
Relative condition number of the fit. Singular values smaller than this relative to the
largest singular value will be ignored. The default value is len(x)*eps, where eps is the
relative precision of the float type, about 2e-16 in most cases.
full : bool, optional

932 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Switch determining nature of return value. When it is False (the default) just the coef-
ficients are returned, when True diagnostic information from the singular value decom-
position is also returned.
Returns
p : ndarray, shape (M,) or (M, K)
Polynomial coefficients, highest power first. If y was 2-D, the coefficients for k-th data
set are in p[:,k].
residuals, rank, singular_values, rcond : present only if full = True
Residuals of the least-squares fit, the effective rank of the scaled Vandermonde coeffi-
cient matrix, its singular values, and the specified value of rcond. For more details, see
linalg.lstsq.
See Also:

polyval
Computes polynomial values.
linalg.lstsq
Computes a least-squares fit.
scipy.interpolate.UnivariateSpline
Computes spline fits.

Notes
Any masked values in x is propagated in y, and vice-versa.

References
[R59], [R60]

Examples

>>> x = np.array([0.0, 1.0, 2.0, 3.0, 4.0, 5.0])

>>> y = np.array([0.0, 0.8, 0.9, 0.1, -0.8, -1.0])
>>> z = np.polyfit(x, y, 3)
>>> z
array([ 0.08703704, -0.81349206, 1.69312169, -0.03968254])

It is convenient to use poly1d objects for dealing with polynomials:

>>> p = np.poly1d(z)
>>> p(0.5)
0.6143849206349179
>>> p(3.5)
-0.34732142857143039
>>> p(10)
22.579365079365115

High-order polynomials may oscillate wildly:

>>> p30 = np.poly1d(np.polyfit(x, y, 30))

/... RankWarning: Polyfit may be poorly conditioned...
>>> p30(4)
-0.80000000000000204
>>> p30(5)

3.20. Masked array operations 933

NumPy Reference, Release 2.0.0.dev8464

-0.99999999999999445
>>> p30(4.5)
-0.10547061179440398

Illustration:

>>> import matplotlib.pyplot as plt

>>> xp = np.linspace(-2, 6, 100)
>>> plt.plot(x, y, ., xp, p(xp), -, xp, p30(xp), --)
[<matplotlib.lines.Line2D object at 0x...>, <matplotlib.lines.Line2D object at 0x...>, <matplotl
>>> plt.ylim(-2,2)
(-2, 2)
>>> plt.show()

2.0
1.5
1.0
0.5
0.0
0.5
1.0
1.5
2.0
2 1 0 1 2 3 4 5 6

Clipping and rounding

ma.around Round an array to the given number of decimals.

ma.clip(a, a_min, a_max[, out]) Clip (limit) the values in an array.
ma.round(a[, decimals, out]) Return a copy of a, rounded to decimals places.
ma.MaskedArray.clip(a_min, a_max[, out]) Return an array whose values are limited to [a_min, a_max].
ma.MaskedArray.round([decimals, out]) Return an array rounded a to the given number of decimals.
around
Round an array to the given number of decimals.
Refer to around for full documentation.
See Also:

around
equivalent function

clip(a, a_min, a_max, out=None)

Clip (limit) the values in an array.
Given an interval, values outside the interval are clipped to the interval edges. For example, if an interval of
[0, 1] is specified, values smaller than 0 become 0, and values larger than 1 become 1.

934 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
a : array_like
Array containing elements to clip.
a_min : scalar or array_like
Minimum value.
a_max : scalar or array_like
Maximum value. If a_min or a_max are array_like, then they will be broadcasted to the
shape of a.
out : ndarray, optional
The results will be placed in this array. It may be the input array for in-place clipping.
out must be of the right shape to hold the output. Its type is preserved.
Returns
clipped_array : ndarray
An array with the elements of a, but where values < a_min are replaced with a_min, and
those > a_max with a_max.
See Also:

numpy.doc.ufuncs
Section Output arguments

Examples

>>> a = np.arange(10)
>>> np.clip(a, 1, 8)
array([1, 1, 2, 3, 4, 5, 6, 7, 8, 8])
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> np.clip(a, 3, 6, out=a)
array([3, 3, 3, 3, 4, 5, 6, 6, 6, 6])
>>> a = np.arange(10)
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> np.clip(a, [3,4,1,1,1,4,4,4,4,4], 8)
array([3, 4, 2, 3, 4, 5, 6, 7, 8, 8])

round(a, decimals=0, out=None)

Return a copy of a, rounded to decimals places.
When decimals is negative, it specifies the number of positions to the left of the decimal point. The real and
imaginary parts of complex numbers are rounded separately. Nothing is done if the array is not of float type and
decimals is greater than or equal to 0.
Parameters
decimals : int
Number of decimals to round to. May be negative.
out : array_like
Existing array to use for output. If not given, returns a default copy of a.

3.20. Masked array operations 935

NumPy Reference, Release 2.0.0.dev8464

Notes
If out is given and does not have a mask attribute, the mask of a is lost!
clip(a_min, a_max, out=None)
Return an array whose values are limited to [a_min, a_max].
Refer to numpy.clip for full documentation.
See Also:

numpy.clip
equivalent function

round(decimals=0, out=None)
Return an array rounded a to the given number of decimals.
Refer to numpy.around for full documentation.
See Also:

numpy.around
equivalent function

Miscellanea

ma.allequal(a, b[, fill_value]) Return True if all entries of a and b are equal, using
ma.allclose(a, b[, masked_equal, rtol, Returns True if two arrays are element-wise equal within a
...]) tolerance.
ma.apply_along_axis(func1d, axis, Apply a function to 1-D slices along the given axis.
arr, ...)
ma.arange([dtype]) Return evenly spaced values within a given interval.
ma.choose(indices, choices[, out, mode]) Use an index array to construct a new array from a set of choices.
ma.ediff1d(arr[, to_end, to_begin]) Compute the differences between consecutive elements of an array.
ma.indices(dimensions[, dtype]) Return an array representing the indices of a grid.
ma.where(condition[, x, y]) Return a masked array with elements from x or y, depending on
condition.
allequal(a, b, fill_value=True)
Return True if all entries of a and b are equal, using fill_value as a truth value where either or both are masked.
Parameters
a, b : array_like
Input arrays to compare.
fill_value : bool, optional
Whether masked values in a or b are considered equal (True) or not (False).
Returns
y : bool
Returns True if the two arrays are equal within the given tolerance, False otherwise. If
either array contains NaN, then False is returned.
See Also:
all, any, numpy.ma.allclose

936 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = ma.array([1e10, 1e-7, 42.0], mask=[0, 0, 1])

>>> a
masked_array(data = [10000000000.0 1e-07 --],
mask = [False False True],
fill_value=1e+20)

>>> b = array([1e10, 1e-7, -42.0])

>>> b
array([ 1.00000000e+10, 1.00000000e-07, -4.20000000e+01])
>>> ma.allequal(a, b, fill_value=False)
False
>>> ma.allequal(a, b)
True

allclose(a, b, masked_equal=True, rtol=1.0000000000000001e-05, atol=1e-08, fill_value=None)

Returns True if two arrays are element-wise equal within a tolerance.
This function is equivalent to allclose except that masked values are treated as equal (default) or unequal,
depending on the masked_equal argument.
Parameters
a, b : array_like
Input arrays to compare.
masked_equal : bool, optional
Whether masked values in a and b are considered equal (True) or not (False). They are
considered equal by default.
rtol : float, optional
Relative tolerance. The relative difference is equal to rtol * b. Default is 1e-5.
atol : float, optional
Absolute tolerance. The absolute difference is equal to atol. Default is 1e-8.
fill_value : bool, optional
Deprecated - Whether masked values in a or b are considered equal (True) or not
(False).
Returns
y : bool
Returns True if the two arrays are equal within the given tolerance, False otherwise. If
either array contains NaN, then False is returned.
See Also:
all, any
numpy.allclose
the non-masked allclose.

Notes
If the following equation is element-wise True, then allclose returns True:

3.20. Masked array operations 937

NumPy Reference, Release 2.0.0.dev8464

absolute(a - b) <= (atol + rtol * absolute(b))

Return True if all elements of a and b are equal subject to given tolerances.

Examples

>>> a = ma.array([1e10, 1e-7, 42.0], mask=[0, 0, 1])

>>> a
masked_array(data = [10000000000.0 1e-07 --],
mask = [False False True],
fill_value = 1e+20)
>>> b = ma.array([1e10, 1e-8, -42.0], mask=[0, 0, 1])
>>> ma.allclose(a, b)
False

>>> a = ma.array([1e10, 1e-8, 42.0], mask=[0, 0, 1])

>>> b = ma.array([1.00001e10, 1e-9, -42.0], mask=[0, 0, 1])
>>> ma.allclose(a, b)
True
>>> ma.allclose(a, b, masked_equal=False)
False

Masked values are not compared directly.

>>> a = ma.array([1e10, 1e-8, 42.0], mask=[0, 0, 1])

>>> b = ma.array([1.00001e10, 1e-9, 42.0], mask=[0, 0, 1])
>>> ma.allclose(a, b)
True
>>> ma.allclose(a, b, masked_equal=False)
False

apply_along_axis(func1d, axis, arr, *args, **kwargs)

Apply a function to 1-D slices along the given axis.
Execute func1d(a, *args) where func1d operates on 1-D arrays and a is a 1-D slice of arr along axis.
Parameters
func1d : function
This function should accept 1-D arrays. It is applied to 1-D slices of arr along the
specified axis.
axis : integer
Axis along which arr is sliced.
arr : ndarray
Input array.
args : any
Additional arguments to func1d.
Returns
outarr : ndarray
The output array. The shape of outarr is identical to the shape of arr, except along the
axis dimension, where the length of outarr is equal to the size of the return value of
func1d. If func1d returns a scalar outarr will have one fewer dimensions than arr.

938 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:

apply_over_axes
Apply a function repeatedly over multiple axes.

Examples

>>> def my_func(a):

... """Average first and last element of a 1-D array"""
... return (a[0] + a[-1]) * 0.5
>>> b = np.array([[1,2,3], [4,5,6], [7,8,9]])
>>> np.apply_along_axis(my_func, 0, b)
array([ 4., 5., 6.])
>>> np.apply_along_axis(my_func, 1, b)
array([ 2., 5., 8.])

For a function that doesnt return a scalar, the number of dimensions in outarr is the same as arr.

>>> def new_func(a):

... """Divide elements of a by 2."""
... return a * 0.5
>>> b = np.array([[1,2,3], [4,5,6], [7,8,9]])
>>> np.apply_along_axis(new_func, 0, b)
array([[ 0.5, 1. , 1.5],
[ 2. , 2.5, 3. ],
[ 3.5, 4. , 4.5]])

arange
Return evenly spaced values within a given interval.
Values are generated within the half-open interval [start, stop) (in other words, the interval including
start but excluding stop). For integer arguments the function is equivalent to the Python built-in range function,
but returns a ndarray rather than a list.
Parameters
start : number, optional
Start of interval. The interval includes this value. The default start value is 0.
stop : number
End of interval. The interval does not include this value.
step : number, optional
Spacing between values. For any output out, this is the distance between two adjacent
values, out[i+1] - out[i]. The default step size is 1. If step is specified, start
must also be given.
dtype : dtype
The type of the output array. If dtype is not given, infer the data type from the other
input arguments.
Returns
out : ndarray
Array of evenly spaced values.

3.20. Masked array operations 939

NumPy Reference, Release 2.0.0.dev8464

For floating point arguments, the length of the result is ceil((stop -

start)/step). Because of floating point overflow, this rule may result in the last
element of out being greater than stop.
See Also:

linspace
Evenly spaced numbers with careful handling of endpoints.
ogrid
Arrays of evenly spaced numbers in N-dimensions
mgrid
Grid-shaped arrays of evenly spaced numbers in N-dimensions

Examples

>>> np.arange(3)
array([0, 1, 2])
>>> np.arange(3.0)
array([ 0., 1., 2.])
>>> np.arange(3,7)
array([3, 4, 5, 6])
>>> np.arange(3,7,2)
array([3, 5])

choose(indices, choices, out=None, mode=raise)

Use an index array to construct a new array from a set of choices.
Given an array of integers and a set of n choice arrays, this method will create a new array that merges each of
the choice arrays. Where a value in a is i, the new array will have the value that choices[i] contains in the same
place.
Parameters
a : ndarray of ints
This array must contain integers in [0, n-1], where n is the number of choices.
choices : sequence of arrays
Choice arrays. The index array and all of the choices should be broadcastable to the
same shape.
out : array, optional
If provided, the result will be inserted into this array. It should be of the appropriate
shape and dtype.
mode : {raise, wrap, clip}, optional
Specifies how out-of-bounds indices will behave.
raise : raise an error
wrap : wrap around
clip : clip to the range
Returns
merged_array : array
See Also:

940 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

choose
equivalent function

Examples

>>> choice = np.array([[1,1,1], [2,2,2], [3,3,3]])

>>> a = np.array([2, 1, 0])
>>> np.ma.choose(a, choice)
masked_array(data = [3 2 1],
mask = False,
fill_value=999999)

ediff1d(arr, to_end=None, to_begin=None)

Compute the differences between consecutive elements of an array.
This function is the equivalent of numpy.ediff1d that takes masked values into account, see
numpy.ediff1d for details.
See Also:

numpy.ediff1d
Equivalent function for ndarrays.

indices(dimensions, dtype=<type int>)

Return an array representing the indices of a grid.
Compute an array where the subarrays contain index values 0,1,... varying only along the corresponding axis.
Parameters
dimensions : sequence of ints
The shape of the grid.
dtype : dtype, optional
Data type of the result.
Returns
grid : ndarray
The array of grid indices, grid.shape = (len(dimensions),) +
tuple(dimensions).
See Also:
mgrid, meshgrid

Notes
The output shape is obtained by prepending the number of dimensions in front of the tuple of dimensions, i.e. if
dimensions is a tuple (r0, ..., rN-1) of length N, the output shape is (N,r0,...,rN-1).
The subarrays grid[k] contains the N-D array of indices along the k-th axis. Explicitly:

grid[k,i0,i1,...,iN-1] = ik

Examples

>>> grid = np.indices((2, 3))

>>> grid.shape
(2, 2, 3)

3.20. Masked array operations 941

NumPy Reference, Release 2.0.0.dev8464

>>> grid[0] # row indices

array([[0, 0, 0],
[1, 1, 1]])
>>> grid[1] # column indices
array([[0, 1, 2],
[0, 1, 2]])

The indices can be used as an index into an array.

>>> x = np.arange(20).reshape(5, 4)
>>> row, col = np.indices((2, 3))
>>> x[row, col]
array([[0, 1, 2],
[4, 5, 6]])

Note that it would be more straightforward in the above example to extract the required elements directly with
x[:2, :3].
where(condition, x=None, y=None)
Return a masked array with elements from x or y, depending on condition.
Returns a masked array, shaped like condition, where the elements are from x when condition is True, and from
y otherwise. If neither x nor y are given, the function returns a tuple of indices where condition is True (the
result of condition.nonzero()).
Parameters
condition : array_like, bool
The condition to meet. For each True element, yield the corresponding element from x,
otherwise from y.
x, y : array_like, optional
Values from which to choose. x and y need to have the same shape as condition, or be
broadcast-able to that shape.
Returns
out : MaskedArray or tuple of ndarrays
The resulting masked array if x and y were given, otherwise the result of
condition.nonzero().
See Also:

numpy.where
Equivalent function in the top-level NumPy module.

Examples

>>> x = np.ma.array(np.arange(9.).reshape(3, 3), mask=[[0, 1, 0],

... [1, 0, 1],
... [0, 1, 0]])
>>> print x
[[0.0 -- 2.0]
[-- 4.0 --]
[6.0 -- 8.0]]
>>> np.ma.where(x > 5) # return the indices where x > 5
(array([2, 2]), array([0, 2]))

942 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> print np.ma.where(x > 5, x, -3.1416)

[[-3.1416 -- -3.1416]
[-- -3.1416 --]
[6.0 -- 8.0]]

3.21 Numpy-specific help functions

3.21.1 Finding help

lookfor(what[, module, import_modules, ...]) Do a keyword search on docstrings.

lookfor(what, module=None, import_modules=True, regenerate=False, output=None)
Do a keyword search on docstrings.
A list of of objects that matched the search is displayed, sorted by relevance. All given keywords need to be
found in the docstring for it to be returned as a result, but the order does not matter.
Parameters
what : str
String containing words to look for.
module : str or list, optional
Name of module(s) whose docstrings to go through.
import_modules : bool, optional
Whether to import sub-modules in packages. Default is True.
regenerate : bool, optional
Whether to re-generate the docstring cache. Default is False.
output : file-like, optional
File-like object to write the output to. If omitted, use a pager.
See Also:
source, info

Notes
Relevance is determined only roughly, by checking if the keywords occur in the function name, at the start of a
docstring, etc.

Examples

>>> np.lookfor(binary representation)

Search results for binary representation
------------------------------------------
numpy.binary_repr
Return the binary representation of the input number as a string.
numpy.core.setup_common.long_double_representation
Given a binary dump as given by GNU od -b, look for long double
numpy.base_repr
Return a string representation of a number in the given base system.
...

3.21. Numpy-specific help functions 943

NumPy Reference, Release 2.0.0.dev8464

3.21.2 Reading help

info([object, maxwidth, output, toplevel]) Get help information for a function, class, or module.
source(object[, output]) Print or write to a file the source code for a Numpy object.
info(object=None, maxwidth=76, output=<open file <stdout>, mode w at 0x2aaaaaac9198>,
toplevel=numpy)
Get help information for a function, class, or module.
Parameters
object : object or str, optional
Input object or name to get information about. If object is a numpy object, its docstring
is given. If it is a string, available modules are searched for matching objects. If None,
information about info itself is returned.
maxwidth : int, optional
Printing width.
output : file like object, optional
File like object that the output is written to, default is stdout. The object has to be
opened in w or a mode.
toplevel : str, optional
Start search at this level.
See Also:
source, lookfor

Notes
When used interactively with an object, np.info(obj) is equivalent to help(obj) on the Python prompt
or obj? on the IPython prompt.

Examples

>>> np.info(np.polyval)
polyval(p, x)
Evaluate the polynomial p at x.
...

When using a string for object it is possible to get multiple results.

>>> np.info(fft)
*** Found in numpy ***
Core FFT routines
...
*** Found in numpy.fft ***
fft(a, n=None, axis=-1)
...
*** Repeat reference found in numpy.fft.fftpack ***
*** Total of 3 references found. ***

source(object, output=<open file <stdout>, mode w at 0x2aaaaaac9198>)

Print or write to a file the source code for a Numpy object.
The source code is only returned for objects written in Python. Many functions and classes are defined in C and
will therefore not return useful information.

944 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
object : numpy object
Input object. This can be any object (function, class, module, ...).
output : file object, optional
If output not supplied then source code is printed to screen (sys.stdout). File object must
be created with either write w or append a modes.
See Also:
lookfor, info

Examples

>>> np.source(np.interp)
In file: /usr/lib/python2.6/dist-packages/numpy/lib/function_base.py
def interp(x, xp, fp, left=None, right=None):
""".... (full docstring printed)"""
if isinstance(x, (float, int, number)):
return compiled_interp([x], xp, fp, left, right).item()
else:
return compiled_interp(x, xp, fp, left, right)

The source code is only returned for objects written in Python.

>>> np.source(np.array)
Not available for this object.

3.22 Miscellaneous routines

3.22.1 Buffer objects

getbuffer() Create a buffer object from the given object referencing a slice of length size starting at offset.
newbuffer(size) Return a new uninitialized buffer object of size bytes
getbuffer(obj, [offset, [size]])
Create a buffer object from the given object referencing a slice of length size starting at offset.
Default is the entire buffer. A read-write buffer is attempted followed by a read-only buffer.
Parameters
obj : object
offset : int, optional
size : int, optional
Returns
buffer_obj : buffer

Examples

>>> buf = np.getbuffer(np.ones(5), 1, 3)

>>> len(buf)
3
>>> buf[0]

3.22. Miscellaneous routines 945

NumPy Reference, Release 2.0.0.dev8464

\x00
>>> buf
<read-write buffer for 0x8af1e70, size 3, offset 1 at 0x8ba4ec0>

newbuffer(size)
Return a new uninitialized buffer object of size bytes

3.22.2 Performance tuning

alterdot() Change dot, vdot, and innerproduct to use accelerated BLAS functions.
restoredot() Restore dot, vdot, and innerproduct to the default non-BLAS
setbufsize(size) Set the size of the buffer used in ufuncs.
getbufsize() Return the size of the buffer used in ufuncs.
alterdot()
Change dot, vdot, and innerproduct to use accelerated BLAS functions.
Typically, as a user of Numpy, you do not explicitly call this function. If Numpy is built with an accelerated
BLAS, this function is automatically called when Numpy is imported.
When Numpy is built with an accelerated BLAS like ATLAS, these functions are replaced to make use of the
faster implementations. The faster implementations only affect float32, float64, complex64, and complex128
arrays. Furthermore, the BLAS API only includes matrix-matrix, matrix-vector, and vector-vector products.
Products of arrays with larger dimensionalities use the built in functions and are not accelerated.
See Also:

restoredot
restoredot undoes the effects of alterdot.

restoredot()
Restore dot, vdot, and innerproduct to the default non-BLAS implementations.
Typically, the user will only need to call this when troubleshooting and installation problem, reproducing the
conditions of a build without an accelerated BLAS, or when being very careful about benchmarking linear
algebra operations.
See Also:

alterdot
restoredot undoes the effects of alterdot.

setbufsize(size)
Set the size of the buffer used in ufuncs.
Parameters
size : int
Size of buffer.
getbufsize()
Return the size of the buffer used in ufuncs.

3.23 Test Support (numpy.testing)

Common test support for all numpy test scripts.

946 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

This single module should provide all the common functionality for numpy tests in a single location, so that test scripts
can just import it and work right away.

3.24 Asserts

assert_almost_equal(actual,Raise an assertion if two items are not equal up to desired precision.

desired[, ...])
assert_approx_equal(actual,Raise an assertion if two items are not equal up to significant digits.
desired[, ...])
assert_array_almost_equal(x, Raise an assertion if two objects are not equal up to desired precision.
y[, decimal, ...])
assert_array_equal(x, Raise an assertion if two array_like objects are not equal.
y[, err_msg, verbose])
assert_array_less(x, y[, Raise an assertion if two array_like objects are not ordered by less than.
err_msg, verbose])
assert_equal(actual, Raise an assertion if two objects are not equal.
desired[, err_msg, verbose])
Fail unless an exception of class exception_class is thrown by callable when
assert_raises(exception_class,
callable, ...) invoked with arguments args and keyword arguments kwargs.
assert_warns(warning_class, Fail unless the given callable throws the specified warning.
func, *args, **kw)
assert_string_equal(actual,Test if two strings are equal.
desired)
assert_almost_equal(actual, desired, decimal=7, err_msg=, verbose=True)
Raise an assertion if two items are not equal up to desired precision.
The test is equivalent to abs(desired-actual) < 0.5 * 10**(-decimal)
Given two objects (numbers or ndarrays), check that all elements of these objects are almost equal. An exception
is raised at conflicting values. For ndarrays this delegates to assert_array_almost_equal
Parameters
actual : number or ndarray
The object to check.
desired : number or ndarray
The expected object.
decimal : integer (decimal=7)
desired precision
err_msg : string
The error message to be printed in case of failure.
verbose : bool
If True, the conflicting values are appended to the error message.
Raises
AssertionError :
If actual and desired are not equal up to specified precision.
See Also:

3.24. Asserts 947

NumPy Reference, Release 2.0.0.dev8464

assert_array_almost_equal
compares array_like objects
assert_equal
tests objects for equality

Examples

>>> import numpy.testing as npt

>>> npt.assert_almost_equal(2.3333333333333, 2.33333334)
>>> npt.assert_almost_equal(2.3333333333333, 2.33333334, decimal=10)
...
<type exceptions.AssertionError>:
Items are not equal:
ACTUAL: 2.3333333333333002
DESIRED: 2.3333333399999998

>>> npt.assert_almost_equal(np.array([1.0,2.3333333333333]),
np.array([1.0,2.33333334]), decimal=9)
...
<type exceptions.AssertionError>:
Arrays are not almost equal
<BLANKLINE>
(mismatch 50.0%)
x: array([ 1. , 2.33333333])
y: array([ 1. , 2.33333334])

assert_approx_equal(actual, desired, significant=7, err_msg=, verbose=True)

Raise an assertion if two items are not equal up to significant digits.
Given two numbers, check that they are approximately equal. Approximately equal is defined as the number of
significant digits that agree.
Parameters
actual : number
The object to check.
desired : number
The expected object.
significant : integer (significant=7)
desired precision
err_msg : string
The error message to be printed in case of failure.
verbose : bool
If True, the conflicting values are appended to the error message.
Raises
AssertionError :
If actual and desired are not equal up to specified precision.
See Also:

948 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

assert_almost_equal
compares objects by decimals
assert_array_almost_equal
compares array_like objects by decimals
assert_equal
tests objects for equality

Examples

>>> np.testing.assert_approx_equal(0.12345677777777e-20, 0.1234567e-20)

>>> np.testing.assert_approx_equal(0.12345670e-20, 0.12345671e-20,
significant=8)
>>> np.testing.assert_approx_equal(0.12345670e-20, 0.12345672e-20,
significant=8)
...
<type exceptions.AssertionError>:
Items are not equal to 8 significant digits:
ACTUAL: 1.234567e-021
DESIRED: 1.2345672000000001e-021

the evaluated condition that raises the exception is

>>> abs(0.12345670e-20/1e-21 - 0.12345672e-20/1e-21) >= 10**-(8-1)

True

assert_array_almost_equal(x, y, decimal=6, err_msg=, verbose=True)

Raise an assertion if two objects are not equal up to desired precision.
The test verifies identical shapes and verifies values with abs(desired-actual) < 0.5 * 10**(-decimal)
Given two array_like objects, check that the shape is equal and all elements of these objects are almost equal.
An exception is raised at shape mismatch or conflicting values. In contrast to the standard usage in numpy,
NaNs are compared like numbers, no assertion is raised if both objects have NaNs in the same positions.
Parameters
x : array_like
The actual object to check.
y : array_like
The desired, expected object.
decimal : integer (decimal=6)
desired precision
err_msg : string
The error message to be printed in case of failure.
verbose : bool
If True, the conflicting values are appended to the error message.
Raises
AssertionError :
If actual and desired are not equal up to specified precision.
See Also:

3.24. Asserts 949

NumPy Reference, Release 2.0.0.dev8464

assert_almost_equal
simple version for comparing numbers
assert_array_equal
tests objects for equality

Examples
the first assert does not raise an exception

>>> np.testing.assert_array_almost_equal([1.0,2.333,np.nan],
[1.0,2.333,np.nan])

>>> np.testing.assert_array_almost_equal([1.0,2.33333,np.nan],
[1.0,2.33339,np.nan], decimal=5)
...
<type exceptions.AssertionError>:
AssertionError:
Arrays are not almost equal
<BLANKLINE>
(mismatch 50.0%)
x: array([ 1. , 2.33333, NaN])
y: array([ 1. , 2.33339, NaN])

>>> np.testing.assert_array_almost_equal([1.0,2.33333,np.nan],
[1.0,2.33333, 5], decimal=5)
<type exceptions.ValueError>:
ValueError:
Arrays are not almost equal
x: array([ 1. , 2.33333, NaN])
y: array([ 1. , 2.33333, 5. ])

assert_array_equal(x, y, err_msg=, verbose=True)

Raise an assertion if two array_like objects are not equal.
Given two array_like objects, check that the shape is equal and all elements of these objects are equal. An
exception is raised at shape mismatch or conflicting values. In contrast to the standard usage in numpy, NaNs
are compared like numbers, no assertion is raised if both objects have NaNs in the same positions.
The usual caution for verifying equality with floating point numbers is advised.
Parameters
x : array_like
The actual object to check.
y : array_like
The desired, expected object.
err_msg : string
The error message to be printed in case of failure.
verbose : bool
If True, the conflicting values are appended to the error message.
Raises
AssertionError :

950 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

If actual and desired objects are not equal.

See Also:

assert_array_almost_equal
test objects for equality up to precision
assert_equal
tests objects for equality

Examples
the first assert does not raise an exception

>>> np.testing.assert_array_equal([1.0,2.33333,np.nan],
[np.exp(0),2.33333, np.nan])

assert fails with numerical inprecision with floats

>>> np.testing.assert_array_equal([1.0,np.pi,np.nan],
[1, np.sqrt(np.pi)**2, np.nan])
...
<type exceptions.ValueError>:
AssertionError:
Arrays are not equal
<BLANKLINE>
(mismatch 50.0%)
x: array([ 1. , 3.14159265, NaN])
y: array([ 1. , 3.14159265, NaN])

use assert_array_almost_equal for these cases instead

>>> np.testing.assert_array_almost_equal([1.0,np.pi,np.nan],
[1, np.sqrt(np.pi)**2, np.nan], decimal=15)

assert_array_less(x, y, err_msg=, verbose=True)

Raise an assertion if two array_like objects are not ordered by less than.
Given two array_like objects, check that the shape is equal and all elements of the first object are strictly smaller
than those of the second object. An exception is raised at shape mismatch or incorrectly ordered values. Shape
mismatch does not raise if an object has zero dimension. In contrast to the standard usage in numpy, NaNs are
compared, no assertion is raised if both objects have NaNs in the same positions.
Parameters
x : array_like
The smaller object to check.
y : array_like
The larger object to compare.
err_msg : string
The error message to be printed in case of failure.
verbose : bool
If True, the conflicting values are appended to the error message.
Raises
AssertionError :

3.24. Asserts 951

NumPy Reference, Release 2.0.0.dev8464

If actual and desired objects are not equal.

See Also:

assert_array_equal
tests objects for equality
assert_array_almost_equal
test objects for equality up to precision

Examples

>>> np.testing.assert_array_less([1.0, 1.0, np.nan], [1.1, 2.0, np.nan])

>>> np.testing.assert_array_less([1.0, 1.0, np.nan], [1, 2.0, np.nan])
...
<type exceptions.ValueError>:
Arrays are not less-ordered
(mismatch 50.0%)
x: array([ 1., 1., NaN])
y: array([ 1., 2., NaN])

>>> np.testing.assert_array_less([1.0, 4.0], 3)

...
<type exceptions.ValueError>:
Arrays are not less-ordered
(mismatch 50.0%)
x: array([ 1., 4.])
y: array(3)

>>> np.testing.assert_array_less([1.0, 2.0, 3.0], [4])

...
<type exceptions.ValueError>:
Arrays are not less-ordered
(shapes (3,), (1,) mismatch)
x: array([ 1., 2., 3.])
y: array([4])

assert_equal(actual, desired, err_msg=, verbose=True)

Raise an assertion if two objects are not equal.
Given two objects (lists, tuples, dictionaries or numpy arrays), check that all elements of these objects are equal.
An exception is raised at the first conflicting values.
Parameters
actual : list, tuple, dict or ndarray
The object to check.
desired : list, tuple, dict or ndarray
The expected object.
err_msg : string
The error message to be printed in case of failure.
verbose : bool
If True, the conflicting values are appended to the error message.

952 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Raises
AssertionError :
If actual and desired are not equal.

Examples

>>> np.testing.assert_equal([4,5], [4,6])

...
<type exceptions.AssertionError>:
Items are not equal:
item=1
ACTUAL: 5
DESIRED: 6

assert_raises(exception_class, callable, *args, **kwargs)

Fail unless an exception of class exception_class is thrown by callable when invoked with arguments args and
keyword arguments kwargs. If a different type of exception is thrown, it will not be caught, and the test case
will be deemed to have suffered an error, exactly as for an unexpected exception.
assert_warns(warning_class, func, *args, **kw)
Fail unless the given callable throws the specified warning.
A warning of class warning_class should be thrown by the callable when invoked with arguments args and
keyword arguments kwargs. If a different type of warning is thrown, it will not be caught, and the test case will
be deemed to have suffered an error.
Parameters
warning_class : class
The class defining the warning that func is expected to throw.
func : callable
The callable to test.
*args : Arguments
Arguments passed to func.
**kwargs : Kwargs
Keyword arguments passed to func.
Returns
None :
assert_string_equal(actual, desired)
Test if two strings are equal.
If the given strings are equal, assert_string_equal does nothing. If they are not equal, an AssertionError is
raised, and the diff between the strings is shown.
Parameters
actual : str
The string to test for equality against the expected string.
desired : str
The expected string.

3.24. Asserts 953

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> np.testing.assert_string_equal(abc, abc)

>>> np.testing.assert_string_equal(abc, abcd)
...
AssertionError: Differences in strings:
- abc+ abcd? +

3.24.1 Decorators

decorators.deprecated([conditional]) Filter deprecation warnings while running the test suite.

decorators.knownfailureif(fail_condition[,Make function raise KnownFailureTest exception if given
msg]) condition is true.
decorators.setastest([tf]) Signals to nose that this function is or is not a test.
decorators.skipif(skip_condition[, msg]) Make function raise SkipTest exception if a given condition is
true.
decorators.slow(t) Label a test as slow.
decorate_methods(cls, decorator[, Apply a decorator to all methods in a class matching a regular
testmatch]) expression.
deprecated(conditional=True)
Filter deprecation warnings while running the test suite.
This decorator can be used to filter DeprecationWarnings, to avoid printing them during the test suite run, while
checking that the test actually raises a DeprecationWarning.
Parameters
conditional : bool or callable, optional
Flag to determine whether to mark test as deprecated or not. If the condition is a callable,
it is used at runtime to dynamically make the decision. Default is True.
Returns
decorator : function
The deprecated decorator itself.

Notes
New in version 1.4.0.
knownfailureif(fail_condition, msg=None)
Make function raise KnownFailureTest exception if given condition is true.
If the condition is a callable, it is used at runtime to dynamically make the decision. This is useful for tests that
may require costly imports, to delay the cost until the test suite is actually executed.
Parameters
fail_condition : bool or callable
Flag to determine whether to mark the decorated test as a known failure (if True) or not
(if False).
msg : str, optional
Message to give on raising a KnownFailureTest exception. Default is None.
Returns
decorator : function

954 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Decorator, which, when applied to a function, causes SkipTest to be raised when

skip_condition is True, and the function to be called normally otherwise.

Notes
The decorator itself is decorated with the nose.tools.make_decorator function in order to transmit
function name, and various other metadata.
setastest(tf=True)
Signals to nose that this function is or is not a test.
Parameters
tf : bool
If True, specifies that the decorated callable is a test. If False, specifies that the decorated
callable is not a test. Default is True.

Notes
This decorator cant use the nose namespace, because it can be called from a non-test module. See also istest
and nottest in nose.tools.

Examples
setastest can be used in the following way:

from numpy.testing.decorators import setastest

@setastest(False)
def func_with_test_in_name(arg1, arg2):
pass

skipif(skip_condition, msg=None)
Make function raise SkipTest exception if a given condition is true.
If the condition is a callable, it is used at runtime to dynamically make the decision. This is useful for tests that
may require costly imports, to delay the cost until the test suite is actually executed.
Parameters
skip_condition : bool or callable
Flag to determine whether to skip the decorated test.
msg : str, optional
Message to give on raising a SkipTest exception. Default is None.
Returns
decorator : function
Decorator which, when applied to a function, causes SkipTest to be raised when
skip_condition is True, and the function to be called normally otherwise.

Notes
The decorator itself is decorated with the nose.tools.make_decorator function in order to transmit
function name, and various other metadata.
slow(t)
Label a test as slow.

3.24. Asserts 955

NumPy Reference, Release 2.0.0.dev8464

The exact definition of a slow test is obviously both subjective and hardware-dependent, but in general any
individual test that requires more than a second or two should be labeled as slow (the whole suite consits of
thousands of tests, so even a second is significant).
Parameters
t : callable
The test to label as slow.
Returns
t : callable
The decorated test t.

Examples
The numpy.testing module includes import decorators as dec. A test can be decorated as slow
like this:

from numpy.testing import *

@dec.slow
def test_big(self):
print Big, slow test

decorate_methods(cls, decorator, testmatch=None)

Apply a decorator to all methods in a class matching a regular expression.
The given decorator is applied to all public methods of cls that are matched by the regular expression testmatch
(testmatch.search(methodname)). Methods that are private, i.e. start with an underscore, are ignored.
Parameters
cls : class
Class whose methods to decorate.
decorator : function
Decorator to apply to methods
testmatch : compiled regexp or str, optional
The regular expression. Default value is None, in which case the nose default
(re.compile(r(?:^|[\b_\.%s-])[Tt]est % os.sep)) is used. If
testmatch is a string, it is compiled to a regular expression first.

3.24.2 Test Running

Tester Nose test runner.

run_module_suite([file_to_run])
rundocs([filename, raise_on_error]) Run doctests found in the given file.
Tester
alias of NoseTester
run_module_suite(file_to_run=None)

rundocs(filename=None, raise_on_error=True)
Run doctests found in the given file.
By default rundocs raises an AssertionError on failure.

956 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
filename : str
The path to the file for which the doctests are run.
raise_on_error : bool
Whether to raise an AssertionError when a doctest fails. Default is True.

Notes
The doctests can be run by the user/developer by adding the doctests argument to the test() call. For
example, to run all tests (including doctests) for numpy.lib:

>>> np.lib.test(doctests=True)

3.25 Mathematical functions with automatic domain (numpy.emath)

Note: numpy.emath is a preferred alias for numpy.lib.scimath, available after numpy is imported.
Wrapper functions to more user-friendly calling of certain math functions whose output data-type is
different than the input data-type in certain domains of the input.
For example, for functions like log() with branch cuts, the versions in this module provide the mathematically valid
answers in the complex plane:

>>> import math

>>> from numpy.lib import scimath
>>> scimath.log(-math.exp(1)) == (1+1j*math.pi)
True

Similarly, sqrt(), other base logarithms, power() and trig functions are correctly handled. See their respective docstrings
for specific examples.

3.26 Matrix library (numpy.matlib)

This module contains all functions in the numpy namespace, with the following replacement functions that return
matrices instead of ndarrays.

3.27 Optionally Scipy-accelerated routines (numpy.dual)

Aliases for functions which may be accelerated by Scipy.

Scipy can be built to use accelerated or otherwise improved libraries for FFTs, linear algebra, and special functions.
This module allows developers to transparently support these accelerated functions when scipy is available but still
support users who have only installed Numpy.

3.25. Mathematical functions with automatic domain (numpy.emath) 957

NumPy Reference, Release 2.0.0.dev8464

3.27.1 Linear algebra

cholesky(a) Cholesky decomposition.

det(a) Compute the determinant of an array.
eig(a) Compute the eigenvalues and right eigenvectors of a square array.
eigh(a[, UPLO]) Return the eigenvalues and eigenvectors of a Hermitian or symmetric matrix.
eigvals(a) Compute the eigenvalues of a general matrix.
eigvalsh(a[, UPLO]) Compute the eigenvalues of a Hermitian or real symmetric matrix.
inv(a) Compute the (multiplicative) inverse of a matrix.
lstsq(a, b[, rcond]) Return the least-squares solution to a linear matrix equation.
norm(x[, ord]) Matrix or vector norm.
pinv(a[, rcond]) Compute the (Moore-Penrose) pseudo-inverse of a matrix.
solve(a, b) Solve a linear matrix equation, or system of linear scalar equations.
svd(a[, full_matrices, compute_uv]) Singular Value Decomposition.

3.27.2 FFT

fft(a[, n, axis]) Compute the one-dimensional discrete Fourier Transform.

fft2(a[, s, axes]) Compute the 2-dimensional discrete Fourier Transform
fftn(a[, s, axes]) Compute the N-dimensional discrete Fourier Transform.
ifft(a[, n, axis]) Compute the one-dimensional inverse discrete Fourier Transform.
ifft2(a[, s, axes]) Compute the 2-dimensional inverse discrete Fourier Transform.
ifftn(a[, s, axes]) Compute the N-dimensional inverse discrete Fourier Transform.

3.27.3 Other

i0(x) Modified Bessel function of the first kind, order 0.

3.28 Numarray compatibility (numpy.numarray)

3.29 Old Numeric compatibility (numpy.oldnumeric)

3.30 C-Types Foreign Function Interface (numpy.ctypeslib)

as_array(*args, **kwds)
Dummy object that raises an ImportError if ctypes is not available.
Raises
ImportError :
If ctypes is not available.
as_ctypes(*args, **kwds)
Dummy object that raises an ImportError if ctypes is not available.
Raises
ImportError :
If ctypes is not available.

958 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

ctypes_load_library(*args, **kwds)
Dummy object that raises an ImportError if ctypes is not available.
Raises
ImportError :
If ctypes is not available.
load_library(*args, **kwds)
Dummy object that raises an ImportError if ctypes is not available.
Raises
ImportError :
If ctypes is not available.
ndpointer(dtype=None, ndim=None, shape=None, flags=None)
Array-checking restype/argtypes.
An ndpointer instance is used to describe an ndarray in restypes and argtypes specifications. This approach
is more flexible than using, for example, POINTER(c_double), since several restrictions can be specified,
which are verified upon calling the ctypes function. These include data type, number of dimensions, shape and
flags. If a given array does not satisfy the specified restrictions, a TypeError is raised.
Parameters
dtype : data-type, optional
Array data-type.
ndim : int, optional
Number of array dimensions.
shape : tuple of ints, optional
Array shape.
flags : str or tuple of str
Array flags; may be one or more of:
C_CONTIGUOUS / C / CONTIGUOUS
F_CONTIGUOUS / F / FORTRAN
OWNDATA / O
WRITEABLE / W
ALIGNED / A
UPDATEIFCOPY / U
Returns
klass : ndpointer type object
A type object, which is an _ndtpr instance containing dtype, ndim, shape and flags
information.
Raises
TypeError :
If a given array does not satisfy the specified restrictions.

3.30. C-Types Foreign Function Interface (numpy.ctypeslib) 959

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> clib.somefunc.argtypes = [np.ctypeslib.ndpointer(dtype=np.float64,

... ndim=1,
... flags=C_CONTIGUOUS)]
...
>>> clib.somefunc(np.array([1, 2, 3], dtype=np.float64))
...

3.31 String operations

This module provides a set of vectorized string operations for arrays of type numpy.string_ or
numpy.unicode_. All of them are based on the string methods in the Python standard library.

960 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

3.31.1 String operations

add(x1, x2) Return (x1 + x2), that is string concatenation, element-wise for a pair of array_likes of str
or unicode.
multiply(a, i) Return (a * i), that is string multiple concatenation, element-wise.
mod(a, values) Return (a % i), that is pre-Python 2.6 string formatting
capitalize(a) Return a copy of a with only the first character of each element capitalized.
center(a, width[, Return a copy of a with its elements centered in a string of length width.
fillchar])
decode(a[, Calls str.decode element-wise.
encoding, errors])
encode(a[, Calls str.encode element-wise.
encoding, errors])
join(sep, seq) Return a string which is the concatenation of the strings in the sequence seq.
ljust(a, width[, Return an array with the elements of a left-justified in a string of length width.
fillchar])
lower(a) Return an array with the elements of a converted to lowercase.
lstrip(a[, For each element in a, return a copy with the leading characters removed.
chars])
partition
replace(a, old, For each element in a, return a copy of the string with all occurrences of substring old
new[, count]) replaced by new.
rjust(a, width[, Return an array with the elements of a right-justified in a string of length width.
fillchar])
rpartition
rsplit(a[, sep, For each element in a, return a list of the words in the string, using sep as the delimiter
maxsplit]) string.
rstrip(a[, For each element in a, return a copy with the trailing characters removed.
chars])
split(a[, sep, For each element in a, return a list of the words in the string, using sep as the delimiter
maxsplit]) string.
splitlines(a[, For each element in a, return a list of the lines in the element, breaking at line boundaries.
keepends])
strip(a[, chars]) For each element in a, return a copy with the leading and trailing characters removed.
swapcase(a) For each element in a, return a copy of the string with uppercase characters converted to
lowercase and vice versa.
title(a) For each element in a, return a titlecased version of the string: words start with uppercase
characters, all remaining cased characters are lowercase.
translate(a, For each element in a, return a copy of the string where all characters occurring in the
table[, optional argument deletechars are removed, and the remaining characters have been
deletechars]) mapped through the given translation table.
upper(a) Return an array with the elements of a converted to uppercase.
zfill(a, width) Return the numeric string left-filled with zeros in a string of length width.
add(x1, x2)
Return (x1 + x2), that is string concatenation, element-wise for a pair of array_likes of str or unicode.
Parameters
x1 : array_like of str or unicode
x2 : array_like of str or unicode
Returns
out : ndarray
Output array of string_ or unicode_, depending on input types

3.31. String operations 961

NumPy Reference, Release 2.0.0.dev8464

multiply(a, i)
Return (a * i), that is string multiple concatenation, element-wise.
Values in i of less than 0 are treated as 0 (which yields an empty string).
Parameters
a : array_like of str or unicode
i : array_like of ints
Returns
out : ndarray
Output array of str or unicode, depending on input types
mod(a, values)
Return (a % i), that is pre-Python 2.6 string formatting (iterpolation), element-wise for a pair of array_likes of
str or unicode.
Parameters
a : array_like of str or unicode
values : array_like of values
These values will be element-wise interpolated into the string.
Returns
out : ndarray
Output array of str or unicode, depending on input types
See Also:
str.__mod__
capitalize(a)
Return a copy of a with only the first character of each element capitalized.
Calls str.capitalize element-wise.
For 8-bit strings, this method is locale-dependent.
Parameters
a : array_like of str or unicode
Returns
out : ndarray
Output array of str or unicode, depending on input types
See Also:
str.capitalize

Examples

>>> c = np.array([a1b2,1b2a,b2a1,2a1b],S4); c
array([a1b2, 1b2a, b2a1, 2a1b],
dtype=|S4)
>>> np.char.capitalize(c)
array([A1b2, 1b2a, B2a1, 2a1b],
dtype=|S4)

center(a, width, fillchar= )

Return a copy of a with its elements centered in a string of length width.

962 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Calls str.center element-wise.

Parameters
a : array_like of str or unicode
width : int
The length of the resulting strings
fillchar : str or unicode, optional
The padding character to use (default is space).
Returns
out : ndarray
Output array of str or unicode, depending on input types
See Also:
str.center
decode(a, encoding=None, errors=None)
Calls str.decode element-wise.
The set of available codecs comes from the Python standard library, and may be extended at runtime. For more
information, see the codecs module.
Parameters
a : array_like of str or unicode
encoding : str, optional
The name of an encoding
errors : str, optional
Specifies how to handle encoding errors
Returns
out : ndarray
See Also:
str.decode

Notes
The type of the result will depend on the encoding specified.

Examples

>>> c = np.array([aAaAaA, aA , abBABba])

>>> c
array([aAaAaA, aA , abBABba],
dtype=|S7)
>>> np.char.encode(c, encoding=cp037)
array([\x81\xc1\x81\xc1\x81\xc1, @@\x81\xc1@@,
\x81\x82\xc2\xc1\xc2\x82\x81],
dtype=|S7)

encode(a, encoding=None, errors=None)

Calls str.encode element-wise.

3.31. String operations 963

NumPy Reference, Release 2.0.0.dev8464

The set of available codecs comes from the Python standard library, and may be extended at runtime. For more
information, see the codecs module.
Parameters
a : array_like of str or unicode
encoding : str, optional
The name of an encoding
errors : str, optional
Specifies how to handle encoding errors
Returns
out : ndarray
See Also:
str.encode

Notes
The type of the result will depend on the encoding specified.
join(sep, seq)
Return a string which is the concatenation of the strings in the sequence seq.
Calls str.join element-wise.
Parameters
sep : array_like of str or unicode
seq : array_like of str or unicode
Returns
out : ndarray
Output array of str or unicode, depending on input types
See Also:
str.join
ljust(a, width, fillchar= )
Return an array with the elements of a left-justified in a string of length width.
Calls str.ljust element-wise.
Parameters
a : array_like of str or unicode
width : int
The length of the resulting strings
fillchar : str or unicode, optional
The character to use for padding
Returns
out : ndarray
Output array of str or unicode, depending on input type
See Also:
str.ljust

964 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

lower(a)
Return an array with the elements of a converted to lowercase.
Call str.lower element-wise.
For 8-bit strings, this method is locale-dependent.
Parameters
a : array-like of str or unicode
Returns
out : ndarray, str or unicode
Output array of str or unicode, depending on input type
See Also:
str.lower

Examples

>>> c = np.array([A1B C, 1BCA, BCA1]); c

array([A1B C, 1BCA, BCA1],
dtype=|S5)
>>> np.char.lower(c)
array([a1b c, 1bca, bca1],
dtype=|S5)

lstrip(a, chars=None)
For each element in a, return a copy with the leading characters removed.
Calls str.lstrip element-wise.
Parameters
a : array-like of str or unicode
chars : str or unicode, optional
The chars argument is a string specifying the set of characters to be removed. If omitted
or None, the chars argument defaults to removing whitespace. The chars argument is
not a prefix; rather, all combinations of its values are stripped.
Returns
out : ndarray, str or unicode
Output array of str or unicode, depending on input type
See Also:
str.lstrip

Examples

>>> c = np.array([aAaAaA, aA , abBABba])

>>> c
array([aAaAaA, aA , abBABba],
dtype=|S7)
>>> np.char.lstrip(c, a) # a unstripped from c[1] because whitespace leading
array([AaAaA, aA , bBABba],
dtype=|S7)
>>> np.char.lstrip(c, A) # leaves c unchanged
array([aAaAaA, aA , abBABba],
dtype=|S7)

3.31. String operations 965

NumPy Reference, Release 2.0.0.dev8464

>>> (np.char.lstrip(c, ) == np.char.lstrip(c, )).all()

... # XXX: is this a regression? this line now returns False
... # np.char.lstrip(c,) does not modify c at all.
True
>>> (np.char.lstrip(c, ) == np.char.lstrip(c, None)).all()
True

replace(a, old, new, count=None)

For each element in a, return a copy of the string with all occurrences of substring old replaced by new.
Calls str.replace element-wise.
Parameters
a : array-like of str or unicode
old, new : str or unicode
count : int, optional
If the optional argument count is given, only the first count occurrences are replaced.
Returns
out : ndarray
Output array of str or unicode, depending on input type
See Also:
str.replace
rjust(a, width, fillchar= )
Return an array with the elements of a right-justified in a string of length width.
Calls str.rjust element-wise.
Parameters
a : array_like of str or unicode
width : int
The length of the resulting strings
fillchar : str or unicode, optional
The character to use for padding
Returns
out : ndarray
Output array of str or unicode, depending on input type
See Also:
str.rjust
rsplit(a, sep=None, maxsplit=None)
For each element in a, return a list of the words in the string, using sep as the delimiter string.
Calls str.rsplit element-wise.
Except for splitting from the right, rsplit behaves like split.
Parameters
a : array_like of str or unicode
sep : str or unicode, optional

966 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

If sep is not specified or None, any whitespace string is a separator.

maxsplit : int, optional
If maxsplit is given, at most maxsplit splits are done, the rightmost ones.
Returns
out : ndarray
Array of list objects
See Also:
str.rsplit, split
rstrip(a, chars=None)
For each element in a, return a copy with the trailing characters removed.
Calls str.rstrip element-wise.
Parameters
a : array-like of str or unicode
chars : str or unicode, optional
The chars argument is a string specifying the set of characters to be removed. If omitted
or None, the chars argument defaults to removing whitespace. The chars argument is
not a suffix; rather, all combinations of its values are stripped.
Returns
out : ndarray
Output array of str or unicode, depending on input type
See Also:
str.rstrip

Examples

>>> c = np.array([aAaAaA, abBABba], dtype=S7); c

array([aAaAaA, abBABba],
dtype=|S7)
>>> np.char.rstrip(c, a)
array([aAaAaA, abBABb],
dtype=|S7)
>>> np.char.rstrip(c, A)
array([aAaAa, abBABba],
dtype=|S7)

split(a, sep=None, maxsplit=None)

For each element in a, return a list of the words in the string, using sep as the delimiter string.
Calls str.rsplit element-wise.
Parameters
a : array_like of str or unicode
sep : str or unicode, optional
If sep is not specified or None, any whitespace string is a separator.
maxsplit : int, optional
If maxsplit is given, at most maxsplit splits are done.

3.31. String operations 967

NumPy Reference, Release 2.0.0.dev8464

Returns
out : ndarray
Array of list objects
See Also:
str.split, rsplit
splitlines(a, keepends=None)
For each element in a, return a list of the lines in the element, breaking at line boundaries.
Calls str.splitlines element-wise.
Parameters
a : array_like of str or unicode
keepends : bool, optional
Line breaks are not included in the resulting list unless keepends is given and true.
Returns
out : ndarray
Array of list objects
See Also:
str.splitlines
strip(a, chars=None)
For each element in a, return a copy with the leading and trailing characters removed.
Calls str.rstrip element-wise.
Parameters
a : array-like of str or unicode
chars : str or unicode, optional
The chars argument is a string specifying the set of characters to be removed. If omitted
or None, the chars argument defaults to removing whitespace. The chars argument is
not a prefix or suffix; rather, all combinations of its values are stripped.
Returns
out : ndarray
Output array of str or unicode, depending on input type
See Also:
str.strip

Examples

>>> c = np.array([aAaAaA, aA , abBABba])

>>> c
array([aAaAaA, aA , abBABba],
dtype=|S7)
>>> np.char.strip(c)
array([aAaAaA, aA, abBABba],
dtype=|S7)
>>> np.char.strip(c, a) # a unstripped from c[1] because whitespace leads
array([AaAaA, aA , bBABb],
dtype=|S7)

968 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

>>> np.char.strip(c, A) # A unstripped from c[1] because (unprinted) ws trails

array([aAaAa, aA , abBABba],
dtype=|S7)

swapcase(a)
For each element in a, return a copy of the string with uppercase characters converted to lowercase and vice
versa.
Calls str.swapcase element-wise.
For 8-bit strings, this method is locale-dependent.
Parameters
a : array-like of str or unicode
Returns
out : ndarray
Output array of str or unicode, depending on input type
See Also:
str.swapcase

Examples

>>> c=np.array([a1B c,1b Ca,b Ca1,cA1b],S5); c

array([a1B c, 1b Ca, b Ca1, cA1b],
dtype=|S5)
>>> np.char.swapcase(c)
array([A1b C, 1B cA, B cA1, Ca1B],
dtype=|S5)

title(a)
For each element in a, return a titlecased version of the string: words start with uppercase characters, all
remaining cased characters are lowercase.
Calls str.title element-wise.
For 8-bit strings, this method is locale-dependent.
Parameters
a : array-like of str or unicode
Returns
out : ndarray
Output array of str or unicode, depending on input type
See Also:
str.title

Examples

>>> c=np.array([a1b c,1b ca,b ca1,ca1b],S5); c

array([a1b c, 1b ca, b ca1, ca1b],
dtype=|S5)
>>> np.char.title(c)
array([A1B C, 1B Ca, B Ca1, Ca1B],
dtype=|S5)

3.31. String operations 969

NumPy Reference, Release 2.0.0.dev8464

translate(a, table, deletechars=None)

For each element in a, return a copy of the string where all characters occurring in the optional argument
deletechars are removed, and the remaining characters have been mapped through the given translation table.
Calls str.translate element-wise.
Parameters
a : array-like of str or unicode
table : str of length 256
deletechars : str
Returns
out : ndarray
Output array of str or unicode, depending on input type
See Also:
str.translate
upper(a)
Return an array with the elements of a converted to uppercase.
Calls str.upper element-wise.
For 8-bit strings, this method is locale-dependent.
Parameters
a : array-like of str or unicode
Returns
out : ndarray
Output array of str or unicode, depending on input type
See Also:
str.upper

Examples

>>> c = np.array([a1b c, 1bca, bca1]); c

array([a1b c, 1bca, bca1],
dtype=|S5)
>>> np.char.upper(c)
array([A1B C, 1BCA, BCA1],
dtype=|S5)

zfill(a, width)
Return the numeric string left-filled with zeros in a string of length width.
Calls str.zfill element-wise.
Parameters
a : array-like of str or unicode
width : int
Returns
out : ndarray
Output array of str or unicode, depending on input type

970 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:
str.zfill

3.31.2 Comparison

Unlike the standard numpy comparison operators, the ones in the char module strip trailing whitespace characters
before performing the comparison.
equal(x1, x2) Return (x1 == x2) element-wise.
not_equal(x1, x2) Return (x1 != x2) element-wise.
greater_equal(x1, x2) Return (x1 >= x2) element-wise.
less_equal(x1, x2) Return (x1 <= x2) element-wise.
greater(x1, x2) Return (x1 > x2) element-wise.
less(x1, x2) Return (x1 < x2) element-wise.
equal(x1, x2)
Return (x1 == x2) element-wise.
Unlike numpy.equal, this comparison is performed by first stripping whitespace characters from the end of
the string. This behavior is provided for backward-compatibility with numarray.
Parameters
x1, x2 : array_like of str or unicode
Input arrays of the same shape.
Returns
out : {ndarray, bool}
Output array of bools, or a single bool if x1 and x2 are scalars.
See Also:
not_equal, greater_equal, less_equal, greater, less
not_equal(x1, x2)
Return (x1 != x2) element-wise.
Unlike numpy.not_equal, this comparison is performed by first stripping whitespace characters from the
end of the string. This behavior is provided for backward-compatibility with numarray.
Parameters
x1, x2 : array_like of str or unicode
Input arrays of the same shape.
Returns
out : {ndarray, bool}
Output array of bools, or a single bool if x1 and x2 are scalars.
See Also:
equal, greater_equal, less_equal, greater, less
greater_equal(x1, x2)
Return (x1 >= x2) element-wise.
Unlike numpy.greater_equal, this comparison is performed by first stripping whitespace characters from
the end of the string. This behavior is provided for backward-compatibility with numarray.
Parameters
x1, x2 : array_like of str or unicode

3.31. String operations 971

NumPy Reference, Release 2.0.0.dev8464

Input arrays of the same shape.

Returns
out : {ndarray, bool}
Output array of bools, or a single bool if x1 and x2 are scalars.
See Also:
equal, not_equal, less_equal, greater, less
less_equal(x1, x2)
Return (x1 <= x2) element-wise.
Unlike numpy.less_equal, this comparison is performed by first stripping whitespace characters from the
end of the string. This behavior is provided for backward-compatibility with numarray.
Parameters
x1, x2 : array_like of str or unicode
Input arrays of the same shape.
Returns
out : {ndarray, bool}
Output array of bools, or a single bool if x1 and x2 are scalars.
See Also:
equal, not_equal, greater_equal, greater, less
greater(x1, x2)
Return (x1 > x2) element-wise.
Unlike numpy.greater, this comparison is performed by first stripping whitespace characters from the end
of the string. This behavior is provided for backward-compatibility with numarray.
Parameters
x1, x2 : array_like of str or unicode
Input arrays of the same shape.
Returns
out : {ndarray, bool}
Output array of bools, or a single bool if x1 and x2 are scalars.
See Also:
equal, not_equal, greater_equal, less_equal, less
less(x1, x2)
Return (x1 < x2) element-wise.
Unlike numpy.greater, this comparison is performed by first stripping whitespace characters from the end
of the string. This behavior is provided for backward-compatibility with numarray.
Parameters
x1, x2 : array_like of str or unicode
Input arrays of the same shape.
Returns
out : {ndarray, bool}
Output array of bools, or a single bool if x1 and x2 are scalars.

972 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:
equal, not_equal, greater_equal, less_equal, greater

3.31.3 String information

count(a, sub[, start, Returns an array with the number of non-overlapping occurrences of substring sub in
end]) the range [start, end].
len
find(a, sub[, start, For each element, return the lowest index in the string where substring sub is found.
end])
index(a, sub[, start, Like find, but raises ValueError when the substring is not found.
end])
isalpha(a) Returns true for each element if all characters in the string are alphabetic and there is at
least one character, false otherwise.
isdecimal(a) For each element in a, return True if there are only decimal
isdigit(a) Returns true for each element if all characters in the string are digits and there is at least
one character, false otherwise.
islower(a) Returns true for each element if all cased characters in the string are lowercase and
there is at least one cased character, false otherwise.
isnumeric(a) For each element in a, return True if there are only numeric
isspace(a) Returns true for each element if there are only whitespace characters in the string and
there is at least one character, false otherwise.
istitle(a) Returns true for each element if the element is a titlecased string and there is at least
one character, false otherwise.
isupper(a) Returns true for each element if all cased characters in the string are uppercase and
there is at least one character, false otherwise.
rfind(a, sub[, start, For each element in a, return the highest index in the string where substring sub is
end]) found, such that sub is contained within [start, end].
rindex(a, sub[, start, Like rfind, but raises ValueError when the substring sub is
end])
startswith(a, Returns a boolean array which is True where the string element
prefix[, start, end])
count(a, sub, start=0, end=None)
Returns an array with the number of non-overlapping occurrences of substring sub in the range [start, end].
Calls str.count element-wise.
Parameters
a : array_like of str or unicode
sub : str or unicode
The substring to search for.
start, end : int, optional
Optional arguments start and end are interpreted as slice notation to specify the range
in which to count.
Returns
out : ndarray
Output array of ints.
See Also:
str.count

3.31. String operations 973

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> c = np.array([aAaAaA, aA , abBABba])

>>> c
array([aAaAaA, aA , abBABba],
dtype=|S7)
>>> np.char.count(c, A)
array([3, 1, 1])
>>> np.char.count(c, aA)
array([3, 1, 0])
>>> np.char.count(c, A, start=1, end=4)
array([2, 1, 1])
>>> np.char.count(c, A, start=1, end=3)
array([1, 0, 0])

find(a, sub, start=0, end=None)

For each element, return the lowest index in the string where substring sub is found.
Calls str.find element-wise.
For each element, return the lowest index in the string where substring sub is found, such that sub is contained
in the range [start, end].
Parameters
a : array_like of str or unicode
sub : str or unicode
start, end : int, optional
Optional arguments start and end are interpreted as in slice notation.
Returns
out : ndarray or int
Output array of ints. Returns -1 if sub is not found.
See Also:
str.find
index(a, sub, start=0, end=None)
Like find, but raises ValueError when the substring is not found.
Calls str.index element-wise.
Parameters
a : array_like of str or unicode
sub : str or unicode
start, end : int, optional
Returns
out : ndarray
Output array of ints. Returns -1 if sub is not found.
See Also:
find, str.find
isalpha(a)
Returns true for each element if all characters in the string are alphabetic and there is at least one character,
false otherwise.

974 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Calls str.isalpha element-wise.

For 8-bit strings, this method is locale-dependent.
Parameters
a : array_like of str or unicode
Returns
out : ndarray
Output array of bools
See Also:
str.isalpha
isdecimal(a)
For each element in a, return True if there are only decimal characters in the element.
Calls unicode.isdecimal element-wise.
Decimal characters include digit characters, and all characters that that can be used to form decimal-radix
numbers, e.g. U+0660, ARABIC-INDIC DIGIT ZERO.
Parameters
a : array-like of unicode
Returns
out : ndarray
Array of booleans
See Also:
unicode.isdecimal
isdigit(a)
Returns true for each element if all characters in the string are digits and there is at least one character, false
otherwise.
Calls str.isdigit element-wise.
For 8-bit strings, this method is locale-dependent.
Parameters
a : array_like of str or unicode
Returns
out : ndarray
Output array of bools
See Also:
str.isdigit
islower(a)
Returns true for each element if all cased characters in the string are lowercase and there is at least one cased
character, false otherwise.
Calls str.islower element-wise.
For 8-bit strings, this method is locale-dependent.
Parameters
a : array_like of str or unicode

3.31. String operations 975

NumPy Reference, Release 2.0.0.dev8464

Returns
out : ndarray
Output array of bools
See Also:
str.islower
isnumeric(a)
For each element in a, return True if there are only numeric characters in the element.
Calls unicode.isnumeric element-wise.
Numeric characters include digit characters, and all characters that have the Unicode numeric value property,
e.g. U+2155, VULGAR FRACTION ONE FIFTH.
Parameters
a : array-like of unicode
Returns
out : ndarray
Array of booleans
See Also:
unicode.isnumeric
isspace(a)
Returns true for each element if there are only whitespace characters in the string and there is at least one
character, false otherwise.
Calls str.isspace element-wise.
For 8-bit strings, this method is locale-dependent.
Parameters
a : array_like of str or unicode
Returns
out : ndarray
Output array of bools
See Also:
str.isspace
istitle(a)
Returns true for each element if the element is a titlecased string and there is at least one character, false
otherwise.
Call str.istitle element-wise.
For 8-bit strings, this method is locale-dependent.
Parameters
a : array_like of str or unicode
Returns
out : ndarray
Output array of bools

976 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:
str.istitle
isupper(a)
Returns true for each element if all cased characters in the string are uppercase and there is at least one character,
false otherwise.
Call str.isupper element-wise.
For 8-bit strings, this method is locale-dependent.
Parameters
a : array_like of str or unicode
Returns
out : ndarray
Output array of bools
See Also:
str.isupper
rfind(a, sub, start=0, end=None)
For each element in a, return the highest index in the string where substring sub is found, such that sub is
contained within [start, end].
Calls str.rfind element-wise.
Parameters
a : array-like of str or unicode
sub : str or unicode
start, end : int, optional
Optional arguments start and end are interpreted as in slice notation.
Returns
out : ndarray
Output array of ints. Return -1 on failure.
See Also:
str.rfind
rindex(a, sub, start=0, end=None)
Like rfind, but raises ValueError when the substring sub is not found.
Calls str.rindex element-wise.
Parameters
a : array-like of str or unicode
sub : str or unicode
start, end : int, optional
Returns
out : ndarray
Output array of ints.
See Also:
rfind, str.rindex

3.31. String operations 977

NumPy Reference, Release 2.0.0.dev8464

startswith(a, prefix, start=0, end=None)

Returns a boolean array which is True where the string element in a starts with prefix, otherwise False.
Calls str.startswith element-wise.
Parameters
a : array_like of str or unicode
suffix : str
start, end : int, optional
With optional start, test beginning at that position. With optional end, stop comparing
at that position.
Returns
out : ndarray
Array of booleans
See Also:
str.startswith

3.31.4 Convenience class

chararray Provides a convenient view on arrays of string and unicode values.

class chararray()
Provides a convenient view on arrays of string and unicode values.
Note: The chararray class exists for backwards compatibility with Numarray, it is not recommended for new
development. Starting from numpy 1.4, if one needs arrays of strings, it is recommended to use arrays of dtype
object_, string_ or unicode_, and use the free functions in the numpy.char module for fast vectorized string
operations.
Versus a regular Numpy array of type str or unicode, this class adds the following functionality:
1.values automatically have whitespace removed from the end when indexed
2.comparison operators automatically remove whitespace from the end when comparing values
3.vectorized string operations are provided as methods (e.g. endswith) and infix operators (e.g. "+", "*",
"%")
chararrays should be created using numpy.char.array or numpy.char.asarray, rather than this con-
structor directly.
This constructor creates the array, using buffer (with offset and strides) if it is not None. If buffer is None, then
constructs a new array with strides in C order, unless both len(shape) >= 2 and order=Fortran,
in which case strides is in Fortran order.
Parameters
shape : tuple
Shape of the array.
itemsize : int, optional
Length of each array element, in number of characters. Default is 1.
unicode : bool, optional
Are the array elements of type unicode (True) or string (False). Default is False.

978 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

buffer : int, optional

Memory address of the start of the array data. Default is None, in which case a new
array is created.
offset : int, optional
Fixed stride displacement from the beginning of an axis? Default is 0. Needs to be >=0.
strides : array_like of ints, optional
Strides for the array (see ndarray.strides for full description). Default is None.
order : {C, F}, optional
The order in which the array data is stored in memory: C -> row major order (the
default), F -> column major (Fortran) order.

Examples

>>> charar = np.chararray((3, 3))

>>> charar[:] = a
>>> charar
chararray([[a, a, a],
[a, a, a],
[a, a, a]],
dtype=|S1)

>>> charar = np.chararray(charar.shape, itemsize=5)

>>> charar[:] = abc
>>> charar
chararray([[abc, abc, abc],
[abc, abc, abc],
[abc, abc, abc]],
dtype=|S5)

Methods

astype(t) Copy of the array, cast

argsort([axis, kind, order])
copy([order]) Return a copy of
count(sub[, start, end]) Returns an array with the number of non-overlapping
decode([encoding, errors]) Calls str.decode elem
dump(file) Dump a pickle of the ar
dumps() Returns the pickle of the
encode([encoding, errors]) Calls str.encode elem
endswith(suffix[, start, end]) Returns a boolean array which
expandtabs([tabsize]) Return a copy of each string element where all
fill(value) Fill the array with
find(sub[, start, end]) For each element, return the lowest index
flatten([order]) Return a copy of the array
getfield(dtype, offset) Returns a field of the given
index(sub[, start, end]) Like find, but raises ValueError wh
isalnum() Returns true for each element if all characters in the string
isalpha() Returns true for each element if all characters in the strin
isdecimal() For each element in self,
isdigit() Returns true for each element if all characters in the st

3.31. String operations 979

NumPy Reference, Release 2.0.0.dev8464

Table 3.3 continued from

islower() Returns true for each element if all cased characters in the st
isnumeric() For each element in self,
isspace() Returns true for each element if there are only whitespace ch
istitle() Returns true for each element if the element is a tit
isupper() Returns true for each element if all cased characters in the
item(*args) Copy an element of an array to
join(seq) Return a string which is the concatenation
ljust(width[, fillchar]) Return an array with the elements of se
lower() Return an array with the elements
lstrip([chars]) For each element in self, return a
nonzero() Return the indices of the
put(indices, values[, mode]) Set a.flat[n] = values[n] for
ravel() Return a flattened
repeat(repeats[, axis]) Repeat elements of
replace(old, new[, count]) For each element in self, return a copy of the
reshape(shape[, order]) Returns an array containing the sam
resize(new_shape[, refcheck]) Change shape and size
rfind(sub[, start, end]) For each element in self, return the highest index in the string
rindex(sub[, start, end]) Like rfind, but raises ValueEr
rjust(width[, fillchar]) Return an array with the elements of se
rsplit([sep, maxsplit]) For each element in self, return a list of the
rstrip([chars]) For each element in self, return a
searchsorted(v[, side]) Find indices where elements of v should
setfield(val, dtype[, offset]) Put a value into a specified place
setflags([write, align, uic]) Set array flags WRITEABLE, ALIGNE
sort([axis, kind, order]) Sort an array,
split([sep, maxsplit]) For each element in self, return a list of the
splitlines([keepends]) For each element in self, return a list of
squeeze() Remove single-dimensional entries from
startswith(prefix[, start, end]) Returns a boolean array which
strip([chars]) For each element in self, return a copy
swapaxes(axis1, axis2) Return a view of the array
swapcase() For each element in self, return a copy of the stri
take(indices[, axis, out, mode]) Return an array formed from the el
title() For each element in self, return a titlecased version of the string:
tofile(fid[, sep, format]) Write array to a file as
tolist() Return the array as a
tostring([order]) Construct a Python string containing the
translate(table[, deletechars]) For each element in self, return a copy of the string where all characters occurring in the option
transpose(*axes) Returns a view of the
upper() Return an array with the elements
view([dtype, type]) New view of array with
zfill(width) Return the numeric string left-filled with

astype(t)
Copy of the array, cast to a specified type.
Parameters
t : string or dtype
Typecode or data-type to which the array is cast.

980 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.array([1, 2, 2.5])

>>> x
array([ 1. , 2. , 2.5])

>>> x.astype(int)
array([1, 2, 2])

argsort(axis=-1, kind=quicksort, order=None)

copy(order=C)
Return a copy of the array.
Parameters
order : {C, F, A}, optional
By default, the result is stored in C-contiguous (row-major) order in memory. If order is
F, the result has Fortran (column-major) order. If order is A (Any), then the result
has the same order as the input.

Examples

>>> x = np.array([[1,2,3],[4,5,6]], order=F)

>>> y = x.copy()

>>> x.fill(0)

>>> x
array([[0, 0, 0],
[0, 0, 0]])

>>> y
array([[1, 2, 3],
[4, 5, 6]])

>>> y.flags[C_CONTIGUOUS]
True

count(sub, start=0, end=None)

Returns an array with the number of non-overlapping occurrences of substring sub in the range [start,
end].
See Also:
char.count
decode(encoding=None, errors=None)
Calls str.decode element-wise.
See Also:
char.decode

3.31. String operations 981

NumPy Reference, Release 2.0.0.dev8464

dump(file)
Dump a pickle of the array to the specified file. The array can be read back with pickle.load or numpy.load.
Parameters
file : str
A string naming the dump file.
dumps()
Returns the pickle of the array as a string. pickle.loads or numpy.loads will convert the string back to an
array.
Parameters
None :
encode(encoding=None, errors=None)
Calls str.encode element-wise.
See Also:
char.encode
endswith(suffix, start=0, end=None)
Returns a boolean array which is True where the string element in self ends with suffix, otherwise False.
See Also:
char.endswith
expandtabs(tabsize=8)
Return a copy of each string element where all tab characters are replaced by one or more spaces.
See Also:
char.expandtabs
fill(value)
Fill the array with a scalar value.
Parameters
value : scalar
All elements of a will be assigned this value.

Examples

>>> a = np.array([1, 2])

>>> a.fill(0)
>>> a
array([0, 0])
>>> a = np.empty(2)
>>> a.fill(1)
>>> a
array([ 1., 1.])

find(sub, start=0, end=None)

For each element, return the lowest index in the string where substring sub is found.
See Also:
char.find
flatten(order=C)
Return a copy of the array collapsed into one dimension.

982 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

Parameters
order : {C, F}, optional
Whether to flatten in C (row-major) or Fortran (column-major) order. The default is C.
Returns
y : ndarray
A copy of the input array, flattened to one dimension.
See Also:

ravel
Return a flattened array.
flat
A 1-D flat iterator over the array.

Examples

>>> a = np.array([[1,2], [3,4]])

>>> a.flatten()
array([1, 2, 3, 4])
>>> a.flatten(F)
array([1, 3, 2, 4])

getfield(dtype, offset)
Returns a field of the given array as a certain type.
A field is a view of the array data with each itemsize determined by the given type and the offset into the
current array, i.e. from offset * dtype.itemsize to (offset+1) * dtype.itemsize.
Parameters
dtype : str
String denoting the data type of the field.
offset : int
Number of dtype.itemsizes to skip before beginning the element view.

Examples

>>> x = np.diag([1.+1.j]*2)
>>> x
array([[ 1.+1.j, 0.+0.j],
[ 0.+0.j, 1.+1.j]])
>>> x.dtype
dtype(complex128)

>>> x.getfield(complex64, 0) # Note how this != x

array([[ 0.+1.875j, 0.+0.j ],
[ 0.+0.j , 0.+1.875j]], dtype=complex64)

>>> x.getfield(complex64,1) # Note how different this is than x

array([[ 0. +5.87173204e-39j, 0. +0.00000000e+00j],
[ 0. +0.00000000e+00j, 0. +5.87173204e-39j]], dtype=complex64)

3.31. String operations 983

NumPy Reference, Release 2.0.0.dev8464

>>> x.getfield(complex128, 0) # == x
array([[ 1.+1.j, 0.+0.j],
[ 0.+0.j, 1.+1.j]])

If the argument dtype is the same as x.dtype, then offset != 0 raises a ValueError:

>>> x.getfield(complex128, 1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: Need 0 <= offset <= 0 for requested type but received offset = 1

>>> x.getfield(float64, 0)
array([[ 1., 0.],
[ 0., 1.]])

>>> x.getfield(float64, 1)
array([[ 1.77658241e-307, 0.00000000e+000],
[ 0.00000000e+000, 1.77658241e-307]])

index(sub, start=0, end=None)

Like find, but raises ValueError when the substring is not found.
See Also:
char.index
isalnum()
Returns true for each element if all characters in the string are alphanumeric and there is at least one
character, false otherwise.
See Also:
char.isalnum
isalpha()
Returns true for each element if all characters in the string are alphabetic and there is at least one character,
false otherwise.
See Also:
char.isalpha
isdecimal()
For each element in self, return True if there are only decimal characters in the element.
See Also:
char.isdecimal
isdigit()
Returns true for each element if all characters in the string are digits and there is at least one character,
false otherwise.
See Also:
char.isdigit
islower()
Returns true for each element if all cased characters in the string are lowercase and there is at least one
cased character, false otherwise.

984 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:
char.islower
isnumeric()
For each element in self, return True if there are only numeric characters in the element.
See Also:
char.isnumeric
isspace()
Returns true for each element if there are only whitespace characters in the string and there is at least one
character, false otherwise.
See Also:
char.isspace
istitle()
Returns true for each element if the element is a titlecased string and there is at least one character, false
otherwise.
See Also:
char.istitle
isupper()
Returns true for each element if all cased characters in the string are uppercase and there is at least one
character, false otherwise.
See Also:
char.isupper
item(*args)
Copy an element of an array to a standard Python scalar and return it.
Parameters
*args : Arguments (variable number and type)
none: in this case, the method only works for arrays with one element (a.size == 1), which
element is copied into a standard Python scalar object and returned.
int_type: this argument is interpreted as a flat index into the array, specifying which ele-
ment to copy and return.
tuple of int_types: functions as does a single int_type argument, except that the argument
is interpreted as an nd-index into the array.
Returns
z : Standard Python scalar object
A copy of the specified element of the array as a suitable Python scalar

Notes
When the data type of a is longdouble or clongdouble, item() returns a scalar array object because there is
no available Python scalar that would not lose information. Void arrays return a buffer object for item(),
unless fields are defined, in which case a tuple is returned.
item is very similar to a[args], except, instead of an array scalar, a standard Python scalar is returned. This
can be useful for speeding up access to elements of the array and doing arithmetic on elements of the array
using Pythons optimized math.

3.31. String operations 985

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> x = np.random.randint(9, size=(3, 3))

>>> x
array([[3, 1, 7],
[2, 8, 3],
[8, 5, 3]])
>>> x.item(3)
2
>>> x.item(7)
5
>>> x.item((0, 1))
1
>>> x.item((2, 2))
3

join(seq)
Return a string which is the concatenation of the strings in the sequence seq.
See Also:
char.join
ljust(width, fillchar= )
Return an array with the elements of self left-justified in a string of length width.
See Also:
char.ljust
lower()
Return an array with the elements of self converted to lowercase.
See Also:
char.lower
lstrip(chars=None)
For each element in self, return a copy with the leading characters removed.
See Also:
char.lstrip
nonzero()
Return the indices of the elements that are non-zero.
Refer to numpy.nonzero for full documentation.
See Also:

numpy.nonzero
equivalent function

put(indices, values, mode=raise)

Set a.flat[n] = values[n] for all n in indices.
Refer to numpy.put for full documentation.
See Also:

numpy.put
equivalent function

986 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

ravel([order])
Return a flattened array.
Refer to numpy.ravel for full documentation.
See Also:

numpy.ravel
equivalent function
ndarray.flat
a flat iterator on the array.

repeat(repeats, axis=None)
Repeat elements of an array.
Refer to numpy.repeat for full documentation.
See Also:

numpy.repeat
equivalent function

replace(old, new, count=None)

For each element in self, return a copy of the string with all occurrences of substring old replaced by new.
See Also:
char.replace
reshape(shape, order=C)
Returns an array containing the same data with a new shape.
Refer to numpy.reshape for full documentation.
See Also:

numpy.reshape
equivalent function

resize(new_shape, refcheck=True)
Change shape and size of array in-place.
Parameters
new_shape : tuple of ints, or n ints
Shape of resized array.
refcheck : bool, optional
If False, reference count will not be checked. Default is True.
Returns
None :
Raises
ValueError :
If a does not own its own data or references or views to it exist, and the data memory
must be changed.
SystemError :

3.31. String operations 987

NumPy Reference, Release 2.0.0.dev8464

If the order keyword argument is specified. This behaviour is a bug in NumPy.

See Also:

resize
Return a new array with the specified shape.

Notes
This reallocates space for the data area if necessary.
Only contiguous arrays (data elements consecutive in memory) can be resized.
The purpose of the reference count check is to make sure you do not use this array as a buffer for another
Python object and then reallocate the memory. However, reference counts can increase in other ways so if
you are sure that you have not shared the memory for this array with another Python object, then you may
safely set refcheck to False.

Examples
Shrinking an array: array is flattened (in the order that the data are stored in memory), resized, and re-
shaped:

>>> a = np.array([[0, 1], [2, 3]], order=C)

>>> a.resize((2, 1))
>>> a
array([[0],
[1]])

>>> a = np.array([[0, 1], [2, 3]], order=F)

>>> a.resize((2, 1))
>>> a
array([[0],
[2]])

Enlarging an array: as above, but missing entries are filled with zeros:

>>> b = np.array([[0, 1], [2, 3]])

>>> b.resize(2, 3) # new_shape parameter doesnt have to be a tuple
>>> b
array([[0, 1, 2],
[3, 0, 0]])

Referencing an array prevents resizing...

>>> c = a
>>> a.resize((1, 1))
...
ValueError: cannot resize an array that has been referenced ...

Unless refcheck is False:

>>> a.resize((1, 1), refcheck=False)

>>> a
array([[0]])
>>> c
array([[0]])

988 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

rfind(sub, start=0, end=None)

For each element in self, return the highest index in the string where substring sub is found, such that sub
is contained within [start, end].
See Also:
char.rfind
rindex(sub, start=0, end=None)
Like rfind, but raises ValueError when the substring sub is not found.
See Also:
char.rindex
rjust(width, fillchar= )
Return an array with the elements of self right-justified in a string of length width.
See Also:
char.rjust
rsplit(sep=None, maxsplit=None)
For each element in self, return a list of the words in the string, using sep as the delimiter string.
See Also:
char.rsplit
rstrip(chars=None)
For each element in self, return a copy with the trailing characters removed.
See Also:
char.rstrip
searchsorted(v, side=left)
Find indices where elements of v should be inserted in a to maintain order.
For full documentation, see numpy.searchsorted
See Also:

numpy.searchsorted
equivalent function

setfield(val, dtype, offset=0)

Put a value into a specified place in a field defined by a data-type.
Place val into as field defined by dtype and beginning offset bytes into the field.
Parameters
val : object
Value to be placed in field.
dtype : dtype object
Data-type of the field in which to place val.
offset : int, optional
The number of bytes into the field at which to place val.
Returns
None :

3.31. String operations 989

NumPy Reference, Release 2.0.0.dev8464

See Also:
getfield

Examples

>>> x = np.eye(3)
>>> x.getfield(np.float64)
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])
>>> x.setfield(3, np.int32)
>>> x.getfield(np.int32)
array([[3, 3, 3],
[3, 3, 3],
[3, 3, 3]])
>>> x
array([[ 1.00000000e+000, 1.48219694e-323, 1.48219694e-323],
[ 1.48219694e-323, 1.00000000e+000, 1.48219694e-323],
[ 1.48219694e-323, 1.48219694e-323, 1.00000000e+000]])
>>> x.setfield(np.eye(3), np.int32)
>>> x
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])

setflags(write=None, align=None, uic=None)

Set array flags WRITEABLE, ALIGNED, and UPDATEIFCOPY, respectively.
These Boolean-valued flags affect how numpy interprets the memory area used by a (see Notes below).
The ALIGNED flag can only be set to True if the data is actually aligned according to the type. The
UPDATEIFCOPY flag can never be set to True. The flag WRITEABLE can only be set to True if the array
owns its own memory, or the ultimate owner of the memory exposes a writeable buffer interface, or is a
string. (The exception for string is made so that unpickling can be done without copying memory.)
Parameters
write : bool, optional
Describes whether or not a can be written to.
align : bool, optional
Describes whether or not a is aligned properly for its type.
uic : bool, optional
Describes whether or not a is a copy of another base array.

Notes
Array flags provide information about how the memory area used for the array is to be interpreted. There
are 6 Boolean flags in use, only three of which can be changed by the user: UPDATEIFCOPY, WRITE-
ABLE, and ALIGNED.
WRITEABLE (W) the data area can be written to;
ALIGNED (A) the data and strides are aligned appropriately for the hardware (as determined by the com-
piler);
UPDATEIFCOPY (U) this array is a copy of some other array (referenced by .base). When this array is
deallocated, the base array will be updated with the contents of this array.

990 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

All flags can be accessed using their first (upper case) letter as well as the full name.

Examples

>>> y
array([[3, 1, 7],
[2, 0, 0],
[8, 5, 9]])
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : True
WRITEABLE : True
ALIGNED : True
UPDATEIFCOPY : False
>>> y.setflags(write=0, align=0)
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : False
OWNDATA : True
WRITEABLE : False
ALIGNED : False
UPDATEIFCOPY : False
>>> y.setflags(uic=1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: cannot set UPDATEIFCOPY flag to True

sort(axis=-1, kind=quicksort, order=None)

Sort an array, in-place.
Parameters
axis : int, optional
Axis along which to sort. Default is -1, which means sort along the last axis.
kind : {quicksort, mergesort, heapsort}, optional
Sorting algorithm. Default is quicksort.
order : list, optional
When a is an array with fields defined, this argument specifies which fields to compare
first, second, etc. Not all fields need be specified.
See Also:

numpy.sort
Return a sorted copy of an array.
argsort
Indirect sort.
lexsort
Indirect stable sort on multiple keys.
searchsorted
Find elements in sorted array.

Notes
See sort for notes on the different sorting algorithms.

3.31. String operations 991

NumPy Reference, Release 2.0.0.dev8464

Examples

>>> a = np.array([[1,4], [3,1]])

>>> a.sort(axis=1)
>>> a
array([[1, 4],
[1, 3]])
>>> a.sort(axis=0)
>>> a
array([[1, 3],
[1, 4]])

Use the order keyword to specify a field to use when sorting a structured array:

>>> a = np.array([(a, 2), (c, 1)], dtype=[(x, S1), (y, int)])

>>> a.sort(order=y)
>>> a
array([(c, 1), (a, 2)],
dtype=[(x, |S1), (y, <i4)])

split(sep=None, maxsplit=None)
For each element in self, return a list of the words in the string, using sep as the delimiter string.
See Also:
char.split
splitlines(keepends=None)
For each element in self, return a list of the lines in the element, breaking at line boundaries.
See Also:
char.splitlines
squeeze()
Remove single-dimensional entries from the shape of a.
Refer to numpy.squeeze for full documentation.
See Also:

numpy.squeeze
equivalent function

startswith(prefix, start=0, end=None)

Returns a boolean array which is True where the string element in self starts with prefix, otherwise False.
See Also:
char.startswith
strip(chars=None)
For each element in self, return a copy with the leading and trailing characters removed.
See Also:
char.strip
swapaxes(axis1, axis2)
Return a view of the array with axis1 and axis2 interchanged.
Refer to numpy.swapaxes for full documentation.

992 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

See Also:

numpy.swapaxes
equivalent function

swapcase()
For each element in self, return a copy of the string with uppercase characters converted to lowercase and
vice versa.
See Also:
char.swapcase
take(indices, axis=None, out=None, mode=raise)
Return an array formed from the elements of a at the given indices.
Refer to numpy.take for full documentation.
See Also:

numpy.take
equivalent function

title()
For each element in self, return a titlecased version of the string: words start with uppercase characters, all
remaining cased characters are lowercase.
See Also:
char.title
tofile(fid, sep="", format="%s")
Write array to a file as text or binary (default).
Data is always written in C order, independent of the order of a. The data produced by this method can
be recovered using the function fromfile().
Parameters
fid : file or str
An open file object, or a string containing a filename.
sep : str
Separator between array items for text output. If (empty), a binary file is written,
equivalent to file.write(a.tostring()).
format : str
Format string for text file output. Each entry in the array is formatted to text by first
converting it to the closest Python type, and then using format % item.

Notes
This is a convenience function for quick storage of array data. Information on endianness and precision
is lost, so this method is not a good choice for files intended to archive data or transport data between
machines with different endianness. Some of these problems can be overcome by outputting the data as
text files, at the expense of speed and file size.
tolist()
Return the array as a (possibly nested) list.

3.31. String operations 993

NumPy Reference, Release 2.0.0.dev8464

Return a copy of the array data as a (nested) Python list. Data items are converted to the nearest compatible
Python type.
Parameters
none :
Returns
y : list
The possibly nested list of array elements.

Notes
The array may be recreated, a = np.array(a.tolist()).

Examples

>>> a = np.array([1, 2])

>>> a.tolist()
[1, 2]
>>> a = np.array([[1, 2], [3, 4]])
>>> list(a)
[array([1, 2]), array([3, 4])]
>>> a.tolist()
[[1, 2], [3, 4]]

tostring(order=C)
Construct a Python string containing the raw data bytes in the array.
Constructs a Python string showing a copy of the raw contents of data memory. The string can be produced
in either C or Fortran, or Any order (the default is C-order). Any order means C-order unless the
F_CONTIGUOUS flag in the array is set, in which case it means Fortran order.
Parameters
order : {C, F, None}, optional
Order of the data for multidimensional arrays: C, Fortran, or the same as for the original
array.
Returns
s : str
A Python string exhibiting a copy of as raw data.

Examples

>>> x = np.array([[0, 1], [2, 3]])

>>> x.tostring()
\x00\x00\x00\x00\x01\x00\x00\x00\x02\x00\x00\x00\x03\x00\x00\x00
>>> x.tostring(C) == x.tostring()
True
>>> x.tostring(F)
\x00\x00\x00\x00\x02\x00\x00\x00\x01\x00\x00\x00\x03\x00\x00\x00

translate(table, deletechars=None)
For each element in self, return a copy of the string where all characters occurring in the optional argument
deletechars are removed, and the remaining characters have been mapped through the given translation
table.
See Also:

994 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

char.translate
transpose(*axes)
Returns a view of the array with axes transposed.
For a 1-D array, this has no effect. (To change between column and row vectors, first cast the 1-D ar-
ray into a matrix object.) For a 2-D array, this is the usual matrix transpose. For an n-D array, if axes
are given, their order indicates how the axes are permuted (see Examples). If axes are not provided
and a.shape = (i[0], i[1], ... i[n-2], i[n-1]), then a.transpose().shape
= (i[n-1], i[n-2], ... i[1], i[0]).
Parameters
axes : None, tuple of ints, or n ints
None or no argument: reverses the order of the axes.
tuple of ints: i in the j-th place in the tuple means as i-th axis becomes a.transpose()s
j-th axis.
n ints: same as an n-tuple of the same ints (this form is intended simply as a convenience
alternative to the tuple form)
Returns
out : ndarray
View of a, with axes suitably permuted.
See Also:

ndarray.T
Array property returning the array transposed.

Examples

>>> a = np.array([[1, 2], [3, 4]])

>>> a
array([[1, 2],
[3, 4]])
>>> a.transpose()
array([[1, 3],
[2, 4]])
>>> a.transpose((1, 0))
array([[1, 3],
[2, 4]])
>>> a.transpose(1, 0)
array([[1, 3],
[2, 4]])

upper()
Return an array with the elements of self converted to uppercase.
See Also:
char.upper
view(dtype=None, type=None)
New view of array with the same data.
Parameters
dtype : data-type
Data-type descriptor of the returned view, e.g. float32 or int16.

3.31. String operations 995

NumPy Reference, Release 2.0.0.dev8464

type : python type

Type of the returned view, e.g. ndarray or matrix.

Notes
a.view() is used two different ways.
a.view(some_dtype) or a.view(dtype=some_dtype) constructs a view of the arrays memory with a different
dtype. This can cause a reinterpretation of the bytes of memory.
a.view(ndarray_subclass), or a.view(type=ndarray_subclass), just returns an instance of ndarray_subclass
that looks at the same array (same shape, dtype, etc.). This does not cause a reinterpretation of the memory.

Examples

>>> x = np.array([(1, 2)], dtype=[(a, np.int8), (b, np.int8)])

Viewing array data using a different type and dtype:

>>> y = x.view(dtype=np.int16, type=np.matrix)

>>> y
matrix([[513]], dtype=int16)
>>> print type(y)
<class numpy.matrixlib.defmatrix.matrix>

Creating a view on a structured array so it can be used in calculations

>>> x = np.array([(1, 2),(3,4)], dtype=[(a, np.int8), (b, np.int8)])

>>> xv = x.view(dtype=np.int8).reshape(-1,2)
>>> xv
array([[1, 2],
[3, 4]], dtype=int8)
>>> xv.mean(0)
array([ 2., 3.])

Making changes to the view changes the underlying array

>>> xv[0,1] = 20
>>> print x
[(1, 20) (3, 4)]

Using a view to convert an array to a record array:

>>> z = x.view(np.recarray)
>>> z.a
array([1], dtype=int8)

Views share data:

>>> x[0] = (9, 10)

>>> z[0]
(9, 10)

zfill(width)
Return the numeric string left-filled with zeros in a string of length width.
See Also:

996 Chapter 3. Routines

 NumPy Reference, Release 2.0.0.dev8464

char.zfill

3.31. String operations 997

NumPy Reference, Release 2.0.0.dev8464

998 Chapter 3. Routines

 CHAPTER

FOUR

PACKAGING (NUMPY.DISTUTILS)

NumPy provides enhanced distutils functionality to make it easier to build and install sub-packages, auto-generate
code, and extension modules that use Fortran-compiled libraries. To use features of NumPy distutils, use the
setup command from numpy.distutils.core. A useful Configuration class is also provided in
numpy.distutils.misc_util that can make it easier to construct keyword arguments to pass to the setup
function (by passing the dictionary obtained from the todict() method of the class). More information is available in
the NumPy Distutils Users Guide in <site-packages>/numpy/doc/DISTUTILS.txt.

4.1 Modules in numpy.distutils

4.1.1 misc_util

Configuration(**attrs[,
package_name, ...])
get_numpy_include_dirs()
dict_append(d, **kws)
appendpath(prefix, path)
allpath(name) Convert a /-separated pathname to one using the OSs path separator.
dot_join(*args)
generate_config_py(target) Generate config.py file containing system_info information used during
building the package.
get_cmd(cmdname[, _cache])
terminal_has_colors()
red_text(s)
green_text(s)
yellow_text(s)
blue_text(s)
cyan_text(s)
cyg2win32(path)
all_strings(lst) Return True if all items in lst are string objects.
has_f_sources(sources) Return True if sources contains Fortran files
has_cxx_sources(sources) Return True if sources contains C++ files
filter_sources(sources) Return four lists of filenames containing
get_dependencies(sources)
is_local_src_dir(directory) Return true if directory is local directory.
get_ext_source_files(ext)
get_script_files(scripts)

999
NumPy Reference, Release 2.0.0.dev8464

class Configuration(package_name=None, parent_name=None, top_path=None, package_path=None,

caller_level=1, setup_name=setup.py, **attrs)

get_numpy_include_dirs()

dict_append(d, **kws)

appendpath(prefix, path)

allpath(name)
Convert a /-separated pathname to one using the OSs path separator.
dot_join(*args)

generate_config_py(target)
Generate config.py file containing system_info information used during building the package.
Usage:
config[py_modules].append((packagename, __config__,generate_config_py))
get_cmd(cmdname, _cache={})

terminal_has_colors()

red_text(s)

green_text(s)

yellow_text(s)

blue_text(s)

cyan_text(s)

cyg2win32(path)

all_strings(lst)
Return True if all items in lst are string objects.
has_f_sources(sources)
Return True if sources contains Fortran files
has_cxx_sources(sources)
Return True if sources contains C++ files
filter_sources(sources)
Return four lists of filenames containing C, C++, Fortran, and Fortran 90 module sources, respectively.
get_dependencies(sources)

1000 Chapter 4. Packaging (numpy.distutils)

 NumPy Reference, Release 2.0.0.dev8464

is_local_src_dir(directory)
Return true if directory is local directory.
get_ext_source_files(ext)

get_script_files(scripts)

class Configuration(package_name=None, parent_name=None, top_path=None, package_path=None, **at-

trs)
Construct a configuration instance for the given package name. If parent_name is not None, then construct the
package as a sub-package of the parent_name package. If top_path and package_path are None then they are
assumed equal to the path of the file this instance was created in. The setup.py files in the numpy distribution
are good examples of how to use the Configuration instance.
todict()
Return a dictionary compatible with the keyword arguments of distutils setup function.

Examples

>>> setup(**config.todict())

get_distribution()
Return the distutils distribution object for self.
get_subpackage(subpackage_name, subpackage_path=None, parent_name=None, caller_level=1)
Return list of subpackage configurations.
Parameters
subpackage_name: str,None :
Name of the subpackage to get the configuration. * in subpackage_name is handled
as a wildcard.
subpackage_path: str :
If None, then the path is assumed to be the local path plus the subpackage_name. If a
setup.py file is not found in the subpackage_path, then a default configuration is used.
parent_name: str :
Parent name.
add_subpackage(subpackage_name, subpackage_path=None, standalone=False)
Add a sub-package to the current Configuration instance.
This is useful in a setup.py script for adding sub-packages to a package.
Parameters
subpackage_name: str :
name of the subpackage
subpackage_path: str :
if given, the subpackage path such as the subpackage is in subpackage_path / subpack-
age_name. If None,the subpackage is assumed to be located in the local path / subpack-
age_name.
standalone: bool :
add_data_files(*files)
Add data files to configuration data_files.

4.1. Modules in numpy.distutils 1001

NumPy Reference, Release 2.0.0.dev8464

Parameters
files: sequence :
Argument(s) can be either
2-sequence (<datadir prefix>,<path to data file(s)>)
paths to data files where python datadir prefix defaults to package dir.

Notes
The form of each element of the files sequence is very flexible allowing many combinations of where to
get the files from the package and where they should ultimately be installed on the system. The most basic
usage is for an element of the files argument sequence to be a simple filename. This will cause that file
from the local path to be installed to the installation path of the self.name package (package path). The file
argument can also be a relative path in which case the entire relative path will be installed into the package
directory. Finally, the file can be an absolute path name in which case the file will be found at the absolute
path name but installed to the package path.
This basic behavior can be augmented by passing a 2-tuple in as the file argument. The first element of the
tuple should specify the relative path (under the package install directory) where the remaining sequence
of files should be installed to (it has nothing to do with the file-names in the source distribution). The
second element of the tuple is the sequence of files that should be installed. The files in this sequence can
be filenames, relative paths, or absolute paths. For absolute paths the file will be installed in the top-level
package installation directory (regardless of the first argument). Filenames and relative path names will be
installed in the package install directory under the path name given as the first element of the tuple.
Rules for installation paths:
1.file.txt -> (., file.txt)-> parent/file.txt
2.foo/file.txt -> (foo, foo/file.txt) -> parent/foo/file.txt
3./foo/bar/file.txt -> (., /foo/bar/file.txt) -> parent/file.txt
4.*.txt -> parent/a.txt, parent/b.txt
5.foo/*.txt -> parent/foo/a.txt, parent/foo/b.txt
6./.txt -> (, */.txt) -> parent/c/a.txt, parent/d/b.txt
7.(sun, file.txt) -> parent/sun/file.txt
8.(sun, bar/file.txt) -> parent/sun/file.txt
9.(sun, /foo/bar/file.txt) -> parent/sun/file.txt
10.(sun, *.txt) -> parent/sun/a.txt, parent/sun/b.txt
11.(sun, bar/*.txt) -> parent/sun/a.txt, parent/sun/b.txt
12.(sun/, */.txt) -> parent/sun/c/a.txt, parent/d/b.txt
An additional feature is that the path to a data-file can actually be a function that takes no arguments and
returns the actual path(s) to the data-files. This is useful when the data files are generated while building
the package.

Examples
Add files to the list of data_files to be included with the package.

>>> self.add_data_files(foo.dat,
... (fun, [gun.dat, nun/pun.dat, /tmp/sun.dat]),

1002 Chapter 4. Packaging (numpy.distutils)

 NumPy Reference, Release 2.0.0.dev8464

... bar/cat.dat,
... /full/path/to/can.dat)

will install these data files to:

<package install directory>/

foo.dat
fun/
gun.dat
nun/
pun.dat
sun.dat
bar/
car.dat
can.dat

where <package install directory> is the package (or sub-package) directory such as
/usr/lib/python2.4/site-packages/mypackage (C: Python2.4 Lib site-packages mypackage) or
/usr/lib/python2.4/site- packages/mypackage/mysubpackage (C: Python2.4 Lib site-packages my-
package mysubpackage).
add_data_dir(data_path)
Recursively add files under data_path to data_files list.
Recursively add files under data_path to the list of data_files to be installed (and distributed). The data_path
can be either a relative path-name, or an absolute path-name, or a 2-tuple where the first argument shows
where in the install directory the data directory should be installed to.
Parameters
data_path: seq,str :
Argument can be either
2-sequence (<datadir suffix>,<path to data directory>)
path to data directory where python datadir suffix defaults to package dir.

Notes
Rules for installation paths:
foo/bar -> (foo/bar, foo/bar) -> parent/foo/bar (gun, foo/bar) -> parent/gun foo/* -> (foo/a, foo/a),
(foo/b, foo/b) -> parent/foo/a, parent/foo/b (gun, foo/) -> (gun, foo/a), (gun, foo/b) -> gun (gun/, foo/)
-> parent/gun/a, parent/gun/b /foo/bar -> (bar, /foo/bar) -> parent/bar (gun, /foo/bar) -> parent/gun
(fun/ /gun/*, sun/foo/bar) -> parent/fun/foo/gun/bar

Examples
For example suppose the source directory contains fun/foo.dat and fun/bar/car.dat:

>>> self.add_data_dir(fun)
>>> self.add_data_dir((sun, fun))
>>> self.add_data_dir((gun, /full/path/to/fun))

Will install data-files to the locations:

<package install directory>/

fun/
foo.dat
bar/

4.1. Modules in numpy.distutils 1003

NumPy Reference, Release 2.0.0.dev8464

car.dat
sun/
foo.dat
bar/
car.dat
gun/
foo.dat
car.dat

add_include_dirs(*paths)
Add paths to configuration include directories.
Add the given sequence of paths to the beginning of the include_dirs list. This list will be visible to all
extension modules of the current package.
add_headers(*files)
Add installable headers to configuration.
Add the given sequence of files to the beginning of the headers list. By default, headers will be installed
under <python- include>/<self.name.replace(.,/)>/ directory. If an item of files is a tuple, then its first
argument specifies the actual installation location relative to the <python-include> path.
Parameters
files: str, seq :
Argument(s) can be either:
2-sequence (<includedir suffix>,<path to header file(s)>)
path(s) to header file(s) where python includedir suffix will default to package name.
add_extension(name, sources, **kw)
Add extension to configuration.
Create and add an Extension instance to the ext_modules list. This method also takes the following optional
keyword arguments that are passed on to the Extension constructor.
Parameters
name: str :
name of the extension
sources: seq :
list of the sources. The list of sources may contain functions (called source generators)
which must take an extension instance and a build directory as inputs and return a source
file or list of source files or None. If None is returned then no sources are generated.
If the Extension instance has no sources after processing all source generators, then no
extension module is built.
include_dirs: :
define_macros: :
undef_macros: :
library_dirs: :
libraries: :
runtime_library_dirs: :
extra_objects: :
extra_compile_args: :

1004 Chapter 4. Packaging (numpy.distutils)

 NumPy Reference, Release 2.0.0.dev8464

extra_link_args: :
export_symbols: :
swig_opts: :
depends: :
The depends list contains paths to files or directories that the sources of the extension
module depend on. If any path in the depends list is newer than the extension module,
then the module will be rebuilt.
language: :
f2py_options: :
module_dirs: :
extra_info: dict,list :
dict or list of dict of keywords to be appended to keywords.

Notes
The self.paths(...) method is applied to all lists that may contain paths.
add_library(name, sources, **build_info)
Add library to configuration.
Parameters
name : str
Name of the extension.
sources : sequence
List of the sources. The list of sources may contain functions (called source generators)
which must take an extension instance and a build directory as inputs and return a source
file or list of source files or None. If None is returned then no sources are generated.
If the Extension instance has no sources after processing all source generators, then no
extension module is built.
build_info : dict, optional
The following keys are allowed:
depends
macros
include_dirs
extra_compiler_args
f2py_options
language
add_scripts(*files)
Add scripts to configuration.
Add the sequence of files to the beginning of the scripts list. Scripts will be installed under the <prefix>/bin/
directory.
add_installed_library(name, sources, install_dir, build_info=None)
Similar to add_library, but the specified library is installed.

4.1. Modules in numpy.distutils 1005

NumPy Reference, Release 2.0.0.dev8464

Most C libraries used with distutils are only used to build python extensions, but libraries built
through this method will be installed so that they can be reused by third-party packages.
Parameters
name : str
Name of the installed library.
sources : sequence
List of the librarys source files. See add_library for details.
install_dir : str
Path to install the library, relative to the current sub-package.
build_info : dict, optional
The following keys are allowed:
depends
macros
include_dirs
extra_compiler_args
f2py_options
language
Returns
None :
See Also:
add_library, add_npy_pkg_config, get_info

Notes
The best way to encode the options required to link against the specified C libraries is to use a libname.ini
file, and use get_info to retrieve the required options (see add_npy_pkg_config for more information).
add_npy_pkg_config(template, install_dir, subst_dict=None)
Generate and install a npy-pkg config file from a template.
The config file generated from template is installed in the given install directory, using subst_dict for
variable substitution.
Parameters
template : str
The path of the template, relatively to the current package path.
install_dir : str
Where to install the npy-pkg config file, relatively to the current package path.
subst_dict : dict, optional
If given, any string of the form @key@ will be replaced by subst_dict[key] in the
template file when installed. The install prefix is always available through the variable
@prefix@, since the install prefix is not easy to get reliably from setup.py.
See Also:
add_installed_library, get_info

1006 Chapter 4. Packaging (numpy.distutils)

 NumPy Reference, Release 2.0.0.dev8464

Notes
This works for both standard installs and in-place builds, i.e. the @prefix@ refer to the source directory
for in-place builds.

Examples

config.add_npy_pkg_config(foo.ini.in, lib, {foo: bar})

Assuming the foo.ini.in file has the following content:

[meta]
Name=@foo@
Version=1.0
Description=dummy description

[default]
Cflags=-I@prefix@/include
Libs=

The generated file will have the following content:

[meta]
Name=bar
Version=1.0
Description=dummy description

[default]
Cflags=-Iprefix_dir/include
Libs=

and will be installed as foo.ini in the lib subpath.

paths(*paths, **kws)
Apply glob to paths and prepend local_path if needed.
Applies glob.glob(...) to each path in the sequence (if needed) and pre-pends the local_path if needed.
Because this is called on all source lists, this allows wildcard characters to be specified in lists of sources
for extension modules and libraries and scripts and allows path-names be relative to the source directory.
get_config_cmd()
Returns the numpy.distutils config command instance.
get_build_temp_dir()
Return a path to a temporary directory where temporary files should be placed.
have_f77c()
Check for availability of Fortran 77 compiler.
Use it inside source generating function to ensure that setup distribution instance has been initialized.

Notes
True if a Fortran 77 compiler is available (because a simple Fortran 77 code was able to be compiled
successfully).
have_f90c()
Check for availability of Fortran 90 compiler.
Use it inside source generating function to ensure that setup distribution instance has been initialized.

4.1. Modules in numpy.distutils 1007

NumPy Reference, Release 2.0.0.dev8464

Notes
True if a Fortran 90 compiler is available (because a simple Fortran 90 code was able to be compiled
successfully)
get_version(version_file=None, version_variable=None)
Try to get version string of a package.
Return a version string of the current package or None if the version information could not be detected.

Notes
This method scans files named __version__.py, <packagename>_version.py, version.py, and
__svn_version__.py for string variables version, __version__, and <packagename>_version, until a ver-
sion number is found.
make_svn_version_py(delete=True)
Appends a data function to the data_files list that will generate __svn_version__.py file to the current
package directory.
Generate package __svn_version__.py file from SVN revision number, it will be removed after python
exits but will be available when sdist, etc commands are executed.

Notes
If __svn_version__.py existed before, nothing is done.
This is intended for working with source directories that are in an SVN repository.
make_config_py(name=__config__)
Generate package __config__.py file containing system_info information used during building the pack-
age.
This file is installed to the package installation directory.
get_info(*names)
Get resources information.
Return information (from system_info.get_info) for all of the names in the argument list in a single dictio-
nary.

4.1.2 Other modules

system_info.get_info(name[, notfound_action]) notfound_action:

system_info.get_standard_file(fname) Returns a list of files named fname from
cpuinfo.cpu
log.set_verbosity(v[, force])
exec_command exec_command
get_info(name, notfound_action=0)

notfound_action:
0 - do nothing 1 - display warning message 2 - raise error
get_standard_file(fname)
Returns a list of files named fname from 1) System-wide directory (directory-location of this module) 2)
Users HOME directory (os.environ[HOME]) 3) Local directory
cpu

1008 Chapter 4. Packaging (numpy.distutils)

 NumPy Reference, Release 2.0.0.dev8464

set_verbosity(v, force=False)

exec_command
Implements exec_command function that is (almost) equivalent to commands.getstatusoutput function but on NT,
DOS systems the returned status is actually correct (though, the returned status values may be different by a factor).
In addition, exec_command takes keyword arguments for (re-)defining environment variables.
Provides functions:

exec_command execute command in a specified directory and

in the modified environment.
find_executable locate a command using info from environment
variable PATH. Equivalent to posix which command.
Author: Pearu Peterson <[email protected]> Created: 11 January 2003
Requires: Python 2.x
Succesfully tested on:
os.name | sys.platform | comments ++- posix | linux2 | Debian (sid) Linux, Python
2.1.3+, 2.2.3+, 2.3.3
PyCrust 0.9.3, Idle 1.0.2
posix | linux2 | Red Hat 9 Linux, Python 2.1.3, 2.2.2, 2.3.2 posix | sunos5 | SunOS 5.9, Python 2.2, 2.3.2 posix |
darwin | Darwin 7.2.0, Python 2.3 nt | win32 | Windows Me
Python 2.3(EE), Idle 1.0, PyCrust 0.7.2 Python 2.1.1 Idle 0.8
nt | win32 | Windows 98, Python 2.1.1. Idle 0.8 nt | win32 | Cygwin 98-4.10, Python 2.1.1(MSC) - echo tests
fail i.e. redefining environment variables may not work. FIXED: dont use cygwin echo! Comment:
also cmd /c echo will not work but redefining environment variables do work.
posix | cygwin | Cygwin 98-4.10, Python 2.3.3(cygming special) nt | win32 | Windows XP, Python 2.3.3
Known bugs: - Tests, that send messages to stderr, fail when executed from MSYS prompt
because the messages are lost at some point.

Functions
exec_command(command, **env[, execute_in, ...]) Return (status,output) of executed command.
find_executable(exe[, path, _cache]) Return full path of a executable or None.
get_exception()
get_pythonexe()
is_sequence(seq)
make_temp_file([suffix, prefix, text])
open_latin1(filename[, mode])
quote_arg(arg)
splitcmdline(line)
temp_file_name()
test(**kws)
test_cl(**kws)
test_execute_in(**kws)
test_nt(**kws)
test_posix(**kws)
test_svn(**kws)

4.1. Modules in numpy.distutils 1009

NumPy Reference, Release 2.0.0.dev8464

4.2 Building Installable C libraries

Conventional C libraries (installed through add_library) are not installed, and are just used during the build (they
are statically linked). An installable C library is a pure C library, which does not depend on the python C runtime,
and is installed such that it may be used by third-party packages. To build and install the C library, you just use
the method add_installed_library instead of add_library, which takes the same arguments except for an additional
install_dir argument:

>>> config.add_installed_library(foo, sources=[foo.c], install_dir=lib)

4.2.1 npy-pkg-config files

To make the necessary build options available to third parties, you could use the npy-pkg-config mechanism imple-
mented in numpy.distutils. This mechanism is based on a .ini file which contains all the options. A .ini file is
very similar to .pc files as used by the pkg-config unix utility:

[meta]
Name: foo
Version: 1.0
Description: foo library

[variables]
prefix = /home/user/local
libdir = ${prefix}/lib
includedir = ${prefix}/include

[default]
cflags = -I${includedir}
libs = -L${libdir} -lfoo

Generally, the file needs to be generated during the build, since it needs some information known at build time only
(e.g. prefix). This is mostly automatic if one uses the Configuration method add_npy_pkg_config. Assuming we have
a template file foo.ini.in as follows:

[meta]
Name: foo
Version: @version@
Description: foo library

[variables]
prefix = @prefix@
libdir = ${prefix}/lib
includedir = ${prefix}/include

[default]
cflags = -I${includedir}
libs = -L${libdir} -lfoo

and the following code in setup.py:

>>> config.add_installed_library(foo, sources=[foo.c], install_dir=lib)

>>> subst = {version: 1.0}
>>> config.add_npy_pkg_config(foo.ini.in, lib, subst_dict=subst)

1010 Chapter 4. Packaging (numpy.distutils)

 NumPy Reference, Release 2.0.0.dev8464

This will install the file foo.ini into the directory package_dir/lib, and the foo.ini file will be generated from foo.ini.in,
where each @version@ will be replaced by subst_dict[version]. The dictionary has an additional prefix
substitution rule automatically added, which contains the install prefix (since this is not easy to get from setup.py).
npy-pkg-config files can also be installed at the same location as used for numpy, using the path returned from
get_npy_pkg_dir function.

4.2.2 Reusing a C library from another package

Info are easily retrieved from the get_info function in numpy.distutils.misc_util:

>>> info = get_info(npymath)

>>> config.add_extension(foo, sources=[foo.c], extra_info=**info)

An additional list of paths to look for .ini files can be given to get_info.

4.3 Conversion of .src files

NumPy distutils supports automatic conversion of source files named <somefile>.src. This facility can be used to
maintain very similar code blocks requiring only simple changes between blocks. During the build phase of setup, if
a template file named <somefile>.src is encountered, a new file named <somefile> is constructed from the template
and placed in the build directory to be used instead. Two forms of template conversion are supported. The first form
occurs for files named named <file>.ext.src where ext is a recognized Fortran extension (f, f90, f95, f77, for, ftn, pyf).
The second form is used for all other cases.

4.3.1 Fortran files

This template converter will replicate all function and subroutine blocks in the file with names that contain <...>
according to the rules in <...>. The number of comma-separated words in <...> determines the number of times
the block is repeated. What these words are indicates what that repeat rule, <...>, should be replaced with in each
block. All of the repeat rules in a block must contain the same number of comma-separated words indicating the
number of times that block should be repeated. If the word in the repeat rule needs a comma, leftarrow, or rightarrow,
then prepend it with a backslash . If a word in the repeat rule matches \<index> then it will be replaced with the
<index>-th word in the same repeat specification. There are two forms for the repeat rule: named and short.

Named repeat rule

A named repeat rule is useful when the same set of repeats must be used several times in a block. It is specified using
<rule1=item1, item2, item3,..., itemN>, where N is the number of times the block should be repeated. On each repeat
of the block, the entire expression, <...> will be replaced first with item1, and then with item2, and so forth until N
repeats are accomplished. Once a named repeat specification has been introduced, the same repeat rule may be used
in the current block by referring only to the name (i.e. <rule1>.

Short repeat rule

A short repeat rule looks like <item1, item2, item3, ..., itemN>. The rule specifies that the entire expression, <...>
should be replaced first with item1, and then with item2, and so forth until N repeats are accomplished.

4.3. Conversion of .src files 1011

NumPy Reference, Release 2.0.0.dev8464

Pre-defined names

The following predefined named repeat rules are available:

<prefix=s,d,c,z>
<_c=s,d,c,z>
<_t=real, double precision, complex, double complex>
<ftype=real, double precision, complex, double complex>
<ctype=float, double, complex_float, complex_double>
<ftypereal=float, double precision, \0, \1>
<ctypereal=float, double, \0, \1>

4.3.2 Other files

Non-Fortran files use a separate syntax for defining template blocks that should be repeated using a variable expansion
similar to the named repeat rules of the Fortran-specific repeats. The template rules for these files are:
1. /**begin repeat on a line by itself marks the beginning of a segment that should be repeated.
2. Named variable expansions are defined using #name=item1, item2, item3, ..., itemN# and placed on successive
lines. These variables are replaced in each repeat block with corresponding word. All named variables in the
same repeat block must define the same number of words.
3. In specifying the repeat rule for a named variable, item*N is short- hand for item, item, ..., item repeated N
times. In addition, parenthesis in combination with *N can be used for grouping several items that should be
repeated. Thus, #name=(item1, item2)*4# is equivalent to #name=item1, item2, item1, item2, item1, item2,
item1, item2#
4. */ on a line by itself marks the end of the the variable expansion naming. The next line is the first line that
will be repeated using the named rules.
5. Inside the block to be repeated, the variables that should be expanded are specified as @name@.
6. /**end repeat**/ on a line by itself marks the previous line as the last line of the block to be repeated.

1012 Chapter 4. Packaging (numpy.distutils)

 CHAPTER

FIVE

NUMPY C-API

Beware of the man who wont be bothered with details.

William Feather, Sr. The truth is out there.
Chris Carter, The X Files
NumPy provides a C-API to enable users to extend the system and get access to the array object for use in other
routines. The best way to truly understand the C-API is to read the source code. If you are unfamiliar with (C) source
code, however, this can be a daunting experience at first. Be assured that the task becomes easier with practice, and
you may be surprised at how simple the C-code can be to understand. Even if you dont think you can write C-code
from scratch, it is much easier to understand and modify already-written source code then create it de novo.
Python extensions are especially straightforward to understand because they all have a very similar structure. Admit-
tedly, NumPy is not a trivial extension to Python, and may take a little more snooping to grasp. This is especially true
because of the code-generation techniques, which simplify maintenance of very similar code, but can make the code a
little less readable to beginners. Still, with a little persistence, the code can be opened to your understanding. It is my
hope, that this guide to the C-API can assist in the process of becoming familiar with the compiled-level work that can
be done with NumPy in order to squeeze that last bit of necessary speed out of your code.

5.1 Python Types and C-Structures

Several new types are defined in the C-code. Most of these are accessible from Python, but a few are not exposed due
to their limited use. Every new Python type has an associated PyObject * with an internal structure that includes
a pointer to a method table that defines how the new object behaves in Python. When you receive a Python object
into C code, you always get a pointer to a PyObject structure. Because a PyObject structure is very generic and
defines only PyObject_HEAD, by itself it is not very interesting. However, different objects contain more details
after the PyObject_HEAD (but you have to cast to the correct type to access them or use accessor functions or
macros).

5.1.1 New Python Types Defined

Python types are the functional equivalent in C of classes in Python. By constructing a new Python type you make
available a new object for Python. The ndarray object is an example of a new type defined in C. New types are defined
in C by two basic steps:
1. creating a C-structure (usually named Py{Name}Object) that is binary- compatible with the PyObject
structure itself but holds the additional information needed for that particular object;
2. populating the PyTypeObject table (pointed to by the ob_type member of the PyObject structure) with
pointers to functions that implement the desired behavior for the type.

1013
NumPy Reference, Release 2.0.0.dev8464

Instead of special method names which define behavior for Python classes, there are function tables which point to
functions that implement the desired results. Since Python 2.2, the PyTypeObject itself has become dynamic which
allows C types that can be sub-typed from other C-types in C, and sub-classed in Python. The children types inherit
the attributes and methods from their parent(s).
There are two major new types: the ndarray ( PyArray_Type ) and the ufunc ( PyUFunc_Type ). Ad-
ditional types play a supportive role: the PyArrayIter_Type, the PyArrayMultiIter_Type, and the
PyArrayDescr_Type . The PyArrayIter_Type is the type for a flat iterator for an ndarray (the object that
is returned when getting the flat attribute). The PyArrayMultiIter_Type is the type of the object returned
when calling broadcast (). It handles iteration and broadcasting over a collection of nested sequences. Also, the
PyArrayDescr_Type is the data-type-descriptor type whose instances describe the data. Finally, there are 21 new
scalar-array types which are new Python scalars corresponding to each of the fundamental data types available for
arrays. An additional 10 other types are place holders that allow the array scalars to fit into a hierarchy of actual
Python types.

PyArray_Type

PyArray_Type
The Python type of the ndarray is PyArray_Type. In C, every ndarray is a pointer to a PyArrayObject
structure. The ob_type member of this structure contains a pointer to the PyArray_Type typeobject.
PyArrayObject
The PyArrayObject C-structure contains all of the required information for an array. All instances of an
ndarray (and its subclasses) will have this structure. For future compatibility, these structure members should
normally be accessed using the provided macros. If you need a shorter name, then you can make use of
NPY_AO which is defined to be equivalent to PyArrayObject.

typedef struct PyArrayObject {

PyObject_HEAD
char *data;
int nd;
npy_intp *dimensions;
npy_intp *strides;
PyObject *base;
PyArray_Descr *descr;
int flags;
PyObject *weakreflist;
} PyArrayObject;

PyObject_HEAD
This is needed by all Python objects. It consists of (at least) a reference count member ( ob_refcnt ) and
a pointer to the typeobject ( ob_type ). (Other elements may also be present if Python was compiled with
special options see Include/object.h in the Python source tree for more information). The ob_type member
points to a Python type object.
char data
A pointer to the first element of the array. This pointer can (and normally should) be recast to the data type of
the array.
int nd
An integer providing the number of dimensions for this array. When nd is 0, the array is sometimes called a
rank-0 array. Such arrays have undefined dimensions and strides and cannot be accessed. NPY_MAXDIMS is
the largest number of dimensions for any array.
npy_intp dimensions
An array of integers providing the shape in each dimension as long as nd 1. The integer is always large
enough to hold a pointer on the platform, so the dimension size is only limited by memory.

1014 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

npy_intp strides
An array of integers providing for each dimension the number of bytes that must be skipped to get to the next
element in that dimension.
PyObject base
This member is used to hold a pointer to another Python object that is related to this array. There are two use
cases: 1) If this array does not own its own memory, then base points to the Python object that owns it (perhaps
another array object), 2) If this array has the NPY_UPDATEIFCOPY flag set, then this array is a working copy
of a misbehaved array. As soon as this array is deleted, the array pointed to by base will be updated with the
contents of this array.
PyArray_Descr descr
A pointer to a data-type descriptor object (see below). The data-type descriptor object is an instance of a new
built-in type which allows a generic description of memory. There is a descriptor structure for each data type
supported. This descriptor structure contains useful information about the type as well as a pointer to a table of
function pointers to implement specific functionality.
int flags
Flags indicating how the memory pointed to by data is to be interpreted. Possible flags are
NPY_C_CONTIGUOUS, NPY_F_CONTIGUOUS, NPY_OWNDATA, NPY_ALIGNED, NPY_WRITEABLE,
and NPY_UPDATEIFCOPY.
PyObject weakreflist
This member allows array objects to have weak references (using the weakref module).

PyArrayDescr_Type

PyArrayDescr_Type
The PyArrayDescr_Type is the built-in type of the data-type-descriptor objects used to describe how the
bytes comprising the array are to be interpreted. There are 21 statically-defined PyArray_Descr objects
for the built-in data-types. While these participate in reference counting, their reference count should never
reach zero. There is also a dynamic table of user-defined PyArray_Descr objects that is also main-
tained. Once a data-type-descriptor object is registered it should never be deallocated either. The function
PyArray_DescrFromType (...) can be used to retrieve a PyArray_Descr object from an enumerated
type-number (either built-in or user- defined).
PyArray_Descr
The format of the PyArray_Descr structure that lies at the heart of the PyArrayDescr_Type is

typedef struct {
PyObject_HEAD
PyTypeObject *typeobj;
char kind;
char type;
char byteorder;
char unused;
int flags;
int type_num;
int elsize;
int alignment;
PyArray_ArrayDescr *subarray;
PyObject *fields;
PyArray_ArrFuncs *f;
} PyArray_Descr;

PyTypeObject typeobj
Pointer to a typeobject that is the corresponding Python type for the elements of this array. For the builtin

5.1. Python Types and C-Structures 1015

NumPy Reference, Release 2.0.0.dev8464

types, this points to the corresponding array scalar. For user-defined types, this should point to a user-defined
typeobject. This typeobject can either inherit from array scalars or not. If it does not inherit from array scalars,
then the NPY_USE_GETITEM and NPY_USE_SETITEM flags should be set in the flags member.
char kind
A character code indicating the kind of array (using the array interface typestring notation). A b represents
Boolean, a i represents signed integer, a u represents unsigned integer, f represents floating point, c
represents complex floating point, S represents 8-bit character string, U represents 32-bit/character unicode
string, and V repesents arbitrary.
char type
A traditional character code indicating the data type.
char byteorder
A character indicating the byte-order: > (big-endian), < (little- endian), = (native), | (irrelevant, ignore).
All builtin data- types have byteorder =.
int flags
A data-type bit-flag that determines if the data-type exhibits object- array like behavior. Each bit in this member
is a flag which are named as:
NPY_ITEM_REFCOUNT

NPY_ITEM_HASOBJECT
Indicates that items of this data-type must be reference counted (using Py_INCREF and Py_DECREF ).
NPY_ITEM_LISTPICKLE
Indicates arrays of this data-type must be converted to a list before pickling.
NPY_ITEM_IS_POINTER
Indicates the item is a pointer to some other data-type
NPY_NEEDS_INIT
Indicates memory for this data-type must be initialized (set to 0) on creation.
NPY_NEEDS_PYAPI
Indicates this data-type requires the Python C-API during access (so dont give up the GIL if array access
is going to be needed).
NPY_USE_GETITEM
On array access use the f->getitem function pointer instead of the standard conversion to an array
scalar. Must use if you dont define an array scalar to go along with the data-type.
NPY_USE_SETITEM
When creating a 0-d array from an array scalar use f->setitem instead of the standard copy from an
array scalar. Must use if you dont define an array scalar to go along with the data-type.
NPY_FROM_FIELDS
The bits that are inherited for the parent data-type if these bits are set in any field of the data-type. Cur-
rently ( NPY_NEEDS_INIT | NPY_LIST_PICKLE | NPY_ITEM_REFCOUNT | NPY_NEEDS_PYAPI
).
NPY_OBJECT_DTYPE_FLAGS
Bits set for the object data-type: ( NPY_LIST_PICKLE | NPY_USE_GETITEM |
NPY_ITEM_IS_POINTER | NPY_REFCOUNT | NPY_NEEDS_INIT | NPY_NEEDS_PYAPI).
PyDataType_FLAGCHK(PyArray_Descr *dtype, int flags)
Return true if all the given flags are set for the data-type object.
PyDataType_REFCHK(PyArray_Descr *dtype)
Equivalent to PyDataType_FLAGCHK (dtype, NPY_ITEM_REFCOUNT).

1016 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

int type_num
A number that uniquely identifies the data type. For new data-types, this number is assigned when the data-type
is registered.
int elsize
For data types that are always the same size (such as long), this holds the size of the data type. For flexible data
types where different arrays can have a different elementsize, this should be 0.
int alignment
A number providing alignment information for this data type. Specifically, it shows how far from the start
of a 2-element structure (whose first element is a char ), the compiler places an item of this type:
offsetof(struct {char c; type v;}, v)
PyArray_ArrayDescr subarray
If this is non- NULL, then this data-type descriptor is a C-style contiguous array of another data-type descriptor.
In other-words, each element that this descriptor describes is actually an array of some other base descriptor.
This is most useful as the data-type descriptor for a field in another data-type descriptor. The fields member
should be NULL if this is non- NULL (the fields member of the base descriptor can be non- NULL however).
The PyArray_ArrayDescr structure is defined using

typedef struct {
PyArray_Descr *base;
PyObject *shape;
} PyArray_ArrayDescr;

The elements of this structure are:

PyArray_Descr base
The data-type-descriptor object of the base-type.
PyObject shape
The shape (always C-style contiguous) of the sub-array as a Python tuple.
PyObject fields
If this is non-NULL, then this data-type-descriptor has fields described by a Python dictionary whose keys
are names (and also titles if given) and whose values are tuples that describe the fields. Recall that a
data-type-descriptor always describes a fixed-length set of bytes. A field is a named sub-region of that total,
fixed-length collection. A field is described by a tuple composed of another data- type-descriptor and a byte
offset. Optionally, the tuple may contain a title which is normally a Python string. These tuples are placed in
this dictionary keyed by name (and also title if given).
PyArray_ArrFuncs f
A pointer to a structure containing functions that the type needs to implement internal features. These functions
are not the same thing as the universal functions (ufuncs) described later. Their signatures can vary arbitrarily.
PyArray_ArrFuncs
Functions implementing internal features. Not all of these function pointers must be defined for a given type.
The required members are nonzero, copyswap, copyswapn, setitem, getitem, and cast. These
are assumed to be non- NULL and NULL entries will cause a program crash. The other functions may be NULL
which will just mean reduced functionality for that data-type. (Also, the nonzero function will be filled in with
a default function if it is NULL when you register a user-defined data-type).

typedef struct {
PyArray_VectorUnaryFunc *cast[PyArray_NTYPES];
PyArray_GetItemFunc *getitem;
PyArray_SetItemFunc *setitem;
PyArray_CopySwapNFunc *copyswapn;
PyArray_CopySwapFunc *copyswap;

5.1. Python Types and C-Structures 1017

NumPy Reference, Release 2.0.0.dev8464

PyArray_CompareFunc *compare;
PyArray_ArgFunc *argmax;
PyArray_DotFunc *dotfunc;
PyArray_ScanFunc *scanfunc;
PyArray_FromStrFunc *fromstr;
PyArray_NonzeroFunc *nonzero;
PyArray_FillFunc *fill;
PyArray_FillWithScalarFunc *fillwithscalar;
PyArray_SortFunc *sort[PyArray_NSORTS];
PyArray_ArgSortFunc *argsort[PyArray_NSORTS];
PyObject *castdict;
PyArray_ScalarKindFunc *scalarkind;
int **cancastscalarkindto;
int *cancastto;
int listpickle
} PyArray_ArrFuncs;

The concept of a behaved segment is used in the description of the function pointers. A behaved segment is one
that is aligned and in native machine byte-order for the data-type. The nonzero, copyswap, copyswapn,
getitem, and setitem functions can (and must) deal with mis-behaved arrays. The other functions require
behaved memory segments.
void cast(void *from, void *to, npy_intp n, void *fromarr,

void *toarr)

An array of function pointers to cast from the current type to all of the other builtin types. Each
function casts a contiguous, aligned, and notswapped buffer pointed at by from to a contiguous,
aligned, and notswapped buffer pointed at by to The number of items to cast is given by n, and the
arguments fromarr and toarr are interpreted as PyArrayObjects for flexible arrays to get itemsize
information.
PyObject getitem
A pointer to a function that returns a standard Python object from a single element of the array object arr
pointed to by data. This function must be able to deal with misbehaved (misaligned and/or swapped)
arrays correctly.
int setitem
A pointer to a function that sets the Python object item into the array, arr, at the position pointed to by data
. This function deals with misbehaved arrays. If successful, a zero is returned, otherwise, a negative one
is returned (and a Python error set).
void copyswapn(void *dest, npy_intp dstride, void *src,

npy_intp sstride, npy_intp n, int swap, void *arr)

void copyswap
These members are both pointers to functions to copy data from src to dest and swap if indicated. The
value of arr is only used for flexible ( NPY_STRING, NPY_UNICODE, and NPY_VOID ) arrays (and is
obtained from arr->descr->elsize ). The second function copies a single value, while the first
loops over n values with the provided strides. These functions can deal with misbehaved src data. If src is
NULL then no copy is performed. If swap is 0, then no byteswapping occurs. It is assumed that dest and
src do not overlap. If they overlap, then use memmove (...) first followed by copyswap(n) with NULL
valued src.
int compare

1018 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

A pointer to a function that compares two elements of the array, arr, pointed to by d1 and d2. This
function requires behaved arrays. The return value is 1 if * d1 > * d2, 0 if * d1 == * d2, and -1 if * d1 <
* d2. The array object arr is used to retrieve itemsize and field information for flexible arrays.
int argmax(void* data, npy_intp n, npy_intp* max_ind,

void* arr)

A pointer to a function that retrieves the index of the largest of n elements in arr beginning at the
element pointed to by data. This function requires that the memory segment be contiguous and
behaved. The return value is always 0. The index of the largest element is returned in max_ind.
void dotfunc(void* ip1, npy_intp is1, void* ip2, npy_intp is2,

void* op, npy_intp n, void* arr)

A pointer to a function that multiplies two n -length sequences together, adds them, and places
the result in element pointed to by op of arr. The start of the two sequences are pointed to by
ip1 and ip2. To get to the next element in each sequence requires a jump of is1 and is2
bytes, respectively. This function requires behaved (though not necessarily contiguous) memory.
int scanfunc
A pointer to a function that scans (scanf style) one element of the corresponding type from the file de-
scriptor fd into the array memory pointed to by ip. The array is assumed to be behaved. If sep is not
NULL, then a separator string is also scanned from the file before returning. The last argument arr is the
array to be scanned into. A 0 is returned if the scan is successful. A negative number indicates something
went wrong: -1 means the end of file was reached before the separator string could be scanned, -4 means
that the end of file was reached before the element could be scanned, and -3 means that the element could
not be interpreted from the format string. Requires a behaved array.
int fromstr
A pointer to a function that converts the string pointed to by str to one element of the corresponding
type and places it in the memory location pointed to by ip. After the conversion is completed, *endptr
points to the rest of the string. The last argument arr is the array into which ip points (needed for
variable-size data- types). Returns 0 on success or -1 on failure. Requires a behaved array.
Bool nonzero
A pointer to a function that returns TRUE if the item of arr pointed to by data is nonzero. This function
can deal with misbehaved arrays.
void fill
A pointer to a function that fills a contiguous array of given length with data. The first two elements of
the array must already be filled- in. From these two values, a delta will be computed and the values from
item 3 to the end will be computed by repeatedly adding this computed delta. The data buffer must be
well-behaved.
void fillwithscalar(void* buffer, npy_intp length,

void* value, void* arr)

A pointer to a function that fills a contiguous buffer of the given length with a single scalar
value whose address is given. The final argument is the array which is needed to get the
itemsize for variable-length arrays.
int sort
An array of function pointers to a particular sorting algorithms. A particular sorting algorithm is obtained

5.1. Python Types and C-Structures 1019

NumPy Reference, Release 2.0.0.dev8464

using a key (so far PyArray_QUICKSORT, :dataPyArray_HEAPSORT, and PyArray_MERGESORT

are defined). These sorts are done in-place assuming contiguous and aligned data.
int argsort(void* start, npy_intp* result, npy_intp length,

void *arr)

An array of function pointers to sorting algorithms for this data type. The same sorting algorithms
as for sort are available. The indices producing the sort are returned in result (which must be
initialized with indices 0 to length-1 inclusive).
PyObject castdict
Either NULL or a dictionary containing low-level casting functions for user- defined data-types. Each
function is wrapped in a PyCObject * and keyed by the data-type number.
PyArray_SCALARKIND scalarkind
A function to determine how scalars of this type should be interpreted. The argument is NULL or a
0-dimensional array containing the data (if that is needed to determine the kind of scalar). The return
value must be of type PyArray_SCALARKIND.
int cancastscalarkindto
Either NULL or an array of PyArray_NSCALARKINDS pointers. These pointers should each be either
NULL or a pointer to an array of integers (terminated by PyArray_NOTYPE) indicating data-types that
a scalar of this data-type of the specified kind can be cast to safely (this usually means without losing
precision).
int cancastto
Either NULL or an array of integers (terminated by PyArray_NOTYPE ) indicated data-types that this
data-type can be cast to safely (this usually means without losing precision).
int listpickle
Unused.
The PyArray_Type typeobject implements many of the features of Python objects including the tp_as_number,
tp_as_sequence, tp_as_mapping, and tp_as_buffer interfaces. The rich comparison (tp_richcompare) is also used
along with new-style attribute lookup for methods (tp_methods) and properties (tp_getset). The PyArray_Type can
also be sub-typed.
Tip: The tp_as_number methods use a generic approach to call whatever function has been registered for handling
the operation. The function PyNumeric_SetOps(..) can be used to register functions to handle particular mathematical
operations (for all arrays). When the umath module is imported, it sets the numeric operations for all arrays to the
corresponding ufuncs. The tp_str and tp_repr methods can also be altered using PyString_SetStringFunction(...).

PyUFunc_Type

PyUFunc_Type
The ufunc object is implemented by creation of the PyUFunc_Type. It is a very simple type that implements
only basic getattribute behavior, printing behavior, and has call behavior which allows these objects to act
like functions. The basic idea behind the ufunc is to hold a reference to fast 1-dimensional (vector) loops for
each data type that supports the operation. These one-dimensional loops all have the same signature and are
the key to creating a new ufunc. They are called by the generic looping code as appropriate to implement the
N-dimensional function. There are also some generic 1-d loops defined for floating and complexfloating arrays
that allow you to define a ufunc using a single scalar function (e.g. atanh).
PyUFuncObject
The core of the ufunc is the PyUFuncObject which contains all the information needed to call the underlying
C-code loops that perform the actual work. It has the following structure:

1020 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

typedef struct {
PyObject_HEAD
int nin;
int nout;
int nargs;
int identity;
PyUFuncGenericFunction *functions;
void **data;
int ntypes;
int check_return;
char *name;
char *types;
char *doc;
void *ptr;
PyObject *obj;
PyObject *userloops;
} PyUFuncObject;

PyObject_HEAD
required for all Python objects.
int nin
The number of input arguments.
int nout
The number of output arguments.
int nargs
The total number of arguments (nin + nout). This must be less than NPY_MAXARGS.
int identity
Either PyUFunc_One, PyUFunc_Zero, or PyUFunc_None to indicate the identity for this operation.
It is only used for a reduce-like call on an empty array.
void PyUFuncObject.functions(char** args, npy_intp* dims,

npy_intp* steps, void* extradata)

An array of function pointers one for each data type supported by the ufunc. This is the vector
loop that is called to implement the underlying function dims [0] times. The first argument, args,
is an array of nargs pointers to behaved memory. Pointers to the data for the input arguments
are first, followed by the pointers to the data for the output arguments. How many bytes must be
skipped to get to the next element in the sequence is specified by the corresponding entry in the
steps array. The last argument allows the loop to receive extra information. This is commonly
used so that a single, generic vector loop can be used for multiple functions. In this case, the
actual scalar function to call is passed in as extradata. The size of this function pointer array is
ntypes.
void data
Extra data to be passed to the 1-d vector loops or NULL if no extra-data is needed. This C-array must be
the same size ( i.e. ntypes) as the functions array. NULL is used if extra_data is not needed. Several C-API
calls for UFuncs are just 1-d vector loops that make use of this extra data to receive a pointer to the actual
function to call.
int ntypes
The number of supported data types for the ufunc. This number specifies how many different 1-d loops
(of the builtin data types) are available.

5.1. Python Types and C-Structures 1021

NumPy Reference, Release 2.0.0.dev8464

int check_return
Obsolete and unused. However, it is set by the corresponding entry in the main ufunc creation routine:
PyUFunc_FromFuncAndData (...).
char name
A string name for the ufunc. This is used dynamically to build the __doc__ attribute of ufuncs.
char types
An array of nargs ntypes 8-bit type_numbers which contains the type signature for the function for
each of the supported (builtin) data types. For each of the ntypes functions, the corresponding set of type
numbers in this array shows how the args argument should be interpreted in the 1-d vector loop. These
type numbers do not have to be the same type and mixed-type ufuncs are supported.
char doc
Documentation for the ufunc. Should not contain the function signature as this is generated dynamically
when __doc__ is retrieved.
void ptr
Any dynamically allocated memory. Currently, this is used for dynamic ufuncs created from a python
function to store room for the types, data, and name members.
PyObject obj
For ufuncs dynamically created from python functions, this member holds a reference to the underlying
Python function.
PyObject userloops
A dictionary of user-defined 1-d vector loops (stored as CObject ptrs) for user-defined types. A loop may
be registered by the user for any user-defined type. It is retrieved by type number. User defined type
numbers are always larger than NPY_USERDEF.

PyArrayIter_Type

PyArrayIter_Type
This is an iterator object that makes it easy to loop over an N-dimensional array. It is the object returned from
the flat attribute of an ndarray. It is also used extensively throughout the implementation internals to loop
over an N-dimensional array. The tp_as_mapping interface is implemented so that the iterator object can be
indexed (using 1-d indexing), and a few methods are implemented through the tp_methods table. This object
implements the next method and can be used anywhere an iterator can be used in Python.
PyArrayIterObject
The C-structure corresponding to an object of PyArrayIter_Type is the PyArrayIterObject. The
PyArrayIterObject is used to keep track of a pointer into an N-dimensional array. It contains associated
information used to quickly march through the array. The pointer can be adjusted in three basic ways:
1) advance to the next position in the array in a C-style contiguous fashion, 2) advance to an arbitrary
N-dimensional coordinate in the array, and 3) advance to an arbitrary one-dimensional index into the array. The
members of the PyArrayIterObject structure are used in these calculations. Iterator objects keep their
own dimension and strides information about an array. This can be adjusted as needed for broadcasting, or to
loop over only specific dimensions.

typedef struct {
PyObject_HEAD
int nd_m1;
npy_intp index;
npy_intp size;
npy_intp coordinates[NPY_MAXDIMS];
npy_intp dims_m1[NPY_MAXDIMS];
npy_intp strides[NPY_MAXDIMS];

1022 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

npy_intp backstrides[NPY_MAXDIMS];
npy_intp factors[NPY_MAXDIMS];
PyArrayObject *ao;
char *dataptr;
Bool contiguous;
} PyArrayIterObject;

int nd_m1
N 1 where N is the number of dimensions in the underlying array.
npy_intp index
The current 1-d index into the array.
npy_intp size
The total size of the underlying array.
npy_intp coordinates
An N -dimensional index into the array.
npy_intp dims_m1
The size of the array minus 1 in each dimension.
npy_intp strides
The strides of the array. How many bytes needed to jump to the next element in each dimension.
npy_intp backstrides
How many bytes needed to jump from the end of a dimension back to its beginning. Note that backstrides
[k]= strides [k]*d ims_m1 [k], but it is stored here as an optimization.
npy_intp factors
This array is used in computing an N-d index from a 1-d index. It contains needed products of the dimen-
sions.
PyArrayObject ao
A pointer to the underlying ndarray this iterator was created to represent.
char dataptr
This member points to an element in the ndarray indicated by the index.
Bool contiguous
This flag is true if the underlying array is NPY_C_CONTIGUOUS. It is used to simplify calculations when
possible.
How to use an array iterator on a C-level is explained more fully in later sections. Typically, you do not need to
concern yourself with the internal structure of the iterator object, and merely interact with it through the use of the
macros PyArray_ITER_NEXT (it), PyArray_ITER_GOTO (it, dest), or PyArray_ITER_GOTO1D (it, index).
All of these macros require the argument it to be a PyArrayIterObject *.

PyArrayMultiIter_Type

PyArrayMultiIter_Type
This type provides an iterator that encapsulates the concept of broadcasting. It allows N arrays to be broadcast
together so that the loop progresses in C-style contiguous fashion over the broadcasted array. The corresponding
C-structure is the PyArrayMultiIterObject whose memory layout must begin any object, obj, passed in
to the PyArray_Broadcast (obj) function. Broadcasting is performed by adjusting array iterators so that
each iterator represents the broadcasted shape and size, but has its strides adjusted so that the correct element
from the array is used at each iteration.

5.1. Python Types and C-Structures 1023

NumPy Reference, Release 2.0.0.dev8464

PyArrayMultiIterObject

typedef struct {
PyObject_HEAD
int numiter;
npy_intp size;
npy_intp index;
int nd;
npy_intp dimensions[NPY_MAXDIMS];
PyArrayIterObject *iters[NPY_MAXDIMS];
} PyArrayMultiIterObject;

PyObject_HEAD
Needed at the start of every Python object (holds reference count and type identification).
int numiter
The number of arrays that need to be broadcast to the same shape.
npy_intp size
The total broadcasted size.
npy_intp index
The current (1-d) index into the broadcasted result.
int nd
The number of dimensions in the broadcasted result.
npy_intp dimensions
The shape of the broadcasted result (only nd slots are used).
PyArrayIterObject iters
An array of iterator objects that holds the iterators for the arrays to be broadcast together. On return, the
iterators are adjusted for broadcasting.

PyArrayNeighborhoodIter_Type

PyArrayNeighborhoodIter_Type
This is an iterator object that makes it easy to loop over an N-dimensional neighborhood.
PyArrayNeighborhoodIterObject
The C-structure corresponding to an object of PyArrayNeighborhoodIter_Type is the
PyArrayNeighborhoodIterObject.

PyArrayFlags_Type

PyArrayFlags_Type
When the flags attribute is retrieved from Python, a special builtin object of this type is constructed. This special
type makes it easier to work with the different flags by accessing them as attributes or by accessing them as if
the object were a dictionary with the flag names as entries.

ScalarArrayTypes

There is a Python type for each of the different built-in data types that can be present in the array Most of these are sim-
ple wrappers around the corresponding data type in C. The C-names for these types are Py{TYPE}ArrType_Type
where {TYPE} can be

1024 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

Bool, Byte, Short, Int, Long, LongLong, UByte, UShort, UInt, ULong, ULongLong, Float, Double,
LongDouble, CFloat, CDouble, CLongDouble, String, Unicode, Void, and Object.
These type names are part of the C-API and can therefore be created in extension C-code. There is also a
PyIntpArrType_Type and a PyUIntpArrType_Type that are simple substitutes for one of the integer types
that can hold a pointer on the platform. The structure of these scalar objects is not exposed to C-code. The function
PyArray_ScalarAsCtype (..) can be used to extract the C-type value from the array scalar and the function
PyArray_Scalar (...) can be used to construct an array scalar from a C-value.

5.1.2 Other C-Structures

A few new C-structures were found to be useful in the development of NumPy. These C-structures are used in at least
one C-API call and are therefore documented here. The main reason these structures were defined is to make it easy
to use the Python ParseTuple C-API to convert from Python objects to a useful C-Object.

PyArray_Dims

PyArray_Dims
This structure is very useful when shape and/or strides information is supposed to be interpreted. The structure
is:

typedef struct {
npy_intp *ptr;
int len;
} PyArray_Dims;

The members of this structure are

npy_intp ptr
A pointer to a list of (npy_intp) integers which usually represent array shape or array strides.
int len
The length of the list of integers. It is assumed safe to access ptr [0] to ptr [len-1].

PyArray_Chunk

PyArray_Chunk
This is equivalent to the buffer object structure in Python up to the ptr member. On 32-bit platforms (i.e. if
NPY_SIZEOF_INT == NPY_SIZEOF_INTP ) or in Python 2.5, the len member also matches an equivalent
member of the buffer object. It is useful to represent a generic single- segment chunk of memory.

typedef struct {
PyObject_HEAD
PyObject *base;
void *ptr;
npy_intp len;
int flags;
} PyArray_Chunk;

The members are

PyObject_HEAD
Necessary for all Python objects. Included here so that the PyArray_Chunk structure matches that of
the buffer object (at least to the len member).

5.1. Python Types and C-Structures 1025

NumPy Reference, Release 2.0.0.dev8464

PyObject base
The Python object this chunk of memory comes from. Needed so that memory can be accounted for
properly.
void ptr
A pointer to the start of the single-segment chunk of memory.
npy_intp len
The length of the segment in bytes.
int flags
Any data flags (e.g. NPY_WRITEABLE ) that should be used to interpret the memory.

PyArrayInterface

See Also:
The Array Interface
PyArrayInterface
The PyArrayInterface structure is defined so that NumPy and other extension modules can use the rapid
array interface protocol. The __array_struct__ method of an object that supports the rapid array interface
protocol should return a PyCObject that contains a pointer to a PyArrayInterface structure with the
relevant details of the array. After the new array is created, the attribute should be DECREFd which will
free the PyArrayInterface structure. Remember to INCREF the object (whose __array_struct__
attribute was retrieved) and point the base member of the new PyArrayObject to this same object. In this
way the memory for the array will be managed correctly.

typedef struct {
int two;
int nd;
char typekind;
int itemsize;
int flags;
npy_intp *shape;
npy_intp *strides;
void *data;
PyObject *descr;
} PyArrayInterface;

int two
the integer 2 as a sanity check.
int nd
the number of dimensions in the array.
char typekind
A character indicating what kind of array is present according to the typestring convention with t ->
bitfield, b -> Boolean, i -> signed integer, u -> unsigned integer, f -> floating point, c -> complex
floating point, O -> object, S -> string, U -> unicode, V -> void.
int itemsize
The number of bytes each item in the array requires.
int flags
Any of the bits NPY_C_CONTIGUOUS (1), NPY_F_CONTIGUOUS (2), NPY_ALIGNED (0x100),
NPY_NOTSWAPPED (0x200), or NPY_WRITEABLE (0x400) to indicate something about the data. The
NPY_ALIGNED, NPY_C_CONTIGUOUS, and NPY_F_CONTIGUOUS flags can actually be determined
from the other parameters. The flag NPY_ARR_HAS_DESCR (0x800) can also be set to indicate to

1026 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

objects consuming the version 3 array interface that the descr member of the structure is present (it will
be ignored by objects consuming version 2 of the array interface).
npy_intp shape
An array containing the size of the array in each dimension.
npy_intp strides
An array containing the number of bytes to jump to get to the next element in each dimension.
void data
A pointer to the first element of the array.
PyObject descr
A Python object describing the data-type in more detail (same as the descr key in
__array_interface__). This can be NULL if typekind and itemsize provide enough informa-
tion. This field is also ignored unless ARR_HAS_DESCR flag is on in flags.

Internally used structures

Internally, the code uses some additional Python objects primarily for memory management. These types are not
accessible directly from Python, and are not exposed to the C-API. They are included here only for completeness and
assistance in understanding the code.
PyUFuncLoopObject
A loose wrapper for a C-structure that contains the information needed for looping. This is useful if you are
trying to understand the ufunc looping code. The PyUFuncLoopObject is the associated C-structure. It is
defined in the ufuncobject.h header.
PyUFuncReduceObject
A loose wrapper for the C-structure that contains the information needed for reduce-like methods of
ufuncs. This is useful if you are trying to understand the reduce, accumulate, and reduce-at code. The
PyUFuncReduceObject is the associated C-structure. It is defined in the ufuncobject.h header.
PyUFunc_Loop1d
A simple linked-list of C-structures containing the information needed to define a 1-d loop for a ufunc for every
defined signature of a user-defined data-type.
PyArrayMapIter_Type
Advanced indexing is handled with this Python type. It is simply a loose wrapper around the C-
structure containing the variables needed for advanced array indexing. The associated C-structure,
PyArrayMapIterObject, is useful if you are trying to understand the advanced-index mapping
code. It is defined in the arrayobject.h header. This type is not exposed to Python and could be replaced
with a C-structure. As a Python type it takes advantage of reference- counted memory management.

5.2 System configuration

When NumPy is built, information about system configuration is recorded, and is made available for extension modules
using Numpys C API. These are mostly defined in numpyconfig.h (included in ndarrayobject.h). The
public symbols are prefixed by NPY_*. Numpy also offers some functions for querying information about the platform
in use.
For private use, Numpy also constructs a config.h in the NumPy include directory, which is not exported by Numpy
(that is a python extension which use the numpy C API will not see those symbols), to avoid namespace pollution.

5.2. System configuration 1027

NumPy Reference, Release 2.0.0.dev8464

5.2.1 Data type sizes

The NPY_SIZEOF_{CTYPE} constants are defined so that sizeof information is available to the pre-processor.
NPY_SIZEOF_SHORT
sizeof(short)
NPY_SIZEOF_INT
sizeof(int)
NPY_SIZEOF_LONG
sizeof(long)
NPY_SIZEOF_LONG_LONG
sizeof(longlong) where longlong is defined appropriately on the platform (A macro defines
NPY_SIZEOF_LONGLONG as well.)
NPY_SIZEOF_PY_LONG_LONG

NPY_SIZEOF_FLOAT
sizeof(float)
NPY_SIZEOF_DOUBLE
sizeof(double)
NPY_SIZEOF_LONG_DOUBLE
sizeof(longdouble) (A macro defines NPY_SIZEOF_LONGDOUBLE as well.)
NPY_SIZEOF_PY_INTPTR_T
Size of a pointer on this platform (sizeof(void *)) (A macro defines NPY_SIZEOF_INTP as well.)

5.2.2 Platform information

NPY_CPU_X86

NPY_CPU_AMD64

NPY_CPU_IA64

NPY_CPU_PPC

NPY_CPU_PPC64

NPY_CPU_SPARC

NPY_CPU_SPARC64

NPY_CPU_S390

NPY_CPU_PARISC
New in version 1.3.0. CPU architecture of the platform; only one of the above is defined.
Defined in numpy/npy_cpu.h

1028 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

NPY_LITTLE_ENDIAN

NPY_BIG_ENDIAN

NPY_BYTE_ORDER
New in version 1.3.0. Portable alternatives to the endian.h macros of GNU Libc. If big endian,
NPY_BYTE_ORDER == NPY_BIG_ENDIAN, and similarly for little endian architectures.
Defined in numpy/npy_endian.h.
PyArray_GetEndianness()
New in version 1.3.0. Returns the endianness of the current platform. One of NPY_CPU_BIG,
NPY_CPU_LITTLE, or NPY_CPU_UNKNOWN_ENDIAN.

5.3 Data Type API

The standard array can have 21 different data types (and has some support for adding your own types). These data
types all have an enumerated type, an enumerated type-character, and a corresponding array scalar Python type object
(placed in a hierarchy). There are also standard C typedefs to make it easier to manipulate elements of the given data
type. For the numeric types, there are also bit-width equivalent C typedefs and named typenumbers that make it easier
to select the precision desired.

Warning: The names for the types in c code follows c naming conventions more closely. The Python names for
these types follow Python conventions. Thus, NPY_FLOAT picks up a 32-bit float in C, but numpy.float_ in
Python corresponds to a 64-bit double. The bit-width names can be used in both Python and C for clarity.

5.3.1 Enumerated Types

There is a list of enumerated types defined providing the basic 21 data types plus some useful generic names. Whenever
the code requires a type number, one of these enumerated types is requested. The types are all called NPY_{NAME}
where {NAME} can be
BOOL, BYTE, UBYTE, SHORT, USHORT, INT, UINT, LONG, ULONG, LONGLONG, ULONG-
LONG, FLOAT, DOUBLE, LONGDOUBLE, CFLOAT, CDOUBLE, CLONGDOUBLE, OBJECT,
STRING, UNICODE, VOID
NTYPES, NOTYPE, USERDEF, DEFAULT_TYPE
The various character codes indicating certain types are also part of an enumerated list. References to type characters
(should they be needed at all) should always use these enumerations. The form of them is NPY_{NAME}LTR where
{NAME} can be
BOOL, BYTE, UBYTE, SHORT, USHORT, INT, UINT, LONG, ULONG, LONGLONG, ULONG-
LONG, FLOAT, DOUBLE, LONGDOUBLE, CFLOAT, CDOUBLE, CLONGDOUBLE, OBJECT,
STRING, VOID
INTP, UINTP
GENBOOL, SIGNED, UNSIGNED, FLOATING, COMPLEX
The latter group of {NAME}s corresponds to letters used in the array interface typestring specification.

5.3. Data Type API 1029

NumPy Reference, Release 2.0.0.dev8464

5.3.2 Defines

Max and min values for integers

NPY_MAX_INT{bits}

NPY_MAX_UINT{bits}

NPY_MIN_INT{bits}
These are defined for {bits} = 8, 16, 32, 64, 128, and 256 and provide the maximum (minimum) value of the
corresponding (unsigned) integer type. Note: the actual integer type may not be available on all platforms (i.e.
128-bit and 256-bit integers are rare).
NPY_MIN_{type}
This is defined for {type} = BYTE, SHORT, INT, LONG, LONGLONG, INTP
NPY_MAX_{type}
This is defined for all defined for {type} = BYTE, UBYTE, SHORT, USHORT, INT, UINT, LONG,
ULONG, LONGLONG, ULONGLONG, INTP, UINTP

Number of bits in data types

All NPY_SIZEOF_{CTYPE} constants have corresponding NPY_BITSOF_{CTYPE} constants defined. The

NPY_BITSOF_{CTYPE} constants provide the number of bits in the data type. Specifically, the available
{CTYPE}s are
BOOL, CHAR, SHORT, INT, LONG, LONGLONG, FLOAT, DOUBLE, LONGDOUBLE

Bit-width references to enumerated typenums

All of the numeric data types (integer, floating point, and complex) have constants that are defined to be a specific
enumerated type number. Exactly which enumerated type a bit-width type refers to is platform dependent. In partic-
ular, the constants available are PyArray_{NAME}{BITS} where {NAME} is INT, UINT, FLOAT, COMPLEX
and {BITS} can be 8, 16, 32, 64, 80, 96, 128, 160, 192, 256, and 512. Obviously not all bit-widths are available on
all platforms for all the kinds of numeric types. Commonly 8-, 16-, 32-, 64-bit integers; 32-, 64-bit floats; and 64-,
128-bit complex types are available.

Integer that can hold a pointer

The constants PyArray_INTP and PyArray_UINTP refer to an enumerated integer type that is large enough to hold
a pointer on the platform. Index arrays should always be converted to PyArray_INTP , because the dimension of the
array is of type npy_intp.

5.3.3 C-type names

There are standard variable types for each of the numeric data types and the bool data type. Some of these are already
available in the C-specification. You can create variables in extension code with these types.

1030 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

Boolean

npy_bool
unsigned char; The constants NPY_FALSE and NPY_TRUE are also defined.

(Un)Signed Integer

Unsigned versions of the integers can be defined by pre-pending a u to the front of the integer name.
npy_(u)byte
(unsigned) char
npy_(u)short
(unsigned) short
npy_(u)int
(unsigned) int
npy_(u)long
(unsigned) long int
npy_(u)longlong
(unsigned long long int)
npy_(u)intp
(unsigned) Py_intptr_t (an integer that is the size of a pointer on the platform).

(Complex) Floating point

npy_(c)float
float
npy_(c)double
double
npy_(c)longdouble
long double
complex types are structures with .real and .imag members (in that order).

Bit-width names

There are also typedefs for signed integers, unsigned integers, floating point, and complex floating point types of
specific bit- widths. The available type names are
npy_int{bits}, npy_uint{bits}, npy_float{bits}, and npy_complex{bits}
where {bits} is the number of bits in the type and can be 8, 16, 32, 64, 128, and 256 for integer types; 16, 32 , 64,
80, 96, 128, and 256 for floating-point types; and 32, 64, 128, 160, 192, and 512 for complex-valued types. Which
bit-widths are available is platform dependent. The bolded bit-widths are usually available on all platforms.

5.3.4 Printf Formatting

For help in printing, the following strings are defined as the correct format specifier in printf and related commands.
NPY_LONGLONG_FMT, NPY_ULONGLONG_FMT, NPY_INTP_FMT, NPY_UINTP_FMT,
NPY_LONGDOUBLE_FMT

5.3. Data Type API 1031

NumPy Reference, Release 2.0.0.dev8464

5.4 Array API

The test of a first-rate intelligence is the ability to hold two

opposed ideas in the mind at the same time, and still retain the
ability to function.
F. Scott Fitzgerald For a successful technology, reality must take precedence over public
relations, for Nature cannot be fooled.
Richard P. Feynman

5.4.1 Array structure and data access

These macros all access the PyArrayObject structure members. The input argument, obj, can be any PyObject
* that is directly interpretable as a PyArrayObject * (any instance of the PyArray_Type and its sub-types).
void * PyArray_DATA(PyObject *obj)

char * PyArray_BYTES(PyObject *obj)

These two macros are similar and obtain the pointer to the data-buffer for the array. The first macro can (and
should be) assigned to a particular pointer where the second is for generic processing. If you have not
guaranteed a contiguous and/or aligned array then be sure you understand how to access the data in the array to
avoid memory and/or alignment problems.
npy_intp * PyArray_DIMS(PyObject *arr)

npy_intp * PyArray_STRIDES(PyObject* arr)

npy_intp PyArray_DIM(PyObject* arr, int n)

Return the shape in the n th dimension.
npy_intp PyArray_STRIDE(PyObject* arr, int n)
Return the stride in the n th dimension.
PyObject * PyArray_BASE(PyObject* arr)

PyArray_Descr * PyArray_DESCR(PyObject* arr)

int PyArray_FLAGS(PyObject* arr)

int PyArray_ITEMSIZE(PyObject* arr)

Return the itemsize for the elements of this array.
int PyArray_TYPE(PyObject* arr)
Return the (builtin) typenumber for the elements of this array.
PyObject * PyArray_GETITEM(PyObject* arr, void* itemptr)
Get a Python object from the ndarray, arr, at the location pointed to by itemptr. Return NULL on failure.
int PyArray_SETITEM(PyObject* arr, void* itemptr, PyObject* obj)
Convert obj and place it in the ndarray, arr, at the place pointed to by itemptr. Return -1 if an error occurs or 0
on success.
npy_intp PyArray_SIZE(PyObject* arr)
Returns the total size (in number of elements) of the array.

1032 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

npy_intp PyArray_Size(PyObject* obj)

Returns 0 if obj is not a sub-class of bigndarray. Otherwise, returns the total number of elements in the array.
Safer version of PyArray_SIZE (obj).
npy_intp PyArray_NBYTES(PyObject* arr)
Returns the total number of bytes consumed by the array.

Data access

These functions and macros provide easy access to elements of the ndarray from C. These work for all arrays. You
may need to take care when accessing the data in the array, however, if it is not in machine byte-order, misaligned, or
not writeable. In other words, be sure to respect the state of the flags unless you know what you are doing, or have
previously guaranteed an array that is writeable, aligned, and in machine byte-order using PyArray_FromAny. If
you wish to handle all types of arrays, the copyswap function for each type is useful for handling misbehaved arrays.
Some platforms (e.g. Solaris) do not like misaligned data and will crash if you de-reference a misaligned pointer.
Other platforms (e.g. x86 Linux) will just work more slowly with misaligned data.
void* PyArray_GetPtr(PyArrayObject* aobj, npy_intp* ind)
Return a pointer to the data of the ndarray, aobj, at the N-dimensional index given by the c-array, ind, (which
must be at least aobj ->nd in size). You may want to typecast the returned pointer to the data type of the ndarray.
void* PyArray_GETPTR1(PyObject* obj, <npy_intp> i)

void* PyArray_GETPTR2(PyObject* obj, <npy_intp> i, <npy_intp> j)

void* PyArray_GETPTR3(PyObject* obj, <npy_intp> i, <npy_intp> j, <npy_intp> k)

void* PyArray_GETPTR4(PyObject* obj, <npy_intp> i, <npy_intp> j, <npy_intp> k, <npy_intp> l)

Quick, inline access to the element at the given coordinates in the ndarray, obj, which must have respectively 1,
2, 3, or 4 dimensions (this is not checked). The corresponding i, j, k, and l coordinates can be any integer but
will be interpreted as npy_intp. You may want to typecast the returned pointer to the data type of the ndarray.

5.4.2 Creating arrays

From scratch

PyObject* PyArray_NewFromDescr(PyTypeObject* subtype, PyArray_Descr* descr, int nd, npy_intp* dims,

npy_intp* strides, void* data, int flags, PyObject* obj)
This is the main array creation function. Most new arrays are created with this flexible function. The re-
turned object is an object of Python-type subtype, which must be a subtype of PyArray_Type. The
array has nd dimensions, described by dims. The data-type descriptor of the new array is descr. If subtype
is not &PyArray_Type (e.g. a Python subclass of the ndarray), then obj is the object to pass to the
__array_finalize__ method of the subclass. If data is NULL, then new memory will be allocated and
flags can be non-zero to indicate a Fortran-style contiguous array. If data is not NULL, then it is assumed to
point to the memory to be used for the array and the flags argument is used as the new flags for the array
(except the state of NPY_OWNDATA and UPDATEIFCOPY flags of the new array will be reset). In addition, if
data is non-NULL, then strides can also be provided. If strides is NULL, then the array strides are computed
as C-style contiguous (default) or Fortran-style contiguous (flags is nonzero for data = NULL or flags &
NPY_F_CONTIGUOUS is nonzero non-NULL data). Any provided dims and strides are copied into newly
allocated dimension and strides arrays for the new array object.

5.4. Array API 1033

NumPy Reference, Release 2.0.0.dev8464

PyObject* PyArray_New(PyTypeObject* subtype, int nd, npy_intp* dims, int type_num, npy_intp* strides,
void* data, int itemsize, int flags, PyObject* obj)
This is similar to PyArray_DescrNew (...) except you specify the data-type descriptor with type_num and
itemsize, where type_num corresponds to a builtin (or user-defined) type. If the type always has the same
number of bytes, then itemsize is ignored. Otherwise, itemsize specifies the particular size of this array.

Warning: If data is passed to PyArray_NewFromDescr or PyArray_New, this memory must not be deal-
located until the new array is deleted. If this data came from another Python object, this can be accomplished
using Py_INCREF on that object and setting the base member of the new array to point to that object. If strides
are passed in they must be consistent with the dimensions, the itemsize, and the data of the array.

PyObject* PyArray_SimpleNew(int nd, npy_intp* dims, int typenum)

Create a new unitialized array of type, typenum, whose size in each of nd dimensions is given by the integer
array, dims. This function cannot be used to create a flexible-type array (no itemsize given).
PyObject* PyArray_SimpleNewFromData(int nd, npy_intp* dims, int typenum, void* data)
Create an array wrapper around data pointed to by the given pointer. The array flags will have a default that the
data area is well-behaved and C-style contiguous. The shape of the array is given by the dims c-array of length
nd. The data-type of the array is indicated by typenum.
PyObject* PyArray_SimpleNewFromDescr(int nd, npy_intp* dims, PyArray_Descr* descr)
Create a new array with the provided data-type descriptor, descr , of the shape deteremined by nd and dims.
PyArray_FILLWBYTE(PyObject* obj, int val)
Fill the array pointed to by obj which must be a (subclass of) bigndarraywith the contents of val (evaluated
as a byte).
PyObject* PyArray_Zeros(int nd, npy_intp* dims, PyArray_Descr* dtype, int fortran)
Construct a new nd -dimensional array with shape given by dims and data type given by dtype. If fortran is
non-zero, then a Fortran-order array is created, otherwise a C-order array is created. Fill the memory with zeros
(or the 0 object if dtype corresponds to PyArray_OBJECT ).
PyObject* PyArray_ZEROS(int nd, npy_intp* dims, int type_num, int fortran)
Macro form of PyArray_Zeros which takes a type-number instead of a data-type object.
PyObject* PyArray_Empty(int nd, npy_intp* dims, PyArray_Descr* dtype, int fortran)
Construct a new nd -dimensional array with shape given by dims and data type given by dtype. If fortran is
non-zero, then a Fortran-order array is created, otherwise a C-order array is created. The array is uninitialized
unless the data type corresponds to PyArray_OBJECT in which case the array is filled with Py_None.
PyObject* PyArray_EMPTY(int nd, npy_intp* dims, int typenum, int fortran)
Macro form of PyArray_Empty which takes a type-number, typenum, instead of a data-type object.
PyObject* PyArray_Arange(double start, double stop, double step, int typenum)
Construct a new 1-dimensional array of data-type, typenum, that ranges from start to stop (exclusive) in incre-
ments of step . Equivalent to arange (start, stop, step, dtype).
PyObject* PyArray_ArangeObj(PyObject* start, PyObject* stop, PyObject* step, PyArray_Descr* descr)
Construct a new 1-dimensional array of data-type determined by descr, that ranges from start to stop
(exclusive) in increments of step. Equivalent to arange( start, stop, step, typenum ).

From other objects

PyObject* PyArray_FromAny(PyObject* op, PyArray_Descr* dtype, int min_depth, int max_depth, int re-
quirements, PyObject* context)
This is the main function used to obtain an array from any nested sequence, or object that exposes the array
interface, op. The parameters allow specification of the required dtype, the minimum (min_depth) and maximum
(max_depth) number of dimensions acceptable, and other requirements for the array. The dtype argument needs

1034 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

to be a PyArray_Descr structure indicating the desired data-type (including required byteorder). The dtype
argument may be NULL, indicating that any data-type (and byteorder) is acceptable. Unless FORCECAST is
present in flags, this call will generate an error if the data type cannot be safely obtained from the object. If
you want to use NULL for the dtype and ensure the array is notswapped then use PyArray_CheckFromAny.
A value of 0 for either of the depth parameters causes the parameter to be ignored. Any of the following
array flags can be added (e.g. using |) to get the requirements argument. If your code can handle general (e.g.
strided, byte-swapped, or unaligned arrays) then requirements may be 0. Also, if op is not already an array (or
does not expose the array interface), then a new array will be created (and filled from op using the sequence
protocol). The new array will have NPY_DEFAULT as its flags member. The context argument is passed to the
__array__ method of op and is only used if the array is constructed that way. Almost always this parameter
is NULL.
NPY_C_CONTIGUOUS
Make sure the returned array is C-style contiguous
NPY_F_CONTIGUOUS
Make sure the returned array is Fortran-style contiguous.
NPY_ALIGNED
Make sure the returned array is aligned on proper boundaries for its data type. An aligned array has the
data pointer and every strides factor as a multiple of the alignment factor for the data-type- descriptor.
NPY_WRITEABLE
Make sure the returned array can be written to.
NPY_ENSURECOPY
Make sure a copy is made of op. If this flag is not present, data is not copied if it can be avoided.
NPY_ENSUREARRAY
Make sure the result is a base-class ndarray or bigndarray. By default, if op is an instance of a subclass of
the bigndarray, an instance of that same subclass is returned. If this flag is set, an ndarray object will be
returned instead.
NPY_FORCECAST
Force a cast to the output type even if it cannot be done safely. Without this flag, a data cast will occur
only if it can be done safely, otherwise an error is reaised.
NPY_UPDATEIFCOPY
If op is already an array, but does not satisfy the requirements, then a copy is made (which will satisfy the
requirements). If this flag is present and a copy (of an object that is already an array) must be made, then
the corresponding NPY_UPDATEIFCOPY flag is set in the returned copy and op is made to be read-only.
When the returned copy is deleted (presumably after your calculations are complete), its contents will be
copied back into op and the op array will be made writeable again. If op is not writeable to begin with,
then an error is raised. If op is not already an array, then this flag has no effect.
NPY_BEHAVED
NPY_ALIGNED | NPY_WRITEABLE
NPY_CARRAY
NPY_C_CONTIGUOUS | NPY_BEHAVED
NPY_CARRAY_RO
NPY_C_CONTIGUOUS | NPY_ALIGNED
NPY_FARRAY
NPY_F_CONTIGUOUS | NPY_BEHAVED
NPY_FARRAY_RO
NPY_F_CONTIGUOUS | NPY_ALIGNED

5.4. Array API 1035

NumPy Reference, Release 2.0.0.dev8464

NPY_DEFAULT
NPY_CARRAY
NPY_IN_ARRAY
NPY_CONTIGUOUS | NPY_ALIGNED
NPY_IN_FARRAY
NPY_F_CONTIGUOUS | NPY_ALIGNED
NPY_INOUT_ARRAY
NPY_C_CONTIGUOUS | NPY_WRITEABLE | NPY_ALIGNED
NPY_INOUT_FARRAY
NPY_F_CONTIGUOUS | NPY_WRITEABLE | NPY_ALIGNED
NPY_OUT_ARRAY
NPY_C_CONTIGUOUS | NPY_WRITEABLE | NPY_ALIGNED | NPY_UPDATEIFCOPY
NPY_OUT_FARRAY
NPY_F_CONTIGUOUS | NPY_WRITEABLE | NPY_ALIGNED | UPDATEIFCOPY
PyObject* PyArray_CheckFromAny(PyObject* op, PyArray_Descr* dtype, int min_depth, int max_depth,
int requirements, PyObject* context)
Nearly identical to PyArray_FromAny (...) except requirements can contain NPY_NOTSWAPPED (over-
riding the specification in dtype) and NPY_ELEMENTSTRIDES which indicates that the array should be
aligned in the sense that the strides are multiples of the element size.
NPY_NOTSWAPPED
Make sure the returned array has a data-type descriptor that is in machine byte-order, over-riding any specifi-
cation in the dtype argument. Normally, the byte-order requirement is determined by the dtype argument. If
this flag is set and the dtype argument does not indicate a machine byte-order descriptor (or is NULL and the
object is already an array with a data-type descriptor that is not in machine byte- order), then a new data-type
descriptor is created and used with its byte-order field set to native.
NPY_BEHAVED_NS
NPY_ALIGNED | NPY_WRITEABLE | NPY_NOTSWAPPED
NPY_ELEMENTSTRIDES
Make sure the returned array has strides that are multiples of the element size.
PyObject* PyArray_FromArray(PyArrayObject* op, PyArray_Descr* newtype, int requirements)
Special case of PyArray_FromAny for when op is already an array but it needs to be of a specific newtype
(including byte-order) or has certain requirements.
PyObject* PyArray_FromStructInterface(PyObject* op)
Returns an ndarray object from a Python object that exposes the __array_struct__ method and fol-
lows the array interface protocol. If the object does not contain this method then a borrowed reference to
Py_NotImplemented is returned.
PyObject* PyArray_FromInterface(PyObject* op)
Returns an ndarray object from a Python object that exposes the __array_shape__ and
__array_typestr__ methods following the array interface protocol. If the object does not contain
one of these method then a borrowed reference to Py_NotImplemented is returned.
PyObject* PyArray_FromArrayAttr(PyObject* op, PyArray_Descr* dtype, PyObject* context)
Return an ndarray object from a Python object that exposes the __array__ method. The __array__
method can take 0, 1, or 2 arguments ([dtype, context]) where context is used to pass information about where
the __array__ method is being called from (currently only used in ufuncs).
PyObject* PyArray_ContiguousFromAny(PyObject* op, int typenum, int min_depth, int max_depth)
This function returns a (C-style) contiguous and behaved function array from any nested sequence or array
interface exporting object, op, of (non-flexible) type given by the enumerated typenum, of minimum depth

1036 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

min_depth, and of maximum depth max_depth. Equivalent to a call to PyArray_FromAny with requirements
set to NPY_DEFAULT and the type_num member of the type argument set to typenum.
PyObject * PyArray_FromObject(PyObject *op, int typenum, int min_depth, int max_depth)
Return an aligned and in native-byteorder array from any nested sequence or array-interface exporting object,
op, of a type given by the enumerated typenum. The minimum number of dimensions the array can have is
given by min_depth while the maximum is max_depth. This is equivalent to a call to PyArray_FromAny
with requirements set to BEHAVED.
PyObject* PyArray_EnsureArray(PyObject* op)
This function steals a reference to op and makes sure that op is a base-class ndarray. It special cases array
scalars, but otherwise calls PyArray_FromAny ( op, NULL, 0, 0, NPY_ENSUREARRAY).
PyObject* PyArray_FromString(char* string, npy_intp slen, PyArray_Descr* dtype, npy_intp num, char*
sep)
Construct a one-dimensional ndarray of a single type from a binary or (ASCII) text string of length slen.
The data-type of the array to-be-created is given by dtype. If num is -1, then copy the entire string and return
an appropriately sized array, otherwise, num is the number of items to copy from the string. If sep is NULL
(or ), then interpret the string as bytes of binary data, otherwise convert the sub-strings separated by sep to
items of data-type dtype. Some data-types may not be readable in text mode and an error will be raised if that
occurs. All errors return NULL.
PyObject* PyArray_FromFile(FILE* fp, PyArray_Descr* dtype, npy_intp num, char* sep)
Construct a one-dimensional ndarray of a single type from a binary or text file. The open file pointer is fp, the
data-type of the array to be created is given by dtype. This must match the data in the file. If num is -1, then
read until the end of the file and return an appropriately sized array, otherwise, num is the number of items to
read. If sep is NULL (or ), then read from the file in binary mode, otherwise read from the file in text mode
with sep providing the item separator. Some array types cannot be read in text mode in which case an error is
raised.
PyObject* PyArray_FromBuffer(PyObject* buf, PyArray_Descr* dtype, npy_intp count, npy_intp offset)
Construct a one-dimensional ndarray of a single type from an object, buf, that exports the (single-segment)
buffer protocol (or has an attribute __buffer__ that returns an object that exports the buffer protocol). A
writeable buffer will be tried first followed by a read- only buffer. The NPY_WRITEABLE flag of the returned
array will reflect which one was successful. The data is assumed to start at offset bytes from the start of the
memory location for the object. The type of the data in the buffer will be interpreted depending on the data-
type descriptor, dtype. If count is negative then it will be determined from the size of the buffer and the
requested itemsize, otherwise, count represents how many elements should be converted from the buffer.
int PyArray_CopyInto(PyArrayObject* dest, PyArrayObject* src)
Copy from the source array, src, into the destination array, dest, performing a data-type conversion if neces-
sary. If an error occurs return -1 (otherwise 0). The shape of src must be broadcastable to the shape of dest.
The data areas of dest and src must not overlap.
int PyArray_MoveInto(PyArrayObject* dest, PyArrayObject* src)
Move data from the source array, src, into the destination array, dest, performing a data-type conversion if
necessary. If an error occurs return -1 (otherwise 0). The shape of src must be broadcastable to the shape of
dest. The data areas of dest and src may overlap.
PyArrayObject* PyArray_GETCONTIGUOUS(PyObject* op)
If op is already (C-style) contiguous and well-behaved then just return a reference, otherwise return a (contigu-
ous and well-behaved) copy of the array. The parameter op must be a (sub-class of an) ndarray and no checking
for that is done.
PyObject* PyArray_FROM_O(PyObject* obj)
Convert obj to an ndarray. The argument can be any nested sequence or object that exports the array interface.
This is a macro form of PyArray_FromAny using NULL, 0, 0, 0 for the other arguments. Your code must be
able to handle any data-type descriptor and any combination of data-flags to use this macro.

5.4. Array API 1037

NumPy Reference, Release 2.0.0.dev8464

PyObject* PyArray_FROM_OF(PyObject* obj, int requirements)

Similar to PyArray_FROM_O except it can take an argument of requirements indicating properties the resulting
array must have. Available requirements that can be enforced are NPY_CONTIGUOUS, NPY_F_CONTIGUOUS,
NPY_ALIGNED, NPY_WRITEABLE, NPY_NOTSWAPPED, NPY_ENSURECOPY, NPY_UPDATEIFCOPY,
NPY_FORCECAST, and NPY_ENSUREARRAY. Standard combinations of flags can also be used:
PyObject* PyArray_FROM_OT(PyObject* obj, int typenum)
Similar to PyArray_FROM_O except it can take an argument of typenum specifying the type-number the
returned array.
PyObject* PyArray_FROM_OTF(PyObject* obj, int typenum, int requirements)
Combination of PyArray_FROM_OF and PyArray_FROM_OT allowing both a typenum and a flags argument
to be provided..
PyObject* PyArray_FROMANY(PyObject* obj, int typenum, int min, int max, int requirements)
Similar to PyArray_FromAny except the data-type is specified using a typenumber.
PyArray_DescrFromType (typenum) is passed directly to PyArray_FromAny. This macro also
adds NPY_DEFAULT to requirements if NPY_ENSURECOPY is passed in as requirements.
PyObject * PyArray_CheckAxis(PyObject* obj, int* axis, int requirements)
Encapsulate the functionality of functions and methods that take the axis= keyword and work properly with
None as the axis argument. The input array is obj, while *axis is a converted integer (so that >=MAXDIMS
is the None value), and requirements gives the needed properties of obj. The output is a converted version
of the input so that requirements are met and if needed a flattening has occurred. On output negative values of
*axis are converted and the new value is checked to ensure consistency with the shape of obj.

5.4.3 Dealing with types

General check of Python Type

PyArray_Check(op)
Evaluates true if op is a Python object whose type is a sub-type of PyArray_Type.
PyArray_CheckExact(op)
Evaluates true if op is a Python object with type PyArray_Type.
PyArray_HasArrayInterface(op, out)
If op implements any part of the array interface, then out will contain a new reference to the newly created
ndarray using the interface or out will contain NULL if an error during conversion occurs. Otherwise, out will
contain a borrowed reference to Py_NotImplemented and no error condition is set.
PyArray_HasArrayInterfaceType(op, type, context, out)
If op implements any part of the array interface, then out will contain a new reference to the newly created
ndarray using the interface or out will contain NULL if an error during conversion occurs. Otherwise, out will
contain a borrowed reference to Py_NotImplemented and no error condition is set. This version allows setting
of the type and context in the part of the array interface that looks for the __array__ attribute.
PyArray_IsZeroDim(op)
Evaluates true if op is an instance of (a subclass of) PyArray_Type and has 0 dimensions.
PyArray_IsScalar(op, cls)
Evaluates true if op is an instance of Py{cls}ArrType_Type.
PyArray_CheckScalar(op)
Evaluates true if op is either an array scalar (an instance of a sub-type of PyGenericArr_Type ), or an
instance of (a sub-class of) PyArray_Type whose dimensionality is 0.

1038 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

PyArray_IsPythonScalar(op)
Evaluates true if op is a builtin Python scalar object (int, float, complex, str, unicode, long, bool).
PyArray_IsAnyScalar(op)
Evaluates true if op is either a Python scalar or an array scalar (an instance of a sub- type of
PyGenericArr_Type ).

Data-type checking

For the typenum macros, the argument is an integer representing an enumerated array data type. For the array type
checking macros the argument must be a PyObject * that can be directly interpreted as a PyArrayObject *.
PyTypeNum_ISUNSIGNED(num)

PyDataType_ISUNSIGNED(descr)

PyArray_ISUNSIGNED(obj)
Type represents an unsigned integer.
PyTypeNum_ISSIGNED(num)

PyDataType_ISSIGNED(descr)

PyArray_ISSIGNED(obj)
Type represents a signed integer.
PyTypeNum_ISINTEGER(num)

PyDataType_ISINTEGER(descr)

PyArray_ISINTEGER(obj)
Type represents any integer.
PyTypeNum_ISFLOAT(num)

PyDataType_ISFLOAT(descr)

PyArray_ISFLOAT(obj)
Type represents any floating point number.
PyTypeNum_ISCOMPLEX(num)

PyDataType_ISCOMPLEX(descr)

PyArray_ISCOMPLEX(obj)
Type represents any complex floating point number.
PyTypeNum_ISNUMBER(num)

PyDataType_ISNUMBER(descr)

5.4. Array API 1039

NumPy Reference, Release 2.0.0.dev8464

PyArray_ISNUMBER(obj)
Type represents any integer, floating point, or complex floating point number.
PyTypeNum_ISSTRING(num)

PyDataType_ISSTRING(descr)

PyArray_ISSTRING(obj)
Type represents a string data type.
PyTypeNum_ISPYTHON(num)

PyDataType_ISPYTHON(descr)

PyArray_ISPYTHON(obj)
Type represents an enumerated type corresponding to one of the standard Python scalar (bool, int, float, or
complex).
PyTypeNum_ISFLEXIBLE(num)

PyDataType_ISFLEXIBLE(descr)

PyArray_ISFLEXIBLE(obj)
Type represents one of the flexible array types ( NPY_STRING, NPY_UNICODE, or NPY_VOID ).
PyTypeNum_ISUSERDEF(num)

PyDataType_ISUSERDEF(descr)

PyArray_ISUSERDEF(obj)
Type represents a user-defined type.
PyTypeNum_ISEXTENDED(num)

PyDataType_ISEXTENDED(descr)

PyArray_ISEXTENDED(obj)
Type is either flexible or user-defined.
PyTypeNum_ISOBJECT(num)

PyDataType_ISOBJECT(descr)

PyArray_ISOBJECT(obj)
Type represents object data type.
PyTypeNum_ISBOOL(num)

PyDataType_ISBOOL(descr)

1040 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

PyArray_ISBOOL(obj)
Type represents Boolean data type.
PyDataType_HASFIELDS(descr)

PyArray_HASFIELDS(obj)
Type has fields associated with it.
PyArray_ISNOTSWAPPED(m)
Evaluates true if the data area of the ndarray m is in machine byte-order according to the arrays data-type
descriptor.
PyArray_ISBYTESWAPPED(m)
Evaluates true if the data area of the ndarray m is not in machine byte-order according to the arrays data-type
descriptor.
Bool PyArray_EquivTypes(PyArray_Descr* type1, PyArray_Descr* type2)
Return NPY_TRUE if type1 and type2 actually represent equivalent types for this platform (the fortran member
of each type is ignored). For example, on 32-bit platforms, NPY_LONG and NPY_INT are equivalent.
Otherwise return NPY_FALSE.
Bool PyArray_EquivArrTypes(PyArrayObject* a1, PyArrayObject * a2)
Return NPY_TRUE if a1 and a2 are arrays with equivalent types for this platform.
Bool PyArray_EquivTypenums(int typenum1, int typenum2)
Special case of PyArray_EquivTypes (...) that does not accept flexible data types but may be easier to call.
int PyArray_EquivByteorders({byteorder} b1, {byteorder} b2)
True if byteorder characters ( NPY_LITTLE, NPY_BIG, NPY_NATIVE, NPY_IGNORE ) are either equal or
equivalent as to their specification of a native byte order. Thus, on a little-endian machine NPY_LITTLE and
NPY_NATIVE are equivalent where they are not equivalent on a big-endian machine.

Converting data types

PyObject* PyArray_Cast(PyArrayObject* arr, int typenum)

Mainly for backwards compatibility to the Numeric C-API and for simple casts to non-flexible types. Return a
new array object with the elements of arr cast to the data-type typenum which must be one of the enumerated
types and not a flexible type.
PyObject* PyArray_CastToType(PyArrayObject* arr, PyArray_Descr* type, int fortran)
Return a new array of the type specified, casting the elements of arr as appropriate. The fortran argument
specifies the ordering of the output array.
int PyArray_CastTo(PyArrayObject* out, PyArrayObject* in)
Cast the elements of the array in into the array out. The output array should be writeable, have an integer-
multiple of the number of elements in the input array (more than one copy can be placed in out), and have a
data type that is one of the builtin types. Returns 0 on success and -1 if an error occurs.
PyArray_VectorUnaryFunc* PyArray_GetCastFunc(PyArray_Descr* from, int totype)
Return the low-level casting function to cast from the given descriptor to the builtin type number. If no casting
function exists return NULL and set an error. Using this function instead of direct access to from ->f->cast will
allow support of any user-defined casting functions added to a descriptors casting dictionary.
int PyArray_CanCastSafely(int fromtype, int totype)
Returns non-zero if an array of data type fromtype can be cast to an array of data type totype without losing
information. An exception is that 64-bit integers are allowed to be cast to 64-bit floating point values even
though this can lose precision on large integers so as not to proliferate the use of long doubles without explict
requests. Flexible array types are not checked according to their lengths with this function.

5.4. Array API 1041

NumPy Reference, Release 2.0.0.dev8464

int PyArray_CanCastTo(PyArray_Descr* fromtype, PyArray_Descr* totype)

Returns non-zero if an array of data type fromtype (which can include flexible types) can be cast safely to
an array of data type totype (which can include flexible types). This is basically a wrapper around
PyArray_CanCastSafely with additional support for size checking if fromtype and totype are
NPY_STRING or NPY_UNICODE.
int PyArray_ObjectType(PyObject* op, int mintype)
This function is useful for determining a common type that two or more arrays can be converted to. It only
works for non-flexible array types as no itemsize information is passed. The mintype argument represents the
minimum type acceptable, and op represents the object that will be converted to an array. The return value is
the enumerated typenumber that represents the data-type that op should have.
void PyArray_ArrayType(PyObject* op, PyArray_Descr* mintype, PyArray_Descr* outtype)
This function works similarly to PyArray_ObjectType (...) except it handles flexible arrays. The mintype
argument can have an itemsize member and the outtype argument will have an itemsize member at least as big
but perhaps bigger depending on the object op.
PyArrayObject** PyArray_ConvertToCommonType(PyObject* op, int* n)
Convert a sequence of Python objects contained in op to an array of ndarrays each having the same data type.
The type is selected based on the typenumber (larger type number is chosen over a smaller one) ignoring objects
that are only scalars. The length of the sequence is returned in n, and an n -length array of PyArrayObject
pointers is the return value (or NULL if an error occurs). The returned array must be freed by the caller of this
routine (using PyDataMem_FREE ) and all the array objects in it DECREF d or a memory-leak will occur.
The example template-code below shows a typically usage:

mps = PyArray_ConvertToCommonType(obj, &n);

if (mps==NULL) return NULL;
{code}
<before return>
for (i=0; i<n; i++) Py_DECREF(mps[i]);
PyDataMem_FREE(mps);
{return}

char* PyArray_Zero(PyArrayObject* arr)

A pointer to newly created memory of size arr ->itemsize that holds the representation of 0 for that type. The
returned pointer, ret, must be freed using PyDataMem_FREE (ret) when it is not needed anymore.
char* PyArray_One(PyArrayObject* arr)
A pointer to newly created memory of size arr ->itemsize that holds the representation of 1 for that type. The
returned pointer, ret, must be freed using PyDataMem_FREE (ret) when it is not needed anymore.
int PyArray_ValidType(int typenum)
Returns NPY_TRUE if typenum represents a valid type-number (builtin or user-defined or character code).
Otherwise, this function returns NPY_FALSE.

New data types

void PyArray_InitArrFuncs(PyArray_ArrFuncs* f )
Initialize all function pointers and members to NULL.
int PyArray_RegisterDataType(PyArray_Descr* dtype)
Register a data-type as a new user-defined data type for arrays. The type must have most of its entries filled
in. This is not always checked and errors can produce segfaults. In particular, the typeobj member of the
dtype structure must be filled with a Python type that has a fixed-size element-size that corresponds to the
elsize member of dtype. Also the f member must have the required functions: nonzero, copyswap, copyswapn,
getitem, setitem, and cast (some of the cast functions may be NULL if no support is desired). To avoid
confusion, you should choose a unique character typecode but this is not enforced and not relied on internally.

1042 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

A user-defined type number is returned that uniquely identifies the type. A pointer to the new structure can then
be obtained from PyArray_DescrFromType using the returned type number. A -1 is returned if an error
occurs. If this dtype has already been registered (checked only by the address of the pointer), then return the
previously-assigned type-number.
int PyArray_RegisterCastFunc(PyArray_Descr* descr, int totype, PyArray_VectorUnaryFunc* castfunc)
Register a low-level casting function, castfunc, to convert from the data-type, descr, to the given data-type
number, totype. Any old casting function is over-written. A 0 is returned on success or a -1 on failure.
int PyArray_RegisterCanCast(PyArray_Descr* descr, int totype, PyArray_SCALARKIND scalar)
Register the data-type number, totype, as castable from data-type object, descr, of the given scalar kind. Use
scalar = NPY_NOSCALAR to register that an array of data-type descr can be cast safely to a data-type whose
type_number is totype.

Special functions for PyArray_OBJECT

int PyArray_INCREF(PyArrayObject* op)

Used for an array, op, that contains any Python objects. It increments the reference count of every object in the
array according to the data-type of op. A -1 is returned if an error occurs, otherwise 0 is returned.
void PyArray_Item_INCREF(char* ptr, PyArray_Descr* dtype)
A function to INCREF all the objects at the location ptr according to the data-type dtype. If ptr is the start of a
record with an object at any offset, then this will (recursively) increment the reference count of all object-like
items in the record.
int PyArray_XDECREF(PyArrayObject* op)
Used for an array, op, that contains any Python objects. It decrements the reference count of every object in the
array according to the data-type of op. Normal return value is 0. A -1 is returned if an error occurs.
void PyArray_Item_XDECREF(char* ptr, PyArray_Descr* dtype)
A function to XDECREF all the object-like items at the loacation ptr as recorded in the data-type, dtype.
This works recursively so that if dtype itself has fields with data-types that contain object-like items, all the
object-like fields will be XDECREF d.
void PyArray_FillObjectArray(PyArrayObject* arr, PyObject* obj)
Fill a newly created array with a single value obj at all locations in the structure with object data-types. No
checking is performed but arr must be of data-type PyArray_OBJECT and be single-segment and uninitial-
ized (no previous objects in position). Use PyArray_DECREF (arr) if you need to decrement all the items in
the object array prior to calling this function.

5.4.4 Array flags

The flags attribute of the PyArrayObject structure contains important information about the memory used by
the array (pointed to by the data member) This flag information must be kept accurate or strange results and even
segfaults may result.
There are 6 (binary) flags that describe the memory area used by the data buffer. These constants are defined in
arrayobject.h and determine the bit-position of the flag. Python exposes a nice attribute- based interface as well
as a dictionary-like interface for getting (and, if appropriate, setting) these flags.
Memory areas of all kinds can be pointed to by an ndarray, necessitating these flags. If you get an arbitrary
PyArrayObject in C-code, you need to be aware of the flags that are set. If you need to guarantee a certain kind
of array (like NPY_CONTIGUOUS and NPY_BEHAVED), then pass these requirements into the PyArray_FromAny
function.

5.4. Array API 1043

NumPy Reference, Release 2.0.0.dev8464

Basic Array Flags

An ndarray can have a data segment that is not a simple contiguous chunk of well-behaved memory you can manip-
ulate. It may not be aligned with word boundaries (very important on some platforms). It might have its data in a
different byte-order than the machine recognizes. It might not be writeable. It might be in Fortan-contiguous order.
The array flags are used to indicate what can be said about data associated with an array.
NPY_C_CONTIGUOUS
The data area is in C-style contiguous order (last index varies the fastest).
NPY_F_CONTIGUOUS
The data area is in Fortran-style contiguous order (first index varies the fastest).
Notice that contiguous 1-d arrays are always both NPY_FORTRAN contiguous and C contiguous. Both of these flags
can be checked and are convenience flags only as whether or not an array is NPY_CONTIGUOUS or NPY_FORTRAN
can be determined by the strides, dimensions, and itemsize attributes.
NPY_OWNDATA
The data area is owned by this array.
NPY_ALIGNED
The data area is aligned appropriately (for all strides).
NPY_WRITEABLE
The data area can be written to.
Notice that the above 3 flags are are defined so that a new, well- behaved array has these flags defined as true.
NPY_UPDATEIFCOPY
The data area represents a (well-behaved) copy whose information should be transferred back to the original
when this array is deleted.
This is a special flag that is set if this array represents a copy made because a user required certain flags in
PyArray_FromAny and a copy had to be made of some other array (and the user asked for this flag to be
set in such a situation). The base attribute then points to the misbehaved array (which is set read_only).
When the array with this flag set is deallocated, it will copy its contents back to the misbehaved array
(casting if necessary) and will reset the misbehaved array to NPY_WRITEABLE. If the misbehaved ar-
ray was not NPY_WRITEABLE to begin with then PyArray_FromAny would have returned an error because
NPY_UPDATEIFCOPY would not have been possible.
PyArray_UpdateFlags (obj, flags) will update the obj->flags for flags which can be any of
NPY_CONTIGUOUS, NPY_FORTRAN, NPY_ALIGNED, or NPY_WRITEABLE.

Combinations of array flags

NPY_BEHAVED
NPY_ALIGNED | NPY_WRITEABLE
NPY_CARRAY
NPY_C_CONTIGUOUS | NPY_BEHAVED
NPY_CARRAY_RO
NPY_C_CONTIGUOUS | NPY_ALIGNED
NPY_FARRAY
NPY_F_CONTIGUOUS | NPY_BEHAVED
NPY_FARRAY_RO
NPY_F_CONTIGUOUS | NPY_ALIGNED

1044 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

NPY_DEFAULT
NPY_CARRAY
NPY_UPDATE_ALL
NPY_C_CONTIGUOUS | NPY_F_CONTIGUOUS | NPY_ALIGNED

Flag-like constants

These constants are used in PyArray_FromAny (and its macro forms) to specify desired properties of the new array.

NPY_FORCECAST
Cast to the desired type, even if it cant be done without losing information.
NPY_ENSURECOPY
Make sure the resulting array is a copy of the original.
NPY_ENSUREARRAY
Make sure the resulting object is an actual ndarray (or bigndarray), and not a sub-class.
NPY_NOTSWAPPED
Only used in PyArray_CheckFromAny to over-ride the byteorder of the data-type object passed in.
NPY_BEHAVED_NS
NPY_ALIGNED | NPY_WRITEABLE | NPY_NOTSWAPPED

Flag checking

For all of these macros arr must be an instance of a (subclass of) PyArray_Type, but no checking is done.
PyArray_CHKFLAGS(arr, flags)
The first parameter, arr, must be an ndarray or subclass. The parameter, flags, should be an integer consisting of
bitwise combinations of the possible flags an array can have: NPY_C_CONTIGUOUS, NPY_F_CONTIGUOUS,
NPY_OWNDATA, NPY_ALIGNED, NPY_WRITEABLE, NPY_UPDATEIFCOPY.
PyArray_ISCONTIGUOUS(arr)
Evaluates true if arr is C-style contiguous.
PyArray_ISFORTRAN(arr)
Evaluates true if arr is Fortran-style contiguous.
PyArray_ISWRITEABLE(arr)
Evaluates true if the data area of arr can be written to
PyArray_ISALIGNED(arr)
Evaluates true if the data area of arr is properly aligned on the machine.
PyArray_ISBEHAVED(arr)
Evalutes true if the data area of arr is aligned and writeable and in machine byte-order according to its descriptor.
PyArray_ISBEHAVED_RO(arr)
Evaluates true if the data area of arr is aligned and in machine byte-order.
PyArray_ISCARRAY(arr)
Evaluates true if the data area of arr is C-style contiguous, and PyArray_ISBEHAVED (arr) is true.
PyArray_ISFARRAY(arr)
Evaluates true if the data area of arr is Fortran-style contiguous and PyArray_ISBEHAVED (arr) is true.
PyArray_ISCARRAY_RO(arr)
Evaluates true if the data area of arr is C-style contiguous, aligned, and in machine byte-order.

5.4. Array API 1045

NumPy Reference, Release 2.0.0.dev8464

PyArray_ISFARRAY_RO(arr)
Evaluates true if the data area of arr is Fortran-style contiguous, aligned, and in machine byte-order .
PyArray_ISONESEGMENT(arr)
Evaluates true if the data area of arr consists of a single (C-style or Fortran-style) contiguous segment.
void PyArray_UpdateFlags(PyArrayObject* arr, int flagmask)
The NPY_C_CONTIGUOUS, NPY_ALIGNED, and NPY_F_CONTIGUOUS array flags can be calculated
from the array object itself. This routine updates one or more of these flags of arr as specified in flagmask by
performing the required calculation.

Warning: It is important to keep the flags updated (using PyArray_UpdateFlags can help) whenever a
manipulation with an array is performed that might cause them to change. Later calculations in NumPy that rely
on the state of these flags do not repeat the calculation to update them.

5.4.5 Array method alternative API

Conversion

PyObject* PyArray_GetField(PyArrayObject* self, PyArray_Descr* dtype, int offset)

Equivalent to ndarray.getfield (self, dtype, offset). Return a new array of the given dtype using the data
in the current array at a specified offset in bytes. The offset plus the itemsize of the new array type must be
less than self ->descr->elsize or an error is raised. The same shape and strides as the original array are used.
Therefore, this function has the effect of returning a field from a record array. But, it can also be used to select
specific bytes or groups of bytes from any array type.
int PyArray_SetField(PyArrayObject* self, PyArray_Descr* dtype, int offset, PyObject* val)
Equivalent to ndarray.setfield (self, val, dtype, offset ). Set the field starting at offset in bytes and of the
given dtype to val. The offset plus dtype ->elsize must be less than self ->descr->elsize or an error is raised.
Otherwise, the val argument is converted to an array and copied into the field pointed to. If necessary, the
elements of val are repeated to fill the destination array, But, the number of elements in the destination must be
an integer multiple of the number of elements in val.
PyObject* PyArray_Byteswap(PyArrayObject* self, Bool inplace)
Equivalent to ndarray.byteswap (self, inplace). Return an array whose data area is byteswapped. If in-
place is non-zero, then do the byteswap inplace and return a reference to self. Otherwise, create a byteswapped
copy and leave self unchanged.
PyObject* PyArray_NewCopy(PyArrayObject* old, NPY_ORDER order)
Equivalent to ndarray.copy (self, fortran). Make a copy of the old array. The returned array is always
aligned and writeable with data interpreted the same as the old array. If order is NPY_CORDER, then a C-style
contiguous array is returned. If order is NPY_FORTRANORDER, then a Fortran-style contiguous array is
returned. If order is NPY_ANYORDER, then the array returned is Fortran-style contiguous only if the old one is;
otherwise, it is C-style contiguous.
PyObject* PyArray_ToList(PyArrayObject* self )
Equivalent to ndarray.tolist (self ). Return a nested Python list from self.
PyObject* PyArray_ToString(PyArrayObject* self, NPY_ORDER order)
Equivalent to ndarray.tostring (self, order). Return the bytes of this array in a Python string.
PyObject* PyArray_ToFile(PyArrayObject* self, FILE* fp, char* sep, char* format)
Write the contents of self to the file pointer fp in C-style contiguous fashion. Write the data as binary bytes
if sep is the string or NULL. Otherwise, write the contents of self as text using the sep string as the item
separator. Each item will be printed to the file. If the format string is not NULL or , then it is a Python print
statement format string showing how the items are to be written.

1046 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

int PyArray_Dump(PyObject* self, PyObject* file, int protocol)

Pickle the object in self to the given file (either a string or a Python file object). If file is a Python string it is
considered to be the name of a file which is then opened in binary mode. The given protocol is used (if protocol
is negative, or the highest available is used). This is a simple wrapper around cPickle.dump(self, file, protocol).
PyObject* PyArray_Dumps(PyObject* self, int protocol)
Pickle the object in self to a Python string and return it. Use the Pickle protocol provided (or the highest
available if protocol is negative).
int PyArray_FillWithScalar(PyArrayObject* arr, PyObject* obj)
Fill the array, arr, with the given scalar object, obj. The object is first converted to the data type of arr, and then
copied into every location. A -1 is returned if an error occurs, otherwise 0 is returned.
PyObject* PyArray_View(PyArrayObject* self, PyArray_Descr* dtype)
Equivalent to ndarray.view (self, dtype). Return a new view of the array self as possibly a different data-
type, dtype. If dtype is NULL, then the returned array will have the same data type as self. The new data-type
must be consistent with the size of self. Either the itemsizes must be identical, or self must be single-segment
and the total number of bytes must be the same. In the latter case the dimensions of the returned array will be
altered in the last (or first for Fortran-style contiguous arrays) dimension. The data area of the returned array
and self is exactly the same.

Shape Manipulation

PyObject* PyArray_Newshape(PyArrayObject* self, PyArray_Dims* newshape)

Result will be a new array (pointing to the same memory location as self if possible), but having a shape given
by newshape . If the new shape is not compatible with the strides of self, then a copy of the array with the new
specified shape will be returned.
PyObject* PyArray_Reshape(PyArrayObject* self, PyObject* shape)
Equivalent to ndarray.reshape (self, shape) where shape is a sequence. Converts shape to a
PyArray_Dims structure and calls PyArray_Newshape internally.
PyObject* PyArray_Squeeze(PyArrayObject* self )
Equivalent to ndarray.squeeze (self ). Return a new view of self with all of the dimensions of length 1
removed from the shape.

Warning: matrix objects are always 2-dimensional. Therefore, PyArray_Squeeze has no effect on arrays of
matrix sub-class.

PyObject* PyArray_SwapAxes(PyArrayObject* self, int a1, int a2)

Equivalent to ndarray.swapaxes (self, a1, a2). The returned array is a new view of the data in self with
the given axes, a1 and a2, swapped.
PyObject* PyArray_Resize(PyArrayObject* self, PyArray_Dims* newshape, int refcheck, NPY_ORDER for-
tran)
Equivalent to ndarray.resize (self, newshape, refcheck = refcheck, order= fortran ). This function only
works on single-segment arrays. It changes the shape of self inplace and will reallocate the memory for self
if newshape has a different total number of elements then the old shape. If reallocation is necessary, then self
must own its data, have self - >base==NULL, have self - >weakrefs==NULL, and (unless refcheck is 0)
not be referenced by any other array. A reference to the new array is returned. The fortran argument can be
NPY_ANYORDER, NPY_CORDER, or NPY_FORTRANORDER. It currently has no effect. Eventually it could be
used to determine how the resize operation should view the data when constructing a differently-dimensioned
array.
PyObject* PyArray_Transpose(PyArrayObject* self, PyArray_Dims* permute)
Equivalent to ndarray.transpose (self, permute). Permute the axes of the ndarray object self according
to the data structure permute and return the result. If permute is NULL, then the resulting array has its axes

5.4. Array API 1047

NumPy Reference, Release 2.0.0.dev8464

reversed. For example if self has shape 10 20 30, and permute .ptr is (0,2,1) the shape of the result is
10 30 20. If permute is NULL, the shape of the result is 30 20 10.
PyObject* PyArray_Flatten(PyArrayObject* self, NPY_ORDER order)
Equivalent to ndarray.flatten (self, order). Return a 1-d copy of the array. If order is
NPY_FORTRANORDER the elements are scanned out in Fortran order (first-dimension varies the fastest).
If order is NPY_CORDER, the elements of self are scanned in C-order (last dimension varies the fastest). If
order NPY_ANYORDER, then the result of PyArray_ISFORTRAN (self ) is used to determine which order to
flatten.
PyObject* PyArray_Ravel(PyArrayObject* self, NPY_ORDER order)
Equivalent to self.ravel(order). Same basic functionality as PyArray_Flatten (self, order) except if order
is 0 and self is C-style contiguous, the shape is altered but no copy is performed.

Item selection and manipulation

PyObject* PyArray_TakeFrom(PyArrayObject* self, PyObject* indices, int axis, PyArrayObject* ret,

NPY_CLIPMODE clipmode)
Equivalent to ndarray.take (self, indices, axis, ret, clipmode) except axis =None in Python is obtained by
setting axis = NPY_MAXDIMS in C. Extract the items from self indicated by the integer-valued indices along
the given axis. The clipmode argument can be NPY_RAISE, NPY_WRAP, or NPY_CLIP to indicate what to
do with out-of-bound indices. The ret argument can specify an output array rather than having one created
internally.
PyObject* PyArray_PutTo(PyArrayObject* self, PyObject* values, PyObject* indices, NPY_CLIPMODE
clipmode)
Equivalent to self.put(values, indices, clipmode ). Put values into self at the corresponding (flattened) indices.
If values is too small it will be repeated as necessary.
PyObject* PyArray_PutMask(PyArrayObject* self, PyObject* values, PyObject* mask)
Place the values in self wherever corresponding positions (using a flattened context) in mask are true. The
mask and self arrays must have the same total number of elements. If values is too small, it will be repeated as
necessary.
PyObject* PyArray_Repeat(PyArrayObject* self, PyObject* op, int axis)
Equivalent to ndarray.repeat (self, op, axis). Copy the elements of self, op times along the given axis.
Either op is a scalar integer or a sequence of length self ->dimensions[ axis ] indicating how many times to
repeat each item along the axis.
PyObject* PyArray_Choose(PyArrayObject* self, PyObject* op, PyArrayObject* ret, NPY_CLIPMODE
clipmode)
Equivalent to ndarray.choose (self, op, ret, clipmode). Create a new array by selecting elements from the
sequence of arrays in op based on the integer values in self. The arrays must all be broadcastable to the same
shape and the entries in self should be between 0 and len(op). The output is placed in ret unless it is NULL in
which case a new output is created. The clipmode argument determines behavior for when entries in self are
not between 0 and len(op).
NPY_RAISE
raise a ValueError;
NPY_WRAP
wrap values < 0 by adding len(op) and values >=len(op) by subtracting len(op) until they are in range;
NPY_CLIP
all values are clipped to the region [0, len(op) ).
PyObject* PyArray_Sort(PyArrayObject* self, int axis)
Equivalent to ndarray.sort (self, axis). Return an array with the items of self sorted along axis.

1048 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

PyObject* PyArray_ArgSort(PyArrayObject* self, int axis)

Equivalent to ndarray.argsort (self, axis). Return an array of indices such that selection of these indices
along the given axis would return a sorted version of self. If self ->descr is a data-type with fields defined,
then self->descr->names is used to determine the sort order. A comparison where the first field is equal will use
the second field and so on. To alter the sort order of a record array, create a new data-type with a different order
of names and construct a view of the array with that new data-type.
PyObject* PyArray_LexSort(PyObject* sort_keys, int axis)
Given a sequence of arrays (sort_keys) of the same shape, return an array of indices (similar to
PyArray_ArgSort (...)) that would sort the arrays lexicographically. A lexicographic sort specifies
that when two keys are found to be equal, the order is based on comparison of subsequent keys. A merge
sort (which leaves equal entries unmoved) is required to be defined for the types. The sort is accomplished
by sorting the indices first using the first sort_key and then using the second sort_key and so forth. This is
equivalent to the lexsort(sort_keys, axis) Python command. Because of the way the merge-sort works, be sure
to understand the order the sort_keys must be in (reversed from the order you would use when comparing two
elements).
If these arrays are all collected in a record array, then PyArray_Sort (...) can also be used to sort the array
directly.
PyObject* PyArray_SearchSorted(PyArrayObject* self, PyObject* values)
Equivalent to ndarray.searchsorted (self, values). Assuming self is a 1-d array in ascending order
representing bin boundaries then the output is an array the same shape as values of bin numbers, giving the bin
into which each item in values would be placed. No checking is done on whether or not self is in ascending
order.
PyObject* PyArray_Diagonal(PyArrayObject* self, int offset, int axis1, int axis2)
Equivalent to ndarray.diagonal (self, offset, axis1, axis2 ). Return the offset diagonals of the 2-d arrays
defined by axis1 and axis2.
PyObject* PyArray_Nonzero(PyArrayObject* self )
Equivalent to ndarray.nonzero (self ). Returns a tuple of index arrays that select elements of self that are
nonzero. If (nd= PyArray_NDIM ( self ))==1, then a single index array is returned. The index arrays have
data type NPY_INTP. If a tuple is returned (nd 6= 1), then its length is nd.
PyObject* PyArray_Compress(PyArrayObject* self, PyObject* condition, int axis, PyArrayObject* out)
Equivalent to ndarray.compress (self, condition, axis ). Return the elements along axis corresponding to
elements of condition that are true.

Calculation

Tip: Pass in NPY_MAXDIMS for axis in order to achieve the same effect that is obtained by passing in axis = None
in Python (treating the array as a 1-d array).
PyObject* PyArray_ArgMax(PyArrayObject* self, int axis)
Equivalent to ndarray.argmax (self, axis). Return the index of the largest element of self along axis.
PyObject* PyArray_ArgMin(PyArrayObject* self, int axis)
Equivalent to ndarray.argmin (self, axis). Return the index of the smallest element of self along axis.
PyObject* PyArray_Max(PyArrayObject* self, int axis, PyArrayObject* out)
Equivalent to ndarray.max (self, axis). Return the largest element of self along the given axis.
PyObject* PyArray_Min(PyArrayObject* self, int axis, PyArrayObject* out)
Equivalent to ndarray.min (self, axis). Return the smallest element of self along the given axis.
PyObject* PyArray_Ptp(PyArrayObject* self, int axis, PyArrayObject* out)
Equivalent to ndarray.ptp (self, axis). Return the difference between the largest element of self along axis
and the smallest element of self along axis.

5.4. Array API 1049

NumPy Reference, Release 2.0.0.dev8464

Note: The rtype argument specifies the data-type the reduction should take place over. This is important if the data-
type of the array is not large enough to handle the output. By default, all integer data-types are made at least as large
as NPY_LONG for the add and multiply ufuncs (which form the basis for mean, sum, cumsum, prod, and cumprod
functions).
PyObject* PyArray_Mean(PyArrayObject* self, int axis, int rtype, PyArrayObject* out)
Equivalent to ndarray.mean (self, axis, rtype). Returns the mean of the elements along the given axis,
using the enumerated type rtype as the data type to sum in. Default sum behavior is obtained using
PyArray_NOTYPE for rtype.
PyObject* PyArray_Trace(PyArrayObject* self, int offset, int axis1, int axis2, int rtype, PyArrayObject* out)
Equivalent to ndarray.trace (self, offset, axis1, axis2, rtype). Return the sum (using rtype as the data
type of summation) over the offset diagonal elements of the 2-d arrays defined by axis1 and axis2 variables. A
positive offset chooses diagonals above the main diagonal. A negative offset selects diagonals below the main
diagonal.
PyObject* PyArray_Clip(PyArrayObject* self, PyObject* min, PyObject* max)
Equivalent to ndarray.clip (self, min, max). Clip an array, self, so that values larger than max are fixed to
max and values less than min are fixed to min.
PyObject* PyArray_Conjugate(PyArrayObject* self )
Equivalent to ndarray.conjugate (self ). Return the complex conjugate of self. If self is not of complex
data type, then return self with an reference.
PyObject* PyArray_Round(PyArrayObject* self, int decimals, PyArrayObject* out)
Equivalent to ndarray.round (self, decimals, out). Returns the array with elements rounded to the nearest
decimal place. The decimal place is defined as the 10decimals digit so that negative decimals cause rounding to
the nearest 10s, 100s, etc. If out is NULL, then the output array is created, otherwise the output is placed in out
which must be the correct size and type.
PyObject* PyArray_Std(PyArrayObject* self, int axis, int rtype, PyArrayObject* out)
Equivalent to ndarray.std (self, axis, rtype). Return the standard deviation using data along axis converted
to data type rtype.
PyObject* PyArray_Sum(PyArrayObject* self, int axis, int rtype, PyArrayObject* out)
Equivalent to ndarray.sum (self, axis, rtype). Return 1-d vector sums of elements in self along axis. Perform
the sum after converting data to data type rtype.
PyObject* PyArray_CumSum(PyArrayObject* self, int axis, int rtype, PyArrayObject* out)
Equivalent to ndarray.cumsum (self, axis, rtype). Return cumulative 1-d sums of elements in self along
axis. Perform the sum after converting data to data type rtype.
PyObject* PyArray_Prod(PyArrayObject* self, int axis, int rtype, PyArrayObject* out)
Equivalent to ndarray.prod (self, axis, rtype). Return 1-d products of elements in self along axis. Perform
the product after converting data to data type rtype.
PyObject* PyArray_CumProd(PyArrayObject* self, int axis, int rtype, PyArrayObject* out)
Equivalent to ndarray.cumprod (self, axis, rtype). Return 1-d cumulative products of elements in self
along axis. Perform the product after converting data to data type rtype.
PyObject* PyArray_All(PyArrayObject* self, int axis, PyArrayObject* out)
Equivalent to ndarray.all (self, axis). Return an array with True elements for every 1-d sub-array of self
defined by axis in which all the elements are True.
PyObject* PyArray_Any(PyArrayObject* self, int axis, PyArrayObject* out)
Equivalent to ndarray.any (self, axis). Return an array with True elements for every 1-d sub-array of self
defined by axis in which any of the elements are True.

1050 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

5.4.6 Functions

Array Functions

int PyArray_AsCArray(PyObject** op, void* ptr, npy_intp* dims, int nd, int typenum, int itemsize)
Sometimes it is useful to access a multidimensional array as a C-style multi-dimensional array so that algorithms
can be implemented using Cs a[i][j][k] syntax. This routine returns a pointer, ptr, that simulates this kind of
C-style array, for 1-, 2-, and 3-d ndarrays.
Parameters

op The address to any Python object. This Python object will be replaced with an equiv-
alent well-behaved, C-style contiguous, ndarray of the given data type specifice by the last
two arguments. Be sure that stealing a reference in this way to the input object is justified.
ptr The address to a (ctype* for 1-d, ctype** for 2-d or ctype*** for 3-d) variable where
ctype is the equivalent C-type for the data type. On return, ptr will be addressable as a 1-d,
2-d, or 3-d array.
dims An output array that contains the shape of the array object. This array gives bound-
aries on any looping that will take place.
nd The dimensionality of the array (1, 2, or 3).
typenum The expected data type of the array.
itemsize This argument is only needed when typenum represents a flexible array. Otherwise
it should be 0.
Note: The simulation of a C-style array is not complete for 2-d and 3-d arrays. For example, the simulated arrays of
pointers cannot be passed to subroutines expecting specific, statically-defined 2-d and 3-d arrays. To pass to functions
requiring those kind of inputs, you must statically define the required array and copy data.
int PyArray_Free(PyObject* op, void* ptr)
Must be called with the same objects and memory locations returned from PyArray_AsCArray (...). This
function cleans up memory that otherwise would get leaked.
PyObject* PyArray_Concatenate(PyObject* obj, int axis)
Join the sequence of objects in obj together along axis into a single array. If the dimensions or types are not
compatible an error is raised.
PyObject* PyArray_InnerProduct(PyObject* obj1, PyObject* obj2)
Compute a product-sum over the last dimensions of obj1 and obj2. Neither array is conjugated.
PyObject* PyArray_MatrixProduct(PyObject* obj1, PyObject* obj)
Compute a product-sum over the last dimension of obj1 and the second-to-last dimension of obj2. For 2-d
arrays this is a matrix-product. Neither array is conjugated.
PyObject* PyArray_CopyAndTranspose(PyObject * op)
A specialized copy and transpose function that works only for 2-d arrays. The returned array is a transposed
copy of op.
PyObject* PyArray_Correlate(PyObject* op1, PyObject* op2, int mode)
Compute the 1-d correlation of the 1-d arrays op1 and op2 . The correlation is computed at each output point
by multiplying op1 by a shifted version of op2 and summing the result. As a result of the shift, needed values
outside of the defined range of op1 and op2 are interpreted as zero. The mode determines how many shifts to
return: 0 - return only shifts that did not need to assume zero- values; 1 - return an object that is the same size
as op1, 2 - return all possible shifts (any overlap at all is accepted).

5.4. Array API 1051

NumPy Reference, Release 2.0.0.dev8464

Notes
This does not compute the usual correlation: if op2 is larger than op1, the arguments are swapped, and the
conjugate is never taken for complex arrays. See PyArray_Correlate2 for the usual signal processing correlation.
PyObject* PyArray_Correlate2(PyObject* op1, PyObject* op2, int mode)
Updated version of PyArray_Correlate, which uses the usual definition of correlation for 1d arrays. The corre-
lation is computed at each output point by multiplying op1 by a shifted version of op2 and summing the result.
As a result of the shift, needed values outside of the defined range of op1 and op2 are interpreted as zero. The
mode determines how many shifts to return: 0 - return only shifts that did not need to assume zero- values; 1 -
return an object that is the same size as op1, 2 - return all possible shifts (any overlap at all is accepted).

Notes
Compute z as follows:

z[k] = sum_n op1[n] * conj(op2[n+k])

PyObject* PyArray_Where(PyObject* condition, PyObject* x, PyObject* y)

If both x and y are NULL, then return PyArray_Nonzero (condition). Otherwise, both x and y must be given
and the object returned is shaped like condition and has elements of x and y where condition is respectively
True or False.

Other functions

Bool PyArray_CheckStrides(int elsize, int nd, npy_intp numbytes, npy_intp* dims, npy_intp* newstrides)
Determine if newstrides is a strides array consistent with the memory of an nd -dimensional array with shape
dims and element-size, elsize. The newstrides array is checked to see if jumping by the provided number of
bytes in each direction will ever mean jumping more than numbytes which is the assumed size of the available
memory segment. If numbytes is 0, then an equivalent numbytes is computed assuming nd, dims, and elsize
refer to a single-segment array. Return NPY_TRUE if newstrides is acceptable, otherwise return NPY_FALSE.
npy_intp PyArray_MultiplyList(npy_intp* seq, int n)

int PyArray_MultiplyIntList(int* seq, int n)

Both of these routines multiply an n -length array, seq, of integers and return the result. No overflow checking
is performed.
int PyArray_CompareLists(npy_intp* l1, npy_intp* l2, int n)
Given two n -length arrays of integers, l1, and l2, return 1 if the lists are identical; otherwise, return 0.

5.4.7 Array Iterators

An array iterator is a simple way to access the elements of an N-dimensional array quickly and efficiently. Section 2
provides more description and examples of this useful approach to looping over an array.
PyObject* PyArray_IterNew(PyObject* arr)
Return an array iterator object from the array, arr. This is equivalent to arr. flat. The array iterator object
makes it easy to loop over an N-dimensional non-contiguous array in C-style contiguous fashion.
PyObject* PyArray_IterAllButAxis(PyObject* arr, int *axis)
Return an array iterator that will iterate over all axes but the one provided in *axis. The returned iterator cannot
be used with PyArray_ITER_GOTO1D. This iterator could be used to write something similar to what ufuncs
do wherein the loop over the largest axis is done by a separate sub-routine. If *axis is negative then *axis will
be set to the axis having the smallest stride and that axis will be used.

1052 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

PyObject * PyArray_BroadcastToShape(PyObject* arr, npy_intp *dimensions, int nd)

Return an array iterator that is broadcast to iterate as an array of the shape provided by dimensions and nd.
int PyArrayIter_Check(PyObject* op)
Evaluates true if op is an array iterator (or instance of a subclass of the array iterator type).
void PyArray_ITER_RESET(PyObject* iterator)
Reset an iterator to the beginning of the array.
void PyArray_ITER_NEXT(PyObject* iterator)
Incremement the index and the dataptr members of the iterator to point to the next element of the array. If the
array is not (C-style) contiguous, also increment the N-dimensional coordinates array.
void * PyArray_ITER_DATA(PyObject* iterator)
A pointer to the current element of the array.
void PyArray_ITER_GOTO(PyObject* iterator, npy_intp* destination)
Set the iterator index, dataptr, and coordinates members to the location in the array indicated by the
N-dimensional c-array, destination, which must have size at least iterator ->nd_m1+1.
PyArray_ITER_GOTO1D(PyObject* iterator, npy_intp index)
Set the iterator index and dataptr to the location in the array indicated by the integer index which points to an
element in the C-styled flattened array.
int PyArray_ITER_NOTDONE(PyObject* iterator)
Evaluates TRUE as long as the iterator has not looped through all of the elements, otherwise it evaluates FALSE.

5.4.8 Broadcasting (multi-iterators)

PyObject* PyArray_MultiIterNew(int num, ...)

A simplified interface to broadcasting. This function takes the number of arrays to broadcast and then num
extra ( PyObject * ) arguments. These arguments are converted to arrays and iterators are created.
PyArray_Broadcast is then called on the resulting multi-iterator object. The resulting, broadcasted
mult-iterator object is then returned. A broadcasted operation can then be performed using a single loop and
using PyArray_MultiIter_NEXT (..)
void PyArray_MultiIter_RESET(PyObject* multi)
Reset all the iterators to the beginning in a multi-iterator object, multi.
void PyArray_MultiIter_NEXT(PyObject* multi)
Advance each iterator in a multi-iterator object, multi, to its next (broadcasted) element.
void * PyArray_MultiIter_DATA(PyObject* multi, int i)
Return the data-pointer of the i th iterator in a multi-iterator object.
void PyArray_MultiIter_NEXTi(PyObject* multi, int i)
Advance the pointer of only the i th iterator.
void PyArray_MultiIter_GOTO(PyObject* multi, npy_intp* destination)
Advance each iterator in a multi-iterator object, multi, to the given N -dimensional destination where N is the
number of dimensions in the broadcasted array.
void PyArray_MultiIter_GOTO1D(PyObject* multi, npy_intp index)
Advance each iterator in a multi-iterator object, multi, to the corresponding location of the index into the flat-
tened broadcasted array.
int PyArray_MultiIter_NOTDONE(PyObject* multi)
Evaluates TRUE as long as the multi-iterator has not looped through all of the elements (of the broadcasted
result), otherwise it evaluates FALSE.

5.4. Array API 1053

NumPy Reference, Release 2.0.0.dev8464

int PyArray_Broadcast(PyArrayMultiIterObject* mit)

This function encapsulates the broadcasting rules. The mit container should already contain iterators for all
the arrays that need to be broadcast. On return, these iterators will be adjusted so that iteration over each
simultaneously will accomplish the broadcasting. A negative number is returned if an error occurs.
int PyArray_RemoveSmallest(PyArrayMultiIterObject* mit)
This function takes a multi-iterator object that has been previously broadcasted, finds the dimension with
the smallest sum of strides in the broadcasted result and adapts all the iterators so as not to iterate over
that dimension (by effectively making them of length-1 in that dimension). The corresponding dimension is
returned unless mit ->nd is 0, then -1 is returned. This function is useful for constructing ufunc-like routines
that broadcast their inputs correctly and then call a strided 1-d version of the routine as the inner-loop. This 1-d
version is usually optimized for speed and for this reason the loop should be performed over the axis that wont
require large stride jumps.

5.4.9 Neighborhood iterator

New in version 1.4.0. Neighborhood iterators are subclasses of the iterator object, and can be used to iter over a
neighborhood of a point. For example, you may want to iterate over every voxel of a 3d image, and for every such
voxel, iterate over an hypercube. Neighborhood iterator automatically handle boundaries, thus making this kind of
code much easier to write than manual boundaries handling, at the cost of a slight overhead.
PyObject* PyArray_NeighborhoodIterNew(PyArrayIterObject* iter, npy_intp bounds, int mode,
PyArrayObject* fill_value)
This function creates a new neighborhood iterator from an existing iterator. The neighborhood will be computed
relatively to the position currently pointed by iter, the bounds define the shape of the neighborhood iterator, and
the mode argument the boundaries handling mode.
The bounds argument is expected to be a (2 * iter->ao->nd) arrays, such as the range bound[2*i]->bounds[2*i+1]
defines the range where to walk for dimension i (both bounds are included in the walked coordinates). The
bounds should be ordered for each dimension (bounds[2*i] <= bounds[2*i+1]).
The mode should be one of:
NPY_NEIGHBORHOOD_ITER_ZERO_PADDING: zero padding. Outside bounds values will be 0.
NPY_NEIGHBORHOOD_ITER_ONE_PADDING: one padding, Outside bounds values will be 1.
NPY_NEIGHBORHOOD_ITER_CONSTANT_PADDING: constant padding. Outside bounds values
will be the same as the first item in fill_value.
NPY_NEIGHBORHOOD_ITER_MIRROR_PADDING: mirror padding. Outside bounds values will be
as if the array items were mirrored. For example, for the array [1, 2, 3, 4], x[-2] will be 2, x[-2] will be 1,
x[4] will be 4, x[5] will be 1, etc...
NPY_NEIGHBORHOOD_ITER_CIRCULAR_PADDING: circular padding. Outside bounds values will
be as if the array was repeated. For example, for the array [1, 2, 3, 4], x[-2] will be 3, x[-2] will be 4, x[4]
will be 1, x[5] will be 2, etc...
If the mode is constant filling (NPY_NEIGHBORHOOD_ITER_CONSTANT_PADDING), fill_value should
point to an array object which holds the filling value (the first item will be the filling value if the array contains
more than one item). For other cases, fill_value may be NULL.
The iterator holds a reference to iter
Return NULL on failure (in which case the reference count of iter is not changed)
iter itself can be a Neighborhood iterator: this can be useful for .e.g automatic boundaries handling
the object returned by this function should be safe to use as a normal iterator

1054 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

If the position of iter is changed, any subsequent call to PyArrayNeighborhoodIter_Next is undefined

behavior, and PyArrayNeighborhoodIter_Reset must be called.

PyArrayIterObject \*iter;
PyArrayNeighborhoodIterObject \*neigh_iter;
iter = PyArray_IterNew(x);

//For a 3x3 kernel

bounds = {-1, 1, -1, 1};
neigh_iter = (PyArrayNeighborhoodIterObject*)PyArrayNeighborhoodIter_New(
iter, bounds, NPY_NEIGHBORHOOD_ITER_ZERO_PADDING, NULL);

for(i = 0; i < iter->size; ++i) {

for (j = 0; j < neigh_iter->size; ++j) {
// Walk around the item currently pointed by iter->dataptr
PyArrayNeighborhoodIter_Next(neigh_iter);
}

// Move to the next point of iter

PyArrayIter_Next(iter);
PyArrayNeighborhoodIter_Reset(neigh_iter);
}

int PyArrayNeighborhoodIter_Reset(PyArrayNeighborhoodIterObject* iter)

Reset the iterator position to the first point of the neighborhood. This should be called whenever the iter argu-
ment given at PyArray_NeighborhoodIterObject is changed (see example)
int PyArrayNeighborhoodIter_Next(PyArrayNeighborhoodIterObject* iter)
After this call, iter->dataptr points to the next point of the neighborhood. Calling this function after every point
of the neighborhood has been visited is undefined.

5.4.10 Array Scalars

PyObject* PyArray_Return(PyArrayObject* arr)

This function checks to see if arr is a 0-dimensional array and, if so, returns the appropriate array scalar. It
should be used whenever 0-dimensional arrays could be returned to Python.
PyObject* PyArray_Scalar(void* data, PyArray_Descr* dtype, PyObject* itemsize)
Return an array scalar object of the given enumerated typenum and itemsize by copying from memory pointed
to by data . If swap is nonzero then this function will byteswap the data if appropriate to the data-type because
array scalars are always in correct machine-byte order.
PyObject* PyArray_ToScalar(void* data, PyArrayObject* arr)
Return an array scalar object of the type and itemsize indicated by the array object arr copied from the memory
pointed to by data and swapping if the data in arr is not in machine byte-order.
PyObject* PyArray_FromScalar(PyObject* scalar, PyArray_Descr* outcode)
Return a 0-dimensional array of type determined by outcode from scalar which should be an array-scalar
object. If outcode is NULL, then the type is determined from scalar.
void PyArray_ScalarAsCtype(PyObject* scalar, void* ctypeptr)
Return in ctypeptr a pointer to the actual value in an array scalar. There is no error checking so scalar must be
an array-scalar object, and ctypeptr must have enough space to hold the correct type. For flexible-sized types, a
pointer to the data is copied into the memory of ctypeptr, for all other types, the actual data is copied into the
address pointed to by ctypeptr.

5.4. Array API 1055

NumPy Reference, Release 2.0.0.dev8464

void PyArray_CastScalarToCtype(PyObject* scalar, void* ctypeptr, PyArray_Descr* outcode)

Return the data (cast to the data type indicated by outcode) from the array-scalar, scalar, into the memory
pointed to by ctypeptr (which must be large enough to handle the incoming memory).
PyObject* PyArray_TypeObjectFromType(int type)
Returns a scalar type-object from a type-number, type . Equivalent to PyArray_DescrFromType (type)-
>typeobj except for reference counting and error-checking. Returns a new reference to the typeobject on
success or NULL on failure.
NPY_SCALARKIND PyArray_ScalarKind(int typenum, PyArrayObject** arr)
Return the kind of scalar represented by typenum and the array in *arr (if arr is not NULL ). The array is
assumed to be rank-0 and only used if typenum represents a signed integer. If arr is not NULL and the first
element is negative then NPY_INTNEG_SCALAR is returned, otherwise NPY_INTPOS_SCALAR is returned.
The possible return values are NPY_{kind}_SCALAR where {kind} can be INTPOS, INTNEG, FLOAT,
COMPLEX, BOOL, or OBJECT. NPY_NOSCALAR is also an enumerated value NPY_SCALARKIND
variables can take on.
int PyArray_CanCoerceScalar(char thistype, char neededtype, NPY_SCALARKIND scalar)
Implements the rules for scalar coercion. Scalars are only silently coerced from thistype to needed-
type if this function returns nonzero. If scalar is NPY_NOSCALAR, then this function is equivalent to
PyArray_CanCastSafely. The rule is that scalars of the same KIND can be coerced into arrays of the
same KIND. This rule means that high-precision scalars will never cause low-precision arrays of the same
KIND to be upcast.

5.4.11 Data-type descriptors

Warning: Data-type objects must be reference counted so be aware of the action on the data-type reference
of different C-API calls. The standard rule is that when a data-type object is returned it is a new reference.
Functions that take PyArray_Descr * objects and return arrays steal references to the data-type their inputs
unless otherwise noted. Therefore, you must own a reference to any data-type object used as input to such a
function.

int PyArrayDescr_Check(PyObject* obj)

Evaluates as true if obj is a data-type object ( PyArray_Descr * ).
PyArray_Descr* PyArray_DescrNew(PyArray_Descr* obj)
Return a new data-type object copied from obj (the fields reference is just updated so that the new object points
to the same fields dictionary if any).
PyArray_Descr* PyArray_DescrNewFromType(int typenum)
Create a new data-type object from the built-in (or user-registered) data-type indicated by typenum. All builtin
types should not have any of their fields changed. This creates a new copy of the PyArray_Descr structure
so that you can fill it in as appropriate. This function is especially needed for flexible data-types which need to
have a new elsize member in order to be meaningful in array construction.
PyArray_Descr* PyArray_DescrNewByteorder(PyArray_Descr* obj, char newendian)
Create a new data-type object with the byteorder set according to newendian. All referenced data-type objects
(in subdescr and fields members of the data-type object) are also changed (recursively). If a byteorder of
NPY_IGNORE is encountered it is left alone. If newendian is NPY_SWAP, then all byte-orders are swapped.
Other valid newendian values are NPY_NATIVE, NPY_LITTLE, and NPY_BIG which all cause the returned
data-typed descriptor (and all its referenced data-type descriptors) to have the corresponding byte- order.
PyArray_Descr* PyArray_DescrFromObject(PyObject* op, PyArray_Descr* mintype)
Determine an appropriate data-type object from the object op (which should be a nested sequence object) and
the minimum data-type descriptor mintype (which can be NULL ). Similar in behavior to array(op).dtype. Dont

1056 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

confuse this function with PyArray_DescrConverter. This function essentially looks at all the objects in
the (nested) sequence and determines the data-type from the elements it finds.
PyArray_Descr* PyArray_DescrFromScalar(PyObject* scalar)
Return a data-type object from an array-scalar object. No checking is done to be sure that scalar is an array
scalar. If no suitable data-type can be determined, then a data-type of NPY_OBJECT is returned by default.
PyArray_Descr* PyArray_DescrFromType(int typenum)
Returns a data-type object corresponding to typenum. The typenum can be one of the enumerated types, a
character code for one of the enumerated types, or a user-defined type.
int PyArray_DescrConverter(PyObject* obj, PyArray_Descr** dtype)
Convert any compatible Python object, obj, to a data-type object in dtype. A large number of Python objects
can be converted to data-type objects. See Data type objects (dtype) for a complete description. This version of
the converter converts None objects to a NPY_DEFAULT_TYPE data-type object. This function can be used
with the O& character code in PyArg_ParseTuple processing.
int PyArray_DescrConverter2(PyObject* obj, PyArray_Descr** dtype)
Convert any compatible Python object, obj, to a data-type object in dtype. This version of the converter converts
None objects so that the returned data-type is NULL. This function can also be used with the O& character in
PyArg_ParseTuple processing.
int Pyarray_DescrAlignConverter(PyObject* obj, PyArray_Descr** dtype)
Like PyArray_DescrConverter except it aligns C-struct-like objects on word-boundaries as the compiler
would.
int Pyarray_DescrAlignConverter2(PyObject* obj, PyArray_Descr** dtype)
Like PyArray_DescrConverter2 except it aligns C-struct-like objects on word-boundaries as the com-
piler would.
PyObject * PyArray_FieldNames(PyObject* dict)
Take the fields dictionary, dict, such as the one attached to a data-type object and construct an ordered-list of
field names such as is stored in the names field of the PyArray_Descr object.

5.4.12 Conversion Utilities

For use with PyArg_ParseTuple

All of these functions can be used in PyArg_ParseTuple (...) with the O& format specifier to automatically
convert any Python object to the required C-object. All of these functions return NPY_SUCCEED if successful and
NPY_FAIL if not. The first argument to all of these function is a Python object. The second argument is the address
of the C-type to convert the Python object to.

Warning: Be sure to understand what steps you should take to manage the memory when using these conversion
functions. These functions can require freeing memory, and/or altering the reference counts of specific objects
based on your use.

int PyArray_Converter(PyObject* obj, PyObject** address)

Convert any Python object to a PyArrayObject. If PyArray_Check (obj) is TRUE then its reference
count is incremented and a reference placed in address. If obj is not an array, then convert it to an array using
PyArray_FromAny . No matter what is returned, you must DECREF the object returned by this routine in
address when you are done with it.
int PyArray_OutputConverter(PyObject* obj, PyArrayObject** address)
This is a default converter for output arrays given to functions. If obj is Py_None or NULL, then *address will

5.4. Array API 1057

NumPy Reference, Release 2.0.0.dev8464

be NULL but the call will succeed. If PyArray_Check ( obj) is TRUE then it is returned in *address without
incrementing its reference count.
int PyArray_IntpConverter(PyObject* obj, PyArray_Dims* seq)
Convert any Python sequence, obj, smaller than NPY_MAXDIMS to a C-array of npy_intp. The Python
object could also be a single number. The seq variable is a pointer to a structure with members ptr and len.
On successful return, seq ->ptr contains a pointer to memory that must be freed to avoid a memory leak.
The restriction on memory size allows this converter to be conveniently used for sequences intended to be
interpreted as array shapes.
int PyArray_BufferConverter(PyObject* obj, PyArray_Chunk* buf )
Convert any Python object, obj, with a (single-segment) buffer interface to a variable with members that detail
the objects use of its chunk of memory. The buf variable is a pointer to a structure with base, ptr, len, and flags
members. The PyArray_Chunk structure is binary compatibile with the Pythons buffer object (through its
len member on 32-bit platforms and its ptr member on 64-bit platforms or in Python 2.5). On return, the base
member is set to obj (or its base if obj is already a buffer object pointing to another object). If you need to
hold on to the memory be sure to INCREF the base member. The chunk of memory is pointed to by buf ->ptr
member and has length buf ->len. The flags member of buf is NPY_BEHAVED_RO with the NPY_WRITEABLE
flag set if obj has a writeable buffer interface.
int PyArray_AxisConverter(PyObject * obj, int* axis)
Convert a Python object, obj, representing an axis argument to the proper value for passing to the functions that
take an integer axis. Specifically, if obj is None, axis is set to NPY_MAXDIMS which is interpreted correctly by
the C-API functions that take axis arguments.
int PyArray_BoolConverter(PyObject* obj, Bool* value)
Convert any Python object, obj, to NPY_TRUE or NPY_FALSE, and place the result in value.
int PyArray_ByteorderConverter(PyObject* obj, char* endian)
Convert Python strings into the corresponding byte-order character: >, <, s, =, or |.
int PyArray_SortkindConverter(PyObject* obj, NPY_SORTKIND* sort)
Convert Python strings into one of NPY_QUICKSORT (starts with q or Q) , NPY_HEAPSORT (starts with
h or H), or NPY_MERGESORT (starts with m or M).
int PyArray_SearchsideConverter(PyObject* obj, NPY_SEARCHSIDE* side)
Convert Python strings into one of NPY_SEARCHLEFT (starts with l or L), or NPY_SEARCHRIGHT (starts
with r or R).

Other conversions

int PyArray_PyIntAsInt(PyObject* op)

Convert all kinds of Python objects (including arrays and array scalars) to a standard integer. On error, -1 is
returned and an exception set. You may find useful the macro:

#define error_converting(x) (((x) == -1) && PyErr_Occurred()

npy_intp PyArray_PyIntAsIntp(PyObject* op)

Convert all kinds of Python objects (including arrays and array scalars) to a (platform-pointer-sized) integer.
On error, -1 is returned and an exception set.
int PyArray_IntpFromSequence(PyObject* seq, npy_intp* vals, int maxvals)
Convert any Python sequence (or single Python number) passed in as seq to (up to) maxvals pointer-sized
integers and place them in the vals array. The sequence can be smaller then maxvals as the number of converted
objects is returned.

1058 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

int PyArray_TypestrConvert(int itemsize, int gentype)

Convert typestring characters (with itemsize) to basic enumerated data types. The typestring character cor-
responding to signed and unsigned integers, floating point numbers, and complex-floating point numbers are
recognized and converted. Other values of gentype are returned. This function can be used to convert, for
example, the string f4 to NPY_FLOAT32.

5.4.13 Miscellaneous

Importing the API

In order to make use of the C-API from another extension module, the import_array () command must be used.
If the extension module is self-contained in a single .c file, then that is all that needs to be done. If, however, the
extension module involves multiple files where the C-API is needed then some additional steps must be taken.
void import_array(void)
This function must be called in the initialization section of a module that will make use of the C-API. It imports
the module where the function-pointer table is stored and points the correct variable to it.
PY_ARRAY_UNIQUE_SYMBOL

NO_IMPORT_ARRAY
Using these #defines you can use the C-API in multiple files for a single extension module. In each file you
must define PY_ARRAY_UNIQUE_SYMBOL to some name that will hold the C-API (e.g. myexten-
sion_ARRAY_API). This must be done before including the numpy/arrayobject.h file. In the module
intialization routine you call import_array (). In addition, in the files that do not have the module
initialization sub_routine define NO_IMPORT_ARRAY prior to including numpy/arrayobject.h.
Suppose I have two files coolmodule.c and coolhelper.c which need to be compiled and linked into a single
extension module. Suppose coolmodule.c contains the required initcool module initialization function (with the
import_array() function called). Then, coolmodule.c would have at the top:

#define PY_ARRAY_UNIQUE_SYMBOL cool_ARRAY_API

#include numpy/arrayobject.h

On the other hand, coolhelper.c would contain at the top:

#define PY_ARRAY_UNIQUE_SYMBOL cool_ARRAY_API

#define NO_IMPORT_ARRAY
#include numpy/arrayobject.h

Checking the API Version

Because python extensions are not used in the same way as usual libraries on most platforms, some errors cannot
be automatically detected at build time or even runtime. For example, if you build an extension using a function
available only for numpy >= 1.3.0, and you import the extension later with numpy 1.2, you will not get an import error
(but almost certainly a segmentation fault when calling the function). Thats why several functions are provided to
check for numpy versions. The macros NPY_VERSION and NPY_FEATURE_VERSION corresponds to the numpy
version used to build the extension, whereas the versions returned by the functions PyArray_GetNDArrayCVersion
and PyArray_GetNDArrayCFeatureVersion corresponds to the runtime numpys version.
The rules for ABI and API compatibilities can be summarized as follows:
Whenever NPY_VERSION != PyArray_GetNDArrayCVersion, the extension has to be recompiled (ABI incom-
patibility).

5.4. Array API 1059

NumPy Reference, Release 2.0.0.dev8464

NPY_VERSION == PyArray_GetNDArrayCVersion and NPY_FEATURE_VERSION <=

PyArray_GetNDArrayCFeatureVersion means backward compatible changes.
ABI incompatibility is automatically detected in every numpys version. API incompatibility detection was added in
numpy 1.4.0. If you want to supported many different numpy versions with one extension binary, you have to build
your extension with the lowest NPY_FEATURE_VERSION as possible.
unsigned int PyArray_GetNDArrayCVersion(void)
This just returns the value NPY_VERSION. NPY_VERSION changes whenever a backward incompatible
change at the ABI level. Because it is in the C-API, however, comparing the output of this function from the
value defined in the current header gives a way to test if the C-API has changed thus requiring a re-compilation
of extension modules that use the C-API. This is automatically checked in the function import_array.
unsigned int PyArray_GetNDArrayCFeatureVersion(void)
New in version 1.4.0. This just returns the value NPY_FEATURE_VERSION. NPY_FEATURE_VERSION
changes whenever the API changes (e.g. a function is added). A changed value does not always require a
recompile.

Internal Flexibility

int PyArray_SetNumericOps(PyObject* dict)

NumPy stores an internal table of Python callable objects that are used to implement arithmetic operations for
arrays as well as certain array calculation methods. This function allows the user to replace any or all of these
Python objects with their own versions. The keys of the dictionary, dict, are the named functions to replace and
the paired value is the Python callable object to use. Care should be taken that the function used to replace an
internal array operation does not itself call back to that internal array operation (unless you have designed the
function to handle that), or an unchecked infinite recursion can result (possibly causing program crash). The
key names that represent operations that can be replaced are:
add, subtract, multiply, divide, remainder, power, square, reciprocal, ones_like, sqrt, negative,
absolute, invert, left_shift, right_shift, bitwise_and, bitwise_xor, bitwise_or, less, less_equal,
equal, not_equal, greater, greater_equal, floor_divide, true_divide, logical_or, logical_and,
floor, ceil, maximum, minimum, rint.
These functions are included here because they are used at least once in the array objects methods. The function
returns -1 (without setting a Python Error) if one of the objects being assigned is not callable.
PyObject* PyArray_GetNumericOps(void)
Return a Python dictionary containing the callable Python objects stored in the the internal arithmetic operation
table. The keys of this dictionary are given in the explanation for PyArray_SetNumericOps.
void PyArray_SetStringFunction(PyObject* op, int repr)
This function allows you to alter the tp_str and tp_repr methods of the array object to any Python function. Thus
you can alter what happens for all arrays when str(arr) or repr(arr) is called from Python. The function to be
called is passed in as op. If repr is non-zero, then this function will be called in response to repr(arr), otherwise
the function will be called in response to str(arr). No check on whether or not op is callable is performed. The
callable passed in to op should expect an array argument and should return a string to be printed.

Memory management

char* PyDataMem_NEW(size_t nbytes)

PyDataMem_FREE(char* ptr)

1060 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

char* PyDataMem_RENEW(void * ptr, size_t newbytes)

Macros to allocate, free, and reallocate memory. These macros are used internally to create arrays.
npy_intp* PyDimMem_NEW(nd)

PyDimMem_FREE(npy_intp* ptr)

npy_intp* PyDimMem_RENEW(npy_intp* ptr, npy_intp newnd)

Macros to allocate, free, and reallocate dimension and strides memory.
PyArray_malloc(nbytes)

PyArray_free(ptr)

PyArray_realloc(ptr, nbytes)
These macros use different memory allocators, depending on the constant NPY_USE_PYMEM. The system
malloc is used when NPY_USE_PYMEM is 0, if NPY_USE_PYMEM is 1, then the Python memory allocator is
used.

Threading support

These macros are only meaningful if NPY_ALLOW_THREADS evaluates True during compilation of the extension
module. Otherwise, these macros are equivalent to whitespace. Python uses a single Global Interpreter Lock (GIL)
for each Python process so that only a single thread may excecute at a time (even on multi-cpu machines). When
calling out to a compiled function that may take time to compute (and does not have side-effects for other threads like
updated global variables), the GIL should be released so that other Python threads can run while the time-consuming
calculations are performed. This can be accomplished using two groups of macros. Typically, if one macro in a group
is used in a code block, all of them must be used in the same code block. Currently, NPY_ALLOW_THREADS is
defined to the python-defined WITH_THREADS constant unless the environment variable NPY_NOSMP is set in which
case NPY_ALLOW_THREADS is defined to be 0.

Group 1
This group is used to call code that may take some time but does not use any Python C-API calls. Thus,
the GIL should be released during its calculation.
NPY_BEGIN_ALLOW_THREADS
Equivalent to Py_BEGIN_ALLOW_THREADS except it uses NPY_ALLOW_THREADS to determine
if the macro if replaced with white-space or not.
NPY_END_ALLOW_THREADS
Equivalent to Py_END_ALLOW_THREADS except it uses NPY_ALLOW_THREADS to determine if
the macro if replaced with white-space or not.
NPY_BEGIN_THREADS_DEF
Place in the variable declaration area. This macro sets up the variable needed for storing the Python
state.
NPY_BEGIN_THREADS
Place right before code that does not need the Python interpreter (no Python C-API calls). This
macro saves the Python state and releases the GIL.
NPY_END_THREADS
Place right after code that does not need the Python interpreter. This macro acquires the GIL and
restores the Python state from the saved variable.

5.4. Array API 1061

NumPy Reference, Release 2.0.0.dev8464

NPY_BEGIN_THREADS_DESCR(PyArray_Descr *dtype)
Useful to release the GIL only if dtype does not contain arbitrary Python objects which may need
the Python interpreter during execution of the loop. Equivalent to
NPY_END_THREADS_DESCR(PyArray_Descr *dtype)
Useful to regain the GIL in situations where it was released using the BEGIN form of this macro.

Group 2
This group is used to re-acquire the Python GIL after it has been released. For example, suppose the
GIL has been released (using the previous calls), and then some path in the code (perhaps in a different
subroutine) requires use of the Python C-API, then these macros are useful to acquire the GIL. These
macros accomplish essentially a reverse of the previous three (acquire the LOCK saving what state it had)
and then re-release it with the saved state.
NPY_ALLOW_C_API_DEF
Place in the variable declaration area to set up the necessary variable.
NPY_ALLOW_C_API
Place before code that needs to call the Python C-API (when it is known that the GIL has already
been released).
NPY_DISABLE_C_API
Place after code that needs to call the Python C-API (to re-release the GIL).
Tip: Never use semicolons after the threading support macros.

Priority

NPY_PRIOIRTY
Default priority for arrays.
NPY_SUBTYPE_PRIORITY
Default subtype priority.
NPY_SCALAR_PRIORITY
Default scalar priority (very small)
double PyArray_GetPriority(PyObject* obj, double def )
Return the __array_priority__ attribute (converted to a double) of obj or def if no attribute of that name
exists. Fast returns that avoid the attribute lookup are provided for objects of type PyArray_Type.

Default buffers

NPY_BUFSIZE
Default size of the user-settable internal buffers.
NPY_MIN_BUFSIZE
Smallest size of user-settable internal buffers.
NPY_MAX_BUFSIZE
Largest size allowed for the user-settable buffers.

Other constants

NPY_NUM_FLOATTYPE
The number of floating-point types

1062 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

NPY_MAXDIMS
The maximum number of dimensions allowed in arrays.
NPY_VERSION
The current version of the ndarray object (check to see if this variable is defined to guarantee the
numpy/arrayobject.h header is being used).
NPY_FALSE
Defined as 0 for use with Bool.
NPY_TRUE
Defined as 1 for use with Bool.
NPY_FAIL
The return value of failed converter functions which are called using the O& syntax in
PyArg_ParseTuple-like functions.
NPY_SUCCEED
The return value of successful converter functions which are called using the O& syntax in
PyArg_ParseTuple-like functions.

Miscellaneous Macros

PyArray_SAMESHAPE(a1, a2)
Evaluates as True if arrays a1 and a2 have the same shape.
PyArray_MAX(a, b)
Returns the maximum of a and b. If (a) or (b) are expressions they are evaluated twice.
PyArray_MIN(a, b)
Returns the minimum of a and b. If (a) or (b) are expressions they are evaluated twice.
PyArray_CLT(a, b)

PyArray_CGT(a, b)

PyArray_CLE(a, b)

PyArray_CGE(a, b)

PyArray_CEQ(a, b)

PyArray_CNE(a, b)
Implements the complex comparisons between two complex numbers (structures with a real and imag member)
using NumPys definition of the ordering which is lexicographic: comparing the real parts first and then the
complex parts if the real parts are equal.
PyArray_REFCOUNT(PyObject* op)
Returns the reference count of any Python object.
PyArray_XDECREF_ERR(PyObject *obj)
DECREFs an array object which may have the NPY_UPDATEIFCOPY flag set without causing the contents to
be copied back into the original array. Resets the NPY_WRITEABLE flag on the base object. This is useful for
recovering from an error condition when NPY_UPDATEIFCOPY is used.

5.4. Array API 1063

NumPy Reference, Release 2.0.0.dev8464

Enumerated Types

NPY_SORTKIND
A special variable-type which can take on the values NPY_{KIND} where {KIND} is
QUICKSORT, HEAPSORT, MERGESORT
NPY_NSORTS
Defined to be the number of sorts.
NPY_SCALARKIND
A special variable type indicating the number of kinds of scalars distinguished in determining scalar-coercion
rules. This variable can take on the values NPY_{KIND} where {KIND} can be
NOSCALAR, BOOL_SCALAR, INTPOS_SCALAR, INTNEG_SCALAR, FLOAT_SCALAR,
COMPLEX_SCALAR, OBJECT_SCALAR
NPY_NSCALARKINDS
Defined to be the number of scalar kinds (not including NPY_NOSCALAR).
NPY_ORDER
A variable type indicating the order that an array should be interpreted in. The value of a variable of this type
can be NPY_{ORDER} where {ORDER} is
ANYORDER, CORDER, FORTRANORDER
NPY_CLIPMODE
A variable type indicating the kind of clipping that should be applied in certain functions. The value of a
variable of this type can be NPY_{MODE} where {MODE} is
CLIP, WRAP, RAISE

5.5 UFunc API

5.5.1 Constants

UFUNC_ERR_{HANDLER}
{HANDLER} can be IGNORE, WARN, RAISE, or CALL
UFUNC_{THING}_{ERR}
{THING} can be MASK, SHIFT, or FPE, and {ERR} can be DIVIDEBYZERO, OVERFLOW, UNDER-
FLOW, and INVALID.
PyUFunc_{VALUE}
{VALUE} can be One (1), Zero (0), or None (-1)

5.5.2 Macros

NPY_LOOP_BEGIN_THREADS
Used in universal function code to only release the Python GIL if loop->obj is not true (i.e. this is not an
OBJECT array loop). Requires use of NPY_BEGIN_THREADS_DEF in variable declaration area.
NPY_LOOP_END_THREADS
Used in universal function code to re-acquire the Python GIL if it was released (because loop->obj was not
true).

1064 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

UFUNC_CHECK_ERROR(loop)
A macro used internally to check for errors and goto fail if found. This macro requires a fail label in the current
code block. The loop variable must have at least members (obj, errormask, and errorobj). If loop ->obj is
nonzero, then PyErr_Occurred () is called (meaning the GIL must be held). If loop ->obj is zero, then if
loop ->errormask is nonzero, PyUFunc_checkfperr is called with arguments loop ->errormask and loop
->errobj. If the result of this check of the IEEE floating point registers is true then the code redirects to the fail
label which must be defined.
UFUNC_CHECK_STATUS(ret)
A macro that expands to platform-dependent code. The ret variable can can be any integer. The
UFUNC_FPE_{ERR} bits are set in ret according to the status of the corresponding error flags of the
floating point processor.

5.5.3 Functions

PyObject* PyUFunc_FromFuncAndData(PyUFuncGenericFunction* func,()

void** data, char* types, int ntypes, int nin, int nout, int identity,()

char* name, char* doc, int check_return)()

Create a new broadcasting universal function from required variables. Each ufunc builds around
the notion of an element-by-element operation. Each ufunc object contains pointers to 1-d loops
implementing the basic functionality for each supported type.
Note: The func, data, types, name, and doc arguments are not copied by
PyUFunc_FromFuncAndData. The caller must ensure that the memory used by these
arrays is not freed as long as the ufunc object is alive.
param func
Must to an array of length ntypes containing PyUFuncGenericFunction items.
These items are pointers to functions that actually implement the underlying (element-
by-element) function N times.
param data
Should be NULL or a pointer to an array of size ntypes . This array may contain arbitrary
extra-data to be passed to the corresponding 1-d loop function in the func array.
param types
Must be of length (nin + nout) * ntypes, and it contains the data-types (built-in only)
that the corresponding function in the func array can deal with.
param ntypes
How many different data-type signatures the ufunc has implemented.
param nin
The number of inputs to this operation.
param nout
The number of outputs
param name
The name for the ufunc. Specifying a name of add or multiply enables a special
behavior for integer-typed reductions when no dtype is given. If the input type is an
integer (or boolean) data type smaller than the size of the int_ data type, it will be
internally upcast to the int_ (or uint) data type.

5.5. UFunc API 1065

NumPy Reference, Release 2.0.0.dev8464

param doc
Allows passing in a documentation string to be stored with the ufunc. The documen-
tation string should not contain the name of the function or the calling signature as
that will be dynamically determined from the object and available when accessing the
__doc__ attribute of the ufunc.
param check_return
Unused and present for backwards compatibility of the C-API. A corresponding
check_return integer does exist in the ufunc structure and it does get set with this value
when the ufunc object is created.
int PyUFunc_RegisterLoopForType(PyUFuncObject* ufunc,()

int usertype, PyUFuncGenericFunction function, int* arg_types, void* data)()

This function allows the user to register a 1-d loop with an already- created ufunc to be used whenever
the ufunc is called with any of its input arguments as the user-defined data-type. This is needed
in order to make ufuncs work with built-in data-types. The data-type must have been previously
registered with the numpy system. The loop is passed in as function. This loop can take arbitrary
data which should be passed in as data. The data-types the loop requires are passed in as arg_types
which must be a pointer to memory at least as large as ufunc->nargs.
int PyUFunc_ReplaceLoopBySignature(PyUFuncObject* ufunc,()

PyUFuncGenericFunction newfunc, int* signature,()

PyUFuncGenericFunction* oldfunc)()

Replace a 1-d loop matching the given signature in the already-created ufunc with the new 1-d loop
newfunc. Return the old 1-d loop function in oldfunc. Return 0 on success and -1 on failure. This
function works only with built-in types (use PyUFunc_RegisterLoopForType for user-defined
types). A signature is an array of data-type numbers indicating the inputs followed by the outputs
assumed by the 1-d loop.
int PyUFunc_GenericFunction(PyUFuncObject* self,()

PyObject* args, PyArrayObject** mps)()

A generic ufunc call. The ufunc is passed in as self, the arguments to the ufunc as args. The mps
argument is an array of PyArrayObject pointers containing the converted input arguments as
well as the ufunc outputs on return. The user is responsible for managing this array and receives a
new reference for each array in mps. The total number of arrays in mps is given by self ->nin + self
->nout.
int PyUFunc_checkfperr(int errmask, PyObject* errobj)
A simple interface to the IEEE error-flag checking support. The errmask argument is a mask of
UFUNC_MASK_{ERR} bitmasks indicating which errors to check for (and how to check for them). The
errobj must be a Python tuple with two elements: a string containing the name which will be used in any
communication of error and either a callable Python object (call-back function) or Py_None. The callable
object will only be used if UFUNC_ERR_CALL is set as the desired error checking method. This routine
manages the GIL and is safe to call even after releasing the GIL. If an error in the IEEE-compatibile hardware
is determined a -1 is returned, otherwise a 0 is returned.
void PyUFunc_clearfperr()
Clear the IEEE error flags.

1066 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

void PyUFunc_GetPyValues(char* name, int* bufsize,()

int* errmask, PyObject** errobj)()

Get the Python values used for ufunc processing from the thread-local storage area unless the defaults
have been set in which case the name lookup is bypassed. The name is placed as a string in the first
element of *errobj. The second element is the looked-up function to call on error callback. The value
of the looked-up buffer-size to use is passed into bufsize, and the value of the error mask is placed
into errmask.

5.5.4 Generic functions

At the core of every ufunc is a collection of type-specific functions that defines the basic functionality for each of the
supported types. These functions must evaluate the underlying function N 1 times. Extra-data may be passed in
that may be used during the calculation. This feature allows some general functions to be used as these basic looping
functions. The general function has all the code needed to point variables to the right place and set up a function call.
The general function assumes that the actual function to call is passed in as the extra data and calls it with the correct
values. All of these functions are suitable for placing directly in the array of functions stored in the functions member
of the PyUFuncObject structure.
void PyUFunc_f_f_As_d_d(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_d_d(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_f_f(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_g_g(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_F_F_As_D_D(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_F_F(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_D_D(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_G_G(char** args, npy_intp* dimensions,()

5.5. UFunc API 1067

NumPy Reference, Release 2.0.0.dev8464

npy_intp* steps, void* func)()

Type specific, core 1-d functions for ufuncs where each calculation is obtained by calling a function
taking one input argument and returning one output. This function is passed in func. The letters
correspond to dtypechars of the supported data types ( f - float, d - double, g - long double, F - cfloat,
D - cdouble, G - clongdouble). The argument func must support the same signature. The _As_X_X
variants assume ndarrays of one data type but cast the values to use an underlying function that takes
a different data type. Thus, PyUFunc_f_f_As_d_d uses ndarrays of data type NPY_FLOAT but
calls out to a C-function that takes double and returns double.
void PyUFunc_ff_f_As_dd_d(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_ff_f(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_dd_d(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_gg_g(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_FF_F_As_DD_D(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_DD_D(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_FF_F(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_GG_G(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

Type specific, core 1-d functions for ufuncs where each calculation is obtained by calling a function
taking two input arguments and returning one output. The underlying function to call is passed in as
func. The letters correspond to dtypechars of the specific data type supported by the general-purpose
function. The argument func must support the corresponding signature. The _As_XX_X variants
assume ndarrays of one data type but cast the values at each iteration of the loop to use the underlying
function that takes a different data type.

1068 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

void PyUFunc_O_O(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

void PyUFunc_OO_O(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

One-input, one-output, and two-input, one-output core 1-d functions for the NPY_OBJECT data
type. These functions handle reference count issues and return early on error. The actual func-
tion to call is func and it must accept calls with the signature (PyObject*) (PyObject*) for
PyUFunc_O_O or (PyObject*)(PyObject *, PyObject *) for PyUFunc_OO_O.
void PyUFunc_O_O_method(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

This general purpose 1-d core function assumes that func is a string representing a method of the
input object. For each iteration of the loop, the Python obejct is extracted from the array and its func
method is called returning the result to the output array.
void PyUFunc_OO_O_method(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

This general purpose 1-d core function assumes that func is a string representing a method of the input
object that takes one argument. The first argument in args is the method whose function is called, the
second argument in args is the argument passed to the function. The output of the function is stored
in the third entry of args.
void PyUFunc_On_Om(char** args, npy_intp* dimensions,()

npy_intp* steps, void* func)()

This is the 1-d core function used by the dynamic ufuncs created by umath.frompyfunc(function, nin,
nout). In this case func is a pointer to a PyUFunc_PyFuncData structure which has definition
PyUFunc_PyFuncData

typedef struct {
int nin;
int nout;
PyObject *callable;
} PyUFunc_PyFuncData;

At each iteration of the loop, the nin input objects are exctracted from their object arrays and placed
into an argument tuple, the Python callable is called with the input arguments, and the nout outputs
are placed into their object arrays.

5.5. UFunc API 1069

NumPy Reference, Release 2.0.0.dev8464

5.5.5 Importing the API

PY_UFUNC_UNIQUE_SYMBOL

NO_IMPORT_UFUNC

void import_ufunc(void)
These are the constants and functions for accessing the ufunc C-API from extension modules in precisely
the same way as the array C-API can be accessed. The import_ufunc () function must always be called
(in the initialization subroutine of the extension module). If your extension module is in one file then that is
all that is required. The other two constants are useful if your extension module makes use of multiple files.
In that case, define PY_UFUNC_UNIQUE_SYMBOL to something unique to your code and then in source
files that do not contain the module initialization function but still need access to the UFUNC API, define
PY_UFUNC_UNIQUE_SYMBOL to the same name used previously and also define NO_IMPORT_UFUNC.
The C-API is actually an array of function pointers. This array is created (and pointed to by a global variable)
by import_ufunc. The global variable is either statically defined or allowed to be seen by other files depending
on the state of Py_UFUNC_UNIQUE_SYMBOL and NO_IMPORT_UFUNC.

5.6 Generalized Universal Function API

There is a general need for looping over not only functions on scalars but also over functions on vectors (or arrays),
as explained on http://scipy.org/scipy/numpy/wiki/GeneralLoopingFunctions. We propose to realize this concept by
generalizing the universal functions (ufuncs), and provide a C implementation that adds ~500 lines to the numpy code
base. In current (specialized) ufuncs, the elementary function is limited to element-by-element operations, whereas the
generalized version supports sub-array by sub-array operations. The Perl vector library PDL provides a similar
functionality and its terms are re-used in the following.
Each generalized ufunc has information associated with it that states what the core dimensionality of the inputs is,
as well as the corresponding dimensionality of the outputs (the element-wise ufuncs have zero core dimensions). The
list of the core dimensions for all arguments is called the signature of a ufunc. For example, the ufunc numpy.add
has signature (),()->() defining two scalar inputs and one scalar output.
Another example is (see the GeneralLoopingFunctions page) the function inner1d(a,b) with a signature of
(i),(i)->(). This applies the inner product along the last axis of each input, but keeps the remaining indices
intact. For example, where a is of shape (3,5,N) and b is of shape (5,N), this will return an output of shape
(3,5). The underlying elementary function is called 3*5 times. In the signature, we specify one core dimension (i)
for each input and zero core dimensions () for the output, since it takes two 1-d arrays and returns a scalar. By using
the same name i, we specify that the two corresponding dimensions should be of the same size (or one of them is of
size 1 and will be broadcasted).
The dimensions beyond the core dimensions are called loop dimensions. In the above example, this corresponds to
(3,5).
The usual numpy broadcasting rules apply, where the signature determines how the dimensions of each input/output
object are split into core and loop dimensions:
1. While an input array has a smaller dimensionality than the corresponding number of core dimensions, 1s are
pre-pended to its shape.
2. The core dimensions are removed from all inputs and the remaining dimensions are broadcasted; defining the
loop dimensions.
3. The output is given by the loop dimensions plus the output core dimensions.

1070 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

5.6.1 Definitions

Elementary Function
Each ufunc consists of an elementary function that performs the most basic operation on the smallest portion of
array arguments (e.g. adding two numbers is the most basic operation in adding two arrays). The ufunc applies
the elementary function multiple times on different parts of the arrays. The input/output of elementary functions
can be vectors; e.g., the elementary function of inner1d takes two vectors as input.
Signature
A signature is a string describing the input/output dimensions of the elementary function of a ufunc. See section
below for more details.
Core Dimension
The dimensionality of each input/output of an elementary function is defined by its core dimensions (zero core
dimensions correspond to a scalar input/output). The core dimensions are mapped to the last dimensions of the
input/output arrays.
Dimension Name
A dimension name represents a core dimension in the signature. Different dimensions may share a name,
indicating that they are of the same size (or are broadcastable).
Dimension Index
A dimension index is an integer representing a dimension name. It enumerates the dimension names according
to the order of the first occurrence of each name in the signature.

5.6.2 Details of Signature

The signature defines core dimensionality of input and output variables, and thereby also defines the contraction of
the dimensions. The signature is represented by a string of the following format:
Core dimensions of each input or output array are represented by a list of dimension names in parentheses,
(i_1,...,i_N); a scalar input/output is denoted by (). Instead of i_1, i_2, etc, one can use any valid
Python variable name.
Dimension lists for different arguments are separated by ",". Input/output arguments are separated by "->".
If one uses the same dimension name in multiple locations, this enforces the same size (or broadcastable size)
of the corresponding dimensions.
The formal syntax of signatures is as follows:

<Signature> ::= <Input arguments> "->" <Output arguments>

<Input arguments> ::= <Argument list>
<Output arguments> ::= <Argument list>
<Argument list> ::= nil | <Argument> | <Argument> "," <Argument list>
<Argument> ::= "(" <Core dimension list> ")"
<Core dimension list> ::= nil | <Dimension name> |
<Dimension name> "," <Core dimension list>
<Dimension name> ::= valid Python variable name

Notes:
1. All quotes are for clarity.
2. Core dimensions that share the same name must be broadcastable, as the two i in our example above. Each
dimension name typically corresponding to one level of looping in the elementary functions implementation.
3. White spaces are ignored.

5.6. Generalized Universal Function API 1071

NumPy Reference, Release 2.0.0.dev8464

Here are some examples of signatures:

add (),()->()
inner1d (i),(i)->()
sum1d (i)->()
dot2d matrix multiplication
(m,n),(n,p)->(m,p)
outer_inner inner over the last dimension, outer over the second to last, and
(i,t),(j,t)->(i,j)
loop/broadcast over the rest.

5.6.3 C-API for implementing Elementary Functions

The current interface remains unchanged, and PyUFunc_FromFuncAndData can still be used to implement (spe-
cialized) ufuncs, consisting of scalar elementary functions.
One can use PyUFunc_FromFuncAndDataAndSignature to declare a more general ufunc. The argument list
is the same as PyUFunc_FromFuncAndData, with an additional argument specifying the signature as C string.
Furthermore, the callback function is of the same type as before, void (*foo)(char **args, intp
*dimensions, intp *steps, void *func). When invoked, args is a list of length nargs containing
the data of all input/output arguments. For a scalar elementary function, steps is also of length nargs, denoting the
strides used for the arguments. dimensions is a pointer to a single integer defining the size of the axis to be looped
over.
For a non-trivial signature, dimensions will also contain the sizes of the core dimensions as well, starting at the
second entry. Only one size is provided for each unique dimension name and the sizes are given according to the first
occurrence of a dimension name in the signature.
The first nargs elements of steps remain the same as for scalar ufuncs. The following elements contain the strides
of all core dimensions for all arguments in order.
For example, consider a ufunc with signature (i,j),(i)->(). In this case, args will contain three pointers to the
data of the input/output arrays a, b, c. Furthermore, dimensions will be [N, I, J] to define the size of N of
the loop and the sizes I and J for the core dimensions i and j. Finally, steps will be [a_N, b_N, c_N, a_i,
a_j, b_i], containing all necessary strides.

5.7 Numpy core libraries

New in version 1.3.0. Starting from numpy 1.3.0, we are working on separating the pure C, computational code
from the python dependent code. The goal is twofolds: making the code cleaner, and enabling code reuse by other
extensions outside numpy (scipy, etc...).

5.7.1 Numpy core math library

The numpy core math library (npymath) is a first step in this direction. This library contains most math-related C99
functionality, which can be used on platforms where C99 is not well supported. The core math functions have the
same API as the C99 ones, except for the npy_* prefix.
The available functions are defined in npy_math.h - please refer to this header in doubt.

Floating point classification

NPY_NAN
This macro is defined to a NaN (Not a Number), and is guaranteed to have the signbit unset (positive NaN).

1072 Chapter 5. Numpy C-API

 NumPy Reference, Release 2.0.0.dev8464

The corresponding single and extension precision macro are available with the suffix F and L.
NPY_INFINITY
This macro is defined to a positive inf. The corresponding single and extension precision macro are available
with the suffix F and L.
NPY_PZERO
This macro is defined to positive zero. The corresponding single and extension precision macro are available
with the suffix F and L.
NPY_NZERO
This macro is defined to negative zero (that is with the sign bit set). The corresponding single and extension
precision macro are available with the suffix F and L.
int npy_isnan(x)
This is a macro, and is equivalent to C99 isnan: works for single, double and extended precision, and return a
non 0 value is x is a NaN.
int npy_isfinite(x)
This is a macro, and is equivalent to C99 isfinite: works for single, double and extended precision, and return a
non 0 value is x is neither a NaN or a infinity.
int npy_isinf(x)
This is a macro, and is equivalent to C99 isinf: works for single, double and extended precision, and return a
non 0 value is x is infinite (positive and negative).
int npy_signbit(x)
This is a macro, and is equivalent to C99 signbit: works for single, double and extended precision, and return a
non 0 value is x has the signbit set (that is the number is negative).
double npy_copysign(double x, double y)
This is a function equivalent to C99 copysign: return x with the same sign as y. Works for any value, including
inf and nan. Single and extended precisions are available with suffix f and l. New in version 1.4.0.

Useful math constants

The following math constants are available in npy_math.h. Single and extended precision are also available by adding
the F and L suffixes respectively.
NPY_E
Base of natural logarithm (e)
NPY_LOG2E
ln(e)
Logarithm to base 2 of the Euler constant ( ln(2) )
NPY_LOG10E
ln(e)
Logarithm to base 10 of the Euler constant ( ln(10) )
NPY_LOGE2
Natural logarithm of 2 (ln(2))
NPY_LOGE10
Natural logarithm of 10 (ln(10))
NPY_PI
Pi ()
NPY_PI_2
Pi divided by 2 ( 2 )

5.7. Numpy core libraries 1073

NumPy Reference, Release 2.0.0.dev8464

NPY_PI_4
Pi divided by 4 ( 4 )
NPY_1_PI
Reciprocal of pi ( 1 )
NPY_2_PI
Two times the reciprocal of pi ( 2 )
NPY_EULER

The Euler constant

Pn 1
limn ( k=1 k ln n)

Low-level floating point manipulation

Those can be useful for precise floating point comparison.

double npy_nextafter(double x, double y)
This is a function equivalent to C99 nextafter: return next representable floating point value from x in the
direction of y. Single and extended precisions are available with suffix f and l. New in version 1.4.0.
double npy_spacing(double x)
This is a function equivalent to Fortran intrinsic. Return distance between x and next representable floating
point value from x, e.g. spacing(1) == eps. spacing of nan and +/- inf return nan. Single and extended precisions
are available with suffix f and l. New in version 1.4.0.

Complex functions

New in version 1.4.0. C99-like complex functions have been added. Those can be used if you wish to implement
portable C extensions. Since we still support platforms without C99 complex type, you need to restrict to C90-
compatible syntax, e.g.:

/* a = 1 + 2i \*/
npy_complex a = npy_cpack(1, 2);
npy_complex b;

b = npy_log(a);

Linking against the core math library in an extension

New in version 1.4.0. To use the core math library in your own extension, you need to add the npymath compile and
link options to your extension in your setup.py:

>>> from numpy.distutils.misc_utils import get_info

>>> info = get_info(npymath)
>>> config.add_extension(foo, sources=[foo.c], extra_info=**info)

In other words, the usage of info is exactly the same as when using blas_info and co.

1074 Chapter 5. Numpy C-API

 CHAPTER

SIX

NUMPY INTERNALS

6.1 Numpy C Code Explanations

Fanaticism consists of redoubling your efforts when you have forgotten your aim. George Santayana
An authority is a person who can tell you more about something than you really care to know. Unknown
This Chapter attempts to explain the logic behind some of the new pieces of code. The purpose behind these explana-
tions is to enable somebody to be able to understand the ideas behind the implementation somewhat more easily than
just staring at the code. Perhaps in this way, the algorithms can be improved on, borrowed from, and/or optimized.

6.1.1 Memory model

One fundamental aspect of the ndarray is that an array is seen as a chunk of memory starting at some location. The
interpretation of this memory depends on the stride information. For each dimension in an N -dimensional array, an
integer (stride) dictates how many bytes must be skipped to get to the next element in that dimension. Unless you have
a single-segment array, this stride information must be consulted when traversing through an array. It is not difficult
to write code that accepts strides, you just have to use (char *) pointers because strides are in units of bytes. Keep
in mind also that strides do not have to be unit-multiples of the element size. Also, remember that if the number of
dimensions of the array is 0 (sometimes called a rank-0 array), then the strides and dimensions variables are NULL.
Besides the structural information contained in the strides and dimensions members of the PyArrayObject, the
flags contain important information about how the data may be accessed. In particular, the NPY_ALIGNED flag is set
when the memory is on a suitable boundary according to the data-type array. Even if you have a contiguous chunk
of memory, you cannot just assume it is safe to dereference a data- type-specific pointer to an element. Only if the
NPY_ALIGNED flag is set is this a safe operation (on some platforms it will work but on others, like Solaris, it will
cause a bus error). The NPY_WRITEABLE should also be ensured if you plan on writing to the memory area of
the array. It is also possible to obtain a pointer to an unwriteable memory area. Sometimes, writing to the memory
area when the NPY_WRITEABLE flag is not set will just be rude. Other times it can cause program crashes ( e.g. a
data-area that is a read-only memory-mapped file).

6.1.2 Data-type encapsulation

The data-type is an important abstraction of the ndarray. Operations will look to the data-type to provide the key
functionality that is needed to operate on the array. This functionality is provided in the list of function pointers
pointed to by the f member of the PyArray_Descr structure. In this way, the number of data-types can be
extended simply by providing a PyArray_Descr structure with suitable function pointers in the f member. For
built-in types there are some optimizations that by-pass this mechanism, but the point of the data- type abstraction is
to allow new data-types to be added.

1075
NumPy Reference, Release 2.0.0.dev8464

One of the built-in data-types, the void data-type allows for arbitrary records containing 1 or more fields as elements
of the array. A field is simply another data-type object along with an offset into the current record. In order to
support arbitrarily nested fields, several recursive implementations of data-type access are implemented for the void
type. A common idiom is to cycle through the elements of the dictionary and perform a specific operation based on
the data-type object stored at the given offset. These offsets can be arbitrary numbers. Therefore, the possibility of
encountering mis- aligned data must be recognized and taken into account if necessary.

6.1.3 N-D Iterators

A very common operation in much of NumPy code is the need to iterate over all the elements of a general, strided,
N-dimensional array. This operation of a general-purpose N-dimensional loop is abstracted in the notion of an iterator
object. To write an N-dimensional loop, you only have to create an iterator object from an ndarray, work with the
dataptr member of the iterator object structure and call the macro PyArray_ITER_NEXT (it) on the iterator object
to move to the next element. The next element is always in C-contiguous order. The macro works by first special
casing the C-contiguous, 1-D, and 2-D cases which work very simply.
For the general case, the iteration works by keeping track of a list of coordinate counters in the iterator object. At each
iteration, the last coordinate counter is increased (starting from 0). If this counter is smaller then one less than the size
of the array in that dimension (a pre-computed and stored value), then the counter is increased and the dataptr member
is increased by the strides in that dimension and the macro ends. If the end of a dimension is reached, the counter
for the last dimension is reset to zero and the dataptr is moved back to the beginning of that dimension by subtracting
the strides value times one less than the number of elements in that dimension (this is also pre-computed and stored
in the backstrides member of the iterator object). In this case, the macro does not end, but a local dimension counter
is decremented so that the next-to-last dimension replaces the role that the last dimension played and the previously-
described tests are executed again on the next-to-last dimension. In this way, the dataptr is adjusted appropriately for
arbitrary striding.
The coordinates member of the PyArrayIterObject structure maintains the current N-d counter unless
the underlying array is C-contiguous in which case the coordinate counting is by-passed. The index mem-
ber of the PyArrayIterObject keeps track of the current flat index of the iterator. It is updated by the
PyArray_ITER_NEXT macro.

6.1.4 Broadcasting

In Numeric, broadcasting was implemented in several lines of code buried deep in ufuncobject.c. In NumPy, the
notion of broadcasting has been abstracted so that it can be performed in multiple places. Broadcasting is handled by
the function PyArray_Broadcast. This function requires a PyArrayMultiIterObject (or something that
is a binary equivalent) to be passed in. The PyArrayMultiIterObject keeps track of the broadcasted number
of dimensions and size in each dimension along with the total size of the broadcasted result. It also keeps track of the
number of arrays being broadcast and a pointer to an iterator for each of the arrays being broadcasted.
The PyArray_Broadcast function takes the iterators that have already been defined and uses them to determine
the broadcast shape in each dimension (to create the iterators at the same time that broadcasting occurs then use the
PyMultiIter_New function). Then, the iterators are adjusted so that each iterator thinks it is iterating over an
array with the broadcasted size. This is done by adjusting the iterators number of dimensions, and the shape in each
dimension. This works because the iterator strides are also adjusted. Broadcasting only adjusts (or adds) length-1
dimensions. For these dimensions, the strides variable is simply set to 0 so that the data-pointer for the iterator over
that array doesnt move as the broadcasting operation operates over the extended dimension.
Broadcasting was always implemented in Numeric using 0-valued strides for the extended dimensions. It
is done in exactly the same way in NumPy. The big difference is that now the array of strides is kept
track of in a PyArrayIterObject, the iterators involved in a broadcasted result are kept track of in a
PyArrayMultiIterObject, and the PyArray_BroadCast call implements the broad-casting rules.

1076 Chapter 6. Numpy internals

 NumPy Reference, Release 2.0.0.dev8464

6.1.5 Array Scalars

The array scalars offer a hierarchy of Python types that allow a one- to-one correspondence between the data-type
stored in an array and the Python-type that is returned when an element is extracted from the array. An exception to
this rule was made with object arrays. Object arrays are heterogeneous collections of arbitrary Python objects. When
you select an item from an object array, you get back the original Python object (and not an object array scalar which
does exist but is rarely used for practical purposes).
The array scalars also offer the same methods and attributes as arrays with the intent that the same code can be
used to support arbitrary dimensions (including 0-dimensions). The array scalars are read-only (immutable) with
the exception of the void scalar which can also be written to so that record-array field setting works more naturally
(a[0][f1] = value ).

6.1.6 Advanced (Fancy) Indexing

The implementation of advanced indexing represents some of the most difficult code to write and explain. In fact, there
are two implementations of advanced indexing. The first works only with 1-D arrays and is implemented to handle
expressions involving a.flat[obj]. The second is general-purpose that works for arrays of arbitrary dimension (up to
a fixed maximum). The one-dimensional indexing approaches were implemented in a rather straightforward fashion,
and so it is the general-purpose indexing code that will be the focus of this section.
There is a multi-layer approach to indexing because the indexing code can at times return an array scalar and at other
times return an array. The functions with _nice appended to their name do this special handling while the function
without the _nice appendage always return an array (perhaps a 0-dimensional array). Some special-case optimizations
(the index being an integer scalar, and the index being a tuple with as many dimensions as the array) are handled in
array_subscript_nice function which is what Python calls when presented with the code a[obj]. These optimizations
allow fast single-integer indexing, and also ensure that a 0-dimensional array is not created only to be discarded as the
array scalar is returned instead. This provides significant speed-up for code that is selecting many scalars out of an
array (such as in a loop). However, it is still not faster than simply using a list to store standard Python scalars, because
that is optimized by the Python interpreter itself.
After these optimizations, the array_subscript function itself is called. This function first checks for field selection
which occurs when a string is passed as the indexing object. Then, 0-D arrays are given special-case consideration.
Finally, the code determines whether or not advanced, or fancy, indexing needs to be performed. If fancy indexing
is not needed, then standard view-based indexing is performed using code borrowed from Numeric which parses the
indexing object and returns the offset into the data-buffer and the dimensions necessary to create a new view of the
array. The strides are also changed by multiplying each stride by the step-size requested along the corresponding
dimension.

Fancy-indexing check

The fancy_indexing_check routine determines whether or not to use standard view-based indexing or new copy-based
indexing. If the indexing object is a tuple, then view-based indexing is assumed by default. Only if the tuple contains
an array object or a sequence object is fancy-indexing assumed. If the indexing object is an array, then fancy indexing
is automatically assumed. If the indexing object is any other kind of sequence, then fancy-indexing is assumed by
default. This is over-ridden to simple indexing if the sequence contains any slice, newaxis, or Ellipsis objects, and no
arrays or additional sequences are also contained in the sequence. The purpose of this is to allow the construction of
slicing sequences which is a common technique for building up code that works in arbitrary numbers of dimensions.

Fancy-indexing implementation

The concept of indexing was also abstracted using the idea of an iterator. If fancy indexing is performed, then a
PyArrayMapIterObject is created. This internal object is not exposed to Python. It is created in order to handle

6.1. Numpy C Code Explanations 1077

NumPy Reference, Release 2.0.0.dev8464

the fancy-indexing at a high-level. Both get and set fancy-indexing operations are implemented using this object. Fancy
indexing is abstracted into three separate operations: (1) creating the PyArrayMapIterObject from the indexing
object, (2) binding the PyArrayMapIterObject to the array being indexed, and (3) getting (or setting) the items
determined by the indexing object. There is an optimization implemented so that the PyArrayIterObject (which
has its own less complicated fancy-indexing) is used for indexing when possible.

Creating the mapping object

The first step is to convert the indexing objects into a standard form where iterators are created for all of the index
array inputs and all Boolean arrays are converted to equivalent integer index arrays (as if nonzero(arr) had been called).
Finally, all integer arrays are replaced with the integer 0 in the indexing object and all of the index-array iterators are
broadcast to the same shape.

Binding the mapping object

When the mapping object is created it does not know which array it will be used with so once the index iterators are
constructed during mapping-object creation, the next step is to associate these iterators with a particular ndarray. This
process interprets any ellipsis and slice objects so that the index arrays are associated with the appropriate axis (the
axis indicated by the iteraxis entry corresponding to the iterator for the integer index array). This information is then
used to check the indices to be sure they are within range of the shape of the array being indexed. The presence of
ellipsis and/or slice objects implies a sub-space iteration that is accomplished by extracting a sub-space view of the
array (using the index object resulting from replacing all the integer index arrays with 0) and storing the information
about where this sub-space starts in the mapping object. This is used later during mapping-object iteration to select
the correct elements from the underlying array.

Getting (or Setting)

After the mapping object is successfully bound to a particular array, the mapping object contains the shape of the
resulting item as well as iterator objects that will walk through the currently-bound array and either get or set its
elements as needed. The walk is implemented using the PyArray_MapIterNext function. This function sets
the coordinates of an iterator object into the current array to be the next coordinate location indicated by all of the
indexing-object iterators while adjusting, if necessary, for the presence of a sub- space. The result of this function is
that the dataptr member of the mapping object structure is pointed to the next position in the array that needs to be
copied out or set to some value.
When advanced indexing is used to extract an array, an iterator for the new array is constructed and advanced in phase
with the mapping object iterator. When advanced indexing is used to place values in an array, a special broadcasted
iterator is constructed from the object being placed into the array so that it will only work if the values used for setting
have a shape that is broadcastable to the shape implied by the indexing object.

6.1.7 Universal Functions

Universal functions are callable objects that take N inputs and produce M outputs by wrapping basic 1-D loops that
work element-by-element into full easy-to use functions that seamlessly implement broadcasting, type-checking and
buffered coercion, and output-argument handling. New universal functions are normally created in C, although there
is a mechanism for creating ufuncs from Python functions (frompyfunc). The user must supply a 1-D loop that
implements the basic function taking the input scalar values and placing the resulting scalars into the appropriate
output slots as explaine n implementation.

Setup

Every ufunc calculation involves some overhead related to setting up the calculation. The practical significance of
this overhead is that even though the actual calculation of the ufunc is very fast, you will be able to write array
and type-specific code that will work faster for small arrays than the ufunc. In particular, using ufuncs to perform
many calculations on 0-D arrays will be slower than other Python-based solutions (the silently-imported scalarmath

1078 Chapter 6. Numpy internals

 NumPy Reference, Release 2.0.0.dev8464

module exists precisely to give array scalars the look-and-feel of ufunc-based calculations with significantly reduced
overhead).
When a ufunc is called, many things must be done. The information collected from these setup operations is stored
in a loop-object. This loop object is a C-structure (that could become a Python object but is not initialized as such
because it is only used internally). This loop object has the layout needed to be used with PyArray_Broadcast so that
the broadcasting can be handled in the same way as it is handled in other sections of code.
The first thing done is to look-up in the thread-specific global dictionary the current values for the buffer-size, the error
mask, and the associated error object. The state of the error mask controls what happens when an error-condiction is
found. It should be noted that checking of the hardware error flags is only performed after each 1-D loop is executed.
This means that if the input and output arrays are contiguous and of the correct type so that a single 1-D loop is
performed, then the flags may not be checked until all elements of the array have been calcluated. Looking up these
values in a thread- specific dictionary takes time which is easily ignored for all but very small arrays.
After checking, the thread-specific global variables, the inputs are evaluated to determine how the ufunc should proceed
and the input and output arrays are constructed if necessary. Any inputs which are not arrays are converted to arrays
(using context if necessary). Which of the inputs are scalars (and therefore converted to 0-D arrays) is noted.
Next, an appropriate 1-D loop is selected from the 1-D loops available to the ufunc based on the input array types. This
1-D loop is selected by trying to match the signature of the data-types of the inputs against the available signatures.
The signatures corresponding to built-in types are stored in the types member of the ufunc structure. The signatures
corresponding to user-defined types are stored in a linked-list of function-information with the head element stored
as a CObject in the userloops dictionary keyed by the data-type number (the first user-defined type in the argument
list is used as the key). The signatures are searched until a signature is found to which the input arrays can all be cast
safely (ignoring any scalar arguments which are not allowed to determine the type of the result). The implication of
this search procedure is that lesser types should be placed below larger types when the signatures are stored. If no
1-D loop is found, then an error is reported. Otherwise, the argument_list is updated with the stored signature in
case casting is necessary and to fix the output types assumed by the 1-D loop.
If the ufunc has 2 inputs and 1 output and the second input is an Object array then a special-case check is performed
so that NotImplemented is returned if the second input is not an ndarray, has the __array_priority__ attribute, and
has an __r{op}__ special method. In this way, Python is signaled to give the other object a chance to complete the
operation instead of using generic object-array calculations. This allows (for example) sparse matrices to override the
multiplication operator 1-D loop.
For input arrays that are smaller than the specified buffer size, copies are made of all non-contiguous, mis-aligned, or
out-of- byteorder arrays to ensure that for small arrays, a single-loop is used. Then, array iterators are created for all
the input arrays and the resulting collection of iterators is broadcast to a single shape.
The output arguments (if any) are then processed and any missing return arrays are constructed. If any provided output
array doesnt have the correct type (or is mis-aligned) and is smaller than the buffer size, then a new output array is
constructed with the special UPDATEIFCOPY flag set so that when it is DECREFd on completion of the function,
its contents will be copied back into the output array. Iterators for the output arguments are then processed.
Finally, the decision is made about how to execute the looping mechanism to ensure that all elements of the input
arrays are combined to produce the output arrays of the correct type. The options for loop execution are one-loop (for
contiguous, aligned, and correct data- type), strided-loop (for non-contiguous but still aligned and correct data-type),
and a buffered loop (for mis-aligned or incorrect data- type situations). Depending on which execution method is
called for, the loop is then setup and computed.

Function call

This section describes how the basic universal function computation loop is setup and executed for each of the three
different kinds of execution possibilities. If NPY_ALLOW_THREADS is defined during compilation, then the Python
Global Interpreter Lock (GIL) is released prior to calling all of these loops (as long as they dont involve object arrays).

6.1. Numpy C Code Explanations 1079

NumPy Reference, Release 2.0.0.dev8464

It is re-acquired if necessary to handle error conditions. The hardware error flags are checked only after the 1-D loop
is calcluated.

One Loop
This is the simplest case of all. The ufunc is executed by calling the underlying 1-D loop exactly once. This is possible
only when we have aligned data of the correct type (including byte-order) for both input and output and all arrays have
uniform strides (either contiguous, 0-D, or 1-D). In this case, the 1-D computational loop is called once to compute
the calculation for the entire array. Note that the hardware error flags are only checked after the entire calculation is
complete.

Strided Loop
When the input and output arrays are aligned and of the correct type, but the striding is not uniform (non-contiguous
and 2-D or larger), then a second looping structure is employed for the calculation. This approach converts all of the
iterators for the input and output arguments to iterate over all but the largest dimension. The inner loop is then handled
by the underlying 1-D computational loop. The outer loop is a standard iterator loop on the converted iterators. The
hardware error flags are checked after each 1-D loop is completed.

Buffered Loop
This is the code that handles the situation whenever the input and/or output arrays are either misaligned or of the wrong
data-type (including being byte-swapped) from what the underlying 1-D loop expects. The arrays are also assumed to
be non-contiguous. The code works very much like the strided loop except for the inner 1-D loop is modified so that
pre-processing is performed on the inputs and post- processing is performed on the outputs in bufsize chunks (where
bufsize is a user-settable parameter). The underlying 1-D computational loop is called on data that is copied over (if it
needs to be). The setup code and the loop code is considerably more complicated in this case because it has to handle:
memory allocation of the temporary buffers
deciding whether or not to use buffers on the input and output data (mis-aligned and/or wrong data-type)
copying and possibly casting data for any inputs or outputs for which buffers are necessary.
special-casing Object arrays so that reference counts are properly handled when copies and/or casts are neces-
sary.
breaking up the inner 1-D loop into bufsize chunks (with a possible remainder).
Again, the hardware error flags are checked at the end of each 1-D loop.

Final output manipulation

Ufuncs allow other array-like classes to be passed seamlessly through the interface in that inputs of a particu-
lar class will induce the outputs to be of that same class. The mechanism by which this works is the following.
If any of the inputs are not ndarrays and define the __array_wrap__ method, then the class with the largest
__array_priority__ attribute determines the type of all the outputs (with the exception of any output arrays
passed in). The __array_wrap__ method of the input array will be called with the ndarray being returned from
the ufunc as its input. There are two calling styles of the __array_wrap__ function supported. The first takes
the ndarray as the first argument and a tuple of context as the second argument. The context is (ufunc, arguments,
output argument number). This is the first call tried. If a TypeError occurs, then the function is called with just the
ndarray as the first argument.

Methods

Their are three methods of ufuncs that require calculation similar to the general-purpose ufuncs. These are reduce,
accumulate, and reduceat. Each of these methods requires a setup command followed by a loop. There are four loop

1080 Chapter 6. Numpy internals

 NumPy Reference, Release 2.0.0.dev8464

styles possible for the methods corresponding to no-elements, one-element, strided-loop, and buffered- loop. These
are the same basic loop styles as implemented for the general purpose function call except for the no-element and one-
element cases which are special-cases occurring when the input array objects have 0 and 1 elements respectively.

Setup
The setup function for all three methods is construct_reduce. This function creates a reducing loop object and
fills it with parameters needed to complete the loop. All of the methods only work on ufuncs that take 2-inputs and
return 1 output. Therefore, the underlying 1-D loop is selected assuming a signature of [ otype, otype, otype
] where otype is the requested reduction data-type. The buffer size and error handling is then retrieved from (per-
thread) global storage. For small arrays that are mis-aligned or have incorrect data-type, a copy is made so that the
un-buffered section of code is used. Then, the looping strategy is selected. If there is 1 element or 0 elements in the
array, then a simple looping method is selected. If the array is not mis-aligned and has the correct data-type, then
strided looping is selected. Otherwise, buffered looping must be performed. Looping parameters are then established,
and the return array is constructed. The output array is of a different shape depending on whether the method is reduce,
accumulate, or reduceat. If an output array is already provided, then its shape is checked. If the output array is not
C-contiguous, aligned, and of the correct data type, then a temporary copy is made with the UPDATEIFCOPY flag set.
In this way, the methods will be able to work with a well-behaved output array but the result will be copied back into
the true output array when the method computation is complete. Finally, iterators are set up to loop over the correct
axis (depending on the value of axis provided to the method) and the setup routine returns to the actual computation
routine.

Reduce
All of the ufunc methods use the same underlying 1-D computational loops with input and output arguments adjusted
so that the appropriate reduction takes place. For example, the key to the functioning of reduce is that the 1-D loop is
called with the output and the second input pointing to the same position in memory and both having a step- size of 0.
The first input is pointing to the input array with a step- size given by the appropriate stride for the selected axis. In
this way, the operation performed is

o= i[0]
o= i[k]<op>o k = 1...N

where N + 1 is the number of elements in the input, i, o is the output, and i[k] is the k th element of i along the selected
axis. This basic operations is repeated for arrays with greater than 1 dimension so that the reduction takes place for
every 1-D sub-array along the selected axis. An iterator with the selected dimension removed handles this looping.
For buffered loops, care must be taken to copy and cast data before the loop function is called because the underlying
loop expects aligned data of the correct data-type (including byte-order). The buffered loop must handle this copying
and casting prior to calling the loop function on chunks no greater than the user-specified bufsize.

Accumulate
The accumulate function is very similar to the reduce function in that the output and the second input both point to the
output. The difference is that the second input points to memory one stride behind the current output pointer. Thus,
the operation performed is

o[0] = i[0]
o[k] = i[k]<op>o[k 1] k = 1 . . . N.

The output has the same shape as the input and each 1-D loop operates over N elements when the shape in the selected
axis is N +1. Again, buffered loops take care to copy and cast the data before calling the underlying 1-D computational
loop.

Reduceat
The reduceat function is a generalization of both the reduce and accumulate functions. It implements a reduce over
ranges of the input array specified by indices. The extra indices argument is checked to be sure that every input

6.1. Numpy C Code Explanations 1081

NumPy Reference, Release 2.0.0.dev8464

is not too large for the input array along the selected dimension before the loop calculations take place. The loop
implementation is handled using code that is very similar to the reduce code repeated as many times as there are
elements in the indices input. In particular: the first input pointer passed to the underlying 1-D computational loop
points to the input array at the correct location indicated by the index array. In addition, the output pointer and the
second input pointer passed to the underlying 1-D loop point to the same position in memory. The size of the 1-D
computational loop is fixed to be the difference between the current index and the next index (when the current index
is the last index, then the next index is assumed to be the length of the array along the selected dimension). In this
way, the 1-D loop will implement a reduce over the specified indices.
Mis-aligned or a loop data-type that does not match the input and/or output data-type is handled using buffered code
where-in data is copied to a temporary buffer and cast to the correct data-type if necessary prior to calling the under-
lying 1-D function. The temporary buffers are created in (element) sizes no bigger than the user settable buffer-size
value. Thus, the loop must be flexible enough to call the underlying 1-D computational loop enough times to complete
the total calculation in chunks no bigger than the buffer-size.

6.2 Internal organization of numpy arrays

It helps to understand a bit about how numpy arrays are handled under the covers to help understand numpy better.
This section will not go into great detail. Those wishing to understand the full details are referred to Travis Oliphants
book Guide to Numpy.
Numpy arrays consist of two major components, the raw array data (from now on, referred to as the data buffer), and
the information about the raw array data. The data buffer is typically what people think of as arrays in C or Fortran,
a contiguous (and fixed) block of memory containing fixed sized data items. Numpy also contains a significant set of
data that describes how to interpret the data in the data buffer. This extra information contains (among other things):
1. The basic data elements size in bytes
2. The start of the data within the data buffer (an offset relative to the beginning of the data buffer).
3. The number of dimensions and the size of each dimension
4. The separation between elements for each dimension (the stride). This does not have to be a multiple of the
element size
5. The byte order of the data (which may not be the native byte order)
6. Whether the buffer is read-only
7. Information (via the dtype object) about the interpretation of the basic data element. The basic data element
may be as simple as a int or a float, or it may be a compound object (e.g., struct-like), a fixed character field, or
Python object pointers.
8. Whether the array is to interpreted as C-order or Fortran-order.
This arrangement allow for very flexible use of arrays. One thing that it allows is simple changes of the metadata
to change the interpretation of the array buffer. Changing the byteorder of the array is a simple change involving no
rearrangement of the data. The shape of the array can be changed very easily without changing anything in the data
buffer or any data copying at all
Among other things that are made possible is one can create a new array metadata object that uses the same data buffer
to create a new view of that data buffer that has a different interpretation of the buffer (e.g., different shape, offset,
byte order, strides, etc) but shares the same data bytes. Many operations in numpy do just this such as slices. Other
operations, such as transpose, dont move data elements around in the array, but rather change the information about
the shape and strides so that the indexing of the array changes, but the data in the doesnt move.
Typically these new versions of the array metadata but the same data buffer are new views into the data buffer. There
is a different ndarray object, but it uses the same data buffer. This is why it is necessary to force copies through use of
the .copy() method if one really wants to make a new and independent copy of the data buffer.

1082 Chapter 6. Numpy internals

 NumPy Reference, Release 2.0.0.dev8464

New views into arrays mean the the object reference counts for the data buffer increase. Simply doing away with the
original array object will not remove the data buffer if other views of it still exist.

6.3 Multidimensional Array Indexing Order Issues

What is the right way to index multi-dimensional arrays? Before you jump to conclusions about the one and true
way to index multi-dimensional arrays, it pays to understand why this is a confusing issue. This section will try to
explain in detail how numpy indexing works and why we adopt the convention we do for images, and when it may be
appropriate to adopt other conventions.
The first thing to understand is that there are two conflicting conventions for indexing 2-dimensional arrays. Matrix
notation uses the first index to indicate which row is being selected and the second index to indicate which column
is selected. This is opposite the geometrically oriented-convention for images where people generally think the first
index represents x position (i.e., column) and the second represents y position (i.e., row). This alone is the source of
much confusion; matrix-oriented users and image-oriented users expect two different things with regard to indexing.
The second issue to understand is how indices correspond to the order the array is stored in memory. In Fortran the first
index is the most rapidly varying index when moving through the elements of a two dimensional array as it is stored
in memory. If you adopt the matrix convention for indexing, then this means the matrix is stored one column at a time
(since the first index moves to the next row as it changes). Thus Fortran is considered a Column-major language. C
has just the opposite convention. In C, the last index changes most rapidly as one moves through the array as stored
in memory. Thus C is a Row-major language. The matrix is stored by rows. Note that in both cases it presumes that
the matrix convention for indexing is being used, i.e., for both Fortran and C, the first index is the row. Note this
convention implies that the indexing convention is invariant and that the data order changes to keep that so.
But thats not the only way to look at it. Suppose one has large two-dimensional arrays (images or matrices) stored
in data files. Suppose the data are stored by rows rather than by columns. If we are to preserve our index convention
(whether matrix or image) that means that depending on the language we use, we may be forced to reorder the data
if it is read into memory to preserve our indexing convention. For example if we read row-ordered data into memory
without reordering, it will match the matrix indexing convention for C, but not for Fortran. Conversely, it will match
the image indexing convention for Fortran, but not for C. For C, if one is using data stored in row order, and one wants
to preserve the image index convention, the data must be reordered when reading into memory.
In the end, which you do for Fortran or C depends on which is more important, not reordering data or preserving the
indexing convention. For large images, reordering data is potentially expensive, and often the indexing convention is
inverted to avoid that.
The situation with numpy makes this issue yet more complicated. The internal machinery of numpy arrays is flexible
enough to accept any ordering of indices. One can simply reorder indices by manipulating the internal stride infor-
mation for arrays without reordering the data at all. Numpy will know how to map the new index order to the data
without moving the data.
So if this is true, why not choose the index order that matches what you most expect? In particular, why not define
row-ordered images to use the image convention? (This is sometimes referred to as the Fortran convention vs the C
convention, thus the C and FORTRAN order options for array ordering in numpy.) The drawback of doing this
is potential performance penalties. Its common to access the data sequentially, either implicitly in array operations
or explicitly by looping over rows of an image. When that is done, then the data will be accessed in non-optimal
order. As the first index is incremented, what is actually happening is that elements spaced far apart in memory are
being sequentially accessed, with usually poor memory access speeds. For example, for a two dimensional image im
defined so that im[0, 10] represents the value at x=0, y=10. To be consistent with usual Python behavior then im[0]
would represent a column at x=0. Yet that data would be spread over the whole array since the data are stored in
row order. Despite the flexibility of numpys indexing, it cant really paper over the fact basic operations are rendered
inefficient because of data order or that getting contiguous subarrays is still awkward (e.g., im[:,0] for the first row,
vs im[0]), thus one cant use an idiom such as for row in im; for col in im does work, but doesnt yield contiguous
column data.

6.3. Multidimensional Array Indexing Order Issues 1083

NumPy Reference, Release 2.0.0.dev8464

As it turns out, numpy is smart enough when dealing with ufuncs to determine which index is the most rapidly varying
one in memory and uses that for the innermost loop. Thus for ufuncs there is no large intrinsic advantage to either
approach in most cases. On the other hand, use of .flat with an FORTRAN ordered array will lead to non-optimal
memory access as adjacent elements in the flattened array (iterator, actually) are not contiguous in memory.
Indeed, the fact is that Python indexing on lists and other sequences naturally leads to an outside-to inside ordering
(the first index gets the largest grouping, the next the next largest, and the last gets the smallest element). Since image
data are normally stored by rows, this corresponds to position within rows being the last item indexed.
If you do want to use Fortran ordering realize that there are two approaches to consider: 1) accept that the first index
is just not the most rapidly changing in memory and have all your I/O routines reorder your data when going from
memory to disk or visa versa, or use numpys mechanism for mapping the first index to the most rapidly varying
data. We recommend the former if possible. The disadvantage of the latter is that many of numpys functions will
yield arrays without Fortran ordering unless you are careful to use the order keyword. Doing this would be highly
inconvenient.
Otherwise we recommend simply learning to reverse the usual order of indices when accessing elements of an array.
Granted, it goes against the grain, but it is more in line with Python semantics and the natural order of the data.

1084 Chapter 6. Numpy internals

 CHAPTER

SEVEN

ACKNOWLEDGEMENTS

Large parts of this manual originate from Travis E. Oliphants book Guide to Numpy (which generously entered Public
Domain in August 2008). The reference documentation for many of the functions are written by numerous contributors
and developers of Numpy, both prior to and during the Numpy Documentation Marathon.
Please help to improve NumPys documentation! Instructions on how to join the ongoing documentation marathon
can be found on the scipy.org website

1085
NumPy Reference, Release 2.0.0.dev8464

1086 Chapter 7. Acknowledgements

 BIBLIOGRAPHY

[R57] : G. H. Golub and C. F. van Loan, Matrix Computations, 3rd ed., Baltimore, MD, Johns Hopkins University
Press, 1996, pg. 8.
[R58] : G. H. Golub and C. F. van Loan, Matrix Computations, 3rd ed., Baltimore, MD, Johns Hopkins University
Press, 1996, pg. 8.
[R61] Wikipedia, Vandermonde matrix, http://en.wikipedia.org/wiki/Vandermonde_matrix
[R59] Wikipedia, Curve fitting, http://en.wikipedia.org/wiki/Curve_fitting
[R60] Wikipedia, Polynomial interpolation, http://en.wikipedia.org/wiki/Polynomial_interpolation
[R269] http://en.wikipedia.org/wiki/IEEE_754
[R276] Wikipedia, Vandermonde matrix, http://en.wikipedia.org/wiki/Vandermonde_matrix
[R1] Press, Teukolsky, Vetterling and Flannery, Numerical Recipes in C++, 2nd ed, Cambridge University Press,
2002, p. 31.
[R268] Format Specification Mini-Language, Python Documentation.
[R23] Wikipedia, Twos complement, http://en.wikipedia.org/wiki/Twos_complement
[CT] Cooley, James W., and John W. Tukey, 1965, An algorithm for the machine calculation of complex Fourier
series, Math. Comput. 19: 297-301.
[CT] Cooley, James W., and John W. Tukey, 1965, An algorithm for the machine calculation of complex Fourier
series, Math. Comput. 19: 297-301.
[NR] Press, W., Teukolsky, S., Vetterline, W.T., and Flannery, B.P., 2007, Numerical Recipes: The Art of Scientific
Computing, ch. 12-13. Cambridge Univ. Press, Cambridge, UK.
[R63] : G. H. Golub and C. F. van Loan, Matrix Computations, 3rd ed., Baltimore, MD, Johns Hopkins University
Press, 1996, pg. 8.
[R47] G. Strang, Linear Algebra and Its Applications, 2nd Ed., Orlando, FL, Academic Press, Inc., 1980, pg. 222.
[R48] G. H. Golub and C. F. Van Loan, Matrix Computations, Baltimore, MD, Johns Hopkins University Press, 1985,
pg. 15
[R46] G. Strang, Linear Algebra and Its Applications, Orlando, FL, Academic Press, Inc., 1980, pg. 285.
[R50] G. Strang, Linear Algebra and Its Applications, 2nd Ed., Orlando, FL, Academic Press, Inc., 1980, pg. 22.
[R49] G. Strang, Linear Algebra and Its Applications, 2nd Ed., Orlando, FL, Academic Press, Inc., 1980, pp. 139-
142.
[R69] Dalgaard, Peter, Introductory Statistics with R, Springer-Verlag, 2002.

1087
NumPy Reference, Release 2.0.0.dev8464

[R70] Glantz, Stanton A. Primer of Biostatistics., McGraw-Hill, Fifth Edition, 2002.

[R71] Lentner, Marvin, Elementary Applied Statistics, Bogden and Quigley, 1972.
[R72] Weisstein, Eric W. Binomial Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/BinomialDistribution.html
[R73] Wikipedia, Binomial-distribution, http://en.wikipedia.org/wiki/Binomial_distribution
[R74] NIST/SEMATECH e-Handbook of Statistical Methods, http://www.itl.nist.gov/div898/handbook/eda/section3/eda3666.htm
[R75] Wikipedia, Chi-square distribution, http://en.wikipedia.org/wiki/Chi-square_distribution
[R235] David McKay, Information Theory, Inference and Learning Algorithms, chapter 23,
http://www.inference.phy.cam.ac.uk/mackay/
[R76] Peyton Z. Peebles Jr., Probability, Random Variables and Random Signal Principles, 4th ed, 2001, p. 57.
[R77] Poisson Process, Wikipedia, http://en.wikipedia.org/wiki/Poisson_process
[R78] Exponential Distribution, Wikipedia, http://en.wikipedia.org/wiki/Exponential_distribution
[R79] Glantz, Stanton A. Primer of Biostatistics., McGraw-Hill, Fifth Edition, 2002.
[R80] Wikipedia, F-distribution, http://en.wikipedia.org/wiki/F-distribution
[R81] Weisstein, Eric W. Gamma Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/GammaDistribution.html
[R82] Wikipedia, Gamma-distribution, http://en.wikipedia.org/wiki/Gamma-distribution
[R83] Gumbel, E.J. (1958). Statistics of Extremes. Columbia University Press.
[R84] Reiss, R.-D. and Thomas M. (2001), Statistical Analysis of Extreme Values, from Insurance, Finance, Hydrol-
ogy and Other Fields, Birkhauser Verlag, Basel: Boston : Berlin.
[R85] Wikipedia, Gumbel distribution, http://en.wikipedia.org/wiki/Gumbel_distribution
[R86] Lentner, Marvin, Elementary Applied Statistics, Bogden and Quigley, 1972.
[R87] Weisstein, Eric W. Hypergeometric Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/HypergeometricDistribution.html
[R88] Wikipedia, Hypergeometric-distribution, http://en.wikipedia.org/wiki/Hypergeometric-distribution
[R89] Abramowitz, M. and Stegun, I. A. (Eds.). Handbook of Mathematical Functions with Formulas, Graphs, and
Mathematical Tables, 9th printing. New York: Dover, 1972.
[R90] The Laplace distribution and generalizations By Samuel Kotz, Tomasz J. Kozubowski, Krzysztof Podgorski,
Birkhauser, 2001.
[R91] Weisstein, Eric W. Laplace Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/LaplaceDistribution.html
[R92] Wikipedia, Laplace distribution, http://en.wikipedia.org/wiki/Laplace_distribution
[R93] Reiss, R.-D. and Thomas M. (2001), Statistical Analysis of Extreme Values, from Insurance, Finance, Hydrol-
ogy and Other Fields, Birkhauser Verlag, Basel, pp 132-133.
[R94] Weisstein, Eric W. Logistic Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/LogisticDistribution.html
[R95] Wikipedia, Logistic-distribution, http://en.wikipedia.org/wiki/Logistic-distribution
[R96] Eckhard Limpert, Werner A. Stahel, and Markus Abbt, Log-normal Distributions across the Sciences: Keys
and Clues, May 2001 Vol. 51 No. 5 BioScience http://stat.ethz.ch/~stahel/lognormal/bioscience.pdf
[R97] Reiss, R.D., Thomas, M.(2001), Statistical Analysis of Extreme Values, Birkhauser Verlag, Basel, pp 31-32.

1088 Bibliography
 NumPy Reference, Release 2.0.0.dev8464

[R98] Wikipedia, Lognormal distribution, http://en.wikipedia.org/wiki/Lognormal_distribution

[R99] Buzas, Martin A.; Culver, Stephen J., Understanding regional species diversity through the log series distribu-
tion of occurrences: BIODIVERSITY RESEARCH Diversity & Distributions, Volume 5, Number 5, September
1999 , pp. 187-195(9).
[R100] Fisher, R.A A.S. Corbet, and C.B. Williams. 1943. The relation between the number of species and the
number of individuals in a random sample of an animal population. Journal of Animal Ecology, 12:42-58.
[R101] D. J. Hand, F. Daly, D. Lunn, E. Ostrowski, A Handbook of Small Data Sets, CRC Press, 1994.
[R102] Wikipedia, Logarithmic-distribution, http://en.wikipedia.org/wiki/Logarithmic-distribution
[R236] A. Papoulis, Probability, Random Variables, and Stochastic Processes, 3rd ed., McGraw-Hill Companies,
1991
[R237] R.O. Duda, P.E. Hart, and D.G. Stork, Pattern Classification, 2nd ed., Wiley, 2001.
[R238] Weisstein, Eric W. Negative Binomial Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/NegativeBinomialDistribution.html
[R239] Wikipedia, Negative binomial distribution, http://en.wikipedia.org/wiki/Negative_binomial_distribution
[R240] Delhi, M.S. Holla, On a noncentral chi-square distribution in the analysis of weapon systems effectiveness,
Metrika, Volume 15, Number 1 / December, 1970.
[R241] Wikipedia, Noncentral chi-square distribution http://en.wikipedia.org/wiki/Noncentral_chi-
square_distribution
[R242] Wikipedia, Normal distribution, http://en.wikipedia.org/wiki/Normal_distribution
[R243] P. R. Peebles Jr., Central Limit Theorem in Probability, Random Variables and Random Signal Principles,
4th ed., 2001, pp. 51, 51, 125.
[R244] Francis Hunt and Paul Johnson, On the Pareto Distribution of Sourceforge projects.
[R245] Pareto, V. (1896). Course of Political Economy. Lausanne.
[R246] Reiss, R.D., Thomas, M.(2001), Statistical Analysis of Extreme Values, Birkhauser Verlag, Basel, pp 23-30.
[R247] Wikipedia, Pareto distribution, http://en.wikipedia.org/wiki/Pareto_distribution
[R248] Weisstein, Eric W. Poisson Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/PoissonDistribution.html
[R249] Wikipedia, Poisson distribution, http://en.wikipedia.org/wiki/Poisson_distribution
[R250] Christian Kleiber, Samuel Kotz, Statistical size distributions in economics and actuarial sciences, Wiley,
2003.
[R251] Heckert, N. A. and Filliben, James J. (2003). NIST Handbook 148: Dataplot Reference Manual, Volume
2: Let Subcommands and Library Functions, National Institute of Standards and Technology Handbook Series,
June 2003. http://www.itl.nist.gov/div898/software/dataplot/refman2/auxillar/powpdf.pdf
[R253] Weisstein, Eric W. Gamma Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/GammaDistribution.html
[R254] Wikipedia, Gamma-distribution, http://en.wikipedia.org/wiki/Gamma-distribution
[R255] Dalgaard, Peter, Introductory Statistics With R, Springer, 2002.
[R256] Wikipedia, Students t-distribution http://en.wikipedia.org/wiki/Students_t-distribution
[R257] Abramowitz, M. and Stegun, I. A. (ed.), Handbook of Mathematical Functions, National Bureau of Standards,
1964; reprinted Dover Publications, 1965.
[R258] von Mises, Richard, 1964, Mathematical Theory of Probability and Statistics (New York: Academic Press).

Bibliography 1089
NumPy Reference, Release 2.0.0.dev8464

[R259] Wikipedia, Von Mises distribution, http://en.wikipedia.org/wiki/Von_Mises_distribution

[R260] Waloddi Weibull, Professor, Royal Technical University, Stockholm, 1939 A Statistical Theory Of The
Strength Of Materials, Ingeniorsvetenskapsakademiens Handlingar Nr 151, 1939, Generalstabens Litografiska
Anstalts Forlag, Stockholm.
[R261] Waloddi Weibull, 1951 A Statistical Distribution Function of Wide Applicability, Journal Of Applied Me-
chanics ASME Paper.
[R262] Wikipedia, Weibull distribution, http://en.wikipedia.org/wiki/Weibull_distribution
[R263] Weisstein, Eric W. Zipf Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/ZipfDistribution.html
[R264] Wikipedia, Zeta distribution, http://en.wikipedia.org/wiki/Zeta_distribution
[R265] Wikipedia, Zipfs Law, http://en.wikipedia.org/wiki/Zipf%27s_law
[R266] Zipf, George Kingsley (1932): Selected Studies of the Principle of Relative Frequency in Language. Cam-
bridge (Mass.).
[R169] Dalgaard, Peter, Introductory Statistics with R, Springer-Verlag, 2002.
[R170] Glantz, Stanton A. Primer of Biostatistics., McGraw-Hill, Fifth Edition, 2002.
[R171] Lentner, Marvin, Elementary Applied Statistics, Bogden and Quigley, 1972.
[R172] Weisstein, Eric W. Binomial Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/BinomialDistribution.html
[R173] Wikipedia, Binomial-distribution, http://en.wikipedia.org/wiki/Binomial_distribution
[R174] NIST/SEMATECH e-Handbook of Statistical Methods, http://www.itl.nist.gov/div898/handbook/eda/section3/eda3666.htm
[R175] Wikipedia, Chi-square distribution, http://en.wikipedia.org/wiki/Chi-square_distribution
[R176] David McKay, Information Theory, Inference and Learning Algorithms, chapter 23,
http://www.inference.phy.cam.ac.uk/mackay/
[R177] Peyton Z. Peebles Jr., Probability, Random Variables and Random Signal Principles, 4th ed, 2001, p. 57.
[R178] Poisson Process, Wikipedia, http://en.wikipedia.org/wiki/Poisson_process
[R179] Exponential Distribution, Wikipedia, http://en.wikipedia.org/wiki/Exponential_distribution
[R180] Glantz, Stanton A. Primer of Biostatistics., McGraw-Hill, Fifth Edition, 2002.
[R181] Wikipedia, F-distribution, http://en.wikipedia.org/wiki/F-distribution
[R182] Weisstein, Eric W. Gamma Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/GammaDistribution.html
[R183] Wikipedia, Gamma-distribution, http://en.wikipedia.org/wiki/Gamma-distribution
[R184] Gumbel, E.J. (1958). Statistics of Extremes. Columbia University Press.
[R185] Reiss, R.-D. and Thomas M. (2001), Statistical Analysis of Extreme Values, from Insurance, Finance, Hy-
drology and Other Fields, Birkhauser Verlag, Basel: Boston : Berlin.
[R186] Wikipedia, Gumbel distribution, http://en.wikipedia.org/wiki/Gumbel_distribution
[R187] Lentner, Marvin, Elementary Applied Statistics, Bogden and Quigley, 1972.
[R188] Weisstein, Eric W. Hypergeometric Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/HypergeometricDistribution.html
[R189] Wikipedia, Hypergeometric-distribution, http://en.wikipedia.org/wiki/Hypergeometric-distribution

1090 Bibliography
 NumPy Reference, Release 2.0.0.dev8464

[R190] Abramowitz, M. and Stegun, I. A. (Eds.). Handbook of Mathematical Functions with Formulas, Graphs, and
Mathematical Tables, 9th printing. New York: Dover, 1972.
[R191] The Laplace distribution and generalizations By Samuel Kotz, Tomasz J. Kozubowski, Krzysztof Podgorski,
Birkhauser, 2001.
[R192] Weisstein, Eric W. Laplace Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/LaplaceDistribution.html
[R193] Wikipedia, Laplace distribution, http://en.wikipedia.org/wiki/Laplace_distribution
[R194] Reiss, R.-D. and Thomas M. (2001), Statistical Analysis of Extreme Values, from Insurance, Finance, Hy-
drology and Other Fields, Birkhauser Verlag, Basel, pp 132-133.
[R195] Weisstein, Eric W. Logistic Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/LogisticDistribution.html
[R196] Wikipedia, Logistic-distribution, http://en.wikipedia.org/wiki/Logistic-distribution
[R197] Eckhard Limpert, Werner A. Stahel, and Markus Abbt, Log-normal Distributions across the Sciences: Keys
and Clues, May 2001 Vol. 51 No. 5 BioScience http://stat.ethz.ch/~stahel/lognormal/bioscience.pdf
[R198] Reiss, R.D., Thomas, M.(2001), Statistical Analysis of Extreme Values, Birkhauser Verlag, Basel, pp 31-32.
[R199] Wikipedia, Lognormal distribution, http://en.wikipedia.org/wiki/Lognormal_distribution
[R200] Buzas, Martin A.; Culver, Stephen J., Understanding regional species diversity through the log series distri-
bution of occurrences: BIODIVERSITY RESEARCH Diversity & Distributions, Volume 5, Number 5, September
1999 , pp. 187-195(9).
[R201] Fisher, R.A A.S. Corbet, and C.B. Williams. 1943. The relation between the number of species and the
number of individuals in a random sample of an animal population. Journal of Animal Ecology, 12:42-58.
[R202] D. J. Hand, F. Daly, D. Lunn, E. Ostrowski, A Handbook of Small Data Sets, CRC Press, 1994.
[R203] Wikipedia, Logarithmic-distribution, http://en.wikipedia.org/wiki/Logarithmic-distribution
[R204] A. Papoulis, Probability, Random Variables, and Stochastic Processes, 3rd ed., McGraw-Hill Companies,
1991
[R205] R.O. Duda, P.E. Hart, and D.G. Stork, Pattern Classification, 2nd ed., Wiley, 2001.
[R206] Weisstein, Eric W. Negative Binomial Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/NegativeBinomialDistribution.html
[R207] Wikipedia, Negative binomial distribution, http://en.wikipedia.org/wiki/Negative_binomial_distribution
[R208] Delhi, M.S. Holla, On a noncentral chi-square distribution in the analysis of weapon systems effectiveness,
Metrika, Volume 15, Number 1 / December, 1970.
[R209] Wikipedia, Noncentral chi-square distribution http://en.wikipedia.org/wiki/Noncentral_chi-
square_distribution
[R210] Wikipedia, Normal distribution, http://en.wikipedia.org/wiki/Normal_distribution
[R211] P. R. Peebles Jr., Central Limit Theorem in Probability, Random Variables and Random Signal Principles,
4th ed., 2001, pp. 51, 51, 125.
[R212] Francis Hunt and Paul Johnson, On the Pareto Distribution of Sourceforge projects.
[R213] Pareto, V. (1896). Course of Political Economy. Lausanne.
[R214] Reiss, R.D., Thomas, M.(2001), Statistical Analysis of Extreme Values, Birkhauser Verlag, Basel, pp 23-30.
[R215] Wikipedia, Pareto distribution, http://en.wikipedia.org/wiki/Pareto_distribution

Bibliography 1091
NumPy Reference, Release 2.0.0.dev8464

[R216] Weisstein, Eric W. Poisson Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/PoissonDistribution.html
[R217] Wikipedia, Poisson distribution, http://en.wikipedia.org/wiki/Poisson_distribution
[R218] Christian Kleiber, Samuel Kotz, Statistical size distributions in economics and actuarial sciences, Wiley,
2003.
[R219] Heckert, N. A. and Filliben, James J. (2003). NIST Handbook 148: Dataplot Reference Manual, Volume
2: Let Subcommands and Library Functions, National Institute of Standards and Technology Handbook Series,
June 2003. http://www.itl.nist.gov/div898/software/dataplot/refman2/auxillar/powpdf.pdf
[R220] M. Matsumoto and T. Nishimura, Mersenne Twister: A 623-dimensionally equidistributed uniform pseu-
dorandom number generator, ACM Trans. on Modeling and Computer Simulation, Vol. 8, No. 1, pp. 3-30, Jan.
1998.
[R221] Weisstein, Eric W. Gamma Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/GammaDistribution.html
[R222] Wikipedia, Gamma-distribution, http://en.wikipedia.org/wiki/Gamma-distribution
[R223] Dalgaard, Peter, Introductory Statistics With R, Springer, 2002.
[R224] Wikipedia, Students t-distribution http://en.wikipedia.org/wiki/Students_t-distribution
[R225] Abramowitz, M. and Stegun, I. A. (ed.), Handbook of Mathematical Functions, National Bureau of Standards,
1964; reprinted Dover Publications, 1965.
[R226] von Mises, Richard, 1964, Mathematical Theory of Probability and Statistics (New York: Academic Press).
[R227] Wikipedia, Von Mises distribution, http://en.wikipedia.org/wiki/Von_Mises_distribution
[R228] Waloddi Weibull, Professor, Royal Technical University, Stockholm, 1939 A Statistical Theory Of The
Strength Of Materials, Ingeniorsvetenskapsakademiens Handlingar Nr 151, 1939, Generalstabens Litografiska
Anstalts Forlag, Stockholm.
[R229] Waloddi Weibull, 1951 A Statistical Distribution Function of Wide Applicability, Journal Of Applied Me-
chanics ASME Paper.
[R230] Wikipedia, Weibull distribution, http://en.wikipedia.org/wiki/Weibull_distribution
[R231] Weisstein, Eric W. Zipf Distribution. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/ZipfDistribution.html
[R232] Wikipedia, Zeta distribution, http://en.wikipedia.org/wiki/Zeta_distribution
[R233] Wikipedia, Zipfs Law, http://en.wikipedia.org/wiki/Zipf%27s_law
[R234] Zipf, George Kingsley (1932): Selected Studies of the Principle of Relative Frequency in Language. Cam-
bridge (Mass.).
[R252] M. Matsumoto and T. Nishimura, Mersenne Twister: A 623-dimensionally equidistributed uniform pseu-
dorandom number generator, ACM Trans. on Modeling and Computer Simulation, Vol. 8, No. 1, pp. 3-30, Jan.
1998.
[R41] Wikipedia, Twos complement, http://en.wikipedia.org/wiki/Twos_complement
[R23] Wikipedia, Twos complement, http://en.wikipedia.org/wiki/Twos_complement
[R6] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions, 10th printing, 1964, pp. 79.
http://www.math.sfu.ca/~cbm/aands/
[R7] Wikipedia, Inverse trigonometric function, http://en.wikipedia.org/wiki/Inverse_trigonometric_function
[R2] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions, 10th printing, 1964, pp. 79.
http://www.math.sfu.ca/~cbm/aands/

1092 Bibliography
 NumPy Reference, Release 2.0.0.dev8464

[R3] Wikipedia, Inverse trigonometric function, http://en.wikipedia.org/wiki/Inverse_trigonometric_function

[R10] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions, 10th printing, 1964, pp. 79.
http://www.math.sfu.ca/~cbm/aands/
[R11] Wikipedia, Inverse trigonometric function, http://en.wikipedia.org/wiki/Arctan
[R12] Wikipedia, atan2, http://en.wikipedia.org/wiki/Atan2
[R13] ISO/IEC standard 9899:1999, Programming language C, 1999.
[R272] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions. New York, NY: Dover, 1972, pg.
83. http://www.math.sfu.ca/~cbm/aands/
[R273] Wikipedia, Hyperbolic function, http://en.wikipedia.org/wiki/Hyperbolic_function
[R8] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions, 10th printing, 1964, pp. 86.
http://www.math.sfu.ca/~cbm/aands/
[R9] Wikipedia, Inverse hyperbolic function, http://en.wikipedia.org/wiki/Arcsinh
[R4] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions, 10th printing, 1964, pp. 86.
http://www.math.sfu.ca/~cbm/aands/
[R5] Wikipedia, Inverse hyperbolic function, http://en.wikipedia.org/wiki/Arccosh
[R14] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions, 10th printing, 1964, pp. 86.
http://www.math.sfu.ca/~cbm/aands/
[R15] Wikipedia, Inverse hyperbolic function, http://en.wikipedia.org/wiki/Arctanh
[R16] Lecture Notes on the Status of IEEE 754, William Kahan, http://www.cs.berkeley.edu/~wkahan/ieee754status/IEEE754.PDF
[R17] How Futile are Mindless Assessments of Roundoff in Floating-Point Computation?, William Kahan,
http://www.cs.berkeley.edu/~wkahan/Mindless.pdf
[R274] Wikipedia page: http://en.wikipedia.org/wiki/Trapezoidal_rule
[R275] Illustration image: http://en.wikipedia.org/wiki/File:Composite_trapezoidal_rule_illustration.png
[R28] Wikipedia, Exponential function, http://en.wikipedia.org/wiki/Exponential_function
[R29] M. Abramovitz and I. A. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathe-
matical Tables, Dover, 1964, p. 69, http://www.math.sfu.ca/~cbm/aands/page_69.htm
[R51] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions, 10th printing, 1964, pp. 67.
http://www.math.sfu.ca/~cbm/aands/
[R52] Wikipedia, Logarithm. http://en.wikipedia.org/wiki/Logarithm
[R53] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions, 10th printing, 1964, pp. 67.
http://www.math.sfu.ca/~cbm/aands/
[R54] Wikipedia, Logarithm. http://en.wikipedia.org/wiki/Logarithm
[R55] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions, 10th printing, 1964, pp. 67.
http://www.math.sfu.ca/~cbm/aands/
[R56] Wikipedia, Logarithm. http://en.wikipedia.org/wiki/Logarithm
[R38] C. W. Clenshaw, Chebyshev series for mathematical functions, in National Physical Laboratory Mathemat-
ical Tables, vol. 5, London: Her Majestys Stationery Office, 1962.
[R39] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions, 10th printing, New York: Dover,
1964, pp. 379. http://www.math.sfu.ca/~cbm/aands/page_379.htm
[R40] http://kobesearch.cpan.org/htdocs/Math-Cephes/Math/Cephes.html

Bibliography 1093
NumPy Reference, Release 2.0.0.dev8464

[R270] Weisstein, Eric W. Sinc Function. From MathWorldA Wolfram Web Resource.
http://mathworld.wolfram.com/SincFunction.html
[R271] Wikipedia, Sinc function, http://en.wikipedia.org/wiki/Sinc_function
[R27] Wikipedia, Convolution, http://en.wikipedia.org/wiki/Convolution.
[R68] I. N. Bronshtein, K. A. Semendyayev, and K. A. Hirsch (Eng. trans. Ed.), Handbook of Mathematics, New
York, Van Nostrand Reinhold Co., 1985, pg. 720.
[R64] M. Sullivan and M. Sullivan, III, Algebra and Trignometry, Enhanced With Graphing Utilities, Prentice-Hall,
pg. 318, 1996.
[R65] G. Strang, Linear Algebra and Its Applications, 2nd Edition, Academic Press, pg. 182, 1980.
[R267] Wikipedia, Companion matrix, http://en.wikipedia.org/wiki/Companion_matrix
[R66] Wikipedia, Curve fitting, http://en.wikipedia.org/wiki/Curve_fitting
[R67] Wikipedia, Polynomial interpolation, http://en.wikipedia.org/wiki/Polynomial_interpolation
[WRW] Wheeler, D. A., E. Rathke, and R. Weir (Eds.) (2009, May). Open Document Format for Office Appli-
cations (OpenDocument)v1.2, Part 2: Recalculated Formula (OpenFormula) Format - Annotated Version, Pre-
Draft 12. Organization for the Advancement of Structured Information Standards (OASIS). Billerica, MA, USA.
[ODT Document]. Available: http://www.oasis-open.org/committees/documents.php?wg_abbrev=office-formula
OpenDocument-formula-20090508.odt
[WRW] Wheeler, D. A., E. Rathke, and R. Weir (Eds.) (2009, May). Open Document Format for Office Appli-
cations (OpenDocument)v1.2, Part 2: Recalculated Formula (OpenFormula) Format - Annotated Version, Pre-
Draft 12. Organization for the Advancement of Structured Information Standards (OASIS). Billerica, MA, USA.
[ODT Document]. Available: http://www.oasis-open.org/committees/documents.php?wg_abbrev=office-formula
OpenDocument-formula-20090508.odt
[G61] L. J. Gitman, Principles of Managerial Finance, Brief, 3rd ed., Addison-Wesley, 2003, pg. 346.
[WRW] Wheeler, D. A., E. Rathke, and R. Weir (Eds.) (2009, May). Open Document Format for Office Ap-
plications (OpenDocument)v1.2, Part 2: Recalculated Formula (OpenFormula) Format - Annotated Version,
Pre-Draft 12. Organization for the Advancement of Structured Information Standards (OASIS). Billerica, MA,
USA. [ODT Document]. Available: http://www.oasis-open.org/committees/documents.php ?wg_abbrev=office-
formulaOpenDocument-formula-20090508.odt
[G41] L. J. Gitman, Principles of Managerial Finance, Brief, 3rd ed., Addison-Wesley, 2003, pg. 348.
[R18] M.S. Bartlett, Periodogram Analysis and Continuous Spectra, Biometrika 37, 1-16, 1950.
[R19] E.R. Kanasewich, Time Sequence Analysis in Geophysics, The University of Alberta Press, 1975, pp. 109-
110.
[R20] A.V. Oppenheim and R.W. Schafer, Discrete-Time Signal Processing, Prentice-Hall, 1999, pp. 468-471.
[R21] Wikipedia, Window function, http://en.wikipedia.org/wiki/Window_function
[R22] W.H. Press, B.P. Flannery, S.A. Teukolsky, and W.T. Vetterling, Numerical Recipes, Cambridge University
Press, 1986, page 429.
[R24] Blackman, R.B. and Tukey, J.W., (1958) The measurement of power spectra, Dover Publications, New York.
[R25] Wikipedia, Window function, http://en.wikipedia.org/wiki/Window_function
[R26] Oppenheim, A.V., and R.W. Schafer. Discrete-Time Signal Processing. Upper Saddle River, NJ: Prentice-Hall,
1999, pp. 468-471.
[R30] Blackman, R.B. and Tukey, J.W., (1958) The measurement of power spectra, Dover Publications, New York.

1094 Bibliography
 NumPy Reference, Release 2.0.0.dev8464

[R31] E.R. Kanasewich, Time Sequence Analysis in Geophysics, The University of Alberta Press, 1975, pp. 109-
110.
[R32] Wikipedia, Window function, http://en.wikipedia.org/wiki/Window_function
[R33] W.H. Press, B.P. Flannery, S.A. Teukolsky, and W.T. Vetterling, Numerical Recipes, Cambridge University
Press, 1986, page 425.
[R34] Blackman, R.B. and Tukey, J.W., (1958) The measurement of power spectra, Dover Publications, New York.
[R35] E.R. Kanasewich, Time Sequence Analysis in Geophysics, The University of Alberta Press, 1975, pp. 106-
108.
[R36] Wikipedia, Window function, http://en.wikipedia.org/wiki/Window_function
[R37] W.H. Press, B.P. Flannery, S.A. Teukolsky, and W.T. Vetterling, Numerical Recipes, Cambridge University
Press, 1986, page 425.
[R43] J. F. Kaiser, Digital Filters - Ch 7 in Systems analysis by digital computer, Editors: F.F. Kuo and J.F.
Kaiser, p 218-285. John Wiley and Sons, New York, (1966).
[R44] E.R. Kanasewich, Time Sequence Analysis in Geophysics, The University of Alberta Press, 1975, pp. 177-
178.
[R45] Wikipedia, Window function, http://en.wikipedia.org/wiki/Window_function
[R269] http://en.wikipedia.org/wiki/IEEE_754
[R57] : G. H. Golub and C. F. van Loan, Matrix Computations, 3rd ed., Baltimore, MD, Johns Hopkins University
Press, 1996, pg. 8.
[R58] : G. H. Golub and C. F. van Loan, Matrix Computations, 3rd ed., Baltimore, MD, Johns Hopkins University
Press, 1996, pg. 8.
[R61] Wikipedia, Vandermonde matrix, http://en.wikipedia.org/wiki/Vandermonde_matrix
[R59] Wikipedia, Curve fitting, http://en.wikipedia.org/wiki/Curve_fitting
[R60] Wikipedia, Polynomial interpolation, http://en.wikipedia.org/wiki/Polynomial_interpolation

Bibliography 1095
NumPy Reference, Release 2.0.0.dev8464

1096 Bibliography
 INDEX

Symbols __ge__() (numpy.ma.MaskedArray method), 235

__abs__() (numpy.ma.MaskedArray method), 236 __ge__() (numpy.ndarray method), 59
__abs__() (numpy.ndarray method), 60 __getitem__() (numpy.ma.MaskedArray method), 241
__add__() (numpy.ma.MaskedArray method), 236 __getitem__() (numpy.ndarray method), 63
__add__() (numpy.ndarray method), 60 __getslice__() (numpy.ma.MaskedArray method), 241
__and__() (numpy.ma.MaskedArray method), 237 __getslice__() (numpy.ndarray method), 63
__and__() (numpy.ndarray method), 61 __getstate__() (numpy.ma.MaskedArray method), 240
__array__() (in module numpy), 97 __gt__() (numpy.ma.MaskedArray method), 235
__array__() (numpy.generic method), 78 __gt__() (numpy.ndarray method), 59
__array__() (numpy.ma.MaskedArray method), 240 __hex__() (numpy.ma.MaskedArray method), 205
__array__() (numpy.ndarray method), 63 __hex__() (numpy.ndarray method), 64
__array_finalize__() (in module numpy), 96 __iadd__() (numpy.ma.MaskedArray method), 238
__array_interface__ (built-in variable), 366 __iadd__() (numpy.ndarray method), 61
__array_interface__ (numpy.generic attribute), 68 __iand__() (numpy.ma.MaskedArray method), 238
__array_prepare__() (in module numpy), 96 __iand__() (numpy.ndarray method), 62
__array_priority__ (in module numpy), 97 __idiv__() (numpy.ma.MaskedArray method), 238
__array_priority__ (numpy.generic attribute), 69 __idiv__() (numpy.ndarray method), 61
__array_struct__ (C variable), 368 __ifloordiv__() (numpy.ma.MaskedArray method), 238
__array_struct__ (numpy.generic attribute), 68 __ifloordiv__() (numpy.ndarray method), 61
__array_wrap__() (in module numpy), 96 __ilshift__() (numpy.ma.MaskedArray method), 238
__array_wrap__() (numpy.generic method), 69, 78 __ilshift__() (numpy.ndarray method), 62
__array_wrap__() (numpy.ma.MaskedArray method), __imod__() (numpy.ma.MaskedArray method), 238
240 __imod__() (numpy.ndarray method), 61
__array_wrap__() (numpy.ndarray method), 63 __imul__() (numpy.ma.MaskedArray method), 238
__contains__() (numpy.ma.MaskedArray method), 241 __imul__() (numpy.ndarray method), 61
__contains__() (numpy.ndarray method), 63 __int__() (numpy.ma.MaskedArray method), 205
__copy__() (numpy.ma.MaskedArray method), 240 __int__() (numpy.ndarray method), 64
__copy__() (numpy.ndarray method), 62 __invert__() (numpy.ndarray method), 60
__deepcopy__() (numpy.ma.MaskedArray method), 240 __ior__() (numpy.ma.MaskedArray method), 238
__deepcopy__() (numpy.ndarray method), 62 __ior__() (numpy.ndarray method), 62
__delitem__() (numpy.ma.MaskedArray method), 241 __ipow__() (numpy.ma.MaskedArray method), 238
__div__() (numpy.ma.MaskedArray method), 236 __ipow__() (numpy.ndarray method), 62
__div__() (numpy.ndarray method), 60 __irshift__() (numpy.ma.MaskedArray method), 238
__divmod__() (numpy.ma.MaskedArray method), 237 __irshift__() (numpy.ndarray method), 62
__divmod__() (numpy.ndarray method), 60 __isub__() (numpy.ma.MaskedArray method), 238
__eq__() (numpy.ma.MaskedArray method), 235 __isub__() (numpy.ndarray method), 61
__eq__() (numpy.ndarray method), 59 __itruediv__() (numpy.ma.MaskedArray method), 238
__float__() (numpy.ma.MaskedArray method), 205 __itruediv__() (numpy.ndarray method), 61
__float__() (numpy.ndarray method), 64 __ixor__() (numpy.ma.MaskedArray method), 238
__floordiv__() (numpy.ma.MaskedArray method), 237 __ixor__() (numpy.ndarray method), 62
__floordiv__() (numpy.ndarray method), 60 __le__() (numpy.ma.MaskedArray method), 235
__le__() (numpy.ndarray method), 59

1097
NumPy Reference, Release 2.0.0.dev8464

__len__() (numpy.ma.MaskedArray method), 241 __setstate__() (numpy.ndarray method), 62

__len__() (numpy.ndarray method), 63 __str__() (numpy.ma.MaskedArray method), 239
__long__() (numpy.ma.MaskedArray method), 205 __str__() (numpy.ndarray method), 64
__long__() (numpy.ndarray method), 64 __sub__() (numpy.ma.MaskedArray method), 236
__lshift__() (numpy.ma.MaskedArray method), 237 __sub__() (numpy.ndarray method), 60
__lshift__() (numpy.ndarray method), 61 __truediv__() (numpy.ma.MaskedArray method), 237
__lt__() (numpy.ma.MaskedArray method), 235 __truediv__() (numpy.ndarray method), 60
__lt__() (numpy.ndarray method), 59 __xor__() (numpy.ma.MaskedArray method), 237
__mod__() (numpy.ma.MaskedArray method), 237 __xor__() (numpy.ndarray method), 61
__mod__() (numpy.ndarray method), 60
__mul__() (numpy.ma.MaskedArray method), 236 A
__mul__() (numpy.ndarray method), 60 A (numpy.matrix attribute), 97
__ne__() (numpy.ma.MaskedArray method), 235 absolute (in module numpy), 766
__ne__() (numpy.ndarray method), 59 abspath() (numpy.DataSource method), 514
__neg__() (numpy.ndarray method), 60 accumulate
__nonzero__() (numpy.ma.MaskedArray method), 235 ufunc methods, 1081
__nonzero__() (numpy.ndarray method), 59 accumulate() (numpy.ufunc method), 381
__oct__() (numpy.ma.MaskedArray method), 205 add (in module numpy), 752
__oct__() (numpy.ndarray method), 64 add() (in module numpy.core.defchararray), 961
__or__() (numpy.ma.MaskedArray method), 237 add_data_dir() (numpy.distutils.misc_util.Configuration
__or__() (numpy.ndarray method), 61 method), 1003
__pos__() (numpy.ndarray method), 60 add_data_files() (numpy.distutils.misc_util.Configuration
__pow__() (numpy.ma.MaskedArray method), 237 method), 1001
__pow__() (numpy.ndarray method), 60 add_extension() (numpy.distutils.misc_util.Configuration
__radd__() (numpy.ma.MaskedArray method), 236 method), 1004
__rand__() (numpy.ma.MaskedArray method), 237 add_headers() (numpy.distutils.misc_util.Configuration
__rdiv__() (numpy.ma.MaskedArray method), 236 method), 1004
__rdivmod__() (numpy.ma.MaskedArray method), 237 add_include_dirs() (numpy.distutils.misc_util.Configuration
__reduce__() (numpy.dtype method), 91 method), 1004
__reduce__() (numpy.generic method), 79 add_installed_library() (numpy.distutils.misc_util.Configuration
__reduce__() (numpy.ma.MaskedArray method), 240 method), 1005
__reduce__() (numpy.ndarray method), 62 add_library() (numpy.distutils.misc_util.Configuration
__repr__() (numpy.ma.MaskedArray method), 239 method), 1005
__repr__() (numpy.ndarray method), 64 add_npy_pkg_config() (numpy.distutils.misc_util.Configuration
__rfloordiv__() (numpy.ma.MaskedArray method), 237 method), 1006
__rlshift__() (numpy.ma.MaskedArray method), 237 add_scripts() (numpy.distutils.misc_util.Configuration
__rmod__() (numpy.ma.MaskedArray method), 237 method), 1005
__rmul__() (numpy.ma.MaskedArray method), 236 add_subpackage() (numpy.distutils.misc_util.Configuration
__ror__() (numpy.ma.MaskedArray method), 237 method), 1001
__rpow__() (numpy.ma.MaskedArray method), 237 alignment (C member), 1017
__rrshift__() (numpy.ma.MaskedArray method), 237 alignment (numpy.dtype attribute), 90
__rshift__() (numpy.ma.MaskedArray method), 237 all (in module numpy.ma), 253, 830
__rshift__() (numpy.ndarray method), 61 all() (in module numpy), 668
__rsub__() (numpy.ma.MaskedArray method), 236 all() (numpy.generic method), 70
__rtruediv__() (numpy.ma.MaskedArray method), 237 all() (numpy.ma.MaskedArray method), 224, 261, 839
__rxor__() (numpy.ma.MaskedArray method), 237 all() (numpy.matrix method), 100
__setitem__() (numpy.ma.MaskedArray method), 241 all() (numpy.ndarray method), 13, 58
__setitem__() (numpy.ndarray method), 63 all() (numpy.recarray method), 150
__setmask__() (numpy.ma.MaskedArray method), 241 all() (numpy.record method), 168
__setslice__() (numpy.ma.MaskedArray method), 241 all_strings() (in module numpy.distutils.misc_util), 1000
__setslice__() (numpy.ndarray method), 63 allclose() (in module numpy), 680
__setstate__() (numpy.dtype method), 91 allclose() (in module numpy.ma), 360, 937
__setstate__() (numpy.generic method), 79 allequal() (in module numpy.ma), 359, 936
__setstate__() (numpy.ma.MaskedArray method), 240 allpath() (in module numpy.distutils.misc_util), 1000

1098 Index
 NumPy Reference, Release 2.0.0.dev8464

alterdot() (in module numpy), 946 argsort() (numpy.ma.MaskedArray method), 215, 344,
amax() (in module numpy), 693 921
amin() (in module numpy), 692 argsort() (numpy.matrix method), 101
angle() (in module numpy), 761 argsort() (numpy.ndarray method), 14, 54
anom (in module numpy.ma), 319, 896 argsort() (numpy.recarray method), 151
anom() (numpy.ma.MaskedArray method), 225, 331, 908 argsort() (numpy.record method), 168
anomalies (in module numpy.ma), 320, 897 argwhere() (in module numpy), 663
any (in module numpy.ma), 253, 831 arithmetic, 59, 235
any() (in module numpy), 669 around (in module numpy.ma), 357, 934
any() (numpy.generic method), 70 around() (in module numpy), 728
any() (numpy.ma.MaskedArray method), 225, 261, 839 array
any() (numpy.matrix method), 100 C-API, 1032
any() (numpy.ndarray method), 13, 59 interface, 366
any() (numpy.recarray method), 150 protocol, 366
any() (numpy.record method), 168 array iterator, 176, 1076
ao (C member), 1023 array scalars, 1077
append() (in module numpy), 449 array() (in module numpy), 395
appendpath() (in module numpy.distutils.misc_util), 1000 array() (in module numpy.core.defchararray), 147, 407
apply_along_axis() (in module numpy), 772 array() (in module numpy.core.records), 406
apply_along_axis() (in module numpy.ma), 361, 938 array() (in module numpy.ma), 181, 244, 822
apply_over_axes() (in module numpy), 773 array_equal() (in module numpy), 680
arange (in module numpy.ma), 362, 939 array_equiv() (in module numpy), 681
arange() (in module numpy), 409 array_repr() (in module numpy), 505
arccos (in module numpy), 716 array_split() (in module numpy), 441
arccosh (in module numpy), 726 array_str() (in module numpy), 506
arcsin (in module numpy), 715 as_array() (in module numpy.ctypeslib), 958
arcsinh (in module numpy), 726 as_ctypes() (in module numpy.ctypeslib), 958
arctan (in module numpy), 717 asanyarray() (in module numpy), 398, 432
arctan2 (in module numpy), 719 asanyarray() (in module numpy.ma), 183, 299, 876
arctanh (in module numpy), 727 asarray() (in module numpy), 396, 431
argmax() (in module numpy), 661 asarray() (in module numpy.core.defchararray), 408
argmax() (in module numpy.ma), 338, 915 asarray() (in module numpy.ma), 182, 298, 875
argmax() (numpy.generic method), 71 ascontiguousarray() (in module numpy), 399
argmax() (numpy.ma.MaskedArray method), 214, 340, asfarray() (in module numpy), 434
917 asfortranarray() (in module numpy), 434
argmax() (numpy.matrix method), 100 asmatrix() (in module numpy), 123, 399, 433
argmax() (numpy.ndarray method), 13, 56 asscalar() (in module numpy), 435
argmax() (numpy.recarray method), 151 assert_almost_equal() (in module numpy.testing), 947
argmax() (numpy.record method), 168 assert_approx_equal() (in module numpy.testing), 948
argmin() (in module numpy), 662 assert_array_almost_equal() (in module numpy.testing),
argmin() (in module numpy.ma), 338, 915 949
argmin() (numpy.generic method), 71 assert_array_equal() (in module numpy.testing), 950
argmin() (numpy.ma.MaskedArray method), 215, 340, assert_array_less() (in module numpy.testing), 951
917 assert_equal() (in module numpy.testing), 952
argmin() (numpy.matrix method), 101 assert_raises() (in module numpy.testing), 953
argmin() (numpy.ndarray method), 14, 56 assert_string_equal() (in module numpy.testing), 953
argmin() (numpy.recarray method), 151 assert_warns() (in module numpy.testing), 953
argmin() (numpy.record method), 168 astype() (numpy.chararray method), 130
argsort() (in module numpy), 658 astype() (numpy.core.defchararray.chararray method),
argsort() (in module numpy.ma), 342, 919 980
argsort() (numpy.chararray method), 131 astype() (numpy.generic method), 71
argsort() (numpy.core.defchararray.chararray method), astype() (numpy.ma.MaskedArray method), 207
981 astype() (numpy.matrix method), 102
argsort() (numpy.generic method), 71 astype() (numpy.ndarray method), 14, 43

Index 1099
NumPy Reference, Release 2.0.0.dev8464

astype() (numpy.recarray method), 151 C

astype() (numpy.record method), 168 C-API
atleast_1d() (in module numpy), 426 array, 1032
atleast_1d() (in module numpy.ma), 271, 849 ndarray, 1032, 1064
atleast_2d() (in module numpy), 427 ufunc, 1064, 1070
atleast_2d() (in module numpy.ma), 272, 849 C-order, 31
atleast_3d() (in module numpy), 427 c_ (in module numpy), 456
atleast_3d() (in module numpy.ma), 272, 850 can_cast() (in module numpy), 481
attributes cancastscalarkindto (C member), 1020
ufunc, 377 cancastto (C member), 1020
average() (in module numpy), 697 capitalize() (in module numpy.core.defchararray), 962
average() (in module numpy.ma), 320, 897 castdict (C member), 1020
axis, 55 casting rules
ufunc, 375
B ceil (in module numpy), 730
backstrides (C member), 1023 center() (in module numpy.core.defchararray), 962
bartlett() (in module numpy), 803 char (numpy.dtype attribute), 87
base, 3 character arrays, 128
base (C member), 1015, 1017, 1025 chararray (class in numpy), 128
base (numpy.generic attribute), 68 chararray (class in numpy.core.defchararray), 978
base (numpy.ma.MaskedArray attribute), 200 check_return (C member), 1021
base (numpy.ndarray attribute), 12, 34 chisquare() (in module numpy.random), 573
base_repr() (in module numpy), 513 chisquare() (numpy.random.mtrand.RandomState
baseclass (numpy.ma.MaskedArray attribute), 199 method), 612
beta() (in module numpy.random), 571 cholesky() (in module numpy.linalg), 545
beta() (numpy.random.mtrand.RandomState method), choose() (in module numpy), 470
610 choose() (in module numpy.ma), 363, 940
binary_repr() (in module numpy), 512, 691 choose() (numpy.generic method), 71
bincount() (in module numpy), 710 choose() (numpy.ma.MaskedArray method), 216
binomial() (in module numpy.random), 572 choose() (numpy.matrix method), 102
binomial() (numpy.random.mtrand.RandomState choose() (numpy.ndarray method), 15, 53
method), 610 choose() (numpy.recarray method), 152
bitwise_and (in module numpy), 684 choose() (numpy.record method), 169
bitwise_or (in module numpy), 685 clip() (in module numpy), 764
bitwise_xor (in module numpy), 686 clip() (in module numpy.ma), 357, 934
blackman() (in module numpy), 805 clip() (numpy.generic method), 72
blue_text() (in module numpy.distutils.misc_util), 1000 clip() (numpy.ma.MaskedArray method), 225, 359, 936
bmat() (in module numpy), 123, 420 clip() (numpy.matrix method), 103
broadcast (class in numpy), 178, 428 clip() (numpy.ndarray method), 15, 57
broadcast_arrays() (in module numpy), 429 clip() (numpy.recarray method), 152
broadcastable, 371 clip() (numpy.record method), 169
broadcasting, 371, 1076 close() (numpy.memmap method), 127, 509
buffers, 372 code generation, 1011
byteorder (C member), 1016 column-major, 31
byteorder (numpy.dtype attribute), 88 column_stack (in module numpy.ma), 274, 280, 852, 858
bytes() (in module numpy.random), 569 column_stack() (in module numpy), 436
bytes() (numpy.random.mtrand.RandomState method), common_fill_value() (in module numpy.ma), 314, 891
611 common_type() (in module numpy), 482
byteswap() (numpy.generic method), 71, 78 compare (C member), 1018
byteswap() (numpy.ma.MaskedArray method), 207 comparison, 59, 235
byteswap() (numpy.matrix method), 102 compress() (in module numpy), 472
byteswap() (numpy.ndarray method), 14, 44 compress() (numpy.generic method), 72
byteswap() (numpy.recarray method), 151 compress() (numpy.ma.MaskedArray method), 217
byteswap() (numpy.record method), 169 compress() (numpy.matrix method), 103

1100 Index
 NumPy Reference, Release 2.0.0.dev8464

compress() (numpy.ndarray method), 15, 54 copyswap (C member), 1018

compress() (numpy.recarray method), 152 corrcoef() (in module numpy), 703
compress() (numpy.record method), 169 corrcoef() (in module numpy.ma), 322, 899
compress_cols() (in module numpy.ma), 308, 885 correlate() (in module numpy), 704
compress_rowcols() (in module numpy.ma), 309, 886 cos (in module numpy), 713
compress_rows() (in module numpy.ma), 309, 886 cosh (in module numpy), 724
compressed() (in module numpy.ma), 309, 886 count() (in module numpy.core.defchararray), 973
compressed() (numpy.ma.MaskedArray method), 207, count() (in module numpy.ma), 254, 831
310, 887 count() (numpy.chararray method), 131
concatenate() (in module numpy), 437 count() (numpy.core.defchararray.chararray method), 981
concatenate() (in module numpy.ma), 275, 281, 852, 859 count() (numpy.ma.MaskedArray method), 243, 262, 839
cond() (in module numpy.linalg), 555 count_masked() (in module numpy.ma), 254, 832
Configuration (class in numpy.distutils.misc_util), 999, cov() (in module numpy), 705
1001 cov() (in module numpy.ma), 322, 899
conj (in module numpy), 762 cpu (in module numpy.distutils.cpuinfo), 1008
conj() (numpy.generic method), 72 cross() (in module numpy), 739
conj() (numpy.ma.MaskedArray method), 226 ctypes (numpy.ma.MaskedArray attribute), 200
conj() (numpy.matrix method), 103 ctypes (numpy.ndarray attribute), 10, 37, 39
conj() (numpy.ndarray method), 15, 57 ctypes_load_library() (in module numpy.ctypeslib), 958
conj() (numpy.recarray method), 152 cumprod (in module numpy.ma), 324, 901
conj() (numpy.record method), 169 cumprod() (in module numpy), 735
conjugate (in module numpy.ma), 321, 898 cumprod() (numpy.generic method), 72
conjugate() (numpy.generic method), 72 cumprod() (numpy.ma.MaskedArray method), 226, 331,
conjugate() (numpy.ma.MaskedArray method), 226 908
conjugate() (numpy.matrix method), 103 cumprod() (numpy.matrix method), 104
conjugate() (numpy.ndarray method), 15 cumprod() (numpy.ndarray method), 16, 58
conjugate() (numpy.recarray method), 153 cumprod() (numpy.recarray method), 153
conjugate() (numpy.record method), 169 cumprod() (numpy.record method), 169
construction cumsum (in module numpy.ma), 323, 900
from dict, dtype, 86 cumsum() (in module numpy), 736
from dtype, dtype, 83 cumsum() (numpy.generic method), 72
from list, dtype, 86 cumsum() (numpy.ma.MaskedArray method), 227, 332,
from None, dtype, 83 909
from string, dtype, 84 cumsum() (numpy.matrix method), 104
from tuple, dtype, 85 cumsum() (numpy.ndarray method), 16, 57
from type, dtype, 83 cumsum() (numpy.recarray method), 153
container (class in numpy.lib.user_array), 176 cumsum() (numpy.record method), 170
container class, 176 cyan_text() (in module numpy.distutils.misc_util), 1000
contiguous, 31 cyg2win32() (in module numpy.distutils.misc_util), 1000
contiguous (C member), 1023
convolve() (in module numpy), 763 D
coordinates (C member), 1023 data (C member), 1014, 1021, 1027
copy (in module numpy.ma), 245, 823 data (numpy.generic attribute), 68
copy() (in module numpy), 400 data (numpy.ma.MaskedArray attribute), 199, 261, 838
copy() (numpy.chararray method), 131 data (numpy.ndarray attribute), 6, 34
copy() (numpy.core.defchararray.chararray method), 981 dataptr (C member), 1023
copy() (numpy.flatiter method), 481 DataSource (class in numpy), 513
copy() (numpy.generic method), 72 decode() (in module numpy.core.defchararray), 963
copy() (numpy.ma.MaskedArray method), 223, 247, 825 decode() (numpy.chararray method), 131
copy() (numpy.matrix method), 103 decode() (numpy.core.defchararray.chararray method),
copy() (numpy.ndarray method), 16, 44 981
copy() (numpy.recarray method), 153 decorate_methods() (in module numpy.testing), 956
copy() (numpy.record method), 169 default_fill_value() (in module numpy.ma), 314, 891
copysign (in module numpy), 751 deg2rad (in module numpy), 722

Index 1101
NumPy Reference, Release 2.0.0.dev8464

degrees (in module numpy), 720 dtype (numpy.ma.MaskedArray attribute), 202

delete() (in module numpy), 447 dtype (numpy.ndarray attribute), 6, 35
deprecated() (in module numpy.testing.decorators), 954 dump() (in module numpy.ma), 313, 890
deriv() (numpy.poly1d method), 779 dump() (numpy.chararray method), 131
descr (C member), 1015, 1027 dump() (numpy.core.defchararray.chararray method), 981
descr (numpy.dtype attribute), 90 dump() (numpy.generic method), 73
det() (in module numpy.linalg), 556 dump() (numpy.ma.MaskedArray method), 223
diag() (in module numpy), 415, 473 dump() (numpy.matrix method), 104
diag() (in module numpy.ma), 347, 924 dump() (numpy.ndarray method), 17, 43
diag_indices() (in module numpy), 464 dump() (numpy.recarray method), 154
diag_indices_from() (in module numpy), 465 dump() (numpy.record method), 170
diagflat() (in module numpy), 416 dumps() (in module numpy.ma), 313, 890
diagonal() (in module numpy), 474 dumps() (numpy.chararray method), 132
diagonal() (numpy.generic method), 72 dumps() (numpy.core.defchararray.chararray method),
diagonal() (numpy.ma.MaskedArray method), 217 982
diagonal() (numpy.matrix method), 104 dumps() (numpy.generic method), 73
diagonal() (numpy.ndarray method), 17, 54 dumps() (numpy.ma.MaskedArray method), 223
diagonal() (numpy.recarray method), 154 dumps() (numpy.matrix method), 104
diagonal() (numpy.record method), 170 dumps() (numpy.ndarray method), 17, 43
dict_append() (in module numpy.distutils.misc_util), dumps() (numpy.recarray method), 154
1000 dumps() (numpy.record method), 170
diff() (in module numpy), 737
digitize() (in module numpy), 711 E
dimensions (C member), 1014, 1024 ediff1d() (in module numpy), 738
dims_m1 (C member), 1023 ediff1d() (in module numpy.ma), 364, 941
dirichlet() (in module numpy.random.mtrand), 574 eig() (in module numpy.linalg), 549
dirichlet() (numpy.random.mtrand.RandomState eigh() (in module numpy.linalg), 550
method), 613 eigvals() (in module numpy.linalg), 552
distutils, 999 eigvalsh() (in module numpy.linalg), 553
divide (in module numpy), 754 ellipsis, 92
doc (C member), 1022 elsize (C member), 1017
dot() (in module numpy), 536 empty (in module numpy.ma), 248, 826
dot() (in module numpy.ma), 347, 924 empty() (in module numpy), 389
dot() (numpy.matrix method), 104 empty_like (in module numpy.ma), 249, 826
dot() (numpy.ndarray method), 17 empty_like() (in module numpy), 390
dot() (numpy.recarray method), 154 encode() (in module numpy.core.defchararray), 963
dot_join() (in module numpy.distutils.misc_util), 1000 encode() (numpy.chararray method), 132
dsplit() (in module numpy), 441 encode() (numpy.core.defchararray.chararray method),
dstack (in module numpy.ma), 276, 282, 853, 859 982
dstack() (in module numpy), 438 endswith() (numpy.chararray method), 132
dtype, 1075 endswith() (numpy.core.defchararray.chararray method),
construction from dict, 86 982
construction from dtype, 83 equal (in module numpy), 683
construction from list, 86 equal() (in module numpy.core.defchararray), 971
construction from None, 83 error handling, 372
construction from string, 84 errstate (class in numpy), 819
construction from tuple, 85 exists() (numpy.DataSource method), 514
construction from type, 83 exp (in module numpy), 742
field, 79 exp2 (in module numpy), 744
record, 79 expand_dims() (in module numpy), 430
scalar, 79 expand_dims() (in module numpy.ma), 273, 851
sub-array, 79, 85 expandtabs() (numpy.chararray method), 132
dtype (class in numpy), 80, 483 expandtabs() (numpy.core.defchararray.chararray
dtype (numpy.generic attribute), 68 method), 982

1102 Index
 NumPy Reference, Release 2.0.0.dev8464

expm1 (in module numpy), 743 flatiter (class in numpy), 480

exponential() (in module numpy.random), 574 flatnonzero() (in module numpy), 665
exponential() (numpy.random.mtrand.RandomState flatnotmasked_contiguous() (in module numpy.ma), 290,
method), 613 867
extract() (in module numpy), 667 flatnotmasked_edges() (in module numpy.ma), 290, 868
eye() (in module numpy), 391 flatten() (numpy.chararray method), 132
flatten() (numpy.core.defchararray.chararray method),
F 982
f (C member), 1017 flatten() (numpy.generic method), 73
f() (in module numpy.random), 575 flatten() (numpy.ma.MaskedArray method), 211, 267, 845
f() (numpy.random.mtrand.RandomState method), 614 flatten() (numpy.matrix method), 105
fabs (in module numpy), 767 flatten() (numpy.ndarray method), 17, 51, 423
factors (C member), 1023 flatten() (numpy.recarray method), 154
fft() (in module numpy.fft), 515 flatten() (numpy.record method), 170
fft2() (in module numpy.fft), 519 fliplr() (in module numpy), 452
fftfreq() (in module numpy.fft), 533 flipud() (in module numpy), 453
fftn() (in module numpy.fft), 521 floor (in module numpy), 730
fftshift() (in module numpy.fft), 533 floor_divide (in module numpy), 757
field flush() (numpy.memmap method), 127, 509
dtype, 79 fmod (in module numpy), 758
field() (numpy.recarray method), 154 format_parser (class in numpy), 485
fields (C member), 1017 Fortran-order, 31
fields (numpy.dtype attribute), 88 frexp (in module numpy), 752
fill (C member), 1019 from dict
fill() (numpy.chararray method), 132 dtype construction, 86
fill() (numpy.core.defchararray.chararray method), 982 from dtype
fill() (numpy.generic method), 73 dtype construction, 83
fill() (numpy.ma.MaskedArray method), 218 from list
fill() (numpy.matrix method), 105 dtype construction, 86
fill() (numpy.ndarray method), 17, 48 from None
fill() (numpy.recarray method), 154 dtype construction, 83
fill() (numpy.record method), 170 from string
fill_diagonal() (in module numpy), 478 dtype construction, 84
fill_value (numpy.ma.MaskedArray attribute), 199, 319, from tuple
896 dtype construction, 85
filled() (in module numpy.ma), 310, 887 from type
filled() (numpy.ma.MaskedArray method), 208, 310, 887 dtype construction, 83
filter_sources() (in module numpy.distutils.misc_util), fromarrays() (in module numpy.core.records), 406
1000 frombuffer (in module numpy.ma), 246, 824
find() (in module numpy.core.defchararray), 974 frombuffer() (in module numpy), 400
find() (numpy.chararray method), 132 fromfile() (in module numpy), 401
find() (numpy.core.defchararray.chararray method), 982 fromfile() (in module numpy.core.records), 407
find_common_type() (in module numpy), 490 fromfunction (in module numpy.ma), 246, 824
finfo (class in numpy), 486 fromfunction() (in module numpy), 402
fix() (in module numpy), 729 fromiter() (in module numpy), 403
fix_invalid() (in module numpy.ma), 184, 299, 876 frompyfunc() (in module numpy), 775
flags (C member), 1015, 1016, 1026 fromrecords() (in module numpy.core.records), 406
flags (numpy.dtype attribute), 89 fromregex() (in module numpy), 502
flags (numpy.generic attribute), 68 fromstr (C member), 1019
flags (numpy.ma.MaskedArray attribute), 202 fromstring() (in module numpy), 403, 503
flags (numpy.ndarray attribute), 7, 32 fromstring() (in module numpy.core.records), 407
flat (numpy.generic attribute), 68 fv() (in module numpy), 790
flat (numpy.ma.MaskedArray attribute), 205
flat (numpy.ndarray attribute), 7, 36, 177, 423

Index 1103
NumPy Reference, Release 2.0.0.dev8464

G getfield() (numpy.ndarray method), 18, 46

gamma() (in module numpy.random), 576 getfield() (numpy.recarray method), 155
gamma() (numpy.random.mtrand.RandomState method), getfield() (numpy.record method), 171
615 getH() (numpy.matrix method), 106
generate_config_py() (in module getI() (numpy.matrix method), 106
numpy.distutils.misc_util), 1000 getitem (C member), 1018
generic (class in numpy), 69 getmask() (in module numpy.ma), 255, 288, 833, 865
genfromtxt() (in module numpy), 500 getmaskarray() (in module numpy.ma), 256, 289, 834,
geometric() (in module numpy.random), 577 866
geometric() (numpy.random.mtrand.RandomState getslice
method), 616 ndarray special methods, 92
getT()
get_build_temp_dir() (numpy.distutils.misc_util.Configuration (numpy.matrix method), 107
method), 1007 gradient() (in module numpy), 738
get_cmd() (in module numpy.distutils.misc_util), 1000 greater (in module numpy), 682
get_config_cmd() (numpy.distutils.misc_util.Configuration greater() (in module numpy.core.defchararray), 972
method), 1007 greater_equal (in module numpy), 682
get_dependencies() (in module greater_equal() (in module numpy.core.defchararray),
numpy.distutils.misc_util), 1000 971
get_distribution() (numpy.distutils.misc_util.Configuration green_text() (in module numpy.distutils.misc_util), 1000
method), 1001 gumbel() (in module numpy.random), 577
get_ext_source_files() (in module gumbel() (numpy.random.mtrand.RandomState method),
numpy.distutils.misc_util), 1001 617
get_fill_value() (numpy.ma.MaskedArray method), 242,
318, 895
H
get_info() (in module numpy.distutils.system_info), 1008 H (numpy.matrix attribute), 97
get_info() (numpy.distutils.misc_util.Configuration hamming() (in module numpy), 807
method), 1008 hanning() (in module numpy), 809
get_numpy_include_dirs() (in module harden_mask (in module numpy.ma), 296, 873
numpy.distutils.misc_util), 1000 harden_mask() (numpy.ma.MaskedArray method), 241,
get_printoptions() (in module numpy), 510 296, 874
get_script_files() (in module numpy.distutils.misc_util), hardmask (numpy.ma.MaskedArray attribute), 200
1001 has_cxx_sources() (in module numpy.distutils.misc_util),
get_standard_file() (in module 1000
numpy.distutils.system_info), 1008 has_f_sources() (in module numpy.distutils.misc_util),
get_state() (in module numpy.random), 653 1000
get_state() (numpy.random.mtrand.RandomState hasobject (numpy.dtype attribute), 89
method), 616 have_f77c() (numpy.distutils.misc_util.Configuration
get_subpackage() (numpy.distutils.misc_util.Configuration method), 1007
method), 1001 have_f90c() (numpy.distutils.misc_util.Configuration
get_version() (numpy.distutils.misc_util.Configuration method), 1007
method), 1008 hfft() (in module numpy.fft), 532
getA() (numpy.matrix method), 105 histogram() (in module numpy), 706
getA1() (numpy.matrix method), 106 histogram2d() (in module numpy), 707
getbuffer() (in module numpy), 945 histogramdd() (in module numpy), 709
getbufsize() (in module numpy), 946 hsplit (in module numpy.ma), 277, 855
getdata() (in module numpy.ma), 257, 835 hsplit() (in module numpy), 442
geterr() (in module numpy), 816 hstack (in module numpy.ma), 276, 283, 854, 860
geterrcall() (in module numpy), 818 hstack() (in module numpy), 439
geterrobj() (in module numpy), 821 hypergeometric() (in module numpy.random), 579
getfield() (numpy.chararray method), 133 hypergeometric() (numpy.random.mtrand.RandomState
getfield() (numpy.core.defchararray.chararray method), method), 618
983 hypot (in module numpy), 718
getfield() (numpy.generic method), 73
getfield() (numpy.matrix method), 107

1104 Index
 NumPy Reference, Release 2.0.0.dev8464

I isbuiltin (numpy.dtype attribute), 90

I (numpy.matrix attribute), 97 iscomplex() (in module numpy), 674
i0() (in module numpy), 748 iscomplexobj() (in module numpy), 675
identity (C member), 1021 iscontiguous() (numpy.ma.MaskedArray method), 239
identity (in module numpy.ma), 348, 925 isdecimal() (in module numpy.core.defchararray), 975
identity (numpy.ufunc attribute), 379 isdecimal() (numpy.chararray method), 134
identity() (in module numpy), 391 isdecimal() (numpy.core.defchararray.chararray method),
ids() (numpy.ma.MaskedArray method), 239 984
ifft() (in module numpy.fft), 517 isdigit() (in module numpy.core.defchararray), 975
ifft2() (in module numpy.fft), 520 isdigit() (numpy.chararray method), 134
ifftn() (in module numpy.fft), 523 isdigit() (numpy.core.defchararray.chararray method),
ifftshift() (in module numpy.fft), 534 984
ihfft() (in module numpy.fft), 532 isfinite (in module numpy), 670
iinfo (class in numpy), 487 isfortran() (in module numpy), 675
imag (numpy.generic attribute), 68 isinf (in module numpy), 671
imag (numpy.ma.MaskedArray attribute), 205 islower() (in module numpy.core.defchararray), 975
imag (numpy.ndarray attribute), 8, 36 islower() (numpy.chararray method), 134
imag() (in module numpy), 762 islower() (numpy.core.defchararray.chararray method),
import_array (C function), 1059 984
import_ufunc (C function), 1070 isnan (in module numpy), 672
in1d() (in module numpy), 800 isnative (numpy.dtype attribute), 90
index (C member), 1023, 1024 isneginf() (in module numpy), 673
index() (in module numpy.core.defchararray), 974 isnumeric() (in module numpy.core.defchararray), 976
index() (numpy.chararray method), 134 isnumeric() (numpy.chararray method), 135
index() (numpy.core.defchararray.chararray method), 984 isnumeric() (numpy.core.defchararray.chararray method),
indexing, 92, 96, 1077 985
indices() (in module numpy), 461 isposinf() (in module numpy), 673
indices() (in module numpy.ma), 364, 941 isreal() (in module numpy), 676
info() (in module numpy), 944 isrealobj() (in module numpy), 677
inner() (in module numpy), 538 isscalar() (in module numpy), 677
inner() (in module numpy.ma), 348, 925 issctype() (in module numpy), 489
innerproduct() (in module numpy.ma), 349, 926 isspace() (in module numpy.core.defchararray), 976
insert() (in module numpy), 447 isspace() (numpy.chararray method), 135
integ() (numpy.poly1d method), 779 isspace() (numpy.core.defchararray.chararray method),
interface 985
array, 366 issubclass_() (in module numpy), 490
interp() (in module numpy), 771 issubdtype() (in module numpy), 490
intersect1d() (in module numpy), 801 issubsctype() (in module numpy), 490
inv() (in module numpy.linalg), 562 istitle() (in module numpy.core.defchararray), 976
invert (in module numpy), 687 istitle() (numpy.chararray method), 135
ipmt() (in module numpy), 796 istitle() (numpy.core.defchararray.chararray method), 985
irfft() (in module numpy.fft), 527 isupper() (in module numpy.core.defchararray), 977
irfft2() (in module numpy.fft), 528 isupper() (numpy.chararray method), 135
irfftn() (in module numpy.fft), 530 isupper() (numpy.core.defchararray.chararray method),
irr() (in module numpy), 796 985
is_local_src_dir() (in module numpy.distutils.misc_util), item() (numpy.chararray method), 135
1000 item() (numpy.core.defchararray.chararray method), 985
isalnum() (numpy.chararray method), 134 item() (numpy.generic method), 73
isalnum() (numpy.core.defchararray.chararray method), item() (numpy.ma.MaskedArray method), 218
984 item() (numpy.matrix method), 108
isalpha() (in module numpy.core.defchararray), 974 item() (numpy.ndarray method), 19, 41
isalpha() (numpy.chararray method), 134 item() (numpy.recarray method), 156
isalpha() (numpy.core.defchararray.chararray method), item() (numpy.record method), 171
984 itemset() (numpy.generic method), 74

Index 1105
NumPy Reference, Release 2.0.0.dev8464

itemset() (numpy.matrix method), 109 logaddexp (in module numpy), 747

itemset() (numpy.ndarray method), 19, 42 logaddexp2 (in module numpy), 748
itemset() (numpy.recarray method), 157 logical_and (in module numpy), 678
itemset() (numpy.record method), 171 logical_not (in module numpy), 679
itemsize (C member), 1026 logical_or (in module numpy), 678
itemsize (numpy.dtype attribute), 87 logical_xor (in module numpy), 679
itemsize (numpy.generic attribute), 68 logistic() (in module numpy.random), 581
itemsize (numpy.ma.MaskedArray attribute), 202 logistic() (numpy.random.mtrand.RandomState method),
itemsize (numpy.ndarray attribute), 8, 34 621
iters (C member), 1024 lognormal() (in module numpy.random), 582
ix_() (in module numpy), 462 lognormal() (numpy.random.mtrand.RandomState
method), 622
J logseries() (in module numpy.random), 585
join() (in module numpy.core.defchararray), 964 logseries() (numpy.random.mtrand.RandomState
join() (numpy.chararray method), 136 method), 624
join() (numpy.core.defchararray.chararray method), 986 logspace() (in module numpy), 411
lookfor() (in module numpy), 943
K lower() (in module numpy.core.defchararray), 964
kaiser() (in module numpy), 812 lower() (numpy.chararray method), 136
keyword arguments lower() (numpy.core.defchararray.chararray method), 986
ufunc, 377 lstrip() (in module numpy.core.defchararray), 965
kind (C member), 1016 lstrip() (numpy.chararray method), 136
kind (numpy.dtype attribute), 87 lstrip() (numpy.core.defchararray.chararray method), 986
knownfailureif() (in module numpy.testing.decorators), lstsq() (in module numpy.linalg), 561
954
kron() (in module numpy), 544 M
MachAr (class in numpy), 488
L make_config_py() (numpy.distutils.misc_util.Configuration
laplace() (in module numpy.random), 580 method), 1008
laplace() (numpy.random.mtrand.RandomState method), make_mask() (in module numpy.ma), 285, 862
620 make_mask_descr() (in module numpy.ma), 287, 865
ldexp (in module numpy), 752 make_mask_none() (in module numpy.ma), 286, 863
left_shift (in module numpy), 688 make_svn_version_py() (numpy.distutils.misc_util.Configuration
len (C member), 1025, 1026 method), 1008
less (in module numpy), 682 mask (numpy.ma.masked_array attribute), 290, 867
less() (in module numpy.core.defchararray), 972 mask (numpy.ma.MaskedArray attribute), 199, 261, 838
less_equal (in module numpy), 683 mask_cols() (in module numpy.ma), 293, 870
less_equal() (in module numpy.core.defchararray), 972 mask_indices() (in module numpy), 465
lexsort() (in module numpy), 656 mask_or() (in module numpy.ma), 287, 293, 864, 871
LinAlgError, 565 mask_rowcols() (in module numpy.ma), 294, 871
linspace() (in module numpy), 410 mask_rows() (in module numpy.ma), 295, 873
listpickle (C member), 1020 masked (in module numpy.ma), 198
ljust() (in module numpy.core.defchararray), 964 masked arrays, 180
ljust() (numpy.chararray method), 136 masked_all() (in module numpy.ma), 249, 827
ljust() (numpy.core.defchararray.chararray method), 986 masked_all_like() (in module numpy.ma), 250, 828
load() (in module numpy), 494 masked_array (in module numpy.ma), 182, 244, 822
load() (in module numpy.ma), 313, 890 masked_equal() (in module numpy.ma), 185, 300, 877
load_library() (in module numpy.ctypeslib), 959 masked_greater() (in module numpy.ma), 185, 301, 878
loads() (in module numpy.ma), 314, 891 masked_greater_equal() (in module numpy.ma), 185,
loadtxt() (in module numpy), 404, 497 301, 878
log (in module numpy), 744 masked_inside() (in module numpy.ma), 186, 301, 878
log10 (in module numpy), 745 masked_invalid() (in module numpy.ma), 186, 302, 879
log1p (in module numpy), 746 masked_less() (in module numpy.ma), 187, 302, 879
log2 (in module numpy), 746

1106 Index
 NumPy Reference, Release 2.0.0.dev8464

masked_less_equal() (in module numpy.ma), 187, 303, min() (numpy.record method), 171
880 minimum (in module numpy), 769
masked_not_equal() (in module numpy.ma), 188, 303, mintypecode() (in module numpy), 493
880 mirr() (in module numpy), 797
masked_object() (in module numpy.ma), 188, 303, 880 mod (in module numpy), 759
masked_outside() (in module numpy.ma), 189, 304, 881 mod() (in module numpy.core.defchararray), 962
masked_print_options (in module numpy.ma), 198 modf (in module numpy), 759
masked_values() (in module numpy.ma), 189, 305, 882 mr_ (in module numpy.ma), 278, 856
masked_where() (in module numpy.ma), 191, 306, 883 msort() (in module numpy), 660
MaskedArray (class in numpy.ma), 198 multinomial() (in module numpy.random), 586
MaskType (in module numpy.ma), 244, 822 multinomial() (numpy.random.mtrand.RandomState
mat() (in module numpy), 419 method), 625
matrix, 97 multiply (in module numpy), 754
matrix (class in numpy), 98 multiply() (in module numpy.core.defchararray), 961
matrix_power() (in module numpy.linalg), 543 multivariate_normal() (in module numpy.random), 587
max (numpy.iinfo attribute), 488 multivariate_normal() (numpy.random.mtrand.RandomState
max() (in module numpy.ma), 339, 916 method), 626
max() (numpy.generic method), 74
max() (numpy.ma.MaskedArray method), 227, 341, 918 N
max() (numpy.matrix method), 109 name (C member), 1022
max() (numpy.ndarray method), 19 name (numpy.dtype attribute), 87
max() (numpy.recarray method), 157 names (numpy.dtype attribute), 89
max() (numpy.record method), 171 nan_to_num() (in module numpy), 769
maximum (in module numpy), 768 nanargmax() (in module numpy), 662
maximum_fill_value() (in module numpy.ma), 315, 316, nanargmin() (in module numpy), 662
892, 893 nanmax() (in module numpy), 694
mean (in module numpy.ma), 324, 901 nanmin() (in module numpy), 695
mean() (in module numpy), 698 nansum() (in module numpy), 734
mean() (numpy.generic method), 74 nargs (C member), 1021
mean() (numpy.ma.MaskedArray method), 228, 332, 909 nargs (numpy.ufunc attribute), 378
mean() (numpy.matrix method), 110 nbytes (numpy.ma.MaskedArray attribute), 203
mean() (numpy.ndarray method), 20, 58 nbytes (numpy.ndarray attribute), 9, 34
mean() (numpy.recarray method), 157 nd (C member), 1014, 1024, 1026
mean() (numpy.record method), 171 nd_m1 (C member), 1023
median() (in module numpy), 699 ndarray, 96
median() (in module numpy.ma), 325, 902 C-API, 1032, 1064
memmap (class in numpy), 124, 506 memory model, 1075
memory maps, 124 special methods getslice, 92
memory model special methods setslice, 92
ndarray, 1075 view, 93
meshgrid() (in module numpy), 413 ndarray (class in numpy), 4
methods ndenumerate (class in numpy), 178, 479
accumulate, ufunc, 1081 ndim (numpy.generic attribute), 68
reduce, ufunc, 1081 ndim (numpy.ma.MaskedArray attribute), 203
reduceat, ufunc, 1081 ndim (numpy.ndarray attribute), 9, 33
ufunc, 380 ndincr() (numpy.ndindex method), 480
mgrid (in module numpy), 414 ndindex (class in numpy), 479
min (numpy.iinfo attribute), 488 ndpointer() (in module numpy.ctypeslib), 959
min() (in module numpy.ma), 339, 916 negative (in module numpy), 753
min() (numpy.generic method), 74 negative_binomial() (in module numpy.random), 588
min() (numpy.ma.MaskedArray method), 229, 341, 918 negative_binomial() (numpy.random.mtrand.RandomState
min() (numpy.matrix method), 110 method), 627
min() (numpy.ndarray method), 20, 56 newaxis, 92
min() (numpy.recarray method), 157 newaxis (in module numpy), 93

Index 1107
NumPy Reference, Release 2.0.0.dev8464

newbuffer() (in module numpy), 946 NPY_BEGIN_THREADS (C macro), 1061

newbyteorder() (numpy.dtype method), 82, 90, 484 NPY_BEGIN_THREADS_DEF (C macro), 1061
newbyteorder() (numpy.generic method), 74 NPY_BEGIN_THREADS_DESCR (C function), 1061
newbyteorder() (numpy.matrix method), 111 NPY_BEHAVED (C variable), 1035, 1044
newbyteorder() (numpy.ndarray method), 20 NPY_BEHAVED_NS (C variable), 1036, 1045
newbyteorder() (numpy.recarray method), 157 NPY_BIG_ENDIAN (C variable), 1029
newbyteorder() (numpy.record method), 171 npy_bool (C type), 1031
next() (numpy.broadcast method), 179, 429 NPY_BUFSIZE (C variable), 1062
next() (numpy.flatiter method), 481 NPY_BYTE_ORDER (C variable), 1029
next() (numpy.ndenumerate method), 178, 479 NPY_C_CONTIGUOUS (C variable), 1035, 1044
next() (numpy.ndindex method), 480 NPY_CARRAY (C variable), 1035, 1044
nin (C member), 1021 NPY_CARRAY_RO (C variable), 1035, 1044
nin (numpy.ufunc attribute), 377 NPY_CLIP (C variable), 1048
NO_IMPORT_ARRAY (C macro), 1059 NPY_CLIPMODE (C type), 1064
NO_IMPORT_UFUNC (C variable), 1070 npy_copysign (C function), 1073
nomask (in module numpy.ma), 198 NPY_CPU_AMD64 (C variable), 1028
non-contiguous, 31 NPY_CPU_IA64 (C variable), 1028
noncentral_chisquare() (in module numpy.random), 589 NPY_CPU_PARISC (C variable), 1028
noncentral_chisquare() (numpy.random.mtrand.RandomStateNPY_CPU_PPC (C variable), 1028
method), 628 NPY_CPU_PPC64 (C variable), 1028
noncentral_f() (in module numpy.random), 591 NPY_CPU_S390 (C variable), 1028
noncentral_f() (numpy.random.mtrand.RandomState NPY_CPU_SPARC (C variable), 1028
method), 630 NPY_CPU_SPARC64 (C variable), 1028
nonzero (C member), 1019 NPY_CPU_X86 (C variable), 1028
nonzero (in module numpy.ma), 258, 836 NPY_DEFAULT (C variable), 1035, 1044
nonzero() (in module numpy), 459, 664 NPY_DISABLE_C_API (C macro), 1062
nonzero() (numpy.chararray method), 136 NPY_E (C variable), 1073
nonzero() (numpy.core.defchararray.chararray method), NPY_ELEMENTSTRIDES (C variable), 1036
986 NPY_END_ALLOW_THREADS (C macro), 1061
nonzero() (numpy.generic method), 75 NPY_END_THREADS (C macro), 1061
nonzero() (numpy.ma.MaskedArray method), 219, 262, NPY_END_THREADS_DESCR (C function), 1062
840 NPY_ENSUREARRAY (C variable), 1035, 1045
nonzero() (numpy.matrix method), 111 NPY_ENSURECOPY (C variable), 1035, 1045
nonzero() (numpy.ndarray method), 20, 54 NPY_EULER (C variable), 1074
nonzero() (numpy.recarray method), 158 NPY_F_CONTIGUOUS (C variable), 1035, 1044
nonzero() (numpy.record method), 172 NPY_FAIL (C variable), 1063
norm() (in module numpy.linalg), 554 NPY_FALSE (C variable), 1063
normal() (in module numpy.random), 592 NPY_FARRAY (C variable), 1035, 1044
normal() (numpy.random.mtrand.RandomState method), NPY_FARRAY_RO (C variable), 1035, 1044
631 NPY_FORCECAST (C variable), 1035, 1045
not_equal (in module numpy), 684 NPY_FROM_FIELDS (C variable), 1016
not_equal() (in module numpy.core.defchararray), 971 NPY_IN_ARRAY (C variable), 1036
notmasked_contiguous() (in module numpy.ma), 291, 869 NPY_IN_FARRAY (C variable), 1036
notmasked_edges() (in module numpy.ma), 292, 869 NPY_INFINITY (C variable), 1073
nout (C member), 1021 NPY_INOUT_ARRAY (C variable), 1036
nout (numpy.ufunc attribute), 377 NPY_INOUT_FARRAY (C variable), 1036
nper() (in module numpy), 797 npy_isfinite (C function), 1073
npv() (in module numpy), 793 npy_isinf (C function), 1073
NPY_1_PI (C variable), 1074 npy_isnan (C function), 1073
NPY_2_PI (C variable), 1074 NPY_ITEM_HASOBJECT (C variable), 1016
NPY_ALIGNED (C variable), 1035, 1044 NPY_ITEM_IS_POINTER (C variable), 1016
NPY_ALLOW_C_API (C macro), 1062 NPY_ITEM_LISTPICKLE (C variable), 1016
NPY_ALLOW_C_API_DEF (C macro), 1062 NPY_ITEM_REFCOUNT (C variable), 1016
NPY_BEGIN_ALLOW_THREADS (C macro), 1061 NPY_LITTLE_ENDIAN (C variable), 1028

1108 Index
 NumPy Reference, Release 2.0.0.dev8464

NPY_LOG10E (C variable), 1073 ntypes (numpy.ufunc attribute), 378

NPY_LOG2E (C variable), 1073 num (numpy.dtype attribute), 87
NPY_LOGE10 (C variable), 1073 numiter (C member), 1024
NPY_LOGE2 (C variable), 1073 numpy (module), 1
NPY_LOOP_BEGIN_THREADS (C macro), 1064 numpy.distutils (module), 999
NPY_LOOP_END_THREADS (C macro), 1064 numpy.distutils.exec_command (module), 1009
NPY_MAX_BUFSIZE (C variable), 1062 numpy.distutils.misc_util (module), 999
NPY_MAXDIMS (C variable), 1062 numpy.doc.internals (module), 1082
NPY_MIN_BUFSIZE (C variable), 1062 numpy.dual (module), 957
NPY_NAN (C variable), 1072 numpy.fft (module), 515
NPY_NEEDS_INIT (C variable), 1016 numpy.lib.scimath (module), 957
NPY_NEEDS_PYAPI (C variable), 1016 numpy.matlib (module), 957
npy_nextafter (C function), 1074 numpy.numarray (module), 958
NPY_NOTSWAPPED (C variable), 1036, 1045 numpy.oldnumeric (module), 958
NPY_NSCALARKINDS (C variable), 1064
NPY_NSORTS (C variable), 1064 O
NPY_NUM_FLOATTYPE (C variable), 1062 obj (C member), 1022
NPY_NZERO (C variable), 1073 obj2sctype() (in module numpy), 482
NPY_OBJECT_DTYPE_FLAGS (C variable), 1016 offset, 31
NPY_ORDER (C type), 1064 ogrid (in module numpy), 414, 462
NPY_OUT_ARRAY (C variable), 1036 ones (in module numpy.ma), 251, 829
NPY_OUT_FARRAY (C variable), 1036 ones() (in module numpy), 392
NPY_OWNDATA (C variable), 1044 ones_like (in module numpy), 392
NPY_PI (C variable), 1073 open() (numpy.DataSource method), 515
NPY_PI_2 (C variable), 1073 operation, 59, 235
NPY_PI_4 (C variable), 1073 operator, 59, 235
NPY_PRIOIRTY (C variable), 1062 outer() (in module numpy), 540
NPY_PZERO (C variable), 1073 outer() (in module numpy.ma), 350, 927
NPY_RAISE (C variable), 1048 outer() (numpy.ufunc method), 384
NPY_SCALAR_PRIORITY (C variable), 1062 outerproduct() (in module numpy.ma), 351, 928
NPY_SCALARKIND (C type), 1064
npy_signbit (C function), 1073 P
NPY_SIZEOF_DOUBLE (C variable), 1028 packbits() (in module numpy), 689
NPY_SIZEOF_FLOAT (C variable), 1028 pareto() (in module numpy.random), 593
NPY_SIZEOF_INT (C variable), 1028 pareto() (numpy.random.mtrand.RandomState method),
NPY_SIZEOF_LONG (C variable), 1028 633
NPY_SIZEOF_LONG_DOUBLE (C variable), 1028 paths() (numpy.distutils.misc_util.Configuration method),
NPY_SIZEOF_LONG_LONG (C variable), 1028 1007
NPY_SIZEOF_PY_INTPTR_T (C variable), 1028 permutation() (in module numpy.random), 570
NPY_SIZEOF_PY_LONG_LONG (C variable), 1028 permutation() (numpy.random.mtrand.RandomState
NPY_SIZEOF_SHORT (C variable), 1028 method), 634
NPY_SORTKIND (C type), 1064 piecewise() (in module numpy), 776
npy_spacing (C function), 1074 pinv() (in module numpy.linalg), 563
NPY_SUBTYPE_PRIORITY (C variable), 1062 place() (in module numpy), 476
NPY_SUCCEED (C variable), 1063 pmt() (in module numpy), 794
NPY_TRUE (C variable), 1063 poisson() (in module numpy.random), 595
NPY_UPDATE_ALL (C variable), 1045 poisson() (numpy.random.mtrand.RandomState method),
NPY_UPDATEIFCOPY (C variable), 1035, 1044 635
NPY_USE_GETITEM (C variable), 1016 poly() (in module numpy), 780
NPY_USE_SETITEM (C variable), 1016 poly1d (class in numpy), 777
NPY_VERSION (C variable), 1063 polyadd() (in module numpy), 787
NPY_WRAP (C variable), 1048 polyder() (in module numpy), 785
NPY_WRITEABLE (C variable), 1035, 1044 polydiv() (in module numpy), 788
ntypes (C member), 1021 polyfit() (in module numpy), 782

Index 1109
NumPy Reference, Release 2.0.0.dev8464

polyfit() (in module numpy.ma), 355, 932 PyArray_AsCArray (C function), 1051

polyint() (in module numpy), 786 PyArray_AxisConverter (C function), 1058
polymul() (in module numpy), 789 PyArray_BASE (C function), 1032
polysub() (in module numpy), 789 PyArray_BoolConverter (C function), 1058
polyval() (in module numpy), 779 PyArray_Broadcast (C function), 1053
power (in module numpy), 755 PyArray_BroadcastToShape (C function), 1052
power() (in module numpy.ma), 326, 903 PyArray_BufferConverter (C function), 1058
power() (in module numpy.random), 596 PyArray_ByteorderConverter (C function), 1058
power() (numpy.random.mtrand.RandomState method), PyArray_BYTES (C function), 1032
636 PyArray_Byteswap (C function), 1046
ppmt() (in module numpy), 795 PyArray_CanCastSafely (C function), 1041
pprint() (numpy.record method), 172 PyArray_CanCastTo (C function), 1041
prod (in module numpy.ma), 326, 903 PyArray_CanCoerceScalar (C function), 1056
prod() (in module numpy), 731 PyArray_Cast (C function), 1041
prod() (numpy.generic method), 75 PyArray_CastScalarToCtype (C function), 1055
prod() (numpy.ma.MaskedArray method), 229, 333, 910 PyArray_CastTo (C function), 1041
prod() (numpy.matrix method), 111 PyArray_CastToType (C function), 1041
prod() (numpy.ndarray method), 21, 58 PyArray_CEQ (C function), 1063
prod() (numpy.recarray method), 158 PyArray_CGE (C function), 1063
prod() (numpy.record method), 172 PyArray_CGT (C function), 1063
product() (numpy.ma.MaskedArray method), 230 PyArray_Check (C function), 1038
protocol PyArray_CheckAxis (C function), 1038
array, 366 PyArray_CheckExact (C function), 1038
ptp() (in module numpy), 696 PyArray_CheckFromAny (C function), 1036
ptp() (in module numpy.ma), 340, 917 PyArray_CheckScalar (C function), 1038
ptp() (numpy.generic method), 75 PyArray_CheckStrides (C function), 1052
ptp() (numpy.ma.MaskedArray method), 231, 342, 919 PyArray_CHKFLAGS (C function), 1045
ptp() (numpy.matrix method), 112 PyArray_Choose (C function), 1048
ptp() (numpy.ndarray method), 21, 56 PyArray_Chunk (C type), 1025
ptp() (numpy.recarray method), 158 PyArray_CLE (C function), 1063
ptp() (numpy.record method), 172 PyArray_Clip (C function), 1050
ptr (C member), 1022, 1025, 1026 PyArray_CLT (C function), 1063
put() (in module numpy), 476 PyArray_CNE (C function), 1063
put() (numpy.chararray method), 136 PyArray_CompareLists (C function), 1052
put() (numpy.core.defchararray.chararray method), 986 PyArray_Compress (C function), 1049
put() (numpy.generic method), 75 PyArray_Concatenate (C function), 1051
put() (numpy.ma.MaskedArray method), 220 PyArray_Conjugate (C function), 1050
put() (numpy.matrix method), 112 PyArray_ContiguousFromAny (C function), 1036
put() (numpy.ndarray method), 21, 52 PyArray_Converter (C function), 1057
put() (numpy.recarray method), 158 PyArray_ConvertToCommonType (C function), 1042
put() (numpy.record method), 172 PyArray_CopyAndTranspose (C function), 1051
putmask() (in module numpy), 477 PyArray_CopyInto (C function), 1037
pv() (in module numpy), 792 PyArray_Correlate (C function), 1051
PY_ARRAY_UNIQUE_SYMBOL (C macro), 1059 PyArray_Correlate2 (C function), 1052
PY_UFUNC_UNIQUE_SYMBOL (C variable), 1070 PyArray_CumProd (C function), 1050
PyArray_All (C function), 1050 PyArray_CumSum (C function), 1050
PyArray_Any (C function), 1050 PyArray_DATA (C function), 1032
PyArray_Arange (C function), 1034 PyArray_DESCR (C function), 1032
PyArray_ArangeObj (C function), 1034 PyArray_Descr (C type), 1015
PyArray_ArgMax (C function), 1049 Pyarray_DescrAlignConverter (C function), 1057
PyArray_ArgMin (C function), 1049 Pyarray_DescrAlignConverter2 (C function), 1057
PyArray_ArgSort (C function), 1048 PyArray_DescrConverter (C function), 1057
PyArray_ArrayType (C function), 1042 PyArray_DescrConverter2 (C function), 1057
PyArray_ArrFuncs (C type), 1017 PyArray_DescrFromObject (C function), 1056

1110 Index
 NumPy Reference, Release 2.0.0.dev8464

PyArray_DescrFromScalar (C function), 1057 PyArray_GETPTR3 (C function), 1033

PyArray_DescrFromType (C function), 1057 PyArray_GETPTR4 (C function), 1033
PyArray_DescrNew (C function), 1056 PyArray_HasArrayInterface (C function), 1038
PyArray_DescrNewByteorder (C function), 1056 PyArray_HasArrayInterfaceType (C function), 1038
PyArray_DescrNewFromType (C function), 1056 PyArray_HASFIELDS (C function), 1041
PyArray_Diagonal (C function), 1049 PyArray_INCREF (C function), 1043
PyArray_DIM (C function), 1032 PyArray_InitArrFuncs (C function), 1042
PyArray_DIMS (C function), 1032 PyArray_InnerProduct (C function), 1051
PyArray_Dims (C type), 1025 PyArray_IntpConverter (C function), 1058
PyArray_Dump (C function), 1046 PyArray_IntpFromSequence (C function), 1058
PyArray_Dumps (C function), 1047 PyArray_ISALIGNED (C function), 1045
PyArray_EMPTY (C function), 1034 PyArray_IsAnyScalar (C function), 1039
PyArray_Empty (C function), 1034 PyArray_ISBEHAVED (C function), 1045
PyArray_EnsureArray (C function), 1037 PyArray_ISBEHAVED_RO (C function), 1045
PyArray_EquivArrTypes (C function), 1041 PyArray_ISBOOL (C function), 1040
PyArray_EquivByteorders (C function), 1041 PyArray_ISBYTESWAPPED (C function), 1041
PyArray_EquivTypenums (C function), 1041 PyArray_ISCARRAY (C function), 1045
PyArray_EquivTypes (C function), 1041 PyArray_ISCARRAY_RO (C function), 1045
PyArray_FieldNames (C function), 1057 PyArray_ISCOMPLEX (C function), 1039
PyArray_FillObjectArray (C function), 1043 PyArray_ISCONTIGUOUS (C function), 1045
PyArray_FILLWBYTE (C function), 1034 PyArray_ISEXTENDED (C function), 1040
PyArray_FillWithScalar (C function), 1047 PyArray_ISFARRAY (C function), 1045
PyArray_FLAGS (C function), 1032 PyArray_ISFARRAY_RO (C function), 1046
PyArray_Flatten (C function), 1048 PyArray_ISFLEXIBLE (C function), 1040
PyArray_Free (C function), 1051 PyArray_ISFLOAT (C function), 1039
PyArray_free (C function), 1061 PyArray_ISFORTRAN (C function), 1045
PyArray_FROM_O (C function), 1037 PyArray_ISINTEGER (C function), 1039
PyArray_FROM_OF (C function), 1037 PyArray_ISNOTSWAPPED (C function), 1041
PyArray_FROM_OT (C function), 1038 PyArray_ISNUMBER (C function), 1039
PyArray_FROM_OTF (C function), 1038 PyArray_ISOBJECT (C function), 1040
PyArray_FROMANY (C function), 1038 PyArray_ISONESEGMENT (C function), 1046
PyArray_FromAny (C function), 1034 PyArray_ISPYTHON (C function), 1040
PyArray_FromArray (C function), 1036 PyArray_IsPythonScalar (C function), 1038
PyArray_FromArrayAttr (C function), 1036 PyArray_IsScalar (C function), 1038
PyArray_FromBuffer (C function), 1037 PyArray_ISSIGNED (C function), 1039
PyArray_FromFile (C function), 1037 PyArray_ISSTRING (C function), 1040
PyArray_FromInterface (C function), 1036 PyArray_ISUNSIGNED (C function), 1039
PyArray_FromObject (C function), 1037 PyArray_ISUSERDEF (C function), 1040
PyArray_FromScalar (C function), 1055 PyArray_ISWRITEABLE (C function), 1045
PyArray_FromString (C function), 1037 PyArray_IsZeroDim (C function), 1038
PyArray_FromStructInterface (C function), 1036 PyArray_Item_INCREF (C function), 1043
PyArray_GetCastFunc (C function), 1041 PyArray_Item_XDECREF (C function), 1043
PyArray_GETCONTIGUOUS (C function), 1037 PyArray_ITEMSIZE (C function), 1032
PyArray_GetEndianness (C function), 1029 PyArray_ITER_DATA (C function), 1053
PyArray_GetField (C function), 1046 PyArray_ITER_GOTO (C function), 1053
PyArray_GETITEM (C function), 1032 PyArray_ITER_GOTO1D (C function), 1053
PyArray_GetNDArrayCFeatureVersion (C function), PyArray_ITER_NEXT (C function), 1053
1060 PyArray_ITER_NOTDONE (C function), 1053
PyArray_GetNDArrayCVersion (C function), 1060 PyArray_ITER_RESET (C function), 1053
PyArray_GetNumericOps (C function), 1060 PyArray_IterAllButAxis (C function), 1052
PyArray_GetPriority (C function), 1062 PyArray_IterNew (C function), 1052
PyArray_GetPtr (C function), 1033 PyArray_LexSort (C function), 1049
PyArray_GETPTR1 (C function), 1033 PyArray_malloc (C function), 1061
PyArray_GETPTR2 (C function), 1033 PyArray_MatrixProduct (C function), 1051

Index 1111
NumPy Reference, Release 2.0.0.dev8464

PyArray_MAX (C function), 1063 PyArray_SimpleNew (C function), 1034

PyArray_Max (C function), 1049 PyArray_SimpleNewFromData (C function), 1034
PyArray_Mean (C function), 1050 PyArray_SimpleNewFromDescr (C function), 1034
PyArray_MIN (C function), 1063 PyArray_SIZE (C function), 1032
PyArray_Min (C function), 1049 PyArray_Size (C function), 1032
PyArray_MoveInto (C function), 1037 PyArray_Sort (C function), 1048
PyArray_MultiIter_DATA (C function), 1053 PyArray_SortkindConverter (C function), 1058
PyArray_MultiIter_GOTO (C function), 1053 PyArray_Squeeze (C function), 1047
PyArray_MultiIter_GOTO1D (C function), 1053 PyArray_Std (C function), 1050
PyArray_MultiIter_NEXT (C function), 1053 PyArray_STRIDE (C function), 1032
PyArray_MultiIter_NEXTi (C function), 1053 PyArray_STRIDES (C function), 1032
PyArray_MultiIter_NOTDONE (C function), 1053 PyArray_Sum (C function), 1050
PyArray_MultiIter_RESET (C function), 1053 PyArray_SwapAxes (C function), 1047
PyArray_MultiIterNew (C function), 1053 PyArray_TakeFrom (C function), 1048
PyArray_MultiplyIntList (C function), 1052 PyArray_ToFile (C function), 1046
PyArray_MultiplyList (C function), 1052 PyArray_ToList (C function), 1046
PyArray_NBYTES (C function), 1033 PyArray_ToScalar (C function), 1055
PyArray_NeighborhoodIterNew (C function), 1054 PyArray_ToString (C function), 1046
PyArray_New (C function), 1033 PyArray_Trace (C function), 1050
PyArray_NewCopy (C function), 1046 PyArray_Transpose (C function), 1047
PyArray_NewFromDescr (C function), 1033 PyArray_TYPE (C function), 1032
PyArray_Newshape (C function), 1047 PyArray_Type (C variable), 1014
PyArray_Nonzero (C function), 1049 PyArray_TypeObjectFromType (C function), 1056
PyArray_ObjectType (C function), 1042 PyArray_TypestrConvert (C function), 1058
PyArray_One (C function), 1042 PyArray_UpdateFlags (C function), 1046
PyArray_OutputConverter (C function), 1057 PyArray_ValidType (C function), 1042
PyArray_Prod (C function), 1050 PyArray_View (C function), 1047
PyArray_Ptp (C function), 1049 PyArray_Where (C function), 1052
PyArray_PutMask (C function), 1048 PyArray_XDECREF (C function), 1043
PyArray_PutTo (C function), 1048 PyArray_XDECREF_ERR (C function), 1063
PyArray_PyIntAsInt (C function), 1058 PyArray_Zero (C function), 1042
PyArray_PyIntAsIntp (C function), 1058 PyArray_ZEROS (C function), 1034
PyArray_Ravel (C function), 1048 PyArray_Zeros (C function), 1034
PyArray_realloc (C function), 1061 PyArrayDescr_Check (C function), 1056
PyArray_REFCOUNT (C function), 1063 PyArrayDescr_Type (C variable), 1015
PyArray_RegisterCanCast (C function), 1043 PyArrayFlags_Type (C variable), 1024
PyArray_RegisterCastFunc (C function), 1043 PyArrayInterface (C type), 1026
PyArray_RegisterDataType (C function), 1042 PyArrayIter_Check (C function), 1053
PyArray_RemoveSmallest (C function), 1054 PyArrayIter_Type (C variable), 1022
PyArray_Repeat (C function), 1048 PyArrayIterObject (C type), 1022
PyArray_Reshape (C function), 1047 PyArrayMapIter_Type (C variable), 1027
PyArray_Resize (C function), 1047 PyArrayMultiIter_Type (C variable), 1023
PyArray_Return (C function), 1055 PyArrayMultiIterObject (C type), 1023
PyArray_Round (C function), 1050 PyArrayNeighborhoodIter_Next (C function), 1055
PyArray_SAMESHAPE (C function), 1063 PyArrayNeighborhoodIter_Reset (C function), 1055
PyArray_Scalar (C function), 1055 PyArrayNeighborhoodIter_Type (C variable), 1024
PyArray_ScalarAsCtype (C function), 1055 PyArrayNeighborhoodIterObject (C type), 1024
PyArray_ScalarKind (C function), 1056 PyArrayObject (C type), 1014
PyArray_SearchsideConverter (C function), 1058 PyDataMem_FREE (C function), 1060
PyArray_SearchSorted (C function), 1049 PyDataMem_NEW (C function), 1060
PyArray_SetField (C function), 1046 PyDataMem_RENEW (C function), 1060
PyArray_SETITEM (C function), 1032 PyDataType_FLAGCHK (C function), 1016
PyArray_SetNumericOps (C function), 1060 PyDataType_HASFIELDS (C function), 1041
PyArray_SetStringFunction (C function), 1060 PyDataType_ISBOOL (C function), 1040

1112 Index
 NumPy Reference, Release 2.0.0.dev8464

PyDataType_ISCOMPLEX (C function), 1039 randn() (in module numpy.random), 566

PyDataType_ISEXTENDED (C function), 1040 randn() (numpy.random.mtrand.RandomState method),
PyDataType_ISFLEXIBLE (C function), 1040 638
PyDataType_ISFLOAT (C function), 1039 random_integers() (in module numpy.random), 567
PyDataType_ISINTEGER (C function), 1039 random_integers() (numpy.random.mtrand.RandomState
PyDataType_ISNUMBER (C function), 1039 method), 639
PyDataType_ISOBJECT (C function), 1040 random_sample() (in module numpy.random), 569
PyDataType_ISPYTHON (C function), 1040 random_sample() (numpy.random.mtrand.RandomState
PyDataType_ISSIGNED (C function), 1039 method), 640
PyDataType_ISSTRING (C function), 1040 RandomState (class in numpy.random.mtrand), 608
PyDataType_ISUNSIGNED (C function), 1039 RankWarning, 790
PyDataType_ISUSERDEF (C function), 1040 rate() (in module numpy), 798
PyDataType_REFCHK (C function), 1016 ravel (in module numpy.ma), 266, 843
PyDimMem_FREE (C function), 1061 ravel() (in module numpy), 422
PyDimMem_NEW (C function), 1061 ravel() (numpy.chararray method), 137
PyDimMem_RENEW (C function), 1061 ravel() (numpy.core.defchararray.chararray method), 987
PyObject_HEAD (C macro), 1014, 1021, 1024, 1025 ravel() (numpy.generic method), 75
Python Enhancement Proposals ravel() (numpy.ma.MaskedArray method), 211, 268, 845
PEP 3118, 366 ravel() (numpy.matrix method), 112
PyTypeNum_ISBOOL (C function), 1040 ravel() (numpy.ndarray method), 21, 51
PyTypeNum_ISCOMPLEX (C function), 1039 ravel() (numpy.recarray method), 158
PyTypeNum_ISEXTENDED (C function), 1040 ravel() (numpy.record method), 172
PyTypeNum_ISFLEXIBLE (C function), 1040 rayleigh() (in module numpy.random), 597
PyTypeNum_ISFLOAT (C function), 1039 rayleigh() (numpy.random.mtrand.RandomState method),
PyTypeNum_ISINTEGER (C function), 1039 641
PyTypeNum_ISNUMBER (C function), 1039 real (numpy.generic attribute), 68
PyTypeNum_ISOBJECT (C function), 1040 real (numpy.ma.MaskedArray attribute), 205
PyTypeNum_ISPYTHON (C function), 1040 real (numpy.ndarray attribute), 8, 36
PyTypeNum_ISSIGNED (C function), 1039 real() (in module numpy), 761
PyTypeNum_ISSTRING (C function), 1040 real_if_close() (in module numpy), 770
PyTypeNum_ISUNSIGNED (C function), 1039 recarray (class in numpy), 148
PyTypeNum_ISUSERDEF (C function), 1040 reciprocal (in module numpy), 753
PyUFunc_checkfperr (C function), 1066 record
PyUFunc_clearfperr (C function), 1066 dtype, 79
PyUFunc_Loop1d (C type), 1027 record (class in numpy), 167
PyUFunc_PyFuncData (C type), 1069 recordmask (numpy.ma.MaskedArray attribute), 199,
PyUFunc_Type (C variable), 1020 261, 839
PyUFuncLoopObject (C type), 1027 red_text() (in module numpy.distutils.misc_util), 1000
PyUFuncObject (C type), 1020 reduce
PyUFuncReduceObject (C type), 1027 ufunc methods, 1081
reduce() (numpy.ufunc method), 380
Q reduceat
qr() (in module numpy.linalg), 546 ufunc methods, 1081
reduceat() (numpy.ufunc method), 382
R remainder (in module numpy), 760
r_ (in module numpy), 457 repeat() (in module numpy), 446
rad2deg (in module numpy), 722 repeat() (numpy.chararray method), 137
radians (in module numpy), 720 repeat() (numpy.core.defchararray.chararray method),
rand() (in module numpy.random), 565 987
rand() (numpy.random.mtrand.RandomState method), repeat() (numpy.generic method), 75
637 repeat() (numpy.ma.MaskedArray method), 221
randint() (in module numpy.random), 566 repeat() (numpy.matrix method), 113
randint() (numpy.random.mtrand.RandomState method), repeat() (numpy.ndarray method), 21, 52
637 repeat() (numpy.recarray method), 159

Index 1113
NumPy Reference, Release 2.0.0.dev8464

repeat() (numpy.record method), 173 round() (numpy.recarray method), 160

replace() (in module numpy.core.defchararray), 966 round() (numpy.record method), 173
replace() (numpy.chararray method), 137 round_() (in module numpy), 729
replace() (numpy.core.defchararray.chararray method), row-major, 31
987 row_stack (in module numpy.ma), 278, 856
require() (in module numpy), 435 rsplit() (in module numpy.core.defchararray), 966
reset() (numpy.broadcast method), 179, 429 rsplit() (numpy.chararray method), 139
reshape() (in module numpy), 421, 453 rsplit() (numpy.core.defchararray.chararray method), 989
reshape() (in module numpy.ma), 266, 844 rstrip() (in module numpy.core.defchararray), 967
reshape() (numpy.chararray method), 137 rstrip() (numpy.chararray method), 139
reshape() (numpy.core.defchararray.chararray method), rstrip() (numpy.core.defchararray.chararray method), 989
987 run_module_suite() (in module numpy.testing), 956
reshape() (numpy.generic method), 75 rundocs() (in module numpy.testing), 956
reshape() (numpy.ma.MaskedArray method), 212, 268,
846 S
reshape() (numpy.matrix method), 113 s_ (in module numpy), 458
reshape() (numpy.ndarray method), 22, 48 save() (in module numpy), 495
reshape() (numpy.recarray method), 159 savetxt() (in module numpy), 498
reshape() (numpy.record method), 173 savez() (in module numpy), 495
resize() (in module numpy), 449 scalar
resize() (in module numpy.ma), 266, 844 dtype, 79
resize() (numpy.chararray method), 137 scalarkind (C member), 1020
resize() (numpy.core.defchararray.chararray method), 987 scanfunc (C member), 1019
resize() (numpy.generic method), 76 sctype2char() (in module numpy), 492
resize() (numpy.ma.MaskedArray method), 213, 269, 846 searchsorted() (in module numpy), 666
resize() (numpy.matrix method), 113 searchsorted() (numpy.chararray method), 139
resize() (numpy.ndarray method), 22, 48 searchsorted() (numpy.core.defchararray.chararray
resize() (numpy.recarray method), 159 method), 989
resize() (numpy.record method), 173 searchsorted() (numpy.generic method), 76
restoredot() (in module numpy), 946 searchsorted() (numpy.ma.MaskedArray method), 221
rfft() (in module numpy.fft), 525 searchsorted() (numpy.matrix method), 114
rfft2() (in module numpy.fft), 528 searchsorted() (numpy.ndarray method), 23, 54
rfftn() (in module numpy.fft), 529 searchsorted() (numpy.recarray method), 160
rfind() (in module numpy.core.defchararray), 977 searchsorted() (numpy.record method), 173
rfind() (numpy.chararray method), 138 seed() (in module numpy.random), 653
rfind() (numpy.core.defchararray.chararray method), 988 seed() (numpy.random.mtrand.RandomState method),
right_shift (in module numpy), 689 642
rindex() (in module numpy.core.defchararray), 977 select() (in module numpy), 475
rindex() (numpy.chararray method), 139 set_fill_value() (in module numpy.ma), 317, 894
rindex() (numpy.core.defchararray.chararray method), set_fill_value() (numpy.ma.MaskedArray method), 243,
989 318, 895
rint (in module numpy), 729 set_printoptions() (in module numpy), 509
rjust() (in module numpy.core.defchararray), 966 set_state() (in module numpy.random), 654
rjust() (numpy.chararray method), 139 set_state() (numpy.random.mtrand.RandomState
rjust() (numpy.core.defchararray.chararray method), 989 method), 642
roll() (in module numpy), 454 set_string_function() (in module numpy), 511
rollaxis() (in module numpy), 424 set_verbosity() (in module numpy.distutils.log), 1008
roots() (in module numpy), 782 setastest() (in module numpy.testing.decorators), 955
rot90() (in module numpy), 455 setbufsize() (in module numpy), 372, 946
round() (in module numpy.ma), 358, 935 setdiff1d() (in module numpy), 802
round() (numpy.generic method), 76 seterr() (in module numpy), 372, 815
round() (numpy.ma.MaskedArray method), 231, 359, 936 seterrcall() (in module numpy), 374, 817
round() (numpy.matrix method), 114 seterrobj() (in module numpy), 820
round() (numpy.ndarray method), 23, 57 setfield() (numpy.chararray method), 139

1114 Index
 NumPy Reference, Release 2.0.0.dev8464

setfield() (numpy.core.defchararray.chararray method), sort() (numpy.core.defchararray.chararray method), 991

989 sort() (numpy.generic method), 76
setfield() (numpy.generic method), 76 sort() (numpy.ma.MaskedArray method), 221, 345, 922
setfield() (numpy.matrix method), 115 sort() (numpy.matrix method), 116
setfield() (numpy.ndarray method), 23 sort() (numpy.ndarray method), 25, 53, 659
setfield() (numpy.recarray method), 161 sort() (numpy.recarray method), 162
setfield() (numpy.record method), 173 sort() (numpy.record method), 173
setflags() (numpy.chararray method), 140 sort_complex() (in module numpy), 660
setflags() (numpy.core.defchararray.chararray method), source() (in module numpy), 944
990 special methods
setflags() (numpy.generic method), 76, 79 getslice, ndarray, 92
setflags() (numpy.matrix method), 115 setslice, ndarray, 92
setflags() (numpy.ndarray method), 24, 47 split() (in module numpy), 443
setflags() (numpy.recarray method), 161 split() (in module numpy.core.defchararray), 967
setflags() (numpy.record method), 173 split() (numpy.chararray method), 142
setitem (C member), 1018 split() (numpy.core.defchararray.chararray method), 992
setslice splitlines() (in module numpy.core.defchararray), 968
ndarray special methods, 92 splitlines() (numpy.chararray method), 142
setxor1d() (in module numpy), 802 splitlines() (numpy.core.defchararray.chararray method),
shape (C member), 1017, 1027 992
shape (numpy.dtype attribute), 89 sqrt (in module numpy), 765
shape (numpy.generic attribute), 68 square (in module numpy), 765
shape (numpy.ma.MaskedArray attribute), 203 squeeze() (in module numpy), 430
shape (numpy.ndarray attribute), 9, 32 squeeze() (in module numpy.ma), 274, 851
shape() (in module numpy.ma), 259, 264, 837, 842 squeeze() (numpy.chararray method), 142
sharedmask (numpy.ma.MaskedArray attribute), 199 squeeze() (numpy.core.defchararray.chararray method),
shrink_mask() (numpy.ma.MaskedArray method), 242, 992
297, 874 squeeze() (numpy.generic method), 77
shuffle() (in module numpy.random), 570 squeeze() (numpy.ma.MaskedArray method), 213, 274,
shuffle() (numpy.random.mtrand.RandomState method), 852
642 squeeze() (numpy.matrix method), 117
sign (in module numpy), 767 squeeze() (numpy.ndarray method), 26, 52
signbit (in module numpy), 751 squeeze() (numpy.recarray method), 163
sin (in module numpy), 712 squeeze() (numpy.record method), 174
sinc() (in module numpy), 749 standard_cauchy() (in module numpy.random), 597
single-segment, 31 standard_cauchy() (numpy.random.mtrand.RandomState
sinh (in module numpy), 723 method), 642
size (C member), 1023, 1024 standard_exponential() (in module numpy.random), 598
size (numpy.generic attribute), 68 standard_exponential() (numpy.random.mtrand.RandomState
size (numpy.ma.MaskedArray attribute), 203 method), 643
size (numpy.ndarray attribute), 8, 34 standard_gamma() (in module numpy.random), 599
size() (in module numpy.ma), 260, 265, 838, 842 standard_gamma() (numpy.random.mtrand.RandomState
skipif() (in module numpy.testing.decorators), 955 method), 644
slicing, 92 standard_normal() (in module numpy.random), 599
slogdet() (in module numpy.linalg), 557 standard_normal() (numpy.random.mtrand.RandomState
slow() (in module numpy.testing.decorators), 955 method), 644
soften_mask (in module numpy.ma), 296, 874 standard_t() (in module numpy.random), 600
soften_mask() (numpy.ma.MaskedArray method), 241, standard_t() (numpy.random.mtrand.RandomState
297, 874 method), 645
solve() (in module numpy.linalg), 559 startswith() (in module numpy.core.defchararray), 977
sort (C member), 1019 startswith() (numpy.chararray method), 142
sort() (in module numpy), 655 startswith() (numpy.core.defchararray.chararray method),
sort() (in module numpy.ma), 343, 920 992
sort() (numpy.chararray method), 141 std (in module numpy.ma), 327, 904

Index 1115
NumPy Reference, Release 2.0.0.dev8464

std() (in module numpy), 700 take() (numpy.core.defchararray.chararray method), 993

std() (numpy.generic method), 77 take() (numpy.generic method), 77
std() (numpy.ma.MaskedArray method), 231, 334, 911 take() (numpy.ma.MaskedArray method), 222
std() (numpy.matrix method), 118 take() (numpy.matrix method), 119
std() (numpy.ndarray method), 26, 58 take() (numpy.ndarray method), 27, 52
std() (numpy.recarray method), 163 take() (numpy.recarray method), 164
std() (numpy.record method), 174 take() (numpy.record method), 174
str (numpy.dtype attribute), 87 tan (in module numpy), 714
stride, 31 tanh (in module numpy), 725
strides (C member), 1015, 1023, 1027 tensordot() (in module numpy), 541
strides (numpy.generic attribute), 68 tensorinv() (in module numpy.linalg), 564
strides (numpy.ma.MaskedArray attribute), 204 tensorsolve() (in module numpy.linalg), 560
strides (numpy.ndarray attribute), 9, 32 terminal_has_colors() (in module
strip() (in module numpy.core.defchararray), 968 numpy.distutils.misc_util), 1000
strip() (numpy.chararray method), 142 Tester (in module numpy.testing), 956
strip() (numpy.core.defchararray.chararray method), 992 tile() (in module numpy), 445
sub-array title() (in module numpy.core.defchararray), 969
dtype, 79, 85 title() (numpy.chararray method), 143
subarray (C member), 1017 title() (numpy.core.defchararray.chararray method), 993
subdtype (numpy.dtype attribute), 89 todict() (numpy.distutils.misc_util.Configuration
subtract (in module numpy), 756 method), 1001
sum (in module numpy.ma), 328, 905 tofile() (numpy.chararray method), 143
sum() (in module numpy), 733 tofile() (numpy.core.defchararray.chararray method), 993
sum() (numpy.generic method), 77 tofile() (numpy.generic method), 77
sum() (numpy.ma.MaskedArray method), 233, 336, 913 tofile() (numpy.ma.MaskedArray method), 208, 311, 888
sum() (numpy.matrix method), 118 tofile() (numpy.matrix method), 119
sum() (numpy.ndarray method), 27, 57 tofile() (numpy.ndarray method), 27, 43, 504
sum() (numpy.recarray method), 164 tofile() (numpy.recarray method), 164
sum() (numpy.record method), 174 tofile() (numpy.record method), 174
svd() (in module numpy.linalg), 547 toflex() (numpy.ma.MaskedArray method), 208
swapaxes (in module numpy.ma), 269, 847 tolist() (numpy.chararray method), 143
swapaxes() (in module numpy), 424 tolist() (numpy.core.defchararray.chararray method), 993
swapaxes() (numpy.chararray method), 142 tolist() (numpy.generic method), 77
swapaxes() (numpy.core.defchararray.chararray method), tolist() (numpy.ma.MaskedArray method), 209, 311, 888
992 tolist() (numpy.matrix method), 119
swapaxes() (numpy.generic method), 77 tolist() (numpy.ndarray method), 27, 42, 505
swapaxes() (numpy.ma.MaskedArray method), 213, 270, tolist() (numpy.recarray method), 164
847 tolist() (numpy.record method), 175
swapaxes() (numpy.matrix method), 118 tomaxint() (numpy.random.mtrand.RandomState
swapaxes() (numpy.ndarray method), 27, 51 method), 645
swapaxes() (numpy.recarray method), 164 torecords() (numpy.ma.MaskedArray method), 209, 312,
swapaxes() (numpy.record method), 174 889
swapcase() (in module numpy.core.defchararray), 969 tostring() (numpy.chararray method), 144
swapcase() (numpy.chararray method), 143 tostring() (numpy.core.defchararray.chararray method),
swapcase() (numpy.core.defchararray.chararray method), 994
993 tostring() (numpy.generic method), 78
tostring() (numpy.ma.MaskedArray method), 210, 312,
T 889
T (numpy.generic attribute), 68 tostring() (numpy.matrix method), 120
T (numpy.ma.MaskedArray attribute), 214 tostring() (numpy.ndarray method), 28, 42
T (numpy.matrix attribute), 97 tostring() (numpy.recarray method), 165
T (numpy.ndarray attribute), 6, 35, 425 tostring() (numpy.record method), 175
take() (in module numpy), 469 trace (in module numpy.ma), 352, 929
take() (numpy.chararray method), 143 trace() (in module numpy), 558

1116 Index
 NumPy Reference, Release 2.0.0.dev8464

trace() (numpy.generic method), 78 methods reduce, 1081

trace() (numpy.ma.MaskedArray method), 233, 353, 930 methods reduceat, 1081
trace() (numpy.matrix method), 120 UFUNC_CHECK_ERROR (C function), 1064
trace() (numpy.ndarray method), 28, 57 UFUNC_CHECK_STATUS (C function), 1065
trace() (numpy.recarray method), 165 uniform() (in module numpy.random), 602
trace() (numpy.record method), 175 uniform() (numpy.random.mtrand.RandomState method),
translate() (in module numpy.core.defchararray), 969 647
translate() (numpy.chararray method), 144 union1d() (in module numpy), 803
translate() (numpy.core.defchararray.chararray method), unique() (in module numpy), 451, 799
994 unpackbits() (in module numpy), 690
transpose() (in module numpy), 425 unravel_index() (in module numpy), 463
transpose() (in module numpy.ma), 269, 352, 847, 929 unshare_mask() (numpy.ma.MaskedArray method), 242,
transpose() (numpy.chararray method), 145 297, 874
transpose() (numpy.core.defchararray.chararray method), unwrap() (in module numpy), 721
995 upper() (in module numpy.core.defchararray), 970
transpose() (numpy.generic method), 78 upper() (numpy.chararray method), 145
transpose() (numpy.ma.MaskedArray method), 213, 270, upper() (numpy.core.defchararray.chararray method), 995
353, 848, 930 user_array, 176
transpose() (numpy.matrix method), 120 userloops (C member), 1022
transpose() (numpy.ndarray method), 29, 50
transpose() (numpy.recarray method), 166 V
transpose() (numpy.record method), 175 vander() (in module numpy), 418
trapz() (in module numpy), 741 vander() (in module numpy.ma), 354, 931
tri() (in module numpy), 417 var (in module numpy.ma), 329, 906
triangular() (in module numpy.random), 600 var() (in module numpy), 701
triangular() (numpy.random.mtrand.RandomState var() (numpy.generic method), 78
method), 646 var() (numpy.ma.MaskedArray method), 233, 336, 913
tril() (in module numpy), 417 var() (numpy.matrix method), 121
tril_indices() (in module numpy), 466 var() (numpy.ndarray method), 29, 58
tril_indices_from() (in module numpy), 467 var() (numpy.recarray method), 166
trim_zeros() (in module numpy), 450 var() (numpy.record method), 175
triu() (in module numpy), 418 vdot() (in module numpy), 538
triu_indices() (in module numpy), 467 vectorize (class in numpy), 774
triu_indices_from() (in module numpy), 469 view, 3
true_divide (in module numpy), 757 ndarray, 93
trunc (in module numpy), 731 view() (numpy.chararray method), 145
two (C member), 1026 view() (numpy.core.defchararray.chararray method), 995
type (C member), 1016 view() (numpy.generic method), 78
type (numpy.dtype attribute), 87 view() (numpy.ma.MaskedArray method), 205
type_num (C member), 1016 view() (numpy.matrix method), 121
typekind (C member), 1026 view() (numpy.ndarray method), 29, 45
typename() (in module numpy), 491 view() (numpy.recarray method), 167
typeobj (C member), 1015 view() (numpy.record method), 175
types (C member), 1022 vonmises() (in module numpy.random), 603
types (numpy.ufunc attribute), 378 vonmises() (numpy.random.mtrand.RandomState
method), 649
U vsplit() (in module numpy), 444
ufunc, 1078, 1081 vstack (in module numpy.ma), 279, 283, 857, 861
attributes, 377 vstack() (in module numpy), 440
C-API, 1064, 1070
casting rules, 375 W
keyword arguments, 377 wald() (in module numpy.random), 604
methods, 380 wald() (numpy.random.mtrand.RandomState method),
methods accumulate, 1081 650

Index 1117
NumPy Reference, Release 2.0.0.dev8464

weakreflist (C member), 1015

weibull() (in module numpy.random), 606
weibull() (numpy.random.mtrand.RandomState method),
651
where() (in module numpy), 460, 665
where() (in module numpy.ma), 365, 942

Y
yellow_text() (in module numpy.distutils.misc_util), 1000

Z
zeros (in module numpy.ma), 251, 829
zeros() (in module numpy), 392
zeros_like() (in module numpy), 393
zfill() (in module numpy.core.defchararray), 970
zfill() (numpy.chararray method), 146
zfill() (numpy.core.defchararray.chararray method), 996
zipf() (in module numpy.random), 607
zipf() (numpy.random.mtrand.RandomState method), 652

1118 Index