PyTorch笔记4----------Tensor的操作

1.裁剪运算

• 常用于对tensor中的一些元素进行范围约束，约束范围有利于参数学习的稳定，防止过拟合；可以防止梯度离散；方便模型进行量化

import torch
a = torch.rand(2,2) * 10
print("a:\n",a)
b = a.clamp(2,5)
print("b:\n",b)

2.索引与数据筛选

• torch.where(condition,x,y)
  • 按照条件(condition)从x和y中选出满足条件的元素组成新的tensor

import torch
a = torch.rand(4,4)
b = torch.rand(4,4)
print("a:\n",a)
print("b:\n",b)
out = torch.where(a>0.5,a,b)
print('out:\n',out)

• torch.index_select(input,dim,index)
  • 按照指定索引输出tensor
  • 参数dim：维度
  • 参数index：要输出的维度数据的索引，按照给出索引顺序，输出新tensor

import torch
a = torch.rand(4,4)
print("a:\n",a)
out = torch.index_select(a,dim=0,index=torch.tensor([0,3,2]))
print('out:\n',out)

•  torch.gather(input,dim,index)
  • 在指定维度上按照索引复制输出tensor
  • 参数dim：维度
  • 参数index：选择数据的具体索引值

import torch
a = torch.linspace(1,16,16).view(4,4)
print("a:\n",a)
out = torch.gather(a,dim=0,index=torch.tensor([[0,1,1,1],[0,1,2,2],[0,1,3,3]]))
print('out:\n',out)

• torch.masked_select(input,mask)
  • 按照mask输出tensor，输出为向量

import torch
a = torch.linspace(1,16,16).view(4,4)
print("a:\n",a)
mask = torch.gt(a,8)
print('mask:\n',mask)
out = torch.masked_select(a,mask)
print('out:\n',out)

• torch.take(input,indices)
  • 将input看成一维tensor，按照indices的值为索引得到输出tensor

import torch
a = torch.linspace(1,16,16).view(4,4)
print("a:\n",a)
out = torch.take(a,index=torch.tensor([0,15,13,10]))
print('out:\n',out)

• torch.nonzero(input)
  • 输出非0元素坐标

import torch
a = torch.tensor([[0,1,2,0],[2,3,0,1]])
print("a:\n",a)
out = torch.nonzero(a)
print('out:\n',out)

3.组合/拼接

• torch.cat(seq,dim)：按照已经存在的维度进行拼接

import torch
a = torch.zeros((2,4))
b = torch.ones((2,4))
print("a:\n",a)
print("b:\n",b)
out = torch.cat((a,b),dim=1)
print('out:\n',out)

• torch.stack(seq,dim)：按照新的维度进行拼接

import torch
a = torch.linspace(1,6,6).view(2,3)
b = torch.linspace(7,12,6).view(2,3)
print("a:\n",a)
print("b:\n",b)
out = torch.stack((a,b),dim=1)
print('out:\n',out)

4.切片 

• torch.chunk(tensor,chunks,dim)
  • 按照摸个维度平均分块，如果无法平均，最后一个维度值的个数会小于平均值
  • 参数chunks：分块的个数

import torch
a = torch.rand((3,4))
print("a:\n",a)
out = torch.chunk(a,2,dim=0)
print('out:\n',out)

• torch.split(tensor,split_size_or_sections,dim)
  • 按照某个维度，依照第二个参数给出的列表或数对tensor进行分割
  • 参数split_size_or_sections：如果传入一个数值，则效果和torch.chunk相同；若是列表，则会按照列表中的值进行对应切分

import torch
a = torch.rand((10,4))
print("a:\n",a)
out = torch.split(a,[1,3,6],dim=0)
print('out:\n',out)

5.变形操作

• torch.reshape(input,shape)
  • 对输入的tensor的shape进行改变

import torch
a = torch.rand(2,3)
print("a:\n",a)
out = torch.reshape(a,(3,2))
print('out:\n',out)

• torch.t(input) 
  • 对2维tensor进行转置

import torch
a = torch.rand(2,3)
print("a:\n",a)
out = torch.t(a)
print('out:\n',out)

• torch.transpose(input,dim0,dim1)
  • 交换两个维度

import torch
a = torch.rand(2,3)
print("a:\n",a)
out = torch.transpose(a, 0, 1)
print('out:\n',out)

• torch.squeeze(input,dim)
  • 去除维度大小为1的维度

import torch
a = torch.rand(1,2,3)
print("a:\n",a)
out = torch.squeeze(a)
print('out:\n',out)

• torch.unbind(tensor,dim)
  • 去除某个维度

import torch
a = torch.rand(1,2,3)
print("a:\n",a)
out = torch.unbind(a,1)
print('out:\n',out)

• torch.unsqueeze(input,dim)
  • 在指定位置添加维度

import torch
a = torch.rand(1,2,3)
print("a:\n",a)
out = torch.unsqueeze(a,-1)
print('out:\n',out)

• torch.flip(input,dim)
  • 按照给定维度翻转张量

import torch
a = torch.rand(1,2,3)
print("a:\n",a)
out = torch.flip(a,dims=[1,2])
print('out:\n',out)

• torch.rot90(input,k,dim)
  • 按照指定维度(dim)和旋转次数(k)进行张量旋转

import torch
a = torch.rand(1,2,3)
print("a:\n",a)
out = torch.rot90(a,2)
print('out:\n',out)

6.填充

• torch.full()

import torch
a = torch.full((2,3),10)
print("a:\n",a)