Proposal: add an optional copy argument to np.reshape

[shoyer]

Inspired by my discussion with @charris in #9810, I would like to propose adding an optional copy argument to np.reshape() and ndarray.reshape().

Proposed semantics:

• copy=None: copy only if necessary (default).
• copy=True: always copy.
• copy=False: never copy. Raise an error if the data cannot be reshaped without a copy.

This would allow users to ensure desired behavior for reshape and alleviate the primary use case for modifying ndarray.shape.

[seberg]

Always agreed with that, the only slight annoyance is that we need yet another C-API version probably.

[eric-wieser]

Yep, this seems like a very good idea to me too.

[mhvk]

Agreed too, the only reason I use assignments to shape is because in some cases I want to be absolutely sure there won't be a copy for a very large array. I can similarly see the use of being able to ensure a copy is in fact made.

[seberg]

@mhvk and the fun thing is, it does not really help you (except in the sense of an assert statement), because the shape assignment actually does the copy, then sees its a copy and raises an error.

EDIT: Just to note, I am not sure if we have to Deprecate it, but we still need to discourage it in the docs, its a bad habit, but dunno if there might be a handful of actual use cases.

[charris]

ISTM that reshape can also replace ravel.

[charris]

Note that we will need to make sure that this doesn't break current subclasses of ndarray that do not implement the new keyword.

[seberg]

True, about ravel, there is a problem with ravel, in that it makes sure that the output array is contiguous, which was a regression we once had and because someone (maybe scipy/some scikit) relied on it I think.

[shoyer]

On the C-API side, I suppose it would make sense to use enums here rather PyObject for the extra argument? e.g., NPY_COPY_DEFAULT, NPY_COPY_TRUE, NPY_COPY_FALSE?

[rgommers]

In gh-11884, @seberg proposed the same thing as this issue. With the extra motivation that users now modify arr.shape instead of using reshape.