ENH: ascontiguousarray leaves negative strides for length-1 dimensions

[shangke1988]

Describe the issue:

np.ascontiguousarray don't process negative strides of single channle CWH data

Reproduce the code example:

import numpy as np
import torch
a=np.zeros((1,10,20))[::-1]
b=np.ascontiguousarray(a)
b.strides
c=torch.from_numpy(b)

Error message:

>>> a=np.zeros((1,10,20))[::-1]
>>> b=np.ascontiguousarray(a)
>>> b.strides
(-1600, 160, 8)
>>> c=torch.from_numpy(b)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ValueError: At least one stride in the given numpy array is negative, and tensors with negative strides are not currently supported. (You can probably work around this by making a 
copy of your array  with array.copy().)

Runtime information:

import sys, numpy; print(numpy.version); print(sys.version)
1.23.5
3.10.9 | packaged by Anaconda, Inc. | (main, Mar 1 2023, 18:18:15) [MSC v.1916 64 bit (AMD64)]

Context for the issue:

No response

[rkern]

While I do think it would be a good idea for ascontiguousarray() to normalize the stride in such cases, torch.from_numpy() could probably check for this case, too. With a 1-length dimension, that dimension's stride can be anything, and it would still validly describe the same contiguous chunk of memory.

[mattip]

@Adesoji1 I don't think that relates to the question at hand. Clearly copy can be used. The issue is a bit more subtle as captured by the comment above. Note that using dlpack does normalize the strides, but the resulting tensor is read-only, where torch.from_numpy creates a read-write tensor.

[seberg]

Note that using dlpack does normalize the strides

As does the buffer protocol, but mainly due to the fact that cython needed transition time and I suspect there is little reason to remove the normalization at this point.

I don't mind a PR to create a new view in NumPy, but really I think this should be fixed in torch. NumPy fixed the definition of contiguous on this one long before torch even existed.