TST/MAINT: ndimage: follow-ups for array API standard support

[lucascolley]

Follow-ups from gh-21150:

• ENH: ndimage: add array API standard support #21150 (comment) (TST: ndimage: unskip test_boundary_spline_accuracy for torch #21282)
• ENH: ndimage: delegate to CuPy #21091 (comment): We should try to add an @xfail_xp_backends decorator, to match @skip_xp_backends. (TST: add xfail_xp_backends #21334)
• ENH: ndimage: add array API standard support #21150 (comment): Tests for array-likes should be added throughout.
• ENH: ndimage: add array API standard support #21150 (comment): We should try to add the ability to say "only run for these backends" to the skip infra. (done in TST: fix GPU failures #21294)
• ENH: ndimage: add array API standard support #21150 (comment): Some functions produce object arrays which are incompatible with alternative backends. We probably want to deprecate and return regular arrays instead.

[shivanimall]

According to #18867 and docs , the following file should also have support for torch, however, it currently does not, is this in works?

scipy/ndimage/_support_alternative_backends.py (specifically, delegate_xp function)

Thanks for the help!

[lucascolley]

hi @shivanimall ! _support_alternative_backends.py delegates to existing implementations for CuPy and JAX where they exist, and converts arrays to NumPy and back otherwise when they are on CPU.

Could you clarify what 'support for torch' you are interested in? Are there existing torch implementations of the https://docs.scipy.org/doc/scipy/reference/ndimage.html APIs which we could delegate to?

[shivanimall]

Thanks @lucascolley, I am hoping to pass in PyTorch tensors (on GPU) to scipy.ndimage.map_coordinates, which currently fails as the function (map_coordinates) still internally tries to convert to numpy array

[shivanimall]

Currently, PyTorch has an existing implementation of grid sample which is similar to map_coordinates however, it changes the range of coordinates to -1 to 1. I would like to avoid this normalisation, if possible. So, I resorted to using scipy's functionality for the geometric transform.

Would this mean using torch.compile with numpy arrays (into scipy function)?

[lucascolley]

If you want to keep your arrays on the GPU, then I don't think torch.compile can help here, given that ndimage.map_coordinates is implemented on NumPy arrays which are CPU-only.

[shivanimall]

thanks @lucascolley