Preemptively fix incompatibilities with an upcoming array-api-strict release #31517
Conversation
This is required by the Array API and is harmless otherwise.
This error originates from NumPy when `__array__` does not produce a numpy array. This happens when array-api-strict moves over to __buffer__.
There was a problem hiding this comment.
Code changes look fine.
One other effect of using the buffer protocol for np.asarray is that testing with array_api_strict needs to use python 3.12 or above.
Taking a look at our update_environments_and_lock_files.py, all array-api-strict envs are pylatest so I think this should be fine also.
|
@ev-br to clarify, scikit-learn/sklearn/utils/_array_api.py Lines 810 to 811 in 3962c28 would still work right? (I was confused by scipy/scipy#21828 though I see that last comment is quite old) |
|
Yes, |
|
Thanks for this ahead of time fix :D Running the CUDA tests just "to be sure to be sure". I agree that the new minimum version of Python shouldn't be a problem. It only effects those running the tests as no real user should be using |
What does this implement/fix? Explain your changes.
array-api-strictplans a release "soon" (tm). Testing it locally smokes out two small failures, which this PR aims to fix:xp.reshape(array, shape)requiresshapeto be a tuple and rejectsshapebeing a list. https://data-apis.org/array-api/latest/API_specification/generated/array_api.reshape.html#reshapea failed conversion to NumPy may emit a ValueError. This comes from RFC: rm __array__, add __buffer__ data-apis/array-api-strict#115 which --- finally! --- replaces
__array__with__buffer__(This was discussed at length in Remove __array__ data-apis/array-api-strict#69 and Should we be implementing __array__? data-apis/array-api-strict#67).Any other comments?
One other effect of using the buffer protocol for
np.asarrayis that testing witharray_api_strictneeds to use python 3.12 or above. The array api discussion is https://hackmd.io/zn5bvdZTQIeJmb3RW1B-8g#Meeting-minutes-14-November-2024