✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: d951d51. Link to the linter CI: here}

Thanks very much for the PR. Let's a first wait for a second review on the r2_score PR (#27904) before reviewing yours to avoid cross/redundant discussions between PRs.

Thanks for the work @lithomas1 , with #27904 now merged we're resuming reviews on the array api side. If you are still available to iterate on the pr could you pull main on your branch please ?

@fcharras

This should be ready for another review.

Sorry for the slow turnaround on this!

BTW, I have run the tests of this PR with torch and cupy on cuda host and torch on a MPS host and everything is green.

OK, so this PR now fails with errors like

_________ [doctest] sklearn.metrics._regression.mean_poisson_deviance __________
[gw0] linux -- Python 3.11.8 /usr/share/miniconda/envs/testvenv/bin/python
1424     -------
1425     loss : float
1426         A non-negative floating point value (the best value is 0.0).
1427 
1428     Examples
1429     --------
1430     >>> from sklearn.metrics import mean_poisson_deviance
1431     >>> y_true = [2, 0, 1, 4]
1432     >>> y_pred = [0.5, 0.5, 2., 2.]
1433     >>> mean_poisson_deviance(y_true, y_pred)
Expected:
    1.4260...
Got:
    array(1.42601513)

A solution would be to call float on the result from _array_api._average,
(but this would not be ideal for e.g. dask, which would prefer to stay lazy if possible).

An alternative solution would be either to update the doctests (and accept this API change),
or we can patch the numpy version of _array_api._average only here to not return an array (i.e. remove the asarray call).
https://github.com/scikit-learn/scikit-learn/blame/3ee60a720aab3598668af3a3d7eb01d6958859be/sklearn/utils/_array_api.py#L598

EDIT:

I've just added the call to float on the result (to match other usages of _average), given that the dask stuff is on hold for now.

@ogrisel @betatim

This should be ready for another look now.

I've just added the call to float on the result (to match other usages of _average), given that the dask stuff is on hold for now.

Yes and this is consistent with what we do for other 1d regression scores (e.g. r2_score).

I am still +1 for merge BTW.

Maybe @OmarManzoor would be interested in reviewing this one as well :)

Updated.

Thanks for the reviews.