ENH Add Array API compatibility to zero_one_loss and accuracy_score

[EdAbati]

Reference Issues/PRs

Towards #26024

What does this implement/fix? Explain your changes.

It makes the zero_one_loss and accuracy_score(since it was a dependency) implementations compatible and tested with the Array API.

Any other comments?

• I have added a test for _weighted_sum. When having sample_weights as floats I got the following from PyTorch:

  >           return float(sample_score @ sample_weight)
  E           RuntimeError: dot : expected both vectors to have same dtype, but found Double and Float
  

• It feels like I am missing some cases in the tests. For example, should I test with a y_true and y_pred that is not from binary classification? 🤔

[github-actions]

✔️ Linting Passed

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

_{Generated for commit: 51258ed. Link to the linter CI: here}

[ogrisel]

Thanks for the PR. See the reply to your question in the comments below:

[OmarManzoor]

@ogrisel I think I'll wait for this PR to be merged before working on other metrics, so that the addition of tests doesn't conflict or become redundant. What do you think?

[EdAbati]

Thank you everyone for the feedback. :) I have updated the PR

@ogrisel , should we test the multilabel case? Or when y is multilabel, is it always transformed into a sparse matrix?
(is_multilabel does not seem fully compatible with the ArrayAPI yet, because of this)

• Is there a better place for _array_api_atol since it is just used for testing?
• likewise, _array_api_for_tests is in utils.estimators_checks, but it is used for metrics checks too. Should we move it to _array_api or somewhere else?

[ogrisel]

If think we can leave the multilabel out of the scope of this PR. Not sure hiw we could tackle it.

I for the atol helper, we can put in the same module as other array api helpers for now. Since it's private we can always change later.

[ogrisel]

Ok for moving _array_api_for_tests, as well.

[betatim]

Looks good to me.

WDYT @ogrisel

[ogrisel]

I think we could improve the generic metric tests to include testing for sample_weight when applicable.

Other than that, LGTM.

[betatim]

I like the sample_weight suggestion, but also fine doing that in a new PR. @EdAbati what is your feeling/time availability?

[ogrisel]

I merged, let's do that in a follow up PR :)

[EdAbati]

Hi both, sorry for the late reply I am a bit busy these days. I will likely make a follow-up PR during the weekend (if it is not done by then)

Thank you for merging :)