FIX Extract estimator objects before aggregating dict of scores

[alfaro96]

Reference Issues/PRs

Coming from #17332.

What does this implement/fix? Explain your changes.

This PR extracts the estimator objects for each cross-validation split before executing the _aggregate_score_dicts in cross_validation function to avoid that nested estimators are converted to arrays.

Any other comments?

WDYT @thomasjpfan about this solution?

Another idea to solve this issue?

[glemaitre]

You are super-efficient. You open a PR before that I found that the CI was broken :)

[glemaitre]

We should probably have a test to trigger this usage. I am thinking that this is maybe the job of the _aggregate_* function to take care to not convert everything to an array (but I have to look more into details).

[alfaro96]

You are super-efficient. You open a PR before that I found that the CI was broken :)

I was looking to the failures in the CRON jobs and I realize that the documentation build was failing. Since I love keep things properly working, I propose the PR :)

[glemaitre]

LGTM. Maybe @thomasjpfan wants to have a look

[glemaitre]

So maybe your solution is better then :)

[thomasjpfan]

I would prefer to dispatch based on the type:

    return {
        key: np.asarray([score[key] for score in scores])
        if isinstance(scores[0][key], numbers.Number)
        else [score[key] for score in scores]
        for key in scores[0]
    }

and then update the docstring:

Aggregate a list of dicts to a dict of ndarray or list.

WTYD @glemaitre ?

[thomasjpfan]

I made the change and merged so that PRs are not failing. We can have a followup PR if we want to go back to checking the key.

[thomasjpfan]

Hmm... the only recent PR that was blocked was #17743

[thomasjpfan]

Thank you @alfaro96 for being so quick with the fix!

[alfaro96]

Thanks @thomasjpfan and @glemaitre for taking care and finalizing this PR!