[sklearn] Make BaseEstimator inherit from _MetadataRequester

[tmke8]

This inheritance was added to sklearn two years ago here: scikit-learn/scikit-learn@62671a7#diff-ac9c452d8c660d2516089c8448596912c22417c7c690f079715593e61ef6ee0dR133

[tmke8]

Oh, and of course, the inheritance is still there today: https://github.com/scikit-learn/scikit-learn/blob/f462edd741c3cbfccc1c6a2d64a2a66a6599d3f8/sklearn/base.py#L156

[tmke8]

@microsoft-github-policy-service agree

[debonte]

stubs/sklearn/base.pyi:28:1: error: Call to untyped function "init_subclass" in typed context [no-untyped-call]
stubs/sklearn/naive_bayes.pyi:36:1: error: Call to untyped function "init_subclass" in typed context [no-untyped-call]
stubs/sklearn/naive_bayes.pyi:42:1: error: Call to untyped function "init_subclass" in typed context [no-untyped-call]
stubs/sklearn/naive_bayes.pyi:69:1: error: Call to untyped function "init_subclass" in typed context [no-untyped-call]

@Avasam, any idea why mypy is complaining here? I did a spot check and didn't see any explicit __init_subclass__ calls at the reported locations.

[Avasam]

I'm omw out of city for the weekend. So I can't give you an exact detailed reason atm.
But that one can be triggered w/o an explicit call as mypy understand it is automatically called (I did have it come up in previous PRs).

[tmke8]

It seems __init_subclass__ on _MetadataRequester creates the set_*_request() methods: https://github.com/scikit-learn/scikit-learn/blob/f462edd741c3cbfccc1c6a2d64a2a66a6599d3f8/sklearn/utils/_metadata_requests.py#L1406

However, the type stubs already have these methods statically on the class:

python-type-stubs/stubs/sklearn/utils/_metadata_requests.pyi

Lines 82 to 91 in d1b4fb3

	def set_fit_request(self, **kwargs): ...
	def set_partial_fit_request(self, **kwargs): ...
	def set_predict_request(self, **kwargs): ...
	def set_predict_proba_request(self, **kwargs): ...
	def set_predict_log_proba_request(self, **kwargs): ...
	def set_decision_function_request(self, **kwargs): ...
	def set_score_request(self, **kwargs): ...
	def set_split_request(self, **kwargs): ...
	def set_transform_request(self, **kwargs): ...
	def set_inverse_transform_request(self, **kwargs): ...

So, I think that means __init_subclass__ can be removed from the stubs? I don't think anyone will call it explicitly.

[tmke8]

(I found another missing thing and added it in a new commit. Hope that's okay; I can remove it if not. It's about the global config: https://scikit-learn.org/stable/modules/generated/sklearn.set_config.html#sklearn.set_config )

[debonte]

Thanks for contribution @tmke8!