FIX (SLEP6) descriptor shouldn't override method #32111
Conversation
Co-authored-by: Omar Salman <[email protected]>
…into slep6/refactor
Co-authored-by: Stefanie Senger <[email protected]>
Co-authored-by: Stefanie Senger <[email protected]>
…learn into slep6/pipeline/score
Co-authored-by: Omar Salman <[email protected]>
Co-authored-by: Omar Salman <[email protected]>
| # metadata with an UNUSED marker, and that the parent class has | ||
| # the method already. In this case, we cannot remove the method | ||
| # since it exists in the parent class, and instead we set it with | ||
| # an empty set of requests. |
There was a problem hiding this comment.
Aren't we simply not setting any RequestMethod for such cases instead of an empty set of requests?
There was a problem hiding this comment.
The issue is "not setting a RequestMethod" means we inherit the one from the parent class, which in the case of LogisticRegressionCV means having one accepting a sample_weight
There was a problem hiding this comment.
So we need to set one but with empty requests?
There was a problem hiding this comment.
Then shouldn't we be setting requests = dict() in this branch rather than simply continuing the loop?
There was a problem hiding this comment.
This branch has not hasattr(cls, set_method_name), otherwise we don't come here
There was a problem hiding this comment.
I think the comment is a bit misleading, since it talks about the reason to add the not hasattr(cls, set_method_name) condition into the code compared to before rather than about the whole condition.
Co-authored-by: Omar Salman <[email protected]>
There was a problem hiding this comment.
Thanks, @adrinjalali!
I have checked the rendered docs and built locally, but I don’t see that LogisticRegressionCV.set_score_request() has a different docstring compared to before. Maybe I misunderstand the PR, but I cannot see how the def set_score_request() in _BaseScorer would be prevented from being overridden while a CV meta-estimator is defined, since there is no inheritance between those two.
And searching the repo, _BaseScorer is the only place that defines their own set_score_request.
Is this PR rather for third party libraries?
I also suggested a few documentation improvements.
| def is_RequestMethod(obj, name: str): | ||
| value = inspect.getattr_static(obj, name) | ||
| return isinstance(value, RequestMethod) |
There was a problem hiding this comment.
Structurally, is_RequestMethod is a bit similar to defines_get_metadata_routing from #31695. Both check whether an estimator defines something itself or inherits it and feel like workarounds for going through MROs.
For maintainability, do you think we could reduce some of these MRO iterations (for instance via class methods, as you mentioned once)?
| try: | ||
| for method in SIMPLE_METHODS: | ||
| set_method_name = f"set_{method}_request" |
There was a problem hiding this comment.
set_method_name could be called set_method_request as a more intuitive variable name.
| # metadata with an UNUSED marker, and that the parent class has | ||
| # the method already. In this case, we cannot remove the method | ||
| # since it exists in the parent class, and instead we set it with | ||
| # an empty set of requests. |
There was a problem hiding this comment.
I think the comment is a bit misleading, since it talks about the reason to add the not hasattr(cls, set_method_name) condition into the code compared to before rather than about the whole condition.
| # The method is not a descriptor, which means it's explicitly | ||
| # defined in the class, therefore we skip overriding it with a | ||
| # descriptor here. | ||
| # This happens, for instance in Scorers where the | ||
| # `set_score_request` method is defined in the class. |
There was a problem hiding this comment.
| # The method is not a descriptor, which means it's explicitly | |
| # defined in the class, therefore we skip overriding it with a | |
| # descriptor here. | |
| # This happens, for instance in Scorers where the | |
| # `set_score_request` method is defined in the class. | |
| # `set_method_name` is not a `RequestMethod` so we skip | |
| # overriding it. For example, in Scorers the `set_score_request` | |
| # method is explicitly defined in the class. |
Maybe a bit more straightforward?
| @@ -1438,14 +1448,35 @@ def __init_subclass__(cls, **kwargs): | |||
| ---------- | |||
| .. [1] https://www.python.org/dev/peps/pep-0487 | |||
| """ | |||
|
|
|||
| def is_RequestMethod(obj, name: str): | |||
There was a problem hiding this comment.
| def is_RequestMethod(obj, name: str): | |
| def _is_RequestMethod(obj, method_name: str): | |
| # Check if a method is a descriptor. |
I think it would be a private method.
If a class implements a
set_{method}_requestexplicitly (like done in scorers), then the descriptor shouldn't be overriding it.xref: #30859 (comment)
cc @antoinebaker @OmarManzoor
This fixes some existing issues. The added tests should represent where things were failing.