MAINT Param validation: decorate all estimators with _fit_context #26473
Conversation
sklearn/calibration.py
Outdated
| @@ -318,6 +319,7 @@ def _get_estimator(self): | |||
|
|
|||
| return estimator | |||
|
|
|||
| @_fit_context(prefer_skip_nested_validation=False) | |||
There was a problem hiding this comment.
we'd need a comment in the code whenever these are False.
Sorry @adrinjalali, I don't understand your question. Redundant with what ? |
|
What I mean is, in functions, |
|
Here we're setting it for estimators. Every estimator validates the parameters in the
|
| # we don't want to validate again for each call to partial_fit | ||
| partial_fit_and_fitted = ( | ||
| fit_method.__name__ == "partial_fit" and _is_fitted(estimator) | ||
| ) | ||
|
|
||
| if not global_skip_validation and not partial_fit_and_fitted: |
There was a problem hiding this comment.
How about computing partial_fit_and_fitted inside if not global_skip_validation:? I don't know if _is_fitted(estimator) is always fast, because it may check all parameter names of the estimator or call estimator.__sklearn_is_fitted__() if present. It might be a good idea to avoid it when not necessary.
Follow-up of #25815
This PR actually applies the decorator to all estimators.
There are situations where we don't want to skip inner validation (
prefer_skip_nested_validation=False):metricandmetric_params). Note that if the user only passes a callable we want to skip inner validation because the args passed to the callable come from us and not from the user.Estimators receiving cv objects can skip inner validation because cv objects are not validated yet. When we decide to validate cv objects we'll need to revisit this.
Something that I missed in #25815 is that for partial_fit we don't want to validate if fit or partial_fit has already been called (which was not done properly for all estimators). I added such a mechanism in the
_fit_contextdecorator.