API deprecate CalibratedClassifierCV(..., cv=prefit) for FrozenEstimator #30171
Conversation
Co-authored-by: Adam Li <adam2392@gmail.com>
There was a problem hiding this comment.
I have mixed feelings about this: asking the user to import a class and wrap an argument feels like a small usability regression.
What about not deprecating cv="prefit" and automatically do the wrapping with FrozenEstimator on the fly in _get_estimator:
def _get_estimator(self):
"""Resolve which estimator to return (default is LinearSVC)"""
if self.cv == "prefit" and self.estimator is None:
raise ValueError(
'estimator cannot be left to None when cv="prefit"'
)
if self.estimator is None:
# we want all classifiers that don't expose a random_state
# to be deterministic (and we don't want to expose this one).
estimator = LinearSVC(random_state=0)
if _routing_enabled():
estimator.set_fit_request(sample_weight=True)
else:
estimator = self.estimator
if self.cv == "prefit":
estimator = FrozenEstimator(estimator)
return estimatorBut then we would have too equivalent ways to do the same thing:
from sklearn.frozen import FrozenEstimator
CalibratedClassifierCV(FrozenEstimator(model)).fit(X_cal, y_cal)and:
CalibratedClassifierCV(model, cv="prefit").fit(X_cal, y_cal)But maybe that's ok?
|
I wouldn't say it's a usability regression. |
I partially agree... But at the same time, calling In that respect, |
There was a problem hiding this comment.
If @thomasjpfan and others think it's better, and I am in the minority, then so be it. It's true that this will also allow a small code simplification once the deprecation period is over.
| @@ -540,7 +554,7 @@ def test_calibration_dict_pipeline(dict_data, dict_data_pipeline): | |||
| """ | |||
| X, y = dict_data | |||
| clf = dict_data_pipeline | |||
| calib_clf = CalibratedClassifierCV(clf, cv="prefit") | |||
| calib_clf = CalibratedClassifierCV(FrozenEstimator(clf), cv=2) | |||
There was a problem hiding this comment.
This change highlights another small usability regression: it's difficult to understand why we would need to adjust the CV strategy when calibrating a frozen estimator on a dataset with less than 5 data-points per class. One could expect that cross-fitting would be entirely disabled in that case.
Again, I can live with it.
There was a problem hiding this comment.
I'd agree with you, but in reality, people with 5 data points shouldn't be doing statistical ML 😁
There was a problem hiding this comment.
In particular, meaningful calibration with 5 data points per class is quite challenging. Still, it reveals a rough edge in my opinion. But I believe we can fix it later if needed.
Towards #29893
This deprecates
CalibratedClassifierCV(..., cv="prefit")in favor ofCalibratedClassifierCV(FrozenEstimator(est)).Also adds
ensemble="auto"option for automatically handlingFrozenEstimatorcorrectly.cc @glemaitre @ogrisel