Selfnote: Needs to be patched to #4294

Ah! This is interesting... So we need it to delegate to either estimator or best_estimator_ based on the fit status of GridSearchCV! :/

We should always delegate to estimator.
This is caused by #4835, which is apparently not the right fix.

Err... what happens if someone searches over probability? hum..

@jnothman always knows what to do in this kind of situations ;)

Maybe we should ignore the issue with people searching over probability and do either of the two:

1. run the check for probA_ and probB_ only if the estimator was fitted.

2. move the check for probA_ and probB_ into _predict_proba, not predict_proba.

The first one is a bit more magic and I don't really see it as being very necessary. If people set probability=True after fitting, I think it is more appropriate to raise an error "not fitted with probability=True" then branching to decision_function in our usual duck-typing.

So I vote for 2.

Unless I'm missing something, that example I provided should always have probability=True, is that not the case?

@adammenges The example you provided is fine! This is an issue with delegation of the meta-estimator (GridSearchCV)'s methods to best_estimator_, so one can directly access those methods from the GridSearchCV instance instead of (grid_svm.best_estimator_....), (if I understood this correctly) :)

And Andreas was talking about setting probability to True after fitting which is not correct...

I've not the time to read all comments right now, but this is the wrong solution. We need to check the unfitted estimator and delegate to the fitted one.