The reason why np.divide(proba, denominator, where=denominator != 0.0) did not work is because of the next line in the code:

# XXX : for some reason all probas can be 0
proba[np.isnan(proba)] = 1. / n_classes

Since now you don't divide by 0, you don't get inf and the probas are not normalized.

Interesting, I did not know about the where parameter of np.divide. I agree with @jeremiedbb that we can collapse the code to remove any np.nan occurrence from the start.

If this is not already the case, we could probably add a specific test to "calibrate" such a badly behaving classifier that always predicts zeros in predict_probas and check that:

• we get uniform 1/n_classes probabilities as a results;
• we do not get a waning using sklearn.utils._testing.assert_no_warnings.

The reason why np.divide(proba, denominator, where=denominator != 0.0) did not work is because of the next line in the code:

# XXX : for some reason all probas can be 0
proba[np.isnan(proba)] = 1. / n_classes

Since now you don't divide by 0, you don't get inf and the probas are not normalized.

Thanks, I overlooked the fact that this process is only for division by zero.

we do not get a waning using sklearn.utils._testing.assert_no_warnings.

@ogrisel I don't think we want to check that no warning is issued. We don't want to check that every line of code doesn't raise a warning. I think that the fact that the warning disappears with this PR is enough.

I think that the fact that the warning disappears with this PR is enough.

Alright, but don't we want to explicitly check that we get uniform probabilities when the based classifier predicts only zeros.

Alright, but don't we want to explicitly check that we get uniform probabilities when the base classifier predicts only zeros.

I didn't argue against that :)

I didn't argue against that :)

OK, now I'm going to add this test.

OK, now I'm going to add this test.

Since it is difficult to control the behavior of the CalibratedClassifierCV class and the test code becomes large, I added test_calibration_zero_probability, which is just a test for _CalibratedClassifier.predict_proba.

Thank you very much for the fix and the nice new test @t-kusanagi.

ping @lucyleeow to keep you in the loop on your favorite meta-estimator :)