Somewhere between scikit-learn 0.18.2 and 0.20.2 the estimator object return value was added to _fit_and_score. Removing this element fixes a bug that occurs later in SearchCV.py, where an array without estimator object is expected.
I've set the scikit-learn version in requirements.txt to 0.20.2. Since _fit_and_score is a private function that should not be called outside scikit-learn there is no guarantee that this function won't change. Hence the best solution, in my opinion, is to affix the version to a version that we know works. This does mean that we have to update the version once in a while and test if it still works to keep up-to-date. This could be done as part of the release process maybe?
Somewhere between scikit-learn 0.18.2 and 0.20.2 the estimator object return value was added to _fit_and_score. Removing this element fixes a bug that occurs later in SearchCV.py, where an array without estimator object is expected.
I've set the scikit-learn version in requirements.txt to 0.20.2. Since _fit_and_score is a private function that should not be called outside scikit-learn there is no guarantee that this function won't change. Hence the best solution, in my opinion, is to affix the version to a version that we know works. This does mean that we have to update the version once in a while and test if it still works to keep up-to-date. This could be done as part of the release process maybe?