Why is it optional? Ie, what does storing train_score when we should be break?

No we have an option return_train_score that can be set to False. In the current master if it is set to False, it raises an error without this bugfix due to my stupidity. Apologies again.

(One example where it fails)

>>> from sklearn.model_selection import GridSearchCV
>>> from sklearn.svm import LinearSVC
>>> from sklearn.datasets import make_classification
>>> X, y = make_classification()
>>> svm = LinearSVC()
>>> gs = GridSearchCV(svm, param_grid={'C': [0.1, 0.001]},
...                   return_train_score=False)
>>> gs.fit(X, y)
UnboundLocalError: local variable 'train_scores' referenced before assignment

On why we even have that option, it was added so users who don't need it can turn it off to keep the cv_results_ concise...

gah so the false case wasn't tested? That's bad. re back-port: it's release, we need to do 0.18.1 for this fix.

LGTM

Hi @raghavrv
In the future, it would be helpful to have a clear message in the PR main message. I can understand what's going on from Gael's question and the code, but it is much much easier to review when the author clearly explains the content of the PR.

Sorry for the reviewing fail :/

@NelleV Thanks for the feedback and merge. Yes I've been skimping on PR descriptions lately. Apologies... I'll try to make a clear description from the next PR.

So are we gonna do a 0.18.1 relatively soon then?