gbm.best_iteration_ is None when early_stopping_round = 3 in sklearn API

[startakovsky]

The code in the image below indicates that in the sklearn API, best_iteration_ is not set even when early_stopping_round > 0.

Is this expected behavior?

[guolinke]

@startakovsky
I think early_stopping_rounds is in fit function: https://github.com/Microsoft/LightGBM/blob/master/python-package/lightgbm/sklearn.py#L295-L300 .
Not the early_stopping_round in Constructor.

ping @StrikerRUS @wxchan
However, I am confused about why early stopping is working in @startakovsky 's case.
BTW, why best_score is set independently ? https://github.com/Microsoft/LightGBM/blob/master/python-package/lightgbm/sklearn.py#L478-L481

updated:
I think the reason is:
https://github.com/Microsoft/LightGBM/blob/master/python-package/lightgbm/engine.py#L101-L105

@wxchan @StrikerRUS Should we have the same conversion in sklearn interface?

[startakovsky]

@guolinke the following does work... and for now I will use this approach to track gbm.best_iteration_. Just hoping it would also work when supplied to params (i.e. lgb.LGBMClassifier(**params).

[StrikerRUS]

@guolinke

BTW, why best_score is set independently ? https://github.com/Microsoft/LightGBM/blob/master/python-package/lightgbm/sklearn.py#L478-L481

The answer is #890.

[StrikerRUS]

@guolinke
Yeah, it seems that in this situation early_stopping_rounds is passed through params into train() and used there.

Should we have the same conversion in sklearn interface?

In my opinion, all users are adult enough to read the docstring and notice that early_stopping_rounds has own parameter and shouldn't be passed via **kwargs.

[guolinke]

@startakovsky I agree with @StrikerRUS . Logically, early_stopping_rounds is used to control the fit(training) process, so using it in fit is more reasonable.