Why is _CVBooster object hidden class?

[matsuken92]

Some of Kaggler use out-of-fold prediction (oof) which is calculated with validation dataset for stacking, blending, evaluation etc. In order to calculate oof with lgb.cv(), we need to return Boosters however it is not returned from the function. If I customize to return _CVBooster object, it can be calculated, but _CVBooster is hidden class, so I think it is not appropriate to return officially.

Is it possible to remove prefix underscore from _CVBooster? Then I can customize to return the object on lgb.cv().

I would appreciate it if you could advise to it.

_CVBooster

https://github.com/Microsoft/LightGBM/blob/master/python-package/lightgbm/engine.py#L248-L267

Example

https://www.kaggle.com/kenmatsu4/using-trained-booster-from-lightgbm-cv-w-callback

[guolinke]

ping @StrikerRUS

[StrikerRUS]

@guolinke When I joined LightGBM _CVBooster had been already implemented as protected class: #166.
I do not mind make this class public. But is it really needed? This class is pretty small and I suppose it's possible to achieve the goal without this class.

@matsuken92 Isn't ordinary Python list of Boosters not enough?
We already have a discussion about enhancement of cv return value here and here. Can you please look through these issue and tell you opinion? Maybe you have an idea of complete cv function redesign which will help to implement new features? Otherwise, please think about simple list or any other appropriate data structure. Because I think that end users do not need a full API of _CVBooster, but just a small part of it.

Anyway, PR from you in any form is very welcome!

[matsuken92]

Thank you for your comment! Minimum requirements of my thought are:

• Returning list of Boosters
• Returning best_iteration to identify which iteration is best for predict when using early stopping.

In addition to above, current _CVBooster.__getattr__ function is useful for batch operation of boosters on each fold. The following link is the part of usage of the _CVBooster.__getattr__ calculating prediction result for test data.

https://github.com/matsuken92/LightGBM_test/blob/master/src/cv_booster_test.py#L117

I will send my idea as PR, so I would appreciate it if you could review it then!

[StrikerRUS]

@matsuken92 I see! Thanks for the clarification!

Of course, I'll be happy to review it.

Can you please also think about backward compatibility in terms of cv function return type? Maybe a new key in results dict or something else, idk...

[StrikerRUS]

Closing in favor of being in #2302. We decided to keep all feature requests in one place.

Welcome to contribute this feature! Please re-open (or post a comment if you are not a topic starter) this issue if you are actively working on implementing this feature.

[StrikerRUS]

Closed via #3204.