Improving compatibility with Scikit-learn

[daviddiazvico]

Summary

Move parameter validation and "_ parameter" setting from __init__ to fit in the sklearn wrapper.

Motivation

Attain full compatibility with sklearn and skopt automated hyper-parameter search.

Description

According to the Scikit-learn development guide, the __init__ method should not do any parameter validation or set any "trailing _" parameter.

In order to have an estimator compatible with the automated hyper-parameter search objects (GridSearchCV, RandomizedSearchCV, skopt.BayesSearchCV, etc.), that should be handled by the fit method.

In particular, the instantiation of _Booster and other parameters should be moved from __init__ to the beginning of fit in the sklearn wrapper.

This is related to (a sub problem of) #2628

Related PR: #2721

Thanks!

References

• Scikit-learn development guide

[StrikerRUS]

Refer to #2721 (comment).