loading lightgbm.basic.Booster

[marsupialtail]

Is there a way to directly convert a LightGBM booster object instead of using the sklearn API?

With the sklearn API for lgb, I think you can only load file to a booster object, and I'm not sure how to initialize a lgb regressor from a booster object.

Currently when I try convert booster object directly I get: hummingbird.ml.exceptions.MissingConverter: Unable to find converter for model type <class 'lightgbm.basic.Booster'>.

[ksaur]

Right now we only support direct conversion from LGBM Classifier/Ranker/Regressor, so the error you get is expected.

I'm not sure how much additional effort it would be to support lightgbm.Booster; I'll add it to feature requests list.

Other thoughts @interesaaat ?

[interesaaat]

Hey! I think it should be doable. I saw that we get all infos from the booster_ object.

@marsupialtail I think it is a matter of adding LightGBM booster among the supported operators. Do you want to give it a shot?

[marsupialtail]

If you point me to a couple places in the code you think would serve as a good reference, I can try to work on this.

[ksaur]

@marsupialtail - PR #173 (LGBMRanker) is a good starting point! For Booster, you'll probably need to add a separate converter function in hummingbird/ml/operator_converters/lightgbm.py. (For the Ranker one in the PR, it was able to reuse the convert_sklearn_lgbm_regressor.)

I'm not sure what else will be needed, but as Matteo showed above, the information will be there in line 99.

Please reach out if you get stuck!

[marsupialtail]

I made a PR. Please review. Thanks.

[alexyar88]

You can create sklearn object from your booster. Something like that:

lgb_clf = lgb.LGBMClassifier(n_estimators=booster.current_iteration())
lgb_clf._Booster = deepcopy(booster)
lgb_clf._n_features = len(booster.feature_name())
lgb_clf._n_classes = len(np.unique(y))
lgb_clf._le = _LGBMLabelEncoder().fit(y)