LGBM Inference Optimization

optimize_and_predict_daal4py.py

import daal4py as d4p
from lightgbm import LGBMClassifier

model = LGBMClassifier()
model.fit(X_data, y_data)
y_pred_orig = model.predict_proba(x_data_dense)[:,1]

daal_model = d4p.get_gbt_model_from_lightgbm(model.booster_)

predictions_container = d4p.gbt_classification_prediction(nClasses=2, resultsToEvaluate='computeClassProbabilities', fptype='float')
y_pred_opti = predictions_container.compute(X_data, daal_model).probabilities[:,1]

np.allclose(y_pred_orig, y_pred_opti)

optimize_and_predict_treelite.py

#https://treelite.readthedocs.io/en/latest/tutorials/import.html#importing-lightgbm-models
import treelite
from lightgbm import LGBMClassifier

model = LGBMClassifier()
model.fit(X_data, y_data)

model.booster_.save_model('my_model.txt')
treelite_model = treelite.Model.load('my_model.txt', model_format='lightgbm')

# for windows libpath='my_model.dll', toolchain = 'msvc'
treelite_model.export_lib(libpath='my_model.so', toolchain = 'gcc', verbose=True)


# Inference
import treelite_runtime
predictor = treelite_runtime.Predictor('./mymodel.so', verbose=True)
y_pred_opti = predictor.predict(X_data)