predictions results do not match original lightgbm

[JeyKJey]

Dependencies:
lightgbm==2.2.2
treelite==0.32,
mac, python3

It seems it is somehow related to this issue as well.

LightGBM was trained with the following parameters:

params = {
            'learning_rate': 0.1,
            'n_estimators': 1000,
            'max_depth': 4,
            'random_state': 1
        }

Save the model and load predictor:

joblib.dump(model, model_path)
treelite_model = treelite.Model.load(model_path, model_format='lightgbm')
treelite_model.export_lib(toolchain='gcc', libpath=str(directory) + '/model.so', verbose=False, params={'parallel_comp':10})
predictor = Predictor(str(directory / 'model.so'), verbose=True)

Predict with original model and treelite:

print(model.predict(sample_data))
print(predictor.predict_instance(sample_data.values[0]))

[2.786864]
2.7558372

The difference looks small, but once it is exponentiated, it starts to be big. I tested it with different number of trees and the less trees I use the less different are the predictions.

[JeyKJey]

For some reason, when I rerun this code in jupyter notebook without restarting, it produces the exact results above. However, if I do restart the whole notebook, it would produce correct results. Therefore, I will close this issue. Seems that something doesnt get set/updated whatevery properly when not restarting.