Different score after compiling a lightgbm model

[AlexQie]

I have a lightgbm model and I use treelite to compile a .so file, but they have different predict results:

and here is my code:

predictor = treelite.runtime.Predictor('./mymodel.so', verbose=True)
data = np.loadtxt('service/sample.txt')

batch = treelite.runtime.Batch.from_npy2d(data)
out_pred = predictor.predict(batch)

lgb_model = lgb.Booster(model_file=tree_file_path)
lgb_pred = lgb_model.predict(data)
print(out_pred)
print(lgb_pred)

and I have the following configuration:

>>> import treelite
>>> treelite.__version__
'0.32'
>>> import lightgbm
>>> lightgbm.__version__
'2.2.3'
>>> 

[hcho3]

This is probably the same issue as #98 and #95. Treelite currently uses single-precision float (float) to store thresholds, whereas LightGBM models require double-precision floats (double). I'm working to fix the issue as part of 1.0 release (#111)

[eddy-avanan]

I have same issue with XGBoost

[shashankg7]

Hi,

I am facing the same issue (with a LightGBM ranking model).

Did we get this issue resolved?

Thanks

[xcTorres]

Also happened

[hcho3]

The recent refactor (#196, #198, #199, #201, #203) added an option to use float64 data type, so that LightGBM models can be faithfully represented.