Converted LGBMClassifier gives inconsistent predictions

[mkoruszowic]

When converting a trained LGBMClassifier to ONNX, the ONNX model produces a different prediction than the original scikit-learn model for the same input.

---

Reproduction Steps

1. Unzip provided model:

lgbm_class.pickle.zip

2. Run the following code:

import pickle

import numpy as np
import onnxruntime
from lightgbm import LGBMClassifier
from onnxmltools.convert.lightgbm.operator_converters.LightGbm import convert_lightgbm
from skl2onnx import to_onnx
from skl2onnx import update_registered_converter
from skl2onnx.common.data_types import guess_data_type
from skl2onnx.common.shape_calculator import (
    calculate_linear_classifier_output_shapes,
)

update_registered_converter(
    LGBMClassifier,
    "LightGbmLGBMClassifier",
    calculate_linear_classifier_output_shapes,
    convert_lightgbm,
    options={"nocl": [True, False], "zipmap": [True, False, "columns"]},
)

X=np.array([[ 
         3.0000000e+00,  0.0000000e+00,  4.0000000e+00,  5.0000000e+00,
         2.5000000e+02,  1.0000000e+00,  1.0000000e+00,  0.0000000e+00,
         0.0000000e+00,  0.0000000e+00,  1.0000000e+00,  0.0000000e+00,
         0.0000000e+00,  1.0000000e+00,  2.0000000e+00,  0.0000000e+00,
         1.0000000e+00,  0.0000000e+00,  0.0000000e+00,  1.0000000e+00,
        -3.2525266e+03, -1.1329592e+01, -3.5146961e-01, -3.7130871e-01,
         2.6571992e-01, -9.1359066e-03, -5.1521581e-01,  7.3314957e-02,
         0.0000000e+00,  6.2500000e+04,  0.0000000e+00,  1.0578929e+07,
         2.7753743e-03,  1.2353089e-01,  2.5566691e-01,  1.1486175e+00,
         7.0607074e-02,  5.3750831e-03,  4.2937987e-05,  9.2572132e-05]],
      dtype=np.float32)


with open("lgbm_class.pickle", "rb") as f:
    sklearn_model = pickle.load(f)

sklearn_pred = sklearn_model.predict(X)

onnx_model = to_onnx(
        sklearn_model,
        target_opset={"": 19, "ai.onnx.ml": 3},
        initial_types=guess_data_type(X),
    )

sess = onnxruntime.InferenceSession(
    onnx_model.SerializeToString(), providers=["CPUExecutionProvider"],
)
input_feed={sess.get_inputs()[0].name: X.tolist()}
onnx_pred = sess.run(None, input_feed)[0]

print(f"{sklearn_pred=}")
print(f"{onnx_pred=}")

---

Output

sk_pred = array(['Risk'], dtype=object)
onnx_pred = array(['No Risk'], dtype=object)

ONNX env:

python - 3.11

lightgbm                  4.2.0                    pypi_0    pypi
onnx                      1.16.0                   pypi_0    pypi
onnxconverter-common      1.15.0                   pypi_0    pypi
onnxmltools               1.14.0                   pypi_0    pypi
onnxruntime               1.16.0                   pypi_0    pypi
onnxruntime-extensions    0.13.0                   pypi_0    pypi
skl2onnx                  1.19.1                   pypi_0    pypi

---

Expected

Both models should return the same prediction.