Convert_xgboost for XGBRegressor with objective="reg:logistic" gives wrong results

[laurentbeaughon]

When converting a XGBRegressor with objective="reg:logistic", the onnx model and the XGBoost model give different results.

The difference seems to be a sigmoid transformation that is not included in the model.

Here is a script to reproduce:

import numpy as np

from onnxmltools import convert_xgboost
from skl2onnx.common.data_types import FloatTensorType

X_train = np.random.rand(20, 10).astype(np.float32)
Y_train = np.random.randint(2, size=20)
X_test = np.random.rand(1, 10).astype(np.float32)

xg_reg = XGBRegressor(objective="reg:logistic").fit(X_train, Y_train)
xg_pred = xg_reg.predict(X_test)
xgb_onnx = convert_xgboost(xg_reg, initial_types=[("float_input", FloatTensorType([None, 10]))])

import onnxruntime
onx_reg = onnxruntime.InferenceSession(xgb_onnx.SerializeToString())
onnx_pred = onx_reg.run(None, {"float_input": X_test})


print(f"xgboost prediction: {xg_pred}")
print(f"onnx prediction: {onnx_pred}")

onnx_pred_fixed = 1 / (np.exp(-onnx_pred[0] + 0.5) + 1)

print(f"onnx prediction fixed: {onnx_pred_fixed}")