ONNX XGBoost classifier gives very different results from the original model

[xiaohk]

Describe the bug

I am converting a simple XGBoost model to ONNX format. However, the results of the ONNX model mismatch the original model, yielding way worse performance.

Urgency

Quite urgent. This bug can cause harm if people deploy their XGBoost model in ONNX format without noticing this error.

System information

• MacOS 12.4 or Google Colab
• ONNX Runtime installed from (source or binary): binary
• ONNX Runtime version: 1.11.1
• Python version: 3.7.10

To Reproduce

1. Train an XGBoost classifier on the Adult dataset
2. Covert the XGBoost model to ONNX using convert_xgboost
3. Inference ONNX model using InferenceSession

This bug is reproduced in this Colab notebook

Expected behavior

I expect the results of the ONNX model to match the original XGBoost model.

Screenshots

See this Colab notebook

[xadupre]

Thanks for sharing the code was. I was able to replicate the issue. The bug happens when base_score is specified. It needs a fix into onnxmltools.

[xiaohk]

Thanks @xadupre! Should I open a new issue in onnxmltools?

[xadupre]

I created a PR on onnxmltools to fix the issue.