Mismatch between XGBoost and converted onnx, with respect to "colsample_bytree"

[chunhaoIcici]

onnxmltools version: 1.11.2

Dear onnxmltools developers,

I met an issue when converting an Booster from XGBoost to onnx. If I train the model by setting "colsample_bytree" to be 1, the prediction difference between XGBoost and onnx can be within 0.1%. However, when I set this hyper-parameter to be other values like 0.8, the difference can be up to 10%. I'm not sure whether such kind of issue has been found or addressed. If you need me to further clarify the issue, I can provide a demo to reproduce my bug.

Best regards,
Chunhao