Predictions of converted LightGbm ONNX model completely mismatched ML.Net equivalent

[mentorfloat]

System information

• OS version: 10.0.19041
• .NET SDKs: 3.1.402, 5.0.100-preview.7.20366.6

Issue

• What did you do?
  Convert three trained ML.Net models (LightGbm, SdcaMaximumEntropy, LbfgsMaximumEntropy) to ONNX models with ConvertToOnnx and make predictions with 100 randomly selected data, using both the original ML.Net models and the converted ONNX models.

• What happened?
  When making predictions with the converted LightGbm ONNX model, none of the predicted values match the predicted values made using the original ML.Net model. To test further, I repeated the pipeline with different trainers, SdcaMaximumEntropy and LbfgsMaximumEntropy. The converted ONNX models for these two returned predicted values that match the predictions made using their ML.Net equivalents.

• What did you expect?
  LightGbm's converted ONNX model generate the same predictions as its ML.Net counterpart. Similar to how the other two trainers' did.

• Side notes:
  Initially I couldn't get LightGbm to convert to ONNX, but that's resolved now. Closed Issue ConvertToOnnx with LightGbm Model Throws KeyNotFoundException #5542 , but likely unrelated to the above.

[luisquintanilla]

ML.NET 3.0 has been upgraded to a new version of LightGBM which may have fixed this issue. Please open a new issue if it persists.