SaveOnnxCommand appears to ignore predictors when saving a model to ONNX format.

[pieths]

System information

• OS version/distro: Windows 10

Steps To Recreate The Issue

1. Create and save a PredictorModel to disk using the entry point api.

2. Try and convert the model to ONNX format using the entry point api.

3. Notice that SaveOnnxCommand.GetPipe only cycles through the transforms and never encounters the logistic regression node.

   This might be happening because ExecuteGraphCommand.GetOutputToPath saves a TlcModule.DataKind.PredictorModel to disk in step (1). And then, ExecuteGraphCommand.SetInputFromPath loads a TlcModule.DataKind.TransformModel from disk in step (2) (apparently a consequence of SaveOnnxCommand.Arguments.Model being of type TransformModel). PredictorModelImpl and TransformModelImpl don't appear to be compatible from a serialization point of view.

Source code / logs

See here for an ml.net test which demonstrates the issue.

[wschin]

Thanks a lot for pointing this out. I just made a solution based on your finding.