tree_implementation parameter gets overwritten with onnx backend

Hello,

I have been trying to use the gemm with onnx backend but the onnx graph did not really look like we were using gemm tree implementation:

model = convert(skl_model, backend="onnx", test_input=X, extra_config={"tree_implementation":"gemm"})

It seems that the tree_implementation parameter gets overwritten deeper in the code:

        if backend == onnx.__name__:
            # vers = LooseVersion(torch.__version__)
            # allowed_min = LooseVersion("1.6.0")
            # Pytorch <= 1.6.0 has a bug with exporting GEMM into ONNX.
            # For the moment only tree_trav is enabled for pytorch <= 1.6.0
            # if vers < allowed_min:
            extra_config[constants.TREE_IMPLEMENTATION] = "tree_trav"

I am not sure why this is done tho? Removing this part gives me back the expected behavior.

Could you give a bit more info on what is the point of this? Can I do anything to fix this?

[jfrery]

Yes of course I will prepare that. I think it is better to raise an exception rather than overwrite the parameter given by the user if the pytorch version is too low.

Does this sound right? I will have a PR soon.

[interesaaat]

It looks that this error is coming from the onnxruntime. To bypass this, we can force onnxml models to not use gemm, and open an issue regarding this.

You can extra_config[constants.TREE_IMPLEMENTATION] = "tree_trav" somewhere in this method. In this way sklearn models will use gemm (eventually), while onnxml models will stil use tree_trav until we understand where this new error is coming from.