[davidatbu]

I'd love to hear about this too: especially after running the model through an onnx optimizer, like this one [0].

[0] https://github.com/daquexian/onnx-simplifier