The code that Imports/Exports a model into ONNX is a separate package for Caffe2 and is part of the ONNX project - https://github.com/onnx/onnx-caffe2.
Whereas for PyTorch the import/export functionality ships as part of the PyTorch package itself.
Any reason for this difference? Were there any design decisions or tradeoff involved while making this call?