Hey folks,
Today AWS announced contributing ONNX-MXNet, an open source Python package that imports ONNX models into MXNet. @roshrini and I (@lupesko) have worked on the code, which is now publicly available [1], and published a blog post demonstrating usage of the package [2]. Special thanks to dmlc/nnvm team, whose ONNX code was used as a reference for this implementation. What is ONNX? ONNX is an open source format to encode deep learning models. ONNX defines a format to store neural network's computational graph, as well as a storage format for operators used within a neural network graph. For more details, check out onnx.ai [3]. Why I think ONNX is important for MXNet? ONNX is an emerging standard, that holds a lot of potential for Deep Learning practitioners. With ONNX, people can create and train a network with framework A, and deploy it for inference with framework B. The blog post we published demonstrates using a Super Res model trained with PyTorch, and importing it into MXNet Symbolic API for inference. I strongly believe that adopting ONNX early on adds value for deep learning practitioners, and thus supporting it adds value for MXNet as well. As for next steps, I was thinking that porting the functionality and code into MXNet is the logical next step. Would love to get the community's feedback and contributions! [1] https://github.com/onnx/onnx-mxnet [2] https://aws.amazon.com/blogs/ai/announcing-onnx-support-for-apache-mxnet/ [3] https://onnx.ai
