Using ONNX to export PyTorch models

There are scenarios in production systems where most of the already deployed machine learning models are written in a certain deep learning library, say, TensorFlow, with its own sophisticated model-serving infrastructure. However, if a certain model is written using PyTorch, we would like it to be runnable using TensorFlow to conform to the serving strategy. This is one among various other use cases where a framework such as ONNX is useful.

ONNX is a universal format where the essential operations of a deep learning model such as matrix multiplications and activations, written differently in different deep learning libraries, are standardized. It enables us to interchangeably use different deep learning libraries, programming languages, and even operating environments to run the same deep learning model.

Here, we will demonstrate how to run a model, trained using PyTorch, in TensorFlow. We will first export the PyTorch model into ONNX format...