The tools and frameworks ecosystem around neural networks have been evolving rapidly. The interoperability between them, however, has been a challenge. It's not uncommon in the ML industry to quickly prototype and train the model in one tool and then deploy it in a different tool for inference. Because the internal format of these tools aren't compatible, we need to implement and maintain messy convertors to make the models compatible. The Open Neural Network Exchange format ONNX addresses this problem. In ONNX, the neural networks are represented as graphs using standard operator specifications, and together with a serialization format for trained weights, neural network models can be transferred from one tool to another. This opens up lots of possibilities, including Model Zoo, a collection of pretrained models in ONNX format.
