In this second push to make AI a reality, several companies are working on their own deep learning models with the user data they have acquired over years. For example, Facebook’s Caffe2, Microsoft’s Cognitive Toolkit, Google’s TensorFlow, PyTorch, and many more.
AI is as real as the amount and quality of data one has to build it. Thus, we find that AI frameworks from one company is good at something, while another company is good at something else. To take great ideas to reality, companies have come together to partner as ONNX.
ONNX is a open standard format for deep Neural Networks. You can train a model in one framework and use it in another. ONNX enables models to inter-operate between different AI frameworks and allows you to develop AI solutions taking best from each framework.
Caffe2, Cognitive Toolkit and PyTorch today support exporting and importing deep learning models in ONNX formats. There are connectors being developed for other AI frameworks.
An interesting article on Medium talks about importing an ONNX model in MXNet. It is a good read and worth trying.