To create a more open artificial intelligence (AI) ecosystem and make AI more accessible and valuable for everyone, Microsoft recently made an announcement to introduce the Open Neural Network Exchange (ONNX) format in association with Facebook.
An open source project co-developed by Microsoft and Facebook, ONNX is an algorithm sharing platform that offers a shared model representation for interoperability and innovation across AI framework ecosystem.
ONNX will have support of Microsoft’s Cognitive Toolkit, Caffe2 and PyTorch, all of which are lightweight, modular, and scalable deep learning frameworks. It will allow trained models in the frameworks to be exported to another, for inference.
It helps the AI developers to choose the optimum framework for the current stage of their project, and let them easily switch between the frameworks as per the project evolvement. Each framework provides particular characteristics like fast training, inferencing on mobile devices, supporting flexible network architectures, etc.
The optimizations need to be frequently integrated separately into each framework, which at times, is a time-consuming process. Using the ONNX, the performance of neural networks with multiple frameworks can be improved.
“In Facebook’s AI teams (FAIR and AML), we are continuously trying to push the frontier of AI and develop better algorithms for learning. When we have a breakthrough, we’d like to make the better technologies available to people as soon as possible in our applications. With ONNX, we are focused on bringing the worlds of AI research and products closer together so that we can innovate and deploy faster.” – Joaquin Quinonero Candela, Facebook’s Director of Applied Machine Learning.
Also read: Microsoft’s new AI project will focus on accelerated and real time AI in the cloud
ONNX code and documentation are available open source on GitHub. Microsoft will be working actively on ONNX and Cognitive Toolkit’s upcoming release will include its support. With Facebook, Microsoft is also planning to contribute examples, reference implementations, a model zoo and tools. Both the companies will continue to develop ONNX, PyTorch, and Caffe2 to have the latest tools of AI.
