Microsoft and Facebook announce the Open Neural Network Exchange format. Both parties’ open source project brings interoperability to their artificial intelligence frameworks, which should lead to broad optimization.
Specifically, Open Neural Network Exchange is about the representation of neural networks. Developers create computation graphs for this, but each framework has its own format for this. If developers choose the properties of one framework during the research phase, but prefer properties of another framework during the production phase, the conversion of the representation takes valuable time.
Frameworks that use the onnx format simplify that conversion. In addition, organizations looking to increase the performance of neural networks can more easily target multiple frameworks by using the onnx representation. Microsoft’s Cognitive Toolkit and Caffe2 and PyTorch, originally developed by Facebook, support the Open Neural Network Exchange format, Microsoft reports. The code and documentation for the project has been posted on GitHub.