ONNX

A shared representation for models

ONNX defines a format and operator vocabulary for representing machine-learning models. It sits between training frameworks and tools that execute a model, reducing the need to rewrite a network for each environment. Compatibility still depends on the operators, versions and execution support involved; a shared format cannot make every model interchangeable automatically. [1]

Microsoft and Facebook announced ONNX in September 2017. Their initial collaboration connected model representations across frameworks including PyTorch, Caffe2 and Microsoft's Cognitive Toolkit. That announcement preceded the separate December version-one release, so the initial milestone is not a claim that the final version-one format shipped in September. [2]

The present project has open community governance. ONNX is also distinct from ONNX Runtime, an execution engine, and from the weights saved in an ONNX file. A format's openness does not replace an individual model's license or establish the accuracy of its predictions. [1]