ONNX
also: open neural network exchange
An open format for exchanging trained models between frameworks, so a model trained in one tool runs in another.
ONNX (Open Neural Network Exchange) is an open, standardized format for representing trained models, so a model built in one framework (PyTorch) can be exported and run in another runtime or on different hardware without a rewrite. It decouples where a model is trained from where it runs.
Worked example: a team trains in PyTorch, exports to ONNX, and serves it with ONNX Runtime optimized for their production CPUs/GPUs — often faster than the original framework, and portable to mobile or the browser. Gotcha: ONNX export is not always seamless — custom or newest operators may be unsupported and need workarounds, and numerical results can differ slightly across runtimes — so re-validate the exported model’s outputs against the original rather than assuming a perfect round-trip.