The ONNX standard defines an open format for representing machine learning models as a computation graph of typed operators, enabling models to move between training and inference frameworks.
Semantic Classification
Content
- ONNX captures a model as a directed graph of operators with defined semantics and tensor types, plus an opset version that fixes operator behaviour. Exporters and runtimes target this format so a model trained in one framework can be executed in another.
- Model interoperability matters for deploying neural networks inside engines and XR runtimes, where the training framework and the inference target differ. ONNX provides a stable intermediate representation for that handoff.