onnx

Why we included this project

ONNX is the format you will most likely run into when a trained model needs to move between frameworks. It defines a computation graph, a standard set of operators, and the data types those operators work on, so a model exported from PyTorch or TensorFlow can be loaded and scored by a different runtime, or deployed to specialized inference hardware. The repo doubles as the reference: it holds the specification, the Python package for building and inspecting ONNX graphs, and reference implementations of the operators, so it is useful both as a dependency and as the place to check when you want to know exactly how the format behaves. If you are comparing runtimes, porting a model across stacks, or targeting edge devices, reading the spec is a practical step, not an academic one. Just keep in mind ONNX is a standard rather than a product, so you will still need an exporter on one side and a runtime on the other.