Open Neural Network Exchange (ONNX)

An open, general-purpose file format for neural networks that makes it easy to move models between frameworks and inference engines.

ONNX is an open model-representation format launched in 2017 by Microsoft and Facebook, now hosted by LF AI & Data under the Linux Foundation. It stores a network's architecture and weights as a unified computation graph, and often serves as the bridge when a model is trained in PyTorch but deployed with TensorRT, ONNX Runtime, OpenVINO, or RKNN. The most common headache during deployment is an unsupported operator or a mismatched export version (opset), which usually requires rewriting part of the model or swapping operators. For newcomers, it's enough to think of it as the “universal file format” of the model world.

ExampleUse torch.onnx.export to export a trained walking policy as policy.onnx, then hand it to ONNX Runtime for inference on the robot.

Also called

ONNX

Related

ONNX Runtime · NVIDIA TensorRT · Intel OpenVINO · Rockchip RKNN-Toolkit · Inference Deployment · PyTorch

Sources

ONNX | Home