ONNX provides a C++ library for performing arbitrary optimizations on ONNX models, as well as a growing list of prepackaged optimization passes.
The primary motivation is to share work between the many ONNX backend implementations. Not all possible optimizations can be directly implemented on ONNX graphs - some will need additional backend-specific information - but many can, and our aim is to provide all such passes along with ONNX so that they can be re-used with a single function call.
You may be interested in invoking the provided passes, or in implementing new ones (or both).
You can install onnxoptimizer from PyPI:
pip3 install onnxoptimizer
Note that you may need to upgrade your pip first if you have trouble:
pip3 install -U pip
If you want to build from source:
git clone --recursive https://github.com/onnx/optimizer onnxoptimizer cd onnxoptimizer pip3 install -e .
Note that you need to install protobuf before building from source.
python3 -m onnxoptimizer model.onnx output.onnx)
onnx-simplifier: A handy and popular tool based on onnxoptimizer
convertmodel.com: onnx optimizer compiled as WebAssembly so that it can be used out-of-the-box