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 onnxoptimizerNote that you may need to upgrade your pip first if you have trouble:
pip3 install -U pipIf 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.
- Command-line API (e.g.
python3 -m onnxoptimizer model.onnx output.onnx) - More built-in pass
- Separate graph rewriting and constant folding (or a pure graph rewriting mode, see issue #9 for the details)
-
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