Development#

To build ONNX from source please follow the instructions listed here.

Then, after you have made changes to Python and C++ files:

  • Python files: the changes are effective immediately in your installation. You don’t need to install these again.

  • C++ files: you need to install these again to trigger the native extension build.

Assuming build succeed in the initial step, simply running

pip install -e .

from onnx root dir should work.

Folder structure#

  • onnx/: the main folder that all code lies under

    • onnx.proto: the protobuf that contains all the structures

    • checker.py: a utility to check whether a serialized ONNX proto is legal

    • shape_inference.py: a utility to infer types and shapes for ONNX models

    • version_converter.py: a utility to upgrade or downgrade version for ONNX models

    • parser.py: a utility to create an ONNX model or graph from a textual representation

    • hub.py: a utility for downloading models from ONNX Model Zoo

    • compose.py: a utility to merge ONNX models

    • helper.py: tools for graph operation

    • defs/: a subfolder that defines the ONNX operators

    • test/: test files

Generated operator documentation#

Operator docs in Operators.md are automatically generated based on C++ operator definitions and backend Python snippets. To refresh these docs, run the following commands from the repo root and commit the results. Note ONNX_ML=0 updates Operators.md whereas ONNX_ML=1 updates Operators-ml.md:

# Windows
set ONNX_ML=0
# UNIX
# export ONNX_ML=0
pip install setup.py
python onnx/defs/gen_doc.py

Adding a new operator#

ONNX is an open standard, and we encourage developers to contribute high quality operators to ONNX specification. Before proposing a new operator, please read the tutorial.

Code style#

We use flake8, mypy, and clang-format for checking code format. Note: You’ll find the versions of these tools in setup.py. You can run these checks by:

pip install -e .[lint]

./tools/style.sh

Testing#

ONNX uses pytest as a test driver. To run tests, you’ll first need to install pytest:

pip install pytest nbval

After installing pytest, run from the root of the repo:

pytest

to begin the tests.

You’ll need to regenerate test coverage too, by running this command from the root of the repo:

python onnx\backend\test\stat_coverage.py

Cpp tests (googletest)#

Some functionalities are tested with googletest. Those tests are listed in test/cpp, and include tests for shape inference, data propagation, parser, and others.

To run them, first build ONNX with -DONNX_BUILD_TESTS=1 or ONNX_BUILD_TESTS=1 pip install -e ..

Linux and MacOS#

The cpp tests require dynamically linking to built libraries.

export LD_LIBRARY_PATH="./.setuptools-cmake-build/:$LD_LIBRARY_PATH"
.setuptools-cmake-build/onnx_gtests

Windows#

# If you set DEBUG=1, use `.setuptools-cmake-build\Debug\onnx_gtests.exe` instead
.setuptools-cmake-build\Release\onnx_gtests.exe

Static typing (mypy)#

We use mypy to run static type checks on the onnx code base. To check that your code passes, you’ll first need to install the mypy type checker. If you’re using python 3, call from your onnx source folder:

pip install -e .[lint]

Note: You’ll find the version we’re currently using in setup.py.

After having installed mypy, you can run the type checks:

python setup.py typecheck

CI Pipelines#

Every PR needs to pass CIs before merge. CI pipelines details are here. Passing Lint Python CI is not required but recommended.

Other developer documentation#

License#

Apache License v2.0

Code of Conduct#

ONNX Open Source Code of Conduct