Development
You will need to install protobuf and numpy to build ONNX. An easy way to get these dependencies is via Anaconda:
# Use conda-forge protobuf, as defaults doesn't come with protoc
conda install -c conda-forge protobuf numpy
During development, it's convenient to install ONNX in development mode (to disable ONNX-ML, set environment variable ONNX_ML=0):
git clone --recursive https://github.com/onnx/onnx.git
pip install -e onnx/
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.
Folder structure
onnx/: the main folder that all code lies underonnx.proto: the protobuf (v2.6.1) that contains all the structureschecker.py: a utility to check whether a serialized ONNX proto is legalhelper.py: tools for graph operationdefs/: a subfolder that defines the ONNX operatorstest/: 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:
set 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.
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
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 .[mypy]
The type checker cannot run in a python 2 environment (but it will check python 2 code). If you're using python 2, you need to install mypy into your system packages instead:
pip3 install mypy==[version]
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
Other developer documentation
- How to implement ONNX backend (ONNX to something converter)
- Backend test infrastructure and how to add tests