yinghai / onnx-fb-universe

ONNX Integration Builds

onnx-fb-builds

Travis	Jenkins

Why Universe Repo?

• It integrates multiple repos, including onnx, onnx-caffe2, onnx-pytorch, pytorch, caffe2, to provide a convenient environment for developers.
• OnnxBot automatically updates the submodules in repos folder, runs the tests, and automatically detects the breaking changes from different repos.

Installation

./install.sh    # or ./install-develop.sh to install the most of the repos in develop mode.
./test.sh

Dockerfile may be helpful for setting up the whole repo.

Repo Structure

• repos: submodules
• test: test scripts
• jenkins: jenkins CI configuration
• travis: travis CI configuration

Test

Test Script

This repo contains several test scripts in test folder.

• test_caffe2.py: PyTorch ==> ONNX ==> Caffe2 end-to-end tests, in which the decimal precision is checked.
• test_models.py: PyTorch ==> ONNX model tests, check the output ONNX graph with expected files.
• test_operators.py: PyTorch ==> ONNX operator tests, check the output ONNX graph with expected files.
• test_verify.py: PyTorch ==> ONNX verification tests, check the exporting raises exceptions in certain conditions.
• test_pytorch_helper.py: PyTorch ==> ONNX helper tests, check the pytorch related helper functions.

To update the expected files, please add flag --accept to the end of the command.

Convert Test from PyTorch

Using convert_pytorch_test.py, we can easily turn some existing PyTorch testing cases to ONNX backend test cases. Just run:

python test/convert_pytorch_test.py

Generate Test from PyTorch Operator

Script test_operators.py is also able to generate ONNX backend test cases based on operator tests. Just run:

python test/test_operators.py --onnx-test

Check the Generated Tests

All the generated ONNX backend test data is stored in repos/onnx/onnx/backend/test/data/model.

But the generated test cases may not pass end-to-end tests, so we have another script filter_generated_test.py to check whether the cases are passed or not, and the failed data will be moved to test/failed/generated.

filter_generated_test.py supports the following flags:

• -v: verbose
• --delete: delete failed test cases, by default, they are moved to test/failed/generated.
• --no-expect: do not generate expect files for each test, by default, expected files are generated in test/expect folder for debugging purpose.

More Docs

• ONNX IR spec
• ONNX operator spec
• ONNX tutorials
• PyTorch docs
• Caffe2 docs