Interested in contributing to TensorFlow Addons? We appreciate all kinds of help and are working to make this guide as comprehensive as possible. Please let us know if you think of something we could do to help lower the barrier to contributing.
We gladly welcome pull requests.
Have you ever done a pull request with GitHub? If not we recommend you to read this guide to get your started.
Before making any changes, we recommend opening an issue (if it doesn't already exist) and discussing your proposed changes. This will let us give you advice on the proposed changes. If the changes are minor, then feel free to make them without discussion.
All submissions, including submissions by project members, require review.
All new components/features to Addons need to first be submitted as a feature
request issue. This will allow the team to check with our counterparts in the TF
ecosystem and ensure it is not roadmapped internally for Keras or TF core. These
feature requests will be labeled with ecosystem-review while we determine if it
should be included in Addons.
The tensorflow/addons repository contains additional functionality fitting the following criteria:
- The functionality is not otherwise available in TensorFlow
- Addons have to be compatible with TensorFlow 2.x.
- The addon conforms to the code and documentation standards
- The addon is impactful to the community (e.g. an implementation used in widely cited paper)
- Lastly, the functionality conforms to the contribution guidelines of its subpackage.
Note: New contributions often require team-members to read a research paper and understand how it fits into the TensorFlow community. This process can take longer than typical commit reviews so please bare with us
- Docker (code formatting / testing)
- Nvidia-docker (for GPU testing, optional)
- Bazel installed locally (to build custom ops locally, optional)
- NVCC/Cuda installed locally (to build custom ops with gpu locally, optional)
- Docker (code formatting / testing)
- Bazel installed locally (to build custom ops locally, optional)
For Windows, you have two options:
WSL 2 is a very light virtual machine running with hyper-V. When running in
WSL 2, you're in a full linux environment, with a real linux kernel.
WSL 2 networking is shared with Windows and your Windows files can be found under
/mnt/c. When working with WSL 2, you can just follow the linux guides and tutorials
and everything will work as in linux, including
Docker (that means you install docker with apt-get), git, ssh...
See the WSL 2 install guide.
This is if you want to stay in Windows world. In this case, you need:
- Git with git bash install on the PATH (meaning that you can run the
shcommand from Powershell). - Docker desktop with Linux containers (code format, testing on linux, etc...)
- A local Python installation
- Bazel (if you want to compile custom ops on Windows, optional)
- Visual Studio build tools 2019 install with chocolatey or install manually (if you want to compile custom ops on windows, optional).
If you develop on Windows and you encounter issues, we'd be happy to have your feedback! This link might help you.
Try these useful commands below, they only use Docker and don't require anything else (not even python installed):
- Format code automatically:
bash tools/pre-commit.sh - Run sanity check:
bash tools/run_sanity_check.sh - Run CPU unit tests:
bash tools/run_cpu_tests.sh - Run GPU unit tests:
bash tools/run_gpu_tests.sh
If you're running Powershell on Windows, use sh instead of bash when typing the commands.
We provide a pre-commit hook to format your code automatically before each commit, so that you don't have to read our style guide. Install it on Linux/MacOS with
cd .git/hooks && ln -s -f ../../tools/pre-commit.sh pre-commitand you're good to go.
On Windows, in powershell, do:
cd .git/hooks
cmd /c mklink pre-commit ..\..\tools\pre-commit.shNote that this pre-commit needs Docker to run. If you have docker 19.03+, it uses Docker buildkit to make the build step much faster.
See our Style Guide for more details.
Nightly CI tests are ran and results can be found on the central README. To subscribe for alerts please join the addons-testing mailing list.
When running outside Docker, you can use your IDE to debug, and use your local tools to work.
If you're just modifying Python code (as opposed to C++/CUDA code), then you don't need to use Bazel to run your tests. And you don't need to compile anything.
If you want to work in a virtualenv:
pip install virtualenv
venv my_dev_environement
source my_dev_environement/bin/activate # Linux/macos/WSL2
.\my_dev_environement\Scripts\activate # PowerShellIf you want to work in a conda environment:
conda create --name my_dev_environement
conda activate my_dev_environementJust run from the root:
pip install tensorflow==2.4.0
# you can use "pip install tensorflow-cpu==2.4.0" too if you're not testing on gpu.
pip install -e ./It's going to install Addons in editable mode without compiling anything.
You can modify source files and changes will be seen at the next Python
interpreter startup. This command needs to be executed only once.
Now, anywhere on your system, if you do import tensorflow_addons, it's
going to import the code in this git repository.
To undo this operation, for example, you want to later on install TensorFlow Addons from PyPI, the release version, do:
pip uninstall tensorflow-addonsIf TensorFlow Addons is installed in editable mode, you can then just run your tests by running Pytest. For example:
pip install -r tools/install_deps/pytest.txt
python -m pytest tensorflow_addons/rnn/tests/cell_test.py
# or even
python -m pytest tensorflow_addons/rnn/
# or even
python -m pytest tensorflow_addons/
# or even if pytest is in the PATH
pytest tensorflow_addons/Pytest has many cool options to help you make great tests:
# Use multiprocessing to run the tests, 3 workers
pytest -n 3 tensorflow_addons/
pytest -n auto tensorflow_addons/
# Run the whole test suite without compiling any custom ops (.so files).
pytest -v --skip-custom-ops tensorflow_addons/
# Open the debugger to inspect variables and execute code when
# an exception is raised.
pytest --pdb tensorflow_addons/
# or if you prefer the Ipython debugger
pytest --pdb --pdbcls=IPython.terminal.debugger:TerminalPdb --capture no tensorflow_addons/
# by defaults print() aren't displayed with pytest
# if you like to debug with prints (you might get
# the output scrambled)
pytest -s tensorflow_addons/
# get the list of functions you ran
pytest -v tensorflow_addons/
# to rerun all previous tests, running the ones that failed first
pytest --ff tensorflow_addons/
# You know which function to execute, but you're too
# lazy to type the file path
pytest -k "test_get_all_shared_objects" ./tensorflow_addons/
# get the 10 slowest functions
pytest --duration=10 tensorflow_addons/Pycharm has a debugger build in the IDE for visual inspection of variables and step by step executions of Python instructions. It can run your test functions from the little green arrows next to it. And you can add breakpoints by just clicking next to a line in the code (a red dot will appear).
But in order for the debugger to run correctly, you need to specify that you use pytest as your main test runner, not unittest (the default one).
For that, go in File -> Settings -> search box -> Default test runner -> Select "Pytest".
If you need a custom C++/Cuda op for your test, compile your ops with
python configure.py
pip install tensorflow==2.4.0 -e ./ -r tools/install_deps/pytest.txt
bash tools/install_so_files.sh # Linux/macos/WSL2
sh tools/install_so_files.sh # PowerShellNote that you need bazel, a C++ compiler and a NVCC compiler (if you want to test Cuda ops). For that reason, we recommend you run inside the custom-op docker containers. This will avoid you the hassle of installing Bazel, GCC/clang... See below.
Running tests interactively in Docker gives you good flexibility and doesn't require to install any additional tools.
CPU Docker:
docker run --rm -it -v ${PWD}:/addons -w /addons tfaddons/dev_container:latest-cpuGPU Docker:
docker run --gpus all --rm -it -v ${PWD}:/addons -w /addons gcr.io/tensorflow-testing/nosla-cuda11.0-cudnn8-ubuntu18.04-manylinux2010-multipythonConfigure:
python3 -m pip install tensorflow==2.4.0
python3 ./configure.py # Links project with TensorFlow dependencyInstall in editable mode
python3 -m pip install -e .
python3 -m pip install -r tools/install_deps/pytest.txtCompile the custom ops
export TF_NEED_CUDA=1 # If GPU is to be used
bash tools/install_so_files.shRun selected tests:
python3 -m pytest path/to/file/or/directory/to/testRun the gpu only tests with pytest -m needs_gpu ./tensorflow_addons.
Run the cpu only tests with pytest -m 'not needs_gpu' ./tensorflow_addons.
Testing with Bazel is still supported but not recommended unless you have prior experience with Bazel, and would like to use it for specific capabilities (Remote execution, etc). This is because pytest offers many more options to run your test suite and has better error reports, timings reports, open-source plugins and documentation online for Python testing.
Internally, Google can use Bazel to test many commits quickly, as Bazel has great support for caching and distributed testing.
To test with Bazel:
python3 -m pip install tensorflow==2.4.0
python3 configure.py
python3 -m pip install -r tools/install_deps/pytest.txt
bazel test -c opt -k \
--test_timeout 300,450,1200,3600 \
--test_output=all \
--run_under=$(readlink -f tools/testing/parallel_gpu_execute.sh) \
//tensorflow_addons/...We use DocTest to test code snippets
in Python docstrings. The snippet must be executable Python code.
To enable testing, prepend the line with >>> (three left-angle brackets).
Available namespace include np for numpy, tf for TensorFlow, and tfa for TensorFlow Addons.
See docs_ref for more details.
To test docstrings locally, run either
bash tools/run_cpu_tests.shon all files, or
pytest -v -n auto --durations=25 --doctest-modules /path/to/pyfileon specific files.
Ideally, we would like all the functions and classes constructors exposed in the public API to be have type hints (adding the return type for class constructors is not necessary).
We do so to improve the user experience. Some users might use IDEs or static type checking, and having types greatly improve productivity with those tools.
If you are not familiar with type hints, you can read the PEP 484.
We also have a runtime type check that we do using typeguard. For an example, see the normalizations.py file. Please add it if you type a class constructor (Note that the decorator doesn't play nice with autograph at the moment, this is why we don't add it to functions. For more context, see this pull request).
You can import some common types from tensorflow_addons/utils/types.py.
We recommend adding types if you add a new class/function to Addons' public API, but we don't enforce it.
Since adding type hints can be hard, especially for people who are not familiar with it, we made a big todo-list of functions/class constructors that need typing. If you want to add a feature to the public API and don't want to bother adding type hints, please add your feature to the todo-list in tools/testing/source_code_test.py.
Help is welcome to make this TODO list smaller!
If you add a new feature, you should add tests to ensure that new code changes doesn't introduce bugs.
If you fix a bug, you should add a tests which fails before your patch and passes after your patch.
We use Pytest to write tests. We encourage you to read the documentation, but you'll find a quick summary here:
- If you're testing code written in
xxx.py, your tests should be inxxx_test.py. - In
xxx_test.py, all functions starting withtest_are collected and run by Pytest. - Tests are run with the TF 2.x behavior, meaning eager mode my default, unless you use a
tf.function. - Ensure something is working by using
assert. For example:assert my_variable in my_list. - When comparing numpy arrays, use
the testing module of numpy.
Note that since TensorFlow ops often run with float32 of float16, you might need to
increase the default
atolandrtol. You can take a look at the default values used in the TensorFlow repository. - Prefer using your code's public API when writing tests. It ensures future refactoring is possible without changing the tests.
- When testing multiple configurations, prefer using parametrize rather than for loops for a clearer error report.
- Running all the tests in a single file should take no more than 5 seconds. You very rarely need to do heavy computation to test things. Your tests should be small and focused on a specific feature/parameter.
- Don't be afraid to write too many tests. This is fine as long as they're fast.
- It is required to contribute a code example in the docstring when adding new features.
- It is strongly suggested to expand or contribute a new tutorial for more complex features that are hard to be expressed in the docstring only.
We provide fixtures to help your write your tests as well as helper functions. Those can be found in test_utils.py.
Will run your test function twice, once normally and once with
tf.config.run_functions_eagerly(True). To use it:
@pytest.mark.usefixtures("maybe_run_functions_eagerly")
def test_something():
assert ...== ...You should use it only if you are using tf.function and running some control flow
on Tensors, if or for for example. Or with TensorArray. In short, when the
conversion to graph is not trivial. No need to use it on all
your tests. Having fast tests is important.
By default, each test is wrapped behind the scenes with a
with tf.device("CPU:0"):
...This is automatic. But it's also possible to ask the test runner to run the test twice, on CPU and on GPU, or only on GPU. Here is how to do it.
import pytest
import tensorflow as tf
from tensorflow_addons.utils import test_utils
@pytest.mark.with_device(["cpu", "gpu"])
def test_something():
# the code here will run twice, once on gpu, once on cpu.
...
@pytest.mark.with_device(["cpu", "gpu"])
def test_something2(device):
# the code here will run twice, once on gpu, once on cpu.
# device will be "cpu:0" or "gpu:0" or "gpu:1" or "gpu:2" ...
if "cpu" in device:
print("do something.")
if "gpu" in device:
print("do something else.")
@pytest.mark.with_device(["cpu", "gpu", tf.distribute.MirroredStrategy])
def test_something3(device):
# the code here will run three times, once on gpu, once on cpu and once with
# a mirror distributed strategy.
# device will be "cpu:0" or "gpu:0" or the strategy.
# with the MirroredStrategy, it's equivalent to:
# strategy = tf.distribute.MirroredStrategy(...)
# with strategy.scope():
# test_function(strategy)
if "cpu" in device:
print("do something.")
if "gpu" in device:
print("do something else.")
if isinstance(device, tf.distribute.Strategy):
device.run(...)
@pytest.mark.with_device(["gpu"])
def test_something_else():
# This test will be only run on gpu.
# The test runner will call with tf.device("GPU:0") behind the scenes.
...
@pytest.mark.with_device(["cpu"])
def test_something_more():
# Don't do that, this is the default behavior.
...
@pytest.mark.with_device(["no_device"])
@pytest.mark.needs_gpu
def test_something_more2():
# When running the function, there will be no `with tf.device` wrapper.
# You are free to do whatever you wish with the devices in there.
# Make sure to use only the cpu, or only gpus available to the current process with
# test_utils.gpu_for_testing() , otherwise, it might not play nice with
# pytest's multiprocessing.
# If you use a gpu, mark the test with @pytest.mark.needs_gpu , otherwise the
# test will fail if no gpu is available on the system.
# for example
...
strategy = tf.distribute.MirroredStrategy(test_utils.gpus_for_testing())
with strategy.scope():
print("I'm doing whatever I want.")
...Note that if a gpu is not detected on the system, the test will be
skipped and not marked as failed. Only the first gpu of the system is used,
even when running pytest in multiprocessing mode. (-n argument).
Beware of the out of cuda memory errors if the number of pytest workers is too high.
When you test custom CUDA code or float16 ops. We can expect other existing TensorFlow ops to behave the same on CPU and GPU.
Will run your test function twice, once with data_format being channels_first and
once with data_format being channels_last. To use it:
def test_something(data_format):
assert my_function_to_test(..., data_format=data_format) == ...When your function has a data_format argument. You'll want to make sure your
function behaves correctly with both data format.
Is the same as tf.test.TestCase.assertAllCloseAccordingToType but doesn't require any subclassing to be done. Can be used as a plain function. To use it:
from tensorflow_addons.utils import test_utils
def test_something():
expected = ...
computed = my_function_i_just_wrote(...).numpy()
test_utils.assert_allclose_according_to_type(computed, expected)When you want to test your function with multiple dtypes. Different dtypes requires different tolerances when comparing values.
All submissions, including submissions by project members, require review. We use Github pull requests for this purpose.
Contributions to this project must be accompanied by a Contributor License Agreement. You (or your employer) retain the copyright to your contribution; this simply gives us permission to use and redistribute your contributions as part of the project. Head over to https://cla.developers.google.com/ to see your current agreements on file or to sign a new one.
You generally only need to submit a CLA once, so if you've already submitted one (even if it was for a different project), you probably don't need to do it again.