CONTRIBUTING.md

How to contribute code

Follow these steps to submit your code contribution.

Step 1. Open an issue

Before making any changes, we recommend opening an issue (if one doesn't already exist) and discussing your proposed changes. This way, we can give you feedback and validate the proposed changes.

If the changes are minor (simple bug fix or documentation fix), then feel free to open a PR without discussion.

Step 2. Make code changes

To make code changes, you need to fork the repository. You will need to setup a development environment and run the unit tests. This is covered in section "Setup environment".

Step 3. Create a pull request

Once the change is ready, open a pull request from your branch in your fork to the master branch in keras-team/keras.

Step 4. Sign the Contributor License Agreement

After creating the pull request, the google-cla bot will comment on your pull request with instructions on signing the Contributor License Agreement (CLA) if you haven't done so. Please follow the instructions to sign the CLA. A cla:yes tag is then added to the pull request.

Step 5. Code review

A reviewer will review the pull request and provide comments. The reviewer may add a kokoro:force-run label to trigger the continuous integration tests.

If the tests fail, look into the error messages and try to fix it.

There may be several rounds of comments and code changes before the pull request gets approved by the reviewer.

Step 6. Merging

Once the pull request is approved, a ready to pull tag will be added to the pull request. A team member will take care of the merging.

Here is an example pull request for your reference.

Setup environment

To setup the development environment, We provide two options. One is to use our Dockerfile, which builds into a container the required dev tools. Another one is to setup a local environment by install the dev tools needed.

Option 1: Use a Docker container

We provide a Dockerfile to build the dev environment. You can build the Dockerfile into a Docker image named keras-dev with the following command at the root directory of your cloned repo.

docker build -t keras-dev .devcontainer

You can launch a Docker container from the image with the following command. The -it option gives you an interactive shell of the container. The -v path/to/repo/:/home/keras/ mounts your cloned repo to the container. Replace path/to/repo with the path to your cloned repo directory.

docker run -it -v path/to/repo/:/home/keras/ keras-dev

In the container shell, you need to install the latest dependencies with the following command.

pip install -r /home/keras/requirements.txt && pip uninstall keras-nightly -y

Now, the environment setup is complete. You are ready to run the tests.

You may modify the Dockerfile to your specific needs, like installing your own dev tools. You may also mount more volumes with the -v option, like your SSH credentials.

Many popular editors today support developing in a container. Here is list of supported editors with setup instructions.

Option 2: Setup a local environment

To setup your local dev environment, you will need the following tools.

1. Bazel is the tool to build and test Keras. See the installation guide for how to install and config bazel for your local environment.
2. git for code repository management.
3. python to build and code in Keras.

The following commands checks the tools above are successfully installed. Note that Keras requires at least Python 3.7 to run.

bazel --version
git --version
python --version

A Python virtual environment (venv) is a powerful tool to create a self-contained environment that isolates any change from the system level config. It is highly recommended to avoid any unexpected dependency or version issue.

With the following commands, you create a new venv, named venv_dir.

mkdir venv_dir
python3 -m venv venv_dir

You can activate the venv with the following command. You should always run the tests with the venv activated. You need to activate the venv every time you open a new shell.

source venv_dir/bin/activate  # for linux or MacOS
venv_dir\Scripts\activate.bat  # for Windows

Clone your forked repo to your local machine. Go to the cloned directory to install the dependencies into the venv. Since tf-nightly uses keras-nightly as a dependency, we need to uninstall keras-nightly so that tests will run against Keras code in local workspace.

git clone https://github.com/YOUR_GITHUB_USERNAME/keras.git
cd keras
pip install -r requirements.txt
pip uninstall keras-nightly

The environment setup is completed. You may need to update the tf-nightly version regularly to keep your environment up-to-date with the following command.

pip install --upgrade tf-nightly

Code style

The Keras codebase uses the PEP 8 Python style conventions -- with the exception that it uses 2 spaces for indentation instead of 4. To check code style, please run the pylint command from the repo's root directory so that the configuration in .pylintrc is taken into account.

pylint path/to/changed_file.py

Please ignore the errors in the rest of the codebase and only fix the ones relevant to your changes.

Run tests

We use Bazel to build and run the tests.

Run a test file

For example, to run the tests in keras/engine/base_layer_test.py, we can run the following command at the root directory of the repo.

bazel test keras/engine:base_layer_test

keras/engine is the relative path to the directory containing the BUILD file defining the test. base_layer_test is the test target name defined with tf_py_test in the BUILD file.

Run a single test case

To run a single test, you can use --test_filter=<your_regex> to use regular expression to match the test you want to run. For example, you can use the following command to run all the tests in activations_test.py, whose names contain test_serialization.

bazel test keras:activations_test --test_filter=*test_serialization*

Run all tests

You can run all the tests locally by running the following command in the repo root directory.

bazel test --test_timeout 300,450,1200,3600 --test_output=errors --keep_going --define=use_fast_cpp_protos=false --build_tests_only --build_tag_filters=-no_oss --test_tag_filters=-no_oss keras/...

Useful configs

Here we provide a list of useful configs you can use with Bazel.

bazel test [CONFIGS] [YOUR_TEST]

To use these configs, just replace [CONFIGS] with the actual config in the command above.

• -c opt enables the optimizations during the build.
• --test_sharding_strategy=disabled disables the sharding so that all the test outputs are in one file. However, it may slow down the tests for not running in parallel and may cause the test to timeout.

Contributing to Keras applications

Contributions to the pre-trained application library are welcome. Code for Keras applications is located in Keras repository in keras/applications. When contributing to Keras applications, please keep following checklist in mind.

• Keras applications must implement an established and widely used model. Applications should include a link to a paper describing the architecture of the model with at least 20 citations.
• Applications should be provided with pre-trained weights.
  • When submitting a pull request for a Keras application, these weights can be provided at any publically available URL (e.g. a personal Cloud Storage bucket). The weights will be uploaded to a Keras storage bucket while merging the pull request.
  • Weights should be downloaded with the get_file() utility function. Be sure to include the file_hash argument, which allows cache invalidation on the downloaded weights. The command line programs shasum and sha256sum can compute a file hash.
• You should help us verify that the accuracy of the model with pre-trained weighted matches the reported results of the cited paper.
• You should add any new applications to the unit tests defined in applications_test.py and applications_load_weight_test.py.
• For backwards compatibility, all applications should provide a preprocess_input() function. For new applciations, you should leave the function empty (pass through inputs unaltered), and write the model so it can handle raw inputs directly. Adding preprocessing layers to the application model may help with this. For image applications, a Rescaling layer at the beginning of the model is often all that is needed.
• Once the PR is approved, you should create a companion PR to the keras.io application page updating the "Available Models" section. The contribution guide for keras.io can be found here.