Contribute to DGL

Any contribution to DGL is welcome. This guide covers everything about how to contribute to DGL.

General development process

A non-inclusive list of types of contribution is as follows:

For features and bugfix, we recommend first raise an issue using the corresponding issue template, so that the change could be fully discussed with the community before implementation. For document improvement and new models, we suggest post a thread in our discussion forum.

Before development, please first read the following sections about coding styles and testing. All the changes need to be reviewed in the form of pull request. Our committors (who have write permission on the repository) will review the codes and suggest the necessary changes. The PR could be merged once the reviewers approve the changes.

Git setup (for developers)

First, fork the DGL github repository. Suppose the forked repo is

Clone your forked repository locally:

git clone --recursive

Setup the upstream to the DGL official repository:

git remote add upstream

You could verify the remote setting by typing git remove -v:

origin (fetch)
origin (push)
upstream (fetch)
upstream (push)

During developing, we suggest work on another branch than the master.

git branch working-branch
git checkout working-branch

Once the changes are done, create a pull request so we could review your codes.

Once the pull request is merged, update your forked repository and delete your working branch:

git checkout master
git pull upstream master
git push origin master  # update your forked repo
git branch -D working-branch  # the local branch could be deleted

Coding styles

For python codes, we generally follow the PEP8 style guide. The python comments follow NumPy style python docstrings.

For C++ codes, we generally follow the Google C++ style guide. The C++ comments should be Doxygen compatible.

Coding styles check is mandatory for every pull requests. To ease the development, please check it locally first (require cpplint and pylint to be installed first):

bash tests/scripts/

The python code style configure file is tests/scripts/pylintrc. We tweak it a little bit from the standard. For example, following variable names are accepted:

  • i,j,k: for loop variables
  • u,v: for representing nodes
  • e: for representing edges
  • g: for representing graph
  • fn: for representing functions
  • n,m: for representing sizes
  • w,x,y: for representing weight, input, output tensors
  • _: for unused variables

Building and Testing

To build DGL locally, follow the steps described in Install from source. However, to ease the development, we suggest NOT install DGL but directly working in the source tree. To achieve this, export following environment variables:

export DGL_HOME=/path/to/your/dgl/clone

If you are working on performance critical part, you may want to turn on Cython build:

You could test the build by running the following command and see the path of your local clone.

python -c 'import dgl; print(dgl.__path__)'

Unit tests

Currently, we use nose for unit tests. The organization goes as follows:

  • backend: Additional unified tensor interface for supported frameworks. The functions there are only used in unit tests, not DGL itself. Note that the code there are not unit tests by themselves. The additional backend can be imported with

    import backend

    The additional backend contains the following files:

    • backend/ stub file for all additional tensor functions.
    • backend/${DGLBACKEND}/ implementations of the stubs for the backend ${DGLBACKEND}.
    • backend/ when imported, it replaces the stub implementations with the framework-specific code, depending on the selected backend. It also changes the signature of some existing backend functions to automatically select dtypes and contexts.
  • compute: All framework-agnostic computation-related unit tests go there. Anything inside should not depend on a specific tensor library. Tensor functions not provided in DGL unified tensor interface (i.e. dgl.backend) should go into backend directory.

  • ${DGLBACKEND} (e.g. pytorch and mxnet): All framework-specific computation-related unit tests go there.

  • graph_index: All unit tests for C++ graph structure implementation go there. The Python API being tested in this directory, if any, should be as minimal as possible (usually simple wrappers of corresponding C++ functions).

  • lint: Pylint-related files.

  • scripts: Automated test scripts for CI.

To run unit tests, run

sh tests/scripts/ <your-backend>

where <your-backend> can be any supported backends (i.e. pytorch or mxnet).

Building documents

If the change is about document improvement, we suggest build the document and render it locally before pull request. See instructions here.

Data hosting

If the change is about new models or applications, it is very common to have some data files. Data files are not allowed to be uploaded to our repository. Instead, they should be hosted on the cloud storage service (e.g. dropbox, Amazon S3) and downloaded on-the-fly. See our dataset APIs for more details. All the dataset of current DGL models are hosted on Amazon S3. If you want your dataset to be hosted as well, please post in our discussion forum.