Graph4NLP

Graph4NLP is the library for the easy use of Graph Neural Networks for NLP

Installation

Users can install Graph4NLP using pip:

pip install graph4nlp         # CPU Version

For GPU supported versions, use the following commands:

pip install graph4nlp-cu92    # CUDA 9.2 Version
pip install graph4nlp-cu101   # CUDA 10.1 Version
pip install graph4nlp-cu102   # CUDA 10.2 Version
pip install graph4nlp-cu110   # CUDA 11.0 Version

Distributing this package

Currently releasing via PyPI is supported. An automated build and distributing shell script is ./build_and_upload_to_pypi.sh. You can directly run this script after proper configuration, which will be elaborated below.

Build the package

To build the package, run setup.py. You need to specify whether the binary distribution file(wheel) or source code distribution(tgz) is wanted. If you want to build the binary version, run

python setup.py bdist_wheel --universal

The bdist_wheel option is to say that the binary distribution of wheel file will be built. And the --universal option means the wheel file will fit all platforms.

If you also want to package the source code for distribution, run

python setup.py sdist

Here sdist means source distribution. In build_and_upload_to_pypi.sh both options will be built.

Inside setup.py

Version

Everytime the package is updated, a newer version number needs to be specified, otherwise the updated package cannot be uploaded to PyPI. To update the version number, you need to change the version parameter in the setup() function call in setup.py

CUDA Compatibility

The only reason why we keep different packages for different CUDA versions is that DGL does so. In fact our current implementation has nothing to do with CUDA since it is basically a Python frontend of DGL. However, CUDA version still needs to be specified in order to make the DGL dependency explicit.

When running setup.py, you will be prompted to input the CUDA version. You may just input 'none' for a CPU distribution, otherwise you need to specify the version number explicitly, e.g. 10.2. Then a prefix will be generated and appended to the package name, distinguishing among versions with different CUDA dependency. For example graph4nlp-cu101.

Note that this is just a temporal solution for the distribution of the library. This part needs to be optimized in the following releases.

Uploading to PyPI

The twine python package is used for uploading the package to PyPI. To use it, first install it using pip:

pip install twine

Then you need to specify your credential, namely to authenticate your account that hosts Graph4NLP on PyPI.

Currently I am using my personal account (swangeh) to host this package and a dedicated account should be used in the future. You can manually input your username and password everytime you upload the package to PyPI by running

twine upload dist/*

In this way there will be prompts to guide you. To save your time you can also generate a credential on PyPI and save it locally in ~/.pypirc. For example inside your ~/.pypirc it will be

[pypi]
username = <your username>
password = <your password>

For details please refer to PyPI.

Putting them all together

Above is the decomposition of what's inside build_and_upload_to_pypi.sh. You can just run this script to save your time for building and releasing the package.