ef13a9fec0 | ||
---|---|---|
.github/workflows | ||
docs | ||
ptgnn | ||
.amlignore | ||
.editorconfig | ||
.gitattributes | ||
.gitignore | ||
.pre-commit-config.yaml | ||
CODE_OF_CONDUCT.md | ||
Dockerfile | ||
LICENSE | ||
README.md | ||
SECURITY.md | ||
azure-pipelines.yml | ||
environment.yml | ||
pyproject.toml | ||
setup.py |
README.md
ptgnn: A PyTorch GNN Library
This is a library containing pyTorch code for creating graph neural network (GNN) models. The library provides some sample implementations.
If you are interested in using this library, please read about its architecture and how to define GNN models or follow this tutorial.
Note that ptgnn
takes care of defining the whole pipeline, including data wrangling tasks, such
as data loading and tensorization. It also defines PyTorch nn.Module
s for
the neural network operations. These are independent of the
AbstractNeuralModel
s and can be used as all other PyTorch's nn.Module
s,
if one wishes to do so.
The library is mainly engineered to be fast for sparse graphs. For example, for the
Graph2Class task (discussed below) on a V100 with the default hyperparameters and architecture
ptgnn
can process about 82 graphs/sec (209k nodes/sec and 1,129k edges/sec) during training
and about 200 graph/sec (470k nodes/sec and 2,527k edges/sec) during testing.
Implemented Tasks
All task implementations can be found in the ptgnn.implementations
package.
Detailed instructions on the data and the training steps can be found here.
We welcome external contributions. The following GNN-based tasks are implemented:
- PPI The PPI task as described by Zitnik and Leskovec, 2017.
- VarMisuse This is a re-implementation of the VariableMisuse task of Allamanis et al., 2018.
- Graph2Sequence This is re-implementation of the GNN->GRU model of Fernandes et. al., 2019.
- Graph2Class Classify (Label) a subset of the input nodes into classes similar to Graph2Class in Typilus.
The tutorial gives a step-by-step example for coding the Graph2Class model.
Installation
This code was tested with PyTorch 1.4 and depends
on pytorch-scatter
.
Please install the appropriate versions of these libraries based
on your CUDA setup following their instructions. (Note
that the pytorch-scatter
binaries built for CUDA 10.1 also work for
CUDA 10.2).
-
To install PyTorch 1.4, use the up-to-date command from PyTorch Get Started, selecting the appropriate options, e.g. for Linux, pip, and CUDA 10.1 it's currently:
pip install torch torchvision
-
To install
pytorch-scatter
, follow the instructions from the GitHub repo, choosing the appropriate CUDA option, e.g., for CUDA 10.1:pip install torch-scatter==2.0.4+cu101 -f https://pytorch-geometric.com/whl/torch-1.4.0.html
-
To install
ptgnn
from pypi, including all other dependencies:pip install ptgnn
If you want to use ptgnn sampels with Azure ML (e.g. the
--aml
flag in the implementation CLIs), install withpip install ptgnn[aml]
or directly from the sources,
cd
into the root directory of the project and runpip install -e .
To check that installation was successful and run the unit tests:
python setup.py test
Contributing
This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.
When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.
Developing ptgnn
To contribute to this library, first follow the next steps to setup your development environment:
- Install the library requirements.
- Install the pre-commit hooks:
- Run
pip3 install pre-commit
- Install the hooks
pre-commit install
- Run
Using Conda
If you are using conda, then download the correct torch-scatter wheel.
If using torch==1.5.0
and Python 3.7, you can use the environment.yml
included in the repo, with the following steps:
$ conda env create -f environment.yml
$ conda activate ptgnn-env
$ pip install torch_scatter-2.0.4+cu102-cp37-cp37m-linux_x86_64.whl
$ pip install -e .
$ python setup.py test
$ pip install pre-commit
$ pre-commit install
Releasing to PyPi
To create a PyPi release push a tag in the form v1.3.4
in this repository (make
sure that you follow semantic versioning). The
Publish on PyPi
GitHub action will automatically upload a new release.