STGraph is a framework designed for deep-learning practitioners to write and train Graph Neural Networks (GNNs) and Temporal Graph Neural Networks (TGNNs). It is built on top of Seastar and utilizes the vertex-centric approach to produce highly efficient fused GPU kernels for forward and backward passes. It achieves better usability, faster computation time and consumes less memory than state-of-the-art graph deep-learning systems like DGL, PyG and PyG-T.
NOTE: If the contents of this repository are used for research work, kindly cite the paper linked above.
The primary goal of Seastar is more natural GNN programming so that the users learning curve is flattened. Our key observation lies in recognizing that the equation governing a GCN layer, as shown above, takes the form of vertex-centric computation and can be implemented succinctly with only one line of code. Moreover, we can see a clear correspondence between the GNN formulas and the vertex-centric implementations. The benefit is two-fold: users can effortlessly implement GNN models, while simultaneously understanding these models by inspecting their direct implementations.
The Seastar system outperforms state-of-the-art GNN frameworks but lacks support for TGNNs. STGraph bridges that gap and enables users to to develop TGNN models through a vertex-centric approach. STGraph has shown to be significantly faster and more memory efficient that state-of-the-art frameworks like PyG-T for training TGNN models.
This guide is tailored for users of the STGraph package, designed for constructing GNN and TGNN models. We recommend creating a new virtual environment with Python version 3.8 and installing stgraph inside that dedicated environment.
Installing STGraph from PyPI
pip install stgraphInstalling PyTorch
In addition, STGraph relies on PyTorch. Ensure it is installed in your virtual environment with the following command
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118Upon completion of the above steps, you have successfully installed STGraph. Proceed to write and train your first GNN model by referring to the provided tutorial.
This guide is intended for those interested in developing and contributing to STGraph.
Download source files from GitHub
git clone https://github.com/bfGraph/STGraph.git
cd STGraphCreate a dedicated virtual environment
Inside the STGraph directory create and activate a dedicated virtual environment named dev-stgraph with Python version 3.8.
python3.8 -m venv dev-stgraph
source dev-stgraph/bin/activateInstall STGraph in editable mode
Make sure to install the STGraph package in editable mode to ease your development process.
pip install -e .[dev]
pip listInstall PyTorch
Ensure to install PyTorch as well for development
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118With this you have successfully installed STGraph locally to make development changes and contribute to the project. Head out to our Pull Requests page and get started with your first contribution.
Please have a look inside the tutorials/ directory to write and train your own GNNs using STGraph
Please read CONTRIBUTING.md for details on our code of conduct, and the process for submitting pull requests, issues, etc to us.
We follow the PEP-8 format. Black is used as the formatter and pycodestyle as the linter. The linter is is configure to work properly with black (set line length to 88)
Tutorial for Python Docstrings can be found here
sphinx-apidoc -o docs/developers_guide/developer_manual/package_reference/ python/stgraph/ -f
cd docs/
make clean
make html
| Author | Bio |
|---|---|
Joel Mathew Cherian |
Computer Science Student at National Institute of Technology Calicut |
Nithin Puthalath Manoj |
Computer Science Student at National Institute of Technology Calicut |
Dr. Unnikrishnan Cheramangalath |
Assistant Professor in CSED at Indian Institue of Technology Palakkad |
Kevin Jude |
Ph.D. in CSED at Indian Institue of Technology Palakkad |
| Author(s) | Title | Link(s) |
|---|---|---|
Wu, Yidi and Ma, Kaihao and Cai, Zhenkun and Jin, Tatiana and Li, Boyang and Zheng, Chenguang and Cheng, James and Yu, Fan |
Seastar: vertex-centric programming for graph neural networks, 2021 |
paper, code |
Wheatman, Brian and Xu, Helen |
Packed Compressed Sparse Row: A Dynamic Graph Representation, 2018 |
paper, code |
Sha, Mo and Li, Yuchen and He, Bingsheng and Tan, Kian-Lee |
Accelerating Dynamic Graph Analytics on GPUs, 2017 |
paper, code |
Benedek Rozemberczki, Paul Scherer, Yixuan He, George Panagopoulos, Alexander Riedel, Maria Astefanoaei, Oliver Kiss, Ferenc Beres, Guzmán López, Nicolas Collignon, Rik Sarkar |
PyTorch Geometric Temporal: Spatiotemporal Signal Processing with Neural Machine Learning Models, 2021 |
paper, code |