stgraph 1.0.0

🌟 Vertex Centric approach for building GNN/TGNNs

Explore the docs »

Installation guide · Report Bug · View Discussions . Paper

🌟 STGraph

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.

Why STGraph

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.

Getting Started

Pre-requisites

Install the python packages by running the following. It is recommended that you create a virtual environment before installing the packages.

Setup a new virtual environment

conda create --name stgraph
conda activate stgraph

Install the python packages

pip install -r requirements.txt

STGraph requires CUDA Version 11.7 or above to run. Versions below that may or may not run, depending on any changes within CUDA.

Installation

You can git clone STGraph into your workspace by running the following command

git clone https://github.com/bfGraph/STGraph.git
cd STGraph

Running your first STGraph Program

In this is quick mini tutorial, we will show you how to train a simple GCN model on the Cora dataset. After installing STGraph and entering the STGraph directory, enter the following commands to reach the GCN benchmarking folder

cd benchmarking/gcn/stgraph

Run the train.py, with 100 epochs and specify the dataset name. For this example, we shall use Cora

python3 train.py --num_epochs 100 --dataset cora

You should get an output like this. The initial prints are truncated.

.
.
.
Epoch 00090 | Time(s) 0.0048 | train_acc 0.303791 | Used_Memory 32.975098 mb
Epoch 00091 | Time(s) 0.0024 | train_acc 0.303791 | Used_Memory 32.975098 mb
Epoch 00092 | Time(s) 0.0029 | train_acc 0.303791 | Used_Memory 32.975098 mb
Epoch 00093 | Time(s) 0.0029 | train_acc 0.303791 | Used_Memory 32.975098 mb
Epoch 00094 | Time(s) 0.0027 | train_acc 0.303791 | Used_Memory 32.975098 mb
Epoch 00095 | Time(s) 0.0030 | train_acc 0.303791 | Used_Memory 32.975098 mb
Epoch 00096 | Time(s) 0.0024 | train_acc 0.303791 | Used_Memory 32.975098 mb
Epoch 00097 | Time(s) 0.0022 | train_acc 0.303791 | Used_Memory 32.975098 mb
Epoch 00098 | Time(s) 0.0022 | train_acc 0.303791 | Used_Memory 32.975098 mb
Epoch 00099 | Time(s) 0.0036 | train_acc 0.303791 | Used_Memory 32.975098 mb

^^^0.032202^^^0.003098

If you don't get this output and have followed every single step in the setting up and installation section, please raise an issue we will look into it.

How to build STGraph

This is for users who want to make changes to the STGraph codebase and get it build each time. Follow the steps mentioned to properly build STGraph.

Compiling the CUDA code

The following steps need to be done if you made any changes to any CUDA files within the stgraph/graph directory for each graph representation.

STGraph supports training dynamic and static graphs. To handle all the graph representations logic, it is written as a PyBind11 module over a CUDA file. As of now the following CUDA code for different graph representations exists

1. csr.cu
2. pcsr.cu
3. gpma.cu

To compile the [name].cu file, run the following command

/usr/local/cuda-11.7/bin/nvcc $(python3 -m pybind11 --includes) -shared -rdc=true --compiler-options '-fPIC' -D__CDPRT_SUPPRESS_SYNC_DEPRECATION_WARNING -o [name].so [name].cu

This would generate the [name].so shared object file, that is used in the STGraph module.

Building STGraph

Make sure to go back to the root directory and run the following to build and install STGraph

 python3 -m build && pip uninstall stgraph -y && pip install dist/stgraph-1.0.0-py3-none-any.whl

Contributing

Please read CONTRIBUTING.md for details on our code of conduct, and the process for submitting pull requests, issues, etc to us.

How to contribute to Documentation

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

Authors

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

Attributions

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	STGraph: 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