Open source code for the GCN training algorithm in the IEEE/IPDPS'19 paper.

Right now, for ease of extending the framework, we have integrated our sampling and feature propagation algorithm into Tensorflow (where the subgraph feature propagation as a self-defined operation).

Hanqing Zeng: zengh@usc.edu, Hongkuan Zhou: hongkuaz@usc.edu

• python >= 3.6.8
• tensorflow >=1.12.0
• numpy >= 1.14.3
• scipy >= 1.1.0
• scikit-learn >= 0.19.1
• openmp >= 4.0
• mkl >= 2018.0.2

There are three datasets (PPI, Reddit and Yelp) available via this Google Drive link. Rename the folder to data at the root directory. The directory structure should be as below:

GCN-IPDPS19/
│   README.md
│   run_gs_subgraph.sh
│   ... 
│
└───gcn_ipdps19/
│   │   models.py
│   │   train.py
│   │   ...
│   
└───data/
│   └───ppi/
│   │   │    adj_train.npz
│   │   │    adj_full.npz
│   │   │    ...
│   │   
│   └───reddit/
│   │   │    ...
│   │
│   └───...
│

There are two options to run the code on CPU with mkl. You can either use the --mkl flag to directly call C++ mkl libraries or use the mkl-built tensorflow.

To run the code on CPU with mkl-built tensorflow:

./run_training.sh <dataset_name> <path to train_config yml>

To run the code on CPU with python-build tensorflow:

./run_training.sh <dataset_name> <path to train_config yml> --mkl

To run the code on GPU:

./run_training.sh <dataset_name> <path to train_config yml> --gpu <GPU number>

For example --gpu 0 will run on the fisrt GPU.

@INPROCEEDINGS{8820993, 
author={H. {Zeng} and H. {Zhou} and A. {Srivastava} and R. {Kannan} and V. {Prasanna}}, 
booktitle={2019 IEEE International Parallel and Distributed Processing Symposium (IPDPS)}, 
title={Accurate, Efficient and Scalable Graph Embedding}, 
year={2019}, 
volume={}, 
number={}, 
pages={462-471}, 
keywords={Graph embedding;Graph convolutional networks;Subgraph sampling}, 
doi={10.1109/IPDPS.2019.00056}, 
ISSN={}, 
month={May},}