Skip to content

SYSU-RoboticsLab/pytorch_DGCNN

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

40 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PyTorch DGCNN

About

PyTorch implementation of DGCNN (Deep Graph Convolutional Neural Network). Check https://github.com/muhanzhang/DGCNN for more information.

Requirements: python 2.7 or python 3.6; pytorch >= 0.4.0

Installation

This implementation is based on Hanjun Dai's structure2vec graph backend. Under the "lib/" directory, type

make -j4

to compile the necessary c++ files.

After that, under the root directory of this repository, type

./run_DGCNN.sh

to run DGCNN on dataset MUTAG with the default setting.

Or type

./run_DGCNN.sh DATANAME FOLD

to run on dataset = DATANAME using fold number = FOLD (1-10, corresponds to which fold to use as test data in the cross-validation experiments).

If you set FOLD = 0, e.g., typing "./run_DGCNN.sh DD 0", then it will run 10-fold cross validation on DD and report the average accuracy.

Alternatively, type

./run_DGCNN.sh DATANAME 1 200

to use the last 200 graphs in the dataset as testing graphs. The fold number 1 will be ignored.

Check "run_DGCNN.sh" for more options.

Datasets

Default graph datasets are stored in "data/DSName/DSName.txt". Check the "data/README.md" for the format.

In addition to the support of discrete node labels (tags), DGCNN now supports multi-dimensional continuous node features. One example dataset with continuous node features is "Synthie". Check "data/Synthie/Synthie.txt" for the format.

There are two preprocessing scripts in MATLAB: "mat2txt.m" transforms .mat graphs (from Weisfeiler-Lehman Graph Kernel Toolbox), "dortmund2txt.m" transforms graph benchmark datasets downloaded from https://ls11-www.cs.tu-dortmund.de/staff/morris/graphkerneldatasets

How to use your own data

The first step is to transform your graphs to the format described in "data/README.md". You should put your testing graphs at the end of the file. Then, there is an option -test_number X, which enables using the last X graphs from the file as testing graphs. You may also pass X as the third argument to "run_DGCNN.sh" by

./run_DGCNN.sh DATANAME 1 X

where the fold number 1 will be ignored.

Reference

If you find the code useful, please cite our paper:

@inproceedings{zhang2018end,
  title={An End-to-End Deep Learning Architecture for Graph Classification.},
  author={Zhang, Muhan and Cui, Zhicheng and Neumann, Marion and Chen, Yixin},
  booktitle={AAAI},
  year={2018}
}

Muhan Zhang, [email protected] 3/19/2018

About

PyTorch implementation of DGCNN

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 59.0%
  • C++ 21.6%
  • MATLAB 10.3%
  • Shell 5.3%
  • Makefile 2.4%
  • C 1.4%