Graph-CNN

Implementation of "Robust Spatial Filtering with Graph Convolutional Neural Networks" - https://arxiv.org/abs/1703.00792. This project is still under development. Email Felipe at fps7806@rit.edu if you have any suggestions or problems.

Requirements:

• Tensorflow >1.0
• Internet Connection (if automatically downloading datasets)

Instructions

To obtain results on NCI1 and Cora run the following commands:

python src/run_nci1.py

python src/run_cora.py

The scripts will automatically download the datasets and run the experiments.

Result are automatically saved in the results folder. python src/parse_results.py results/Cora.txt will parse the results and display useful information.

Note that the results will not be exactly the same every run.

Scripts will print useful information. Can be disabled by adding exp.silent = True.

Checkpoints and Summaries that can be used with Tensorboard will be generated if the line self.debug = True is removed.

There is an example on how to run experiments on ImageNet under src/run_alexnet.py. The training was not optimized so the final accuracy might not match the expected.

NCI1 - Results

Model	File	Accuracy
64-64-GEP(32)-32-GEP(8)-FC256-FC2	run_nci1.py	84.45 +- 0.94

Cora - Results

Model	File	Accuracy
Embed256-48-Embed32-48-Embed32-7	run_cora.py	89.18 +- 1.96

Citation

Please cite us in your publications if it helps your research:

@article{graphcnn2017,
  Author = {Petroski Such, Felipe, and Sah, Shagan, and Dominguez, Miguel, and Pillai, Suhas, and Zhang, Chao, and Michael,  Andrew, and Cahill, Nathan, and Ptucha, Raymond},
  Journal = {arXiv preprint arXiv:1703.00792},
  Title = {Robust Spatial Filtering with Graph Convolutional Neural Networks},
  Year = {2017}
}