Official implementation of Self-supervised Graph Attention Networks (SuperGAT). This model is presented at How to Find Your Friendly Neighborhood: Graph Attention Design with Self-Supervision, International Conference on Learning Representations (ICLR), 2021.

• The documented SuperGATConv layer with an example has been merged to the PyTorch Geometric's main branch.
• The RandomPartitionGraph is now available at PyTorch Geometric.
• This repository is based on torch==1.4.0+cu100 and torch-geometric==1.4.3, which are somewhat outdated at this point (Feb 2021). If you are using recent PyTorch/CUDA/PyG, we would recommend using the PyG's. If you want to run codes in this repository, please follow #installation.

@inproceedings{
    kim2021how,
    title={How to Find Your Friendly Neighborhood: Graph Attention Design with Self-Supervision},
    author={Dongkwan Kim and Alice Oh},
    booktitle={International Conference on Learning Representations},
    year={2021},
    url={https://openreview.net/forum?id=Wi5KUNlqWty}
}

# In SuperGAT/
bash install.sh ${CUDA, default is cu100}

• If you have any trouble installing PyTorch Geometric, please install PyG's dependencies manually.
• Codes are tested with python 3.7.6 and nvidia/cuda:10.0-cudnn7-devel-ubuntu16.04 image.
• PYG's FAQ might be helpful.

• The main train/test code is in SuperGAT/main.py.
• If you want to see the SuperGAT layer in PyTorch Geometric MessagePassing grammar, refer to SuperGAT/layer.py.
• If you want to see hyperparameter settings, refer to SuperGAT/args.yaml and SuperGAT/arguments.py.

python3 SuperGAT/main.py \
    --dataset-class Planetoid \
    --dataset-name Cora \
    --custom-key EV13NSO8-ES
 
...

## RESULTS SUMMARY ##
best_test_perf: 0.853 +- 0.003
best_test_perf_at_best_val: 0.851 +- 0.004
best_val_perf: 0.825 +- 0.003
test_perf_at_best_val: 0.849 +- 0.004
## RESULTS DETAILS ##
best_test_perf: [0.851, 0.853, 0.857, 0.852, 0.858, 0.852, 0.847]
best_test_perf_at_best_val: [0.851, 0.849, 0.855, 0.852, 0.858, 0.848, 0.844]
best_val_perf: [0.82, 0.824, 0.83, 0.826, 0.828, 0.824, 0.822]
test_perf_at_best_val: [0.851, 0.844, 0.853, 0.849, 0.857, 0.848, 0.844]
Time for runs (s): 173.85422565042973

The default setting is 7 runs with different random seeds. If you want to change this number, change num_total_runs in the main block of SuperGAT/main.py.

For ogbn-arxiv, use SuperGAT/main_ogb.py.

There are three arguments for GPU settings (--num-gpus-total, --num-gpus-to-use, --gpu-deny-list). Default values are from the author's machine, so we recommend you modify these values from SuperGAT/args.yaml or by the command line.

• --num-gpus-total (default 4): The total number of GPUs in your machine.
• --num-gpus-to-use (default 1): The number of GPUs you want to use.
• --gpu-deny-list (default: [1, 2, 3]): The ids of GPUs you want to not use.

If you have four GPUs and want to use the first (cuda:0),

python3 SuperGAT/main.py \
    --dataset-class Planetoid \
    --dataset-name Cora \
    --custom-key EV13NSO8-ES \
    --num-gpus-total 4 \
    --gpu-deny-list 1 2 3

See SuperGAT/args.yaml or run $ python3 SuperGAT/main.py --help.

• https://github.com/rusty1s/pytorch_geometric/blob/master/examples
• https://github.com/Diego999/pyGAT/blob/master/layers.py