EIGNN: Efficient Infinite-Depth Graph Neural Networks

This repository is the official PyTorch implementation of "EIGNN: Efficient Infinite-Depth Graph Neural Networks" (NeurIPS 2021).

Requirements

The script has been tested running under Python 3.6.9, with the following packages installed (along with their dependencies):

• pytorch (tested on 1.6.0)
• torch_geometric (tested on 1.6.3)
• scipy (tested on 1.5.2)
• numpy (tested on 1.19.2)

Synthetic experiments

Here is an example to train the model on synthetic experiments:

python train_EIGNN_chains.py --num_chains 20 --chain_len 100 --seed 14

num_chains and chain_len should be changed accordingly.

Experiments on real-world datasets

Evaluation and trained Models

We provide a few trained models on cornell datasets for demonstration purpose. See saved_model. To evaluate, run the command:

python eval_EIGNN_heterophilic.py --dataset cornell --epoch 10000 --lr 0.8 --weight_decay 5e-06 --gamma 0.8 --idx_split 1 

idx_split should be changed accordingly. There are 10 data splits as used in Geom-GCN.

Train the model

Here is an example to train the model on university datasets,

python train_EIGNN_heterophilic.py --dataset cornell --epoch 10000 --patience 500 --lr 0.8 --weight_decay 5e-06 --gamma 0.8 --idx_split 0 

This implementation is developed based on the original implementation of IGNN. We thank them for their useful implementation.

If you find our implementation useful in your research, please consider citing our paper:

@inproceedings{liu2021eignn,
 author = {Liu, Juncheng and Kawaguchi, Kenji and Hooi, Bryan and Wang, Yiwei and Xiao, Xiaokui},
 booktitle = {Advances in Neural Information Processing Systems},
 title = {EIGNN: Efficient Infinite-Depth Graph Neural Networks},
 year = {2021}
}