Compressing Deep Graph Neural Networks via Adversarial Knowledge Distillation

This is the code of paper Compressing Deep Graph Neural Networks via Adversarial Knowledge Distillation. Huarui He, Jie Wang, Zhanqiu Zhang, Feng Wu. SIGKDD 2022. [arXiv]

Requirements

python 3.7.3
torch 1.9.1
dgl 0.9.1
ogb 1.3.4
torch-geometric 2.1.0
gdown 4.5.1

Reproduce the Results

First, download teacher knowledge from Google Drive

python download_teacher_knowledge.py --data_name=<dataset>
python download_teacher_knowledge.py --data_name=cora

Second, pleaes run the commands in node-level/README.md or graph-level/README.md to reproduce the results.

File tree

GraphAKD
├─ README.md
├─ download_teacher_knowledge.py
├─ datasets
│  └─ ...
├─ distilled
│  ├─ cora-knowledge.pth.tar
│  └─ ...
├─ graph-level
│  ├─ README.md
│  └─ stu-gnn
│     ├─ conv.py
│     ├─ gnn.py
│     └─ main.py
└─ node-level
   ├─ README.md
   ├─ stu-cluster-gcn
   │  ├─ dataset
   │  │  ├─ ogbn-products_160.npy
   │  │  └─ yelp_120.npy
   │  ├─ gcnconv.py
   │  ├─ models.py
   │  ├─ sampler.py
   │  └─ train.py
   └─ stu-gcn
      ├─ gcn.py
      ├─ gcnconv.py
      └─ train.py

Citation

If you find this code useful, please consider citing the following paper.

@inproceedings{KDD22_GraphAKD,
  author={Huarui He and Jie Wang and Zhanqiu Zhang and Feng Wu},
  booktitle={Proc. of SIGKDD},
  title={Compressing Deep Graph Neural Networks via Adversarial Knowledge Distillation},
  year={2022}
}

Acknowledgement

We refer to the code of DGL. Thanks for their contributions.

GraphAKD

Install / Use