GraphNorm: A Principled Approach to Accelerating Graph Neural Network Training

Tianle Cai*, Shengjie Luo*, Keyulu Xu, Di He, Tie-Yan Liu, Liwei Wang

This repository is the official implementation of GraphNorm: A Principled Approach to Accelerating Graph Neural Network Training, based on the PyTorch and DGL library.

Contact

Tianle Cai (caitianle1998@pku.edu.cn), Shengjie Luo (luosj@stu.pku.edu.cn)

Sincerely appreciate your suggestions on our work!

Overview

GraphNorm is a principled normalization method that accelerates the GNNs training on graph classification tasks, where the key idea is to normalize all nodes for each individual graph with a learnable shift. Theoretically, we show that GraphNorm serves as a preconditioner that smooths the distribution of the graph aggregation's spectrum, and the learnable shift is used to improve the expressiveness of the networks. Empirically, we conduct experiments on several popular benckmark datasets, including the recently released Open Graph Benchmark. Results on datasets with different scale consistently show that GNNs with GraphNorm converge much faster and achieve better generalization performance.

Installation

1. Clone this repository

git clone https://github.com/lsj2408/GraphNorm.git

2. Install the dependencies (Python 3.6.8)

pip install -r requirements.txt

Examples

Results for Bioinformatics and Social Network datasets

Training

To reproduce Figure 4 in our paper, run as follows:

cd ./GraphNorm_ws/gnn_ws/gnn_example/scripts/example-train-comparison

• Bioinformatics datasets: MUTAG, PTC, PROTEINS, NCI1

  ./gin-train-bioinformatics.sh
  ./gcn-train-bioinformatics.sh

• Social Network datasets

  • REDDIT-BINARY

    ./gin-train-rdtb.sh
    ./gcn-train-rdtb.sh

  • IMDB-BINARY

    ./gin-train-imdbb.sh
    ./gcn-train-imdbb.sh

  • COLLAB

    ./gin-train-collab.sh
    ./gcn-train-collab.sh

Results are stored in ./gnn_ws/log/Example-train-performance/, you can use the recorded metric numbers to plot the training curves as Figure 4 in our paper.

Testing

Here we provide examples to reproduce the test results on Bioinformatics and Social Network datasets.

cd ./GraphNorm_ws/gnn_ws/gnn_example/scripts/example-test-comparison

• PROTEINS:

  ./gin-train-proteins.sh

• REDDIT-BINARY:

  ./gin-train-rdtb.sh

Results are stored in ./gnn_ws/log/Example-test-performance/. For further results on other datasets, follow the configurations on Appendix C.

Results for Ogbg-molhiv

Training

Training process is performed with 10 different random seeds

• GCN

  cd ./GraphNorm_ws/ogbg_ws/scripts/Example-gcn-test-performance
  ./seed-1-5-gcn_run.sh
  ./seed-6-10-gcn_run.sh

• GIN

  cd ./GraphNorm_ws/ogbg_ws/scripts/Example-gin-test-performance
  ./seed-1-5-gin_run.sh
  ./seed-6-10-gin_run.sh

Results can be found in :

• ./GraphNorm_ws/ogbg_ws/log/: recorded metric values along training process,
• ./GraphNorm_ws/ogbg_ws/model: model checkpoints, which has the maximum validation metric values along training process.

Notes:

The number of epoch as 20~30 is fine for GIN with GraphNorm.

Evaluation

For evaluation, we use the dumped model checkpoints and report the mean and standard deviation metric values.

• Use your own trained models

  Set the MODEL_PATH in ./ogbg_ws/scripts/evaluate_gin.sh / ./ogbg_ws/scripts/evaluate_gcn.sh to the desired item.

• Use provided pre-trained models

  We provide the TOP-1 model on OGBG-MOLHIV datasets here. The training and evaluation command is provided below.

  #!/usr/bin/env bash
  
  set -e
  
  GPU=0
  NORM=gn
  BS=128
  DP=0.1
  EPOCH=50
  Layer=6
  LR=0.0001
  HIDDEN=300
  DS=ogbg-molhiv
  LOG_PATH=../../log/"$NORM"-BS-"$BS"-EPOCH-"$EPOCH"-L-"$Layer"-HIDDEN-"$HIDDEN"-LR-"$LR"-decay/
  MODEL_PATH=../../model/"$NORM"-BS-"$BS"-EPOCH-"$EPOCH"-L-"$Layer"-HIDDEN-"$HIDDEN"-LR-"$LR"-decay/
  DATA_PATH=../../data/dataset/
  
  
  for seed in {1..10}; do
      FILE_NAME=learn-"$DS"-gcn-seed-"$seed"
      python ../../src/train_dgl_ogb.py \
          --gpu $GPU \
          --epoch $EPOCH \
          --dropout $DP \
          --model GCN_dp \
          --batch_size $BS \
          --n_layers $Layer \
          --lr $LR \
          --n_hidden $HIDDEN \
          --seed $seed \
          --dataset $DS \
          --log_dir $LOG_PATH \
          --model_path $MODEL_PATH \
          --data_dir $DATA_PATH \
          --exp $FILE_NAME \
          --norm_type $NORM \
          --log_norm
  done

  #!/usr/bin/env bash
  
  set -e
  
  GPU=0
  L=6
  NORM=gn
  MODEL=GCN_dp
  DS=ogbg-molhiv
  BS=128
  MODEL_PATH=../../model/[The path name of the Pre-trained Model]/
  DATA_PATH=../../data/dataset/
  
  python ../../src/evaluate_ogb.py \
          --gpu $GPU \
          --n_layers $L \
          --dataset $DS \
          --model_path $MODEL_PATH \
          --data_dir $DATA_PATH \
          --norm_type $NORM \
          --model $MODEL \
          --batch_size $BS

Results

• Training Performance

• Test Performance

GCN with GraphNorm outperforms several sophisticated GNNs on OGBG-MOLHIV datasets.

Rank	Method	Test ROC-AUC
1	GCN+GraphNorm	$0.7883\pm0.0100$
2	HIMP	$0.7880\pm0.0082$
3	DeeperGCN	$0.7858\pm0.0117$
4	WEGL	$0.7757\pm0.0111$
5	GIN+virtual node	$0.7707\pm0.0149$
6	GCN	$0.7606\pm0.0097$
7	GCN+virtual node	$0.7599\pm0.0119$
8	GIN	$0.7558\pm0.0140$

• Ablation Study

• Visualizations of Singular value distribution

• Visualizations of Noisy Batch-level Statistics

Citation

@misc{cai2020graphnorm,
    title={GraphNorm: A Principled Approach to Accelerating Graph Neural Network Training},
    author={Tianle Cai and Shengjie Luo and Keyulu Xu and Di He and Tie-yan Liu and Liwei Wang},
    year={2020},
    eprint={2009.03294},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}