GCC: Graph Contrastive Coding for Graph Neural Network Pre-Training @ KDD 2020

Last update: Dec 27, 2022

Overview

GCC: Graph Contrastive Coding for Graph Neural Network Pre-Training

Original implementation for paper GCC: Graph Contrastive Coding for Graph Neural Network Pre-Training.

GCC is a contrastive learning framework that implements unsupervised structural graph representation pre-training and achieves state-of-the-art on 10 datasets on 3 graph mining tasks.

GCC: Graph Contrastive Coding for Graph Neural Network Pre-Training

Installation

Requirements

Linux with Python ≥ 3.6
PyTorch ≥ 1.4.0
0.5 > DGL ≥ 0.4.3
pip install -r requirements.txt
Install RDKit with conda install -c conda-forge rdkit=2019.09.2.

Quick Start

Pretraining

Pre-training datasets

python scripts/download.py --url https://drive.google.com/open?id=1JCHm39rf7HAJSp-1755wa32ToHCn2Twz --path data --fname small.bin
# For regions where Google is not accessible, use
# python scripts/download.py --url https://cloud.tsinghua.edu.cn/f/b37eed70207c468ba367/?dl=1 --path data --fname small.bin

E2E

Pretrain E2E with K = 255:

bash scripts/pretrain.sh <gpu> --batch-size 256

MoCo

Pretrain MoCo with K = 16384; m = 0.999:

bash scripts/pretrain.sh <gpu> --moco --nce-k 16384

Download Pretrained Models

Instead of pretraining from scratch, you can download our pretrained models.

python scripts/download.py --url https://drive.google.com/open?id=1lYW_idy9PwSdPEC7j9IH5I5Hc7Qv-22- --path saved --fname pretrained.tar.gz
# For regions where Google is not accessible, use
# python scripts/download.py --url https://cloud.tsinghua.edu.cn/f/cabec37002a9446d9b20/?dl=1 --path saved --fname pretrained.tar.gz

Downstream Tasks

Downstream datasets

python scripts/download.py --url https://drive.google.com/open?id=12kmPV3XjVufxbIVNx5BQr-CFM9SmaFvM --path data --fname downstream.tar.gz
# For regions where Google is not accessible, use
# python scripts/download.py --url https://cloud.tsinghua.edu.cn/f/2535437e896c4b73b6bb/?dl=1 --path data --fname downstream.tar.gz

Generate embeddings on multiple datasets with

bash scripts/generate.sh <gpu> <load_path> <dataset_1> <dataset_2> ...

For example:

bash scripts/generate.sh 0 saved/Pretrain_moco_True_dgl_gin_layer_5_lr_0.005_decay_1e-05_bsz_32_hid_64_samples_2000_nce_t_0.07_nce_k_16384_rw_hops_256_restart_prob_0.8_aug_1st_ft_False_deg_16_pos_32_momentum_0.999/current.pth usa_airport kdd imdb-binary

Node Classification

Unsupervised (Table 2 freeze)

Run baselines on multiple datasets with bash scripts/node_classification/baseline.sh <hidden_size> <baseline:prone/graphwave> usa_airport h-index.

Evaluate GCC on multiple datasets:

bash scripts/generate.sh <gpu> <load_path> usa_airport h-index
bash scripts/node_classification/ours.sh <load_path> <hidden_size> usa_airport h-index

Supervised (Table 2 full)

Finetune GCC on multiple datasets:

bash scripts/finetune.sh <load_path> <gpu> usa_airport

Note this finetunes the whole network and will take much longer than the freezed experiments above.

Graph Classification

Unsupervised (Table 3 freeze)

bash scripts/generate.sh <gpu> <load_path> imdb-binary imdb-multi collab rdt-b rdt-5k
bash scripts/graph_classification/ours.sh <load_path> <hidden_size> imdb-binary imdb-multi collab rdt-b rdt-5k

Supervised (Table 3 full)

bash scripts/finetune.sh <load_path> <gpu> imdb-binary

Similarity Search (Table 4)

Run baseline (graphwave) on multiple datasets with bash scripts/similarity_search/baseline.sh <hidden_size> graphwave kdd_icdm sigir_cikm sigmod_icde.

Run GCC:

bash scripts/generate.sh <gpu> <load_path> kdd icdm sigir cikm sigmod icde
bash scripts/similarity_search/ours.sh <load_path> <hidden_size> kdd_icdm sigir_cikm sigmod_icde

❗ Common Issues

"XXX file not found" when running pretraining/downstream tasks.

Please make sure you've downloaded the pretraining dataset or downstream task datasets according to GETTING_STARTED.md.

Server crashes/hangs after launching pretraining experiments.

In addition to GPU, our pretraining stage requires a lot of computation resources, including CPU and RAM. If this happens, it usually means the CPU/RAM is exhausted on your machine. You can decrease `--num-workers` (number of dataloaders using CPU) and `--num-copies` (number of datasets copies residing in RAM). With the lowest profile, try `--num-workers 1 --num-copies 1`.

If this still fails, please upgrade your machine :). In the meanwhile, you can still download our pretrained model and evaluate it on downstream tasks.

Having difficulty installing RDKit.

See the P.S. section in [this](https://github.com/THUDM/GCC/issues/12#issue-752080014) post.

Citing GCC

If you use GCC in your research or wish to refer to the baseline results, please use the following BibTeX.

@article{qiu2020gcc,
  title={GCC: Graph Contrastive Coding for Graph Neural Network Pre-Training},
  author={Qiu, Jiezhong and Chen, Qibin and Dong, Yuxiao and Zhang, Jing and Yang, Hongxia and Ding, Ming and Wang, Kuansan and Tang, Jie},
  journal={arXiv preprint arXiv:2006.09963},
  year={2020}
}

Acknowledgements

Part of this code is inspired by Yonglong Tian et al.'s CMC: Contrastive Multiview Coding.

GCC: Graph Contrastive Coding for Graph Neural Network Pre-Training @ KDD 2020

Related tags

Overview

GCC: Graph Contrastive Coding for Graph Neural Network Pre-Training

Installation

Requirements

Quick Start

Pretraining

Pre-training datasets

E2E

MoCo

Download Pretrained Models

Downstream Tasks

Downstream datasets

Node Classification

Unsupervised (Table 2 freeze)

Supervised (Table 2 full)

Graph Classification

Unsupervised (Table 3 freeze)

Supervised (Table 3 full)

Similarity Search (Table 4)

❗ Common Issues

Citing GCC

Acknowledgements

Owner

THUDM

BraTs-VNet - BraTS(Brain Tumour Segmentation) using V-Net

Multi-Task Deep Neural Networks for Natural Language Understanding

[NeurIPS 2021] Introspective Distillation for Robust Question Answering

Lunar is a neural network aimbot that uses real-time object detection accelerated with CUDA on Nvidia GPUs.

OpenIPDM is a MATLAB open-source platform that stands for infrastructures probabilistic deterioration model

Jetson Nano-based smart camera system that measures crowd face mask usage in real-time.

AdelaiDepth is an open source toolbox for monocular depth prediction.

Implementation for NeurIPS 2021 Submission: SparseFed

Understanding Hyperdimensional Computing for Parallel Single-Pass Learning

PyTorch implementation of 'Gen-LaneNet: a generalized and scalable approach for 3D lane detection'

RoadMap and preparation material for Machine Learning and Data Science - From beginner to expert.

Code and training data for our ECCV 2016 paper on Unsupervised Learning

Towards Calibrated Model for Long-Tailed Visual Recognition from Prior Perspective

Source code for CVPR 2021 paper "Riggable 3D Face Reconstruction via In-Network Optimization"

Weakly Supervised Dense Event Captioning in Videos, i.e. generating multiple sentence descriptions for a video in a weakly-supervised manner.

YOLOv5 detection interface - PyQt5 implementation

All course materials for the Zero to Mastery Machine Learning and Data Science course.

OpenPCDet Toolbox for LiDAR-based 3D Object Detection.

Toward Spatially Unbiased Generative Models (ICCV 2021)

3DV 2021: Synergy between 3DMM and 3D Landmarks for Accurate 3D Facial Geometry