This repository is an implementation of paper : Improving the Training of Graph Neural Networks with Consistency Regularization

Related tags

Deep Learning CRGNN
Overview

CRGNN

Paper ： Improving the Training of Graph Neural Networks with Consistency Regularization

Environments

Implementing environment: GeForce RTX™ 3090 24GB (GPU)

Requirements

pytorch>=1.8.1

ogb=1.3.2

numpy=1.21.2

cogdl (latest version)

Training

GAMLP+RLU+SCR

For ogbn-products:

Params: 3335831
python pre_processing.py --num_hops 5 --dataset ogbn-products

python main.py --use-rlu --method R_GAMLP_RLU --stages 400 300 300 300 300 300 --train-num-epochs 0 0 0 0 0 0 --threshold 0.85 --input-drop 0.2 --att-drop 0.5 --label-drop 0 --pre-process --residual --dataset ogbn-products --num-runs 10 --eval 10 --act leaky_relu --batch_size 50000 --patience 300 --n-layers-1 4 --n-layers-2 4 --bns --gama 0.1 --consis --tem 0.5 --lam 0.1 --hidden 512 --ema

GAMLP+MCR

For ogbn-products:

Params: 3335831
python pre_processing.py --num_hops 5 --dataset ogbn-products

python main.py --use-rlu --method R_GAMLP_RLU --stages 800 --train-num-epochs 0 --input-drop 0.2 --att-drop 0.5 --label-drop 0 --pre-process --residual --dataset ogbn-products --num-runs 10 --eval 10 --act leaky_relu --batch_size 100000 --patience 300 --n-layers-1 4 --n-layers-2 4 --bns --gama 0.1 --tem 0.5 --lam 0.5 --ema --mean_teacher --ema_decay 0.999 --lr 0.001 --adap --gap 10 --warm_up 150 --top 0.9 --down 0.8 --kl --kl_lam 0.2 --hidden 512

GIANT-XRT+GAMLP+MCR

Please follow the instruction in GIANT to get the GIANT-XRT node features.

For ogbn-products:

Params: 2144151
python pre_processing.py --num_hops 5 --dataset ogbn-products --giant_path " "

python main.py --use-rlu --method R_GAMLP_RLU --stages 800 --train-num-epochs 0 --input-drop 0.2 --att-drop 0.5 --label-drop 0 --pre-process --residual --dataset ogbn-products --num-runs 10 --eval 10 --act leaky_relu --batch_size 100000 --patience 300 --n-layers-1 4 --n-layers-2 4 --bns --gama 0.1 --tem 0.5 --lam 0.5 --ema --mean_teacher --ema_decay 0.99 --lr 0.001 --adap --gap 10 --warm_up 150 --kl --kl_lam 0.2 --hidden 256 --down 0.7 --top 0.9 --giant

SAGN+MCR

For ogbn-products:

Params: 2179678
python pre_processing.py --num_hops 3 --dataset ogbn-products

python main.py --method SAGN --stages 1000 --train-num-epochs 0 --input-drop 0.2 --att-drop 0.4 --pre-process --residual --dataset ogbn-products --num-runs 10 --eval 10 --batch_size 100000 --patience 300 --tem 0.5 --lam 0.5 --ema --mean_teacher --ema_decay 0.99 --lr 0.001 --adap --gap 20 --warm_up 150 --top 0.85 --down 0.75 --kl --kl_lam 0.01 --hidden 512 --zero-inits --dropout 0.5 --num-heads 1  --label-drop 0.5  --mlp-layer 2 --num_hops 3 --label_num_hops 14

GIANT-XRT+SAGN+MCR

Please follow the instruction in GIANT to get the GIANT-XRT node features.

For ogbn-products:

Params: 1154654
python pre_processing.py --num_hops 3 --dataset ogbn-products --giant_path " "

python main.py --method SAGN --stages 1000 --train-num-epochs 0 --input-drop 0.2 --att-drop 0.4 --pre-process --residual --dataset ogbn-products --num-runs 10 --eval 10 --batch_size 50000 --patience 300 --tem 0.5 --lam 0.5 --ema --mean_teacher --ema_decay 0.99 --lr 0.001 --adap --gap 20 --warm_up 100 --top 0.85 --down 0.75 --kl --kl_lam 0.02 --hidden 256 --zero-inits --dropout 0.5 --num-heads 1  --label-drop 0.5  --mlp-layer 1 --num_hops 3 --label_num_hops 9 --giant

Use Optuna to search for C&S hyperparameters

We searched hyperparameters using Optuna on validation set.

python post_processing.py --file_name --search

GAMLP+RLU+SCR+C&S

python post_processing.py --file_name --correction_alpha 0.4780826957236622 --smoothing_alpha 0.40049734940262954

GIANT-XRT+SAGN+MCR+C&S

python post_processing.py --file_name --correction_alpha 0.42299283241438157 --smoothing_alpha 0.4294212449832242

Node Classification Results:

Performance on ogbn-products(10 runs):

Methods Validation accuracy Test accuracy
SAGN+MCR 0.9325±0.0004 0.8441±0.0005
GAMLP+MCR 0.9319±0.0003 0.8462±0.0003
GAMLP+RLU+SCR 0.9292±0.0005 0.8505±0.0009
GAMLP+RLU+SCR+C&S 0.9304±0.0005 0.8520±0.0008
GIANT-XRT+GAMLP+MCR 0.9402±0.0004 0.8591±0.0008
GIANT-XRT+SAGN+MCR 0.9389±0.0002 0.8651±0.0009
GIANT-XRT+SAGN+MCR+C&S 0.9387±0.0002 0.8673±0.0008

Citation

Our paper:

@misc{zhang2021improving,
      title={Improving the Training of Graph Neural Networks with Consistency Regularization}, 
      author={Chenhui Zhang and Yufei He and Yukuo Cen and Zhenyu Hou and Jie Tang},
      year={2021},
      eprint={2112.04319},
      archivePrefix={arXiv},
      primaryClass={cs.SI}
}

GIANT paper:

@article{chien2021node,
  title={Node Feature Extraction by Self-Supervised Multi-scale Neighborhood Prediction},
  author={Eli Chien and Wei-Cheng Chang and Cho-Jui Hsieh and Hsiang-Fu Yu and Jiong Zhang and Olgica Milenkovic and Inderjit S Dhillon},
  journal={arXiv preprint arXiv:2111.00064},
  year={2021}
}

GAMLP paper:

@article{zhang2021graph,
  title={Graph attention multi-layer perceptron},
  author={Zhang, Wentao and Yin, Ziqi and Sheng, Zeang and Ouyang, Wen and Li, Xiaosen and Tao, Yangyu and Yang, Zhi and Cui, Bin},
  journal={arXiv preprint arXiv:2108.10097},
  year={2021}
}

SAGN paper:

@article{sun2021scalable,
  title={Scalable and Adaptive Graph Neural Networks with Self-Label-Enhanced training},
  author={Sun, Chuxiong and Wu, Guoshi},
  journal={arXiv preprint arXiv:2104.09376},
  year={2021}
}

C&S paper:

@inproceedings{
huang2021combining,
title={Combining Label Propagation and Simple Models out-performs Graph Neural Networks},
author={Qian Huang and Horace He and Abhay Singh and Ser-Nam Lim and Austin Benson},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=8E1-f3VhX1o}
}
You might also like...
PyTorch implementation of
PyTorch implementation of "ContextNet: Improving Convolutional Neural Networks for Automatic Speech Recognition with Global Context" (INTERSPEECH 2020)

ContextNet ContextNet has CNN-RNN-transducer architecture and features a fully convolutional encoder that incorporates global context information into

Sangchun Ha 24 Nov 24, 2022
Stochastic Downsampling for Cost-Adjustable Inference and Improved Regularization in Convolutional Networks
Stochastic Downsampling for Cost-Adjustable Inference and Improved Regularization in Convolutional Networks

Stochastic Downsampling for Cost-Adjustable Inference and Improved Regularization in Convolutional Networks (SDPoint) This repository contains the cod

Jason Kuen 17 Jul 4, 2022
This repository contains the official implementation code of the paper Improving Multimodal Fusion with Hierarchical Mutual Information Maximization for Multimodal Sentiment Analysis, accepted at EMNLP 2021.
This repository contains the official implementation code of the paper Improving Multimodal Fusion with Hierarchical Mutual Information Maximization for Multimodal Sentiment Analysis, accepted at EMNLP 2021.

MultiModal-InfoMax This repository contains the official implementation code of the paper Improving Multimodal Fusion with Hierarchical Mutual Informa

Deep Cognition and Language Research (DeCLaRe) Lab 89 Dec 26, 2022
A static analysis library for computing graph representations of Python programs suitable for use with graph neural networks.

python_graphs This package is for computing graph representations of Python programs for machine learning applications. It includes the following modu

Google Research 258 Dec 29, 2022
Some tentative models that incorporate label propagation to graph neural networks for graph representation learning in nodes, links or graphs.

Some tentative models that incorporate label propagation to graph neural networks for graph representation learning in nodes, links or graphs.

zshicode 1 Nov 18, 2021
On Size-Oriented Long-Tailed Graph Classification of Graph Neural Networks

On Size-Oriented Long-Tailed Graph Classification of Graph Neural Networks We provide the code (in PyTorch) and datasets for our paper "On Size-Orient

Zemin Liu 4 Jun 18, 2022
This repository contains notebook implementations of the following Neural Process variants: Conditional Neural Processes (CNPs), Neural Processes (NPs), Attentive Neural Processes (ANPs).

The Neural Process Family This repository contains notebook implementations of the following Neural Process variants: Conditional Neural Processes (CN

DeepMind 892 Dec 28, 2022
Degree-Quant: Quantization-Aware Training for Graph Neural Networks.

Degree-Quant This repo provides a clean re-implementation of the code associated with the paper Degree-Quant: Quantization-Aware Training for Graph Ne

null 35 Oct 7, 2022
[Preprint]
[Preprint] "Bag of Tricks for Training Deeper Graph Neural Networks A Comprehensive Benchmark Study" by Tianlong Chen*, Kaixiong Zhou*, Keyu Duan, Wenqing Zheng, Peihao Wang, Xia Hu, Zhangyang Wang

Bag of Tricks for Training Deeper Graph Neural Networks: A Comprehensive Benchmark Study Codes for [Preprint] Bag of Tricks for Training Deeper Graph

VITA 101 Dec 29, 2022
Comments
  • RuntimeError: CUDA out of memory.

    RuntimeError: CUDA out of memory.

    I tried to reproduce on 2 Tesla V100s. Only one GPU core in use. In ubuntu 18.04. with ogb = 1.3.3, numpy = 1.20.3, cogdl = 0.5.2

    After the 1st training staged ended, the error message came out below: """ Best Epoch 350, Val 0.9310, Test 0.8349 This history model Train ACC is 0.9655824835338097 This history model Valid ACC is 0.9309564377082115 This history model Test ACC is 0.8348915611694232 Stage: 1, confident nodes: 1766768 Stage: 1, confident_cons nodes: 981983 use teacher label GAMLP R_GAMLP_RLU

    Params: 3335831

    Traceback (most recent call last): File "CRGNN-main/main.py", line 413, in main(args) File "CRGNN-main/main.py", line 297, in main best_val, best_test = run(args, device) File "CRGNN-main/main.py", line 223, in run loss, acc = train_rlu_consis(model, train_loader, enhance_loader, optimizer, evaluator, device, feats, labels, File "CRGNN-main/utils.py", line 102, in train_rlu_consis output_att_f = model(batch_feats, label_emb[idx_3].to(device)) File "/home/anaconda3/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl return forward_call(*input, **kwargs) File "CRGNN-main/model.py", line 340, in forward input_list.append(self.processi) File "/home/anaconda3/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl return forward_call(*input, **kwargs) File "/CRGNN-main/layer.py", line 93, in forward x = self.dropout(self.prelu(self.bnslayer_id)) File "/home/anaconda3/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl return forward_call(*input, **kwargs) File "/home/anaconda3/lib/python3.9/site-packages/torch/nn/modules/batchnorm.py", line 168, in forward return F.batch_norm( File "/home/anaconda3/lib/python3.9/site-packages/torch/nn/functional.py", line 2282, in batch_norm return torch.batch_norm( RuntimeError: CUDA out of memory. Tried to allocate 50.00 MiB (GPU 0; 31.75 GiB total capacity; 6.97 GiB already allocated; 42.12 MiB free; 7.01 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF """ It seems like there is still available sapce in GPU, but this error just came out. Searching online lots of people met this error. Some solved it by modifying the batch size, some tried in the same way but failed. Have you met this problem before in the training stage? Thank you very much.

    opened by wyk201722 2
Owner
THUDM
Data Mining Research Group at Tsinghua University
THUDM
GitHub
Consistency Regularization for Adversarial Robustness

Consistency Regularization for Adversarial Robustness Official PyTorch implementation of Consistency Regularization for Adversarial Robustness by Jiho

null 40 Dec 17, 2022
PyTorch framework, for reproducing experiments from the paper Implicit Regularization in Hierarchical Tensor Factorization and Deep Convolutional Neural Networks

Implicit Regularization in Hierarchical Tensor Factorization and Deep Convolutional Neural Networks. Code, based on the PyTorch framework, for reprodu

Asaf 3 Dec 27, 2022
Official PyTorch implementation of the paper: Improving Graph Neural Network Expressivity via Subgraph Isomorphism Counting.

Improving Graph Neural Network Expressivity via Subgraph Isomorphism Counting Official PyTorch implementation of the paper: Improving Graph Neural Net

Giorgos Bouritsas 58 Dec 31, 2022
Code for the paper Relation Prediction as an Auxiliary Training Objective for Improving Multi-Relational Graph Representations (AKBC 2021).

Relation Prediction as an Auxiliary Training Objective for Knowledge Base Completion This repo provides the code for the paper Relation Prediction as

Facebook Research 85 Jan 2, 2023
Measuring and Improving Consistency in Pretrained Language Models

ParaRel ?? This repository contains the code and data for the paper: Measuring and Improving Consistency in Pretrained Language Models as well as the

Yanai Elazar 26 Dec 2, 2022
This is the repository for the AAAI 21 paper [Contrastive and Generative Graph Convolutional Networks for Graph-based Semi-Supervised Learning].

CG3 This is the repository for the AAAI 21 paper [Contrastive and Generative Graph Convolutional Networks for Graph-based Semi-Supervised Learning]. R

null 12 Oct 28, 2022
We have implemented shaDow-GNN as a general and powerful pipeline for graph representation learning. For more details, please find our paper titled Deep Graph Neural Networks with Shallow Subgraph Samplers, available on arXiv (https//arxiv.org/abs/2012.01380).

Deep GNN, Shallow Sampling Hanqing Zeng, Muhan Zhang, Yinglong Xia, Ajitesh Srivastava, Andrey Malevich, Rajgopal Kannan, Viktor Prasanna, Long Jin, R

Facebook Research 117 Dec 20, 2022
The source code of the paper "Understanding Graph Neural Networks from Graph Signal Denoising Perspectives"

GSDN-F and GSDN-EF This repository provides a reference implementation of GSDN-F and GSDN-EF as described in the paper "Understanding Graph Neural Net

Guoji Fu 18 Nov 14, 2022
Code for CoMatch: Semi-supervised Learning with Contrastive Graph Regularization

CoMatch: Semi-supervised Learning with Contrastive Graph Regularization (Salesforce Research) This is a PyTorch implementation of the CoMatch paper [B

Salesforce 107 Dec 14, 2022
Training vision models with full-batch gradient descent and regularization

Stochastic Training is Not Necessary for Generalization -- Training competitive vision models without stochasticity This repository implements trainin

Jonas Geiping 32 Jan 6, 2023