This repository contains a PyTorch implementation of "SCGC : Self-Supervised Contrastive Graph Clustering".(https://arxiv.org/abs/2204.12656)
The repo, inlcuding data sets and pretrained models are, has been forked initially from SDCN. We use also use the model code from AGCN and portions of contrastive loss code from Graph-MLP.
- Our code was tested on CUDA 11.3.0, python 3.6.9, pytorch 1.3.1.
pip install -q munkresis needed for the hungarian algorithem, for the evaluation metrics- The code also run on Google colab (2022 April) with no modifications, with munkres installed.
Note : SCGC is able to run with no GPU if the GPU timing code is commented out.
The dataset contains 2 folders, data and graph. Please obtain them from the dataset Google drive links. You will need to set --data_path to the parent folder containing data and graph. Please note that the data folder contains the pre-trained .pkl models. We directly use these pre-trained models from SDCN.
-
All parameters are defined in train.py with comments and explanations.
-
To run SCGC on the 6 datasets, for 10 iterations, use the following. This code has GPU time and memory profiling enabled, which can be turned off by commenting relevant code. Our published ACC,NMI,ARI and F1 was run with profiling commented.
python train.py --name usps --iterations 10 --epochs 200 --model SCGC --verbosity 0 --alpha 1 --beta 0.1 --order 4 --tau 0.5 --lr 0.001
python train.py --name hhar --iterations 10 --epochs 200 --model SCGC --verbosity 0 --alpha 1 --beta 10 --order 4 --tau 2.25 --lr 0.001
python train.py --name reut --iterations 10 --epochs 200 --model SCGC --verbosity 0 --alpha 3 --beta 0.1 --order 3 --tau 1 --lr 0.001
python train.py --name acm --iterations 10 --epochs 200 --model SCGC --verbosity 0 --alpha 0.5 --beta 0.1 --order 2 --tau 0.25 --lr 0.001
python train.py --name dblp --iterations 10 --epochs 200 --model SCGC --verbosity 0 --alpha 0.5 --beta 0.1 --order 1 --tau 0.25 --lr 0.001
python train.py --name cite --iterations 10 --epochs 200 --model SCGC --verbosity 0 --alpha 1 --beta 0.1 --order 1 --tau 0.25 --lr 0.0001
- To replicate the SCGC*, which uses Augmented Contrastive (IAC) loss and 50% less model parameters, run the following.
python train.py --name usps --iterations 10 --epochs 200 --model SCGC_TRIM --verbosity 0 --alpha 4 --beta 0.1 --order 4 --tau 0.25 --lr 0.001 --influence
python train.py --name hhar --iterations 10 --epochs 200 --model SCGC_TRIM --verbosity 0 --alpha 1 --beta 10 --order 3 --tau 2.25 --lr 0.001 --influence
python train.py --name reut --iterations 10 --epochs 200 --model SCGC_TRIM --verbosity 0 --alpha 0.5 --beta 0.1 --order 3 --tau 0.25 --lr 0.001 --influence
python train.py --name acm --iterations 10 --epochs 200 --model SCGC_TRIM --verbosity 0 --alpha 1 --beta 0.1 --order 1 --tau 0.25 --lr 0.001 --influence
python train.py --name dblp --iterations 10 --epochs 200 --model SCGC_TRIM --verbosity 0 --alpha 1 --beta 0.1 --order 1 --tau 0.25 --lr 0.001 --influence
python train.py --name cite --iterations 10 --epochs 200 --model SCGC_TRIM --verbosity 0 --alpha 1 --beta 0.1 --order 1 --tau 0.25 --lr 0.0001 --influence
The usps output is
Namespace(alpha=4.0, batch_size=2048, beta=0.1, cuda=True, data_path='/content/drive/MyDrive/001_Clustering/_Dataset_SDCN', epochs=200, influence=True, iterations=1, k=3, lr=0.001, mode='full', model='SCGC_TRIM', n_clusters=10, n_input=256, n_z=10, name='usps', note='-', order=4, seed=42, tau=0.25, verbosity=0)
---------------PROFILING CODE--------------
Loaded PAE acc:0.7098 nmi:0.6748 ari:0.5874 f1:0.6968
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg Self CUDA Self CUDA % CUDA total CUDA time avg CPU Mem Self CPU Mem CUDA Mem Self CUDA Mem # of Calls Total GFLOPs
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
_MODEL_TRAIN_ALL 27.73% 50.714s 99.25% 181.531s 181.531s 0.000us 0.00% 26.670s 26.670s -4 b -2.41 Gb 0 b -367.99 Gb 1 --
_MODEL_TRAIN 0.06% 109.462ms 0.23% 421.977ms 2.110ms 0.000us 0.00% 3.386s 16.931ms 2.34 Kb -800 b 876.65 Mb -42.76 Gb 200 --
_MODEL_KL 0.01% 19.598ms 0.07% 118.949ms 594.745us 0.000us 0.00% 11.177ms 55.885us 800 b -2.34 Kb 71.00 Mb -199.50 Kb 200 --
_MODEL_DIST 1.03% 1.888s 14.81% 27.096s 135.478ms 0.000us 0.00% 15.566s 77.831ms -800 b -64.41 Gb 258.14 Gb -193.11 Gb 200 --
_MODEL_CONTRASTIVE 0.03% 48.284ms 0.08% 149.847ms 749.235us 0.000us 0.00% 4.008s 20.040ms 800 b -2.34 Kb 64.48 Gb -128.92 Gb 200 --
------------------------------------------------------- ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------ ------------
Self CPU time total: 182.898s
Self CUDA time total: 50.388s
Z:acc-nmi-ari-F1-gpu-clock: 0.8489, 0.8489, 0.8489, 0.0000,|, 0.8411, 0.8411, 0.8411, 0.0000,|, 0.7941, 0.7941, 0.7941, 0.0000,|, 0.8152, 0.8152, 0.8152, 0.0000,|,106441.4219, 106441.4219, 106441.4219, 0.0000,|,106.4405, 106.4405, 106.4405, 0.0000,||, Namespace(alpha=4.0, batch_size=2048, beta=0.1, cuda=True, data_path='/content/drive/MyDrive/001_Clustering/_Dataset_SDCN', epochs=200, influence=True, iterations=1, k=3, lr=0.001, mode='full', model='SCGC_TRIM', n_clusters=10, n_input=256, n_z=10, name='usps', note='-', order=4, seed=42, tau=0.25, verbosity=0)
Namespace(alpha=1.0, batch_size=2048, beta=10.0, cuda=True, data_path='/content/drive/MyDrive/001_Clustering/_Dataset_SDCN', epochs=200, influence=True, iterations=1, k=5, lr=0.001, mode='full', model='SCGC_TRIM', n_clusters=6, n_input=561, n_z=10, name='hhar', note='-', order=3, seed=42, tau=2.25, verbosity=0)
The line Z:acc-nmi-ari-F1-gpu-clock gives the min,max,avg,std of ACC,NMI,ARI,F1,GPU,CPU followed by || and all the args.
Profiling _MODEL_XXX contexts capture logical model functions and training. Please see the code for more information.
- To reun the prfiling for AGCN and SGCN we used the following. These use the SDCN and AGCN defaults.
python train.py --name usps --iterations 1 --epochs 200 --model SDCN --verbosity 0 --lr 0.001 --alpha 0.01 --beta 0.1
python train.py --name hhar --iterations 1 --epochs 200 --model SDCN --verbosity 0 --lr 0.001 --alpha 0.01 --beta 0.1
python train.py --name reut --iterations 1 --epochs 200 --model SDCN --verbosity 0 --lr 0.0001 --alpha 0.01 --beta 0.1
python train.py --name acm --iterations 1 --epochs 200 --model SDCN --verbosity 0 --lr 0.001 --alpha 0.01 --beta 0.1
python train.py --name dblp --iterations 1 --epochs 200 --model SDCN --verbosity 0 --lr 0.001 --alpha 0.01 --beta 0.1
python train.py --name cite --iterations 1 --epochs 200 --model SDCN --verbosity 0 --lr 0.0001 --alpha 0.01 --beta 0.1
python train.py --name usps --iterations 1 --epochs 200 --model AGCN --verbosity 0 --lr 0.001 --alpha 1000 --beta 1000
python train.py --name hhar --iterations 1 --epochs 200 --model AGCN --verbosity 0 --lr 0.001 --alpha 0.1 --beta 1
python train.py --name reut --iterations 1 --epochs 200 --model AGCN --verbosity 0 --lr 0.0001 --alpha 10 --beta 10
python train.py --name acm --iterations 1 --epochs 200 --model AGCN --verbosity 0 --lr 0.001 --alpha 0.01 --beta 0.1
python train.py --name dblp --iterations 1 --epochs 200 --model AGCN --verbosity 0 --lr 0.001 --alpha 0.01 --beta 0.1
python train.py --name cite --iterations 1 --epochs 200 --model AGCN --verbosity 0 --lr 0.0001 --alpha 0.01 --beta 0.1
Datasets and code is forked from SDCN. We use also use the model code from AGCN AGCN and portions of contrastive loss code from Graph-MLP. We acknowledge and thank the authors of these works for sharing their code.
@article{kulatilleke2022scgc,
title={SCGC : Self-Supervised Contrastive Graph Clustering},
author={Kulatilleke, Gayan K and Portmann, Marius and Chandra, Shekhar S},
journal={arXiv preprint arXiv:2204.12656},
year={2022}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent