Introduction

Code for Representation Learning for Clustering via Building Consensus

If you find this work useful to your, please cite :

@article{deshmukh2021representation,
  title={Representation Learning for Clustering via Building Consensus},
  author={Deshmukh, Aniket Anand and Regatti, Jayanth Reddy and Manavoglu, Eren and Dogan, Urun},
  journal={Springer Machine Learning Journal arXiv preprint arXiv:2105.01289},
  year={2021}
}

Conda Environment

conda env create -f concurl.yml

Reproduce Results

Download the models from saved_models and store them in saved_models folder. The datasets need to be downloaded prior to executing the following command.

python extract_features_clustering_eval.py

Sample Commands

A sample command to run the algorithm is as follows. The commands for the best models are provided in run.sh.

CUDA_VISIBLE_DEVICES=0 python main.py \
--use-no-grad \
--hidden-mlp 2048 \
--eval-freq 10 \
--trial 2 \
--workers 8 \
--use-consensus \
--use-rp \
--alpha 0.0 \
--beta 1.0 \
--gamma 1.0 \
--projection-dim 136 \
--n-transforms 64 \
--use-torch-resnet \
--arch resnet50 \
--use-slightly-diff-views \
--lr 0.06 \
--optim SGD \
--batch-size 128 \
--num-epochs 1 \
--nce-temp 0.5 \
--nce-k 4096 \
--n_clusters 15 \
--datapath $DATAPATH \
--logdir $LOGPATH \
--image-size 160 

Results

Dataset	Acc	NMI	ARI
ImageNet-10	0.958	0.907	0.909
ImageNet-Dogs	0.695	0.63	0.531
STL10	0.749	0.636	0.566
CIFAR10	0.846	0.762	0.715
CIFAR100	0.479	0.468	0.303

Acknowledgements

The code for the paper is built using the following two repositories: https://github.com/facebookresearch/swav, https://github.com/zhirongw/lemniscate.pytorch