This repo is the PyTorch codes for "Improving Unsupervised Image Clustering With Robust Learning (RUC)"

Improving Unsupervised Image Clustering With Robust Learning

Sungwon Park, Sungwon Han, Sundong Kim, Danu Kim, Sungkyu Park, Seunghoon Hong, Meeyoung Cha.

• Accepted at CVPR 2021.
• 🏆 SOTA on 4 benchmarks. Check out Papers With Code for Image Clustering or Unsup. Classification.

1. RUC is an add-on module to enhance the performance of any off-the-shelf unsupervised learning algorithms. RUC is inspired by robust learning. It first divides clustered data points into clean and noisy set, then refine the clustering results. With RUC, state-of-the-art unsupervised clustering methods; SCAN and TSUC showed showed huge performance improvements. (STL-10 : 86.7%, CIFAR-10 : 90.3%, CIFAR-20 : 54.3%, CIFAR-100 : 36.5 %, ImageNet-50 : 78.5)

2. Prediction results of existing unsupervised learning algorithms were overconfident. RUC can make the prediction of existing algorithms softer with better calibration.

3. Robust to adversarially crafted samples. ERM-based unsupervised clustering algorithms can be prone to adversarial attack. Adding RUC to the clustering models improves robustness against adversarial noise.

• python == 3.6.10
• pytorch == 1.1.0
• scikit-learn == 0.21.2
• scipy == 1.3.0
• numpy == 1.18.5
• pillow == 7.1.2

usage: main_ruc_[dataset].py [-h] [--lr LR] [--momentum M] [--weight_decay W]
                         [--epochs EPOCHS] [--batch_size B] [--s_thr S_THR]
                         [--n_num N_NUM] [--o_model O_MODEL]
                         [--e_model E_MODEL] [--seed SEED]

config for RUC

optional arguments:
  -h, --help            show this help message and exit
  --lr LR               initial learning rate
  --momentum M          momentum
  --weight_decay        weight decay
  --epochs EPOCHS       max epoch per round. (default: 200)
  --batch_size B        training batch size
  --s_thr S_THR         confidence sampling threshold
  --n_num N_NUM         the number of neighbor for metric sampling
  --o_model O_MODEL     original model path
  --e_model E_MODEL     embedding model path
  --seed SEED           random seed

Currently, we support the pretrained model for our model. We used the pretrained SCAN and SimCLR model from SCAN github. (o_model : SCAN , e_model : SimCLR)

If you find this repo useful for your research, please consider citing our paper:

@inproceedings{park2021improving,
  title={Improving Unsupervised Image Clustering With Robust Learning},
  author={Park, Sungwon and Han, Sungwon and Kim, Sundong and Kim, Danu and Park, Sungkyu and Hong, Seunghoon and Cha, Meeyoung},
  booktitle={CVPR},
  year={2021}
}