DivClust: Controlling Diversity in Deep Clustering

This is the official implementation of our CVPR 23' paper DivClust: Controlling Diversity in Deep Clustering.

Overview

Overview of DivClust

DivClust is a method for controlling inter-clustering diversity in deep clustering frameworks. It consists of a novel loss that can be incorporated in most modern deep clustering frameworks in a straightforward way during their training, and which allows the user to specify their desired degree of inter-clustering diversity, which is then enforced in the form of an upper bound threshold.

As shown in our paper, DivClust adds minimal computational overhead and can significantly increase their performance with the use of off-the-shelf consensus clustering algorithms.

Illustration of the inter-clustering similarity between 20 learned clusterings for various diversity targets. We report the diversity target D^T and measured diversity D^R, which are expressed in the NMI metric.

Usage

To create an environment, execute the following commands:

conda create -n divclust_env python=3.8 &&
conda activate divclust_env &&
conda config --add channels conda-forge &&
conda install pytorch==1.12.1 torchvision torchaudio cudatoolkit=11.3 -c pytorch &&
pip install scipy wandb PyYAML scikit-learn termcolor matplotlib opencv-contrib-python

To run experiments, this repository uses the .yaml files in the configs directory. For each run, arguments are read from configs/main_config.yaml, and are then supplemented (and in case of conflicts overwritten) by the secondary config files in that dir, which is identified with the preset argument. Arguments can further be provided in bash.

For example, to run an experiment with the CC deep clustering framework with 2 clusterings with a diversity target of $D^T=0.8$, one would run the following command:

python main.py --preset cc_cifar10 --clusterings 2 --NMI_target 0.8

See also

Original implementation of CC

Original implementation of PICA

Original implementation of IIC

Code for the consensus clustering algorithm SCCBG

We thank the authors of the above for making their code public.

Citation

If you use this repository, please cite:

@inproceedings{divclust2023,
  title={DivClust: Controlling Diversity in Deep Clustering},
  author={Metaxas, Ioannis Maniadis and Tzimiropoulos, Georgios and Patras, Ioannis},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={3418--3428},
  year={2023}
}

License

This project is licensed under the MIT License

Acknowledgement

This work was supported by the EU H2020 AI4Media No. 951911 project.