Code for the paper Hard Regularization to Prevent Deep Online Clustering Collapse without Data Augmentation.
Create a conda environment (tested with python 3.12), and run the script that installs the needed libraries:
. install_requirements.sh.
The experiments in the paper can be replicated with the command
python train.py -d {dataset-name} --n_epochs 10
The dataset names are 'c10', 'c100', 'fashmnist', 'stl' and 'realdisp', meaning Cifar10, Cifar100, Fashion MNIST, STL and RealDisp, respectively. The comparison models can be run by adding the flags '--var', '--ent', '--sinkhorn'' or '--ckm'', which correspond to the names used for these models in the paper.
If you use or refer to this in your work, please cite
title={Hard Regularization to Prevent Deep Online Clustering Collapse without Data Augmentation},
author={Mahon, Louis and Lukasiewicz, Thomas},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={38},
number={13},
pages={14281--14288},
year={2024}
}
Any questions or problems with the code, you can contact lmahon@ed.ec.uk.