PengxinZeng/2023-CVPR-FCMI

PyTorch implementation for Deep Fair Clustering via Maximizing and Minimizing Mutual Information: Theory, Algorithm and Metric (FCMI), CVPR 2023.

PyTorch implementation for

Deep Fair Clustering via Maximizing and Minimizing Mutual Information: Theory, Algorithm and Metric

CVPR 2023

[Paper] [Discussion] [More Information]

Introduction

FCMI framework

Requirements

• Python 3.9
• PyTorch 1.11.0
• faiss

conda install -c pytorch faiss-gpu

Datasets

Color Reverse MNIST, Office-31, MTFL, and MNIST-USPS

We follow DFC to obtain datasets.

HAR

We follow DFDC to obtain dataset.

Training

Modify the ./Utils/PathPresettingOperator.get_dataset_path, then train the model(s):

# Color Reverse MNIST
python main.py --dataset ReverseMNIST --seed 0  
  
# Office-31  
python main.py --dataset Office --seed 0  
  
# MTFL  
python main.py --dataset MTFL --seed 9116  
  
# HAR  
python main.py --dataset HAR --seed 9116  
  
# MNIST-USPS  
python main.py --dataset MNISTUSPS --seed 9116  

Model Zoo

The pre-trained models are available here:

Dataset	Model	Results
Color Reverse MNIST	Model	Results
Office-31	Model	Results
MTFL	Model	Results
HAR	Model	Results
MNIST-USPS	Model	Results

Download the models, then:

python main.py --dataset dataset --seed seed --resume PathToYourModel

Experiment Results:

Citation

If FCMI is useful for your research, please cite the following paper:

@inproceedings{zeng2023deep,
  title={Deep Fair Clustering via Maximizing and Minimizing Mutual Information: Theory, Algorithm and Metric},
  author={Zeng, Pengxin and Li, Yunfan and Hu, Peng and Peng, Dezhong and Lv, Jiancheng and Peng, Xi},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={23986--23995},
  year={2023}
}