README.md

FATDM - Fairness and Accuracy Transfer by Density Matching

---

Code by Thai-Hoang Pham at Ohio State University.

1. Introduction

This repository contains source code (FATDM) and data for paper "Fairness and Accuracy under Domain Generalization" (ICLR 2023)

FATDM is a Pytorch implementation of the two-stage network (see Figure 1) which achieves fair and accurate predictions in unseen target domain (domain generalization) via invariant representation learnings.

Figure 1: Two-stage learning

2. Installation

FATDM depends on pytorch (CUDA toolkit if use GPU), torchvision, numpy, scipy, tqdm, pandas, scikit-learn. You must have them installed before using FATDM. The simple way to install them is using conda.

    # Using GPU
    $ conda install pytorch torchvision pytorch-cuda=11.7 -c pytorch -c nvidia
    $ conda install numpy scipy tqdm pandas scikit-learn
    # Using CPU
    $ conda install pytorch torchvision cpuonly -c pytorch
    $ conda install numpy scipy tqdm pandas scikit-learn

3. Usage

3.1. Data

The datasets used to train and evaluate FATDM is processed from MIMIC-CXR-JPG (chest radiographs with structured labels) dataset retrieved from PhysioNet. MIMIC-CXR-JPG dataset is restricted-access resource. To access this dataset, user must sign the data use agreement in the project website link.

3.2. Training FATDM

The training script for FATDM are as follows:

• main_stargan.py: Training script for StarGAN to learn density mapping functions.
• main_cyclegan.py: Training script for CycleGAN to learn density mapping functions.
• main_fatdm.py: Training script for FATDM to transfer fairness and accuracy to new domains.

4. References

@inproceedings{
pham2023fairness,
title={Fairness and Accuracy under Domain Generalization},
author={Thai-Hoang Pham and Xueru Zhang and Ping Zhang},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=jBEXnEMdNOL}
}

5. Contact

Thai-Hoang Pham < [email protected] >

Department of Computer Science and Engineering, Ohio State University, USA