This repository is reproduction of the work done in paper: Towards Fairness in Visual Recognition: Effective Strategies for Bias Mitigation

Original Authors: Zeyu Wang, Klint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, Olga Russakovsky

Requirements

• Python 3.6+
• PyTorch 1.0+
• h5py
• tensorboardX

Data Preparation

First download and unzip the CIFAR-10 and CINIC-10 by running the script download.sh

Then manually download the CelebA dataset, put Anno into data/celeba/Anno, Eval into data/celeba/Eval, put all align and cropped images to data/celeba/images

Run the preprocess_data.py to generate data for all experiments (this step involves creating h5py file for CelebA images, so would take some time 1~2 hours)

Changes

• In the original code, the certificate to download cinic data had expired. This issue has been resolved here by replacing it with chained certificate.
• There are exisitng issues in getting the complete CelebA dataset bcause of their size. This issue will be resolved in next update.

Run Experiments

To conduct experiments, run main.py with corresponding arguments (experiment specifies which experiment to run, experiment_name specifies a name to this experiment for saving the model and result). For example:

python main.py --experiment cifar_color --experiment_name e1 --random_seed 1

After running, the experiment result will be saved under record/experiment/experiment_name