Bias Mimicking: A simple sampling approach for Bias Mitigation

Official Pytorch implementation of Bias Mimicking: A simple sampling approach for Bias Mitigation.

Setup

Set up conda environment

conda create -n bias_mimicking python=3.8
conda activate bias_mimicking

Install packages

• pytorch=1.10.1
• scipy
• tqdm
• scikit-learn

Prepare dataset.

• CelebA
  Download CelebA dataset under data/celeba

• UTKFace
  Download UTKFace dataset under data/utk_face

• CIFAR10
  Download CIFAR10 dataset under data/cifar10

As discussed in the paper, we train on subsampled versions of CelebA and UTKFace. The information required to reproduce the the splits are in data/[DATASET]/pickles. The code will automatically load the right splits.

Train.

From the main directory, run:

python train_[DATASET]/train_[DATASET]_[METHOD].py --seed [SEED]

To train our method on celeba, run:

python train_celeba/train_celeba_bm.py --mode [none/us/uw/os] --seed [SEED]

where mode refers to whether the distribution is left as is/undersampled/upweighted/oversampled when training the predictive linear layer.

Acknowledgements

The code for non sampling methods builds on this work. Furthermore, the code for GroupDRO is obtained from this work

Citation

If you find this repository useful please give it a star and cite as follows! :) :

    @inproceedings{qraitem2023bias,
      title={Bias Mimicking: A Simple Sampling Approach for Bias Mitigation},
      author={Qraitem, Maan and Saenko, Kate and Plummer, Bryan A},
      booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
      pages={20311--20320},
      year={2023}
    }