Echoes: Unsupervised Debiasing via Pseudo-bias Labeling in an Echo Chamber

Setup

pip install -r requirements.txt

Datasets

CelebA

• Generate CelebA with the following script.

bash ./scripts/prepare_dataset/create_celeba.sh

UrbanCars

• Generate UrbanCars dataset with the following repository.
• Locate it under the path ./data.

BFFHQ

• Download BFFHQ from this link and locate it under the path ./data.
• Unzip the BFFHQ.zip

Training with Echoes

• You can train the model of Echoes using the following commands.
• [Optional] Turn on --wandb to use wandb for logging.
  • Please complete entity information before using WandB, otherwise you will get an error.

CelebA

# Smiling
python trainers/lancher.py --dataset=celeba --target_id=31 --method=echoes --t_error=0.5 --alpha=0.5 --seed=1 --amp
# Narrow Eyes
python trainers/lancher.py --dataset=celeba --target_id=23 --method=echoes --t_error=0.5 --alpha=0.5 --seed=1 --amp
# Arched Eyebrows
python trainers/lancher.py --dataset=celeba --target_id=1 --method=echoes --t_error=0.5 --alpha=0.5 --seed=1 --amp

UrbanCars

python trainers/lancher.py --dataset=urbancars --method=echoes --t_error=0.3 --alpha=0.5 --seed=1 --amp

BFFHQ

# 0.5 pct
python single_bias_trainers/lancher.py --dataset=bffhq_0.5pct --method=echoes --t_error=0.5 --alpha=0.5 --seed=1 --amp
# 1.0 pct
python single_bias_trainers/lancher.py --dataset=bffhq_1pct --method=echoes --t_error=0.5 --alpha=0.5 --seed=1 --amp
# 2.0 pct
python single_bias_trainers/lancher.py --dataset=bffhq_2pct --method=echoes --t_error=0.5 --alpha=0.5 --seed=1 --amp

Acknowledgments

• https://github.com/kakaoenterprise/BiasEnsemble
• https://github.com/facebookresearch/Whac-A-Mole