MMD-B-Fair: Learning Fair Representations with Statistical Testing

Official implementations of MMD-B-Fair published in AISTATS 2023 by Namrata Deka and Danica J. Sutherland.

Environment

• Python 3.8.10
• PyTorch 1.12.1
• Torchvision 0.13.1
• Wandb 0.13.3

Datasets

• UCI Adult
• COMPAS
• Heritage Health

Training

1. Create a .yml config file and store it in config (see available examples, VERY IMPORTANT). The file should specify all data, model and trainer settings/hyperparameters.

2. Then execute python main.py -v <relative-path-to-yml-from-config> --seed <seed> -w

   Example: python main.py -v eq/adult/lambda_1.yml --seed 42 -w

If you do not wish to sync to wandb while training add the option -m offline and sync anytime later with the wandb sync command.

If seed is not specified it will default to 0.

Trained models are saved in the location specified in experiment.output_location in a subfolder named as per the seed. In wandb, experiments are logged under <config file name>/<seed> in the mmd-b-fair workspace.

Adding more datasets/models/trainers

This repository heavily uses the factory design pattern for increased modularity. To add new datasets, models and/or trainers follow the steps below:

1. Create new data/model/trainer class under the appropriate directories. All trainer classes must inherit BaseTrainer and models must inherit BaseModel.
2. Create corresponding builders for new classes.
3. Register all builder objects to the respective factories in data/data.py, model/model.py and trainer/trainer.py.
4. data.data_key, model.model_key and model.trainer_key in the config files must match the registered factory keys.
5. Specify class-specific arguments in the config file. Example, dataset arguements must go in data.args.