Addressing Attribute Bias with Adversarial Support-Matching

Code for the paper Addressing Attribute Bias with Adversarial Support-Matching.

Requires Python 3.10+.

Installation

One of the dependencies is PyTorch. If your system is not compatible with the default torch installation (for example if you require a specific CUDA installation), then please install this from pytorch.org as required. If so, it's recommended that you do this first.

You can install this package, with dependencies, with pip install ..

Running Experiments

ACS

This dataset will be downloaded on first use.

python -m src.run.supmatch +experiment=acs/fcn

NICO++

This dataset has to be downloaded separately.

python -m src.run.supmatch +experiment=nicopp/rn50/pretrained_enc ds.root=/path/to/dataset

Colored MNIST

This dataset will be downloaded automatically.

python -m src.run.supmatch +experiment=cmnist/2d2c labeller=gt

CelebA

The code will try to download this, but the download quota is often saturated, so it might not work immediately.

Without smiling males

python -m src.run.supmatch +experiment=celeba/sm/pt split=celeba/artifact/no_smiling_males

Without smiling females

python -m src.run.supmatch +experiment=celeba/sm/pt split=celeba/artifact/no_smiling_females

Without unsmiling males

python -m src.run.supmatch +experiment=celeba/sm/pt split=celeba/artifact/no_unsmiling_males

Without unsmiling females

python -m src.run.supmatch +experiment=celeba/sm/pt split=celeba/artifact/no_unsmiling_females