Implementation of the bias-to-text (B2T) algorithm described in "Bias-to-Text: Debiasing Unknown Visual Biases through Language Interpretation." B2T identifies and mitigates visual biases in image classifiers and text-to-image generative models using language descriptions.

See the link for more detailed results in bias discovery and debiasing.

Download datasets.

• CelebA
• Waterbirds (direct downloadable link), formed from Caltech-UCSD Birds 200 Places

Clone our repository.

$ git clone https://github.com/Erena-Kim/b2t.git

Run below to create virtual environment for b2t and install all prerequisites.

$ pip install pipenv
$ pipenv --python 3.8
$ pipenv install

To run our code, you need to place datasets and model checkpoints to right directory.
You can download the ClipCap pretrained model here and place the model to [root_dir]/function. (Note that our paper uses the model that trained on Conceptual Captions)
The main point of entry to the code is b2t.py

• dataset: celeba or waterbird
• model: download pretrained checkpoints of CelebA and Waterbirds here and put them into [root_dir]/model
  • best_model_CelebA_erm.pth
  • best_model_CelebA_dro.pth
  • best_model_Waterbirds_erm.pth
  • best_model_Waterbirds_dro.pth
• extract_caption: True or False
  • If set True, automatically generate [root_dir]/data/[dataset]/caption/ and store extracted captions there
• save_result: True or False
  • If set True, automatically generate [root_dir]/diff/ and store csv file of results there

Our code expects the following files/folders in the [root_dir]/data/celebA directory:

• data/list_eval_partition.csv
• data/list_attr_celeba.csv
• data/image_align_celeba/

A sample command to run b2t on CelebA with pretrained erm model is:

$ python b2t.py --dataset celeba --model best_model_CelebA_erm.pth

Our code expects the following files/folders in the [root_dir]/data/cub directory:

• data/waterbird_complete95_forest2water2/

A sample command to run b2t on Waterbirds with pretrained erm model is:

$ python b2t.py --dataset waterbird --model best_model_Waterbirds_erm.pth

To reproduce the debiasing classifier experiments, see b2t_debias.

To reproduce the diffusion model experiments, see b2t_diffusion.