liuguoyou / gan-debiasing

Fair Attribute Classification through Latent Space De-biasing (CVPR 2021)

Fair Attribute Classification through Latent Space De-biasing

Project Page | Paper | Colab Notebook

This repo provides the code for our CVPR 2021 paper "Fair Attribute Classification through Latent Space De-biasing."

@inproceedings{ramaswamy2020debiasing,
author = {Vikram V. Ramaswamy and Sunnie S. Y. Kim and Olga Russakovsky},
title = {Fair Attribute Classification through Latent Space De-biasing},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2021}
}

Our work is featured in Coursera's Generative Adversarial Networks (GANs) Specialization course. Check out the colab notebook linked above for details.

Main experiments

Data processing:

• Download the CelebA dataset and put it in data/celeba.
• Run crop_images.py to crop the aligned & cropped 178×218 images to 128x128.

Baseline:

• Run main.py --experiment baseline to train a standard attribute classifier for each target attribute.

GAN:

• Option 1: Train a (Progressive) GAN on the CelebA training set (162,770 images).
• Option 2: Set pretrained=True in generate_images.py to use a GAN trained by Facebook Research.

Our model:

• Run generate_images.py --experiment orig to sample random latent vectors z and generated images.
• Run get_scores.py to hallucinate labels for the generated images with the trained baseline models.
• Run linear.py to estimate hyperplanes and compute complementary latent vectors z' (our augmentation).
• Run generate_images.py --experiment pair to generate images from z'. Set image output directory and latent vector filename.
• Run main.py --experiment model to train our models (i.e. target classifiers trained with our augmented data).

Extensions of our method

Using domain-dependent hyperplanes:

• Run linear_dom_dep.py to estimate domain-dependent hyperplanes and compute z' with them.
• Run generate_images.py --experiment pair to generate images from z' and train a classifier with these images.

Augmenting real-images with GAN-inversion:

• Train a GAN with an inversion module. We used the in-domain GAN inversion method by Zhu et al.
• Invert CelebA images to latent vectors z_inv.
• Run linear_inv.py to estimate hyperplanes and compute complementary latent vectors z_inv' (our augmentation).
• Run generate_images_inv.py to generate images from z_inv'. This is the only script that requires TensorFlow as the GAN with an inversion module we've trained was implemented in TensorFlow.
• Run main.py --experiment model_inv to train target classifiers trained with data augmented from real images.

Augmenting two protected attributes:

• Run linear_multi_sgd.py to estimate domain-dependent hyperplanes and compute z' with them.
• Run generate_images.py --experiment pair to generate images from z' and train a classifier with these images.

Additional experiments

• full_skew_tests.py: Code for running experiments on the discriminability of attributes.
• linear_underrep.py: Code for estimating hyperplanes with different fractions of positive/negative samples.

Acknowledgements

This work is supported by the National Science Foundation under Grant No. 1763642 and the Princeton First Year Fellowship to SK. We also thank Arvind Narayanan, Deniz Oktay, Angelina Wang, Zeyu Wang, Felix Yu, Sharon Zhang, as well as the Bias in AI reading group for helpful comments and suggestions.