Fair-Vision

This repository consists the source coude for the project fairness in vision conducted at ETH SRI lab.

Data prepration

First, put the CelebA data inside ./data/celeba/. Then run the following scripts to get the LMDB dataset for Celeba. The data will be saved in ./data/celeba/celeba_lmdb.

python scripts/convert_celeba_lmdb.py --split train
python scripts/convert_celeba_lmdb.py --split valid
python scripts/convert_celeba_lmdb.py --split test

Training the VAE

For training the VAE, you can use the train_vae() function inside vae.py. vae.py also contains a number of different functinos for visualizing the vae outputs. The model should be saved in ./final-models/

Training the encoder

To train the base encoder, run

python3 train_encoder.py --base true

To train the encoder with LCIFR, run

python3 train_encoder.py --base false

Training the encoder

To train the base classifier, run

python3 train_classifier.py --robust false

To train the classifier with adversarial training, run

python3 train_classifier.py --robust true

Producing the certification resluts

To get the certification results for the base model as well as the fair model, run

python3 certify.py --robust false
python3 certify.py --robust true

The whole end to end pipeline can be run using

sh run.sh

Changing the experiments

The experiment configurations can be modified using metadata.json.