kkteru/unbiased-gender-classifier

Reducing bias in classifiers by disentangling the input attributes

Reducing bias in classifiers by disentangling the input attributes

Getting started

• Navigate into the data folder and run source get_started.sh. This will download, extract and preprocess the alligned images from UTKFaces dataset.
• Navigate to models folder and run source get_autoencoders.sh to download pretrained vanilla autoencoder and adversarially trained autoencoder.
• The gender_clssifier.py is the main file to train/evaluate the different classifiers.

Training the gender classifier

python gender_classifier.py --help would list all the available parameters along with their default value. The defaults should work out of the box for classifier with vanilla AE. python gender_classifier.py --remove_race should train a classifier with features from adversarially trained autoencoder.

Evaluating the gender classifiers

python gender_classifier.py --eval should start the evalutaion with features derived from vanilla autoencoder. python gender_classifier.py --eval --remove_race should start the evalutaion with features derived from adversarially trained autoencoder.