David Madras, Elliot Creager, Toni Pitassi, Richard Zemel https://arxiv.org/abs/1802.06309

Code represents equal contribution with David Madras. Thanks to Jake Snell and James Lucas for contributing the experiment sweep code.

mkdir ~/venv 
python3 -m venv ~/venv/laftr

where python3 points to python 3.6.X. Then

source ~/venv/laftr/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

or

pip install -r requirements-gpu.txt

for GPU support

source simple_example.sh

The bash script first trains LAFTR and then evaluates by training a naive classifier on the LAFTR representations (encoder outputs). See the paper for further details.

python src/generate_sweep.py sweeps/small_sweep_adult/sweep.json
source sweeps/small_sweep_adult/command.sh

The above script is a small sweep which only trains for a few epochs. It is basically just for making sure everything runs smoothly. For a bigger sweep call src/generate_sweep with sweeps/full_sweep_adult/sweep.json, or design your own sweep config.

The (post-processed) adult dataset is provided in data/adult/adult.npz