To prepare the settings for each experiment, a config file with name format config_[config_id].py needs to be added to configs directory. As shown in the provided template, the the path to the dataset, and also all hyper-parameters can be set in this file. Later on, to run the main python scripts, the id of the config file will be passed to specifiy the experiment configuration.
To train the model using subgroup-aware VAE, use the following command:
python train --config_id [config_id]
To generate latent representation of the input data:
python generate-latent-reps.py --config_id [config_id]
To mine subgroups based on the latent representation:
python mine_subgroups.py --config_id [config_id]
To generate latent traversal images:
python traverse_latents.py --config_id [config_id]
To generate average image of whole population, and each subgroup:
python visualize_sgds.py --config_id [config_id]