This folder contains the implementation of Double InfoGAN for Contrastive Analysis.
It is originally a clone of the repo MM-cVAE
It uses pytorh 1.9 together with pytorch lightning.
In folder preprocess you can find all the codes to create the datasets (celeba, cifar+mnist, etc...). The codes create datasets for training, evaluation and tests, and the created files must be moved in /datasets folder. To run the preprocess, run the following code with
python -m preprocess.preprocess_datasetName
The training code is run_experiment.py. To launch the code, select a config file and chose a directory to store the results and run :
python run_experiment.py --config_file config/config_file.yml --logdir ./results_directory/
In the config file you can modify every hyperparameter of the training.
Evaluation codes is also provide in /eval directory. Codes are provided to measure the target class separation for celeba and cifar/mnist dataset, as well as fvae metrics for mnist/dsprite dataset.
Tu run the code (here target class separation for infogan), chose the directory you wan to evaluate inside the code, then launch :
python -m eval.eval_target_sep_celeba_infogan