Training

One should not change anything in train.py prior to training. For that purpose, adjust all parameters in train_config.json.

• num_epochs - number of epochs for training.
• batch_size - size of batch during training.
• img_size - size of images for both discriminator and generator.
• checkpoint_freq - at how many epochs to dump GAN state.
• log_freq - at how many batches per epoch to dump log.
• imagery_freq - at how many batches per poech to save intermediate results.
• n_power_iterations - number of iterations used for calculating l₂ norm of weight matrix.
• type - kind of GAN architecture. Either DCGAN or SN_DCGAN.

CelebA will be downloaded in folder ./data in case it's not already. Log and images will be saved in proper ./runs subdirectory.

Testing

• is.py - calculating Inception score. See comments in the file for more info.
• interpolation.py - interpolate between given images proper number of times. Saves images to ./data/interpolation.
• inference.py- generated given number of fake images for provided pre-trained model.

Checkpoints

All pretrained models are in ./runs directory. One can check training parameters in proper train_config.json files.