WGAN-GP

Example of greyscale WGAN-GP (GAN with Wasserstein Gradient Penalty Loss). To train the GAN run train_WGAN script.

Edit params file to personalise:

• params['save_name'] = path to save directory
• params['data_path'] = dataset path. IT NEEDS TO BE A NPZ FILE [SIZE_DATASETxHEIGHT*WIDTHx1].
• params['buffer_size'] = dataset size (default 60000)
• params['batch_size'] = bacth size (default 32)
• params['noise_dim'] = noise (default 100)
• params['epochs'] = number of epochs (default 3000)
• params['num_exmaple_to_generate'] = number of image generated and saved every 25 epochs to track improvements.
• params['disc_iterations'] = number of iteration of discriminator per iteration of the generator

Result after 700 epochs on a dataset of 6000 images (random height) of abdominal MRIs:

Not physiologically plausible but surprisingly accurate in the detection of some of the structures (spine, galbladder, aorta, IVC).