Implementation of CapsGAN based on Raeid Saqur et al paper.
The purpose of the repository is to be able to generate high quality faces. The dataset used for the purpose is fhhq dataset. The model was trained on 128x128 thumbnail images.
The model was trained on 56k training images for 160 epochs (1 epoch is 1 pass over the entire train dataset).
Evaluation is based on the FID score. For baseline model DCGAN was used. A DCGAN trained for the same time provides FID score of 72.34 while the CapsGAN provides a much better FID score at 31.25. The FID score was calculated on 50k test images.
Models: DCGAN discriminator, DCGAN generator, CapsGAN discriminator and CapsGAN generator
DCGAN:
CapsGAN:
