samples.md

Training on the CelebA dataset

GANs

• DCGAN (sigmoid activation on the last layer)
• MAD-GAN
• WGAN and GoGAN
• BEGAN

Datasets

• CelebA (http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html)

Comments

• All of the GANs are trained with a fixed LR=1-e5 for 100k iterations.
• Disclaimer: no hyper-parameter search has been done yet.

Generated samples

DCGAN

After 4k, 10k, 50k iterations

Samples from the last iterations

MADGAN

After 4k, 10k, 50k iterations

Samples from the last iterations

WGAN

After 4k, 10k, 50k iterations

Samples from the last iterations

GoGAN

Stage 1 samples after 4k, 10k, 50k iterations

Stage 2 samples from the last iterations

BEGAN

After 4k, 10k, 50k iterations

Samples from the last iterations