Performance-comparison-of-GAN-on-cifar-10

Performance comparison of ACGAN, BEGAN, CGAN, DRAGAN, EBGAN, GAN, infoGAN, LSGAN, VAE, WGAN, WGAN_GP on cifar-10

Reference：https://github.com/hwalsuklee/tensorflow-generative-model-collections

The original code is for data MNIST, we changed the network structure to apply to cifar-10 and test Inception Score.

The net structures are almost same.

The following results can be reproduced with command:

python main.py —dataset cifar-10 —gan_type —epoch 60 —batch_size 64

ACGAN

BEGAN

The result is not well,we don’t pay much time to to adjust the super-parameters.

CGAN

DRAGAN

Stable, robust, fast convergent.

EBGAN

The net structure is the same as BEGAN, but collapse.

GAN

infoGAN

LSGAN (Least Squares GAN)

WGAN

Not as well as paper. The net structure is the same as GAN, but converges too slowly.

WGAN_GP

There are total 300 epochs for Discriminator, but only 60 epochs for generator (The same times as other models). Converges slowly.

VAE

Collapsed. We also try to add or subtract bn layers, but it doesn’t work.