generative_adversarial

• Various Generative Adversarial Networks using tensorflow
• Main Reference: https://github.com/shekkizh/EBGAN.tensorflow'
  • copied main optimizer code and most setup codes
  • refactored model build-up and variable-maintaing codes
  • modified some calculations to follow the details of the original paper

EBGAN

• Generate face images based on CelebA data

DCGAN

• Generate face images based on CelebA data

• Under epoch #1

||||

• Under epoch #2

||||

• Under epoch #3

||||

• Cherry-picked results

||||

InfoGAN

• Generate hand-written number images based on mnist data
• Need to fine-tune parameters
  • It is delicate to make InfoGan converge. It is easy to make simple GAN converge. But fine-tuning is needed when applying latent codes.
• Use simple adaptive generator optimization.
  • accuracy 0.4
• Use categorical latent code only (continuous latent code)

Result

• After 14 epochs

• After 15 epochs

• comments
  • '8' to class #0, '1' to class #1, '3' to class #2, and so on.
  • '7', '9', '4' are not distinguishable in this result, this means that encoding of '7', '9', '4' into distinct codes failed.
    • But network can encode these 3 numbers into separate codes in another trial.
    • Whenever I try it, different result comes. Sometimes I succed all, sometime fail 2 numbers.

Adversarial Autoencoders