isVoid/simple_GAN_practice

Practice project of Generative Adversarial Network

Generative Adversarial Network

To run the model:

python model.py

Require: see requirements.txt

Folder structure:

• GAN/

  • log/--------Where tensorboard log is saved

  • mnist/------Where mnist dataset is placed

  • model/------Where model and checkpoint file is saved

  • output/-----Where generated images from each epoch is saved

  • model.py---defines GAN model and solver

  • ops.py-----defines ops used by GAN

  • utils.py---miscellaneous helper functions

Noteworthy details

• Generator:

  • fc1024 bn lrelu -> fc128*7*7 bn lrelu -> conv_transpose 14*14*64 bn lrelu -> conv_transpose 28*28*1 tanh

• Discriminator:

  • convf64 bn lrelu -> convf128 bn lrelu -> fc1024 bn lrelu -> fc1 sigmoid

• Dataset:

  • MNIST input normalized to [-1, 1]

• Initialization:

  • Xavier Initialization (Makes a huge difference)

• Activation:

  • leaky relu 0.02

• Used batch_norm layer from tf.contrib.layers