DCGAN with TensorFlow

• TensorFLow implementation of Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks which is a stabilize GAN by certain architectural constraints

• Architecture guidelines for stable Deep Convolutional GANs:

  • Replace any pooling layers with strided convolutions(discriminator) and fractional-strided convolutions(generator)
  • Use BatchNorm in both the generator and the discriminator
  • Remove fully connected hidden layers for deeper architectures
  • Use Relu activation in generator for all layers except for the output, which uses Tanh
  • Use LeakyReLU activation in the discriminator for all layers

• Generator architecture

  

• Usage

  • Download dataset CelebA from here
  • Train the model: python main.py

• Training details

  

  

  

• To ensure that d_loss do not go to zero during the training, I run generator 5 times every discriminator updating. Although the d_loss will not be close to 0, the quality of the generated images is very bad.

• Because I do not have GPUs to train all the training set now, so I just test the code with several samples on my MacBook Air. More training results will be updated later.

• Result

  • After 5th epoch

    

  • After 10th epoch

    

  • After 20th epoch

    

• The code borrowed from here