togheppi / dcgan

PyTorch implementation of DCGAN

DCGAN

PyTorch implementation of Deep Convolutional Generative Adversarial Networks (DCGAN)

Network architecture

• Generator

  • hidden layers: Four 4x4 strided convolutional layers (1024, 512, 256, and 128 kernels, respectively) with ReLU
  • output layer: 4x4 strided convolutional layer (4096 nodes = 64x64 size image) with Tanh
  • Batch normalization is used except for output layer

• Discriminator

  • hidden layers: Four 4x4 convolutional layers (128, 256, 512, and 1024 kernels, respectively) with Leaky ReLU
  • output layer: 4x4 convolutional layer (1 node) with Sigmoid
  • Batch normalization is used except for 1st hidden layer & output layer

Generating MNIST dataset

• MNIST image is resized to 64x64 size image

Results

• For learning rate = 0.0002 (Adam optimizer), batch size = 128, # of epochs = 20:

GAN losses	Generated images

Generating CelebA dataset

• 108x108 center region of CelebA image is cropped, and resized to 64x64 size image

Results

• For learning rate = 0.0002 (Adam optimizer), batch size = 128, # of epochs = 20:

GAN losses	Generated images

References

1. https://github.com/yunjey/pytorch-tutorial/blob/master/tutorials/03-advanced/deep_convolutional_gan
2. https://github.com/znxlwm/pytorch-MNIST-CelebA-GAN-DCGAN
3. https://github.com/moono/moo-dl-practice/tree/master/Work-place/DCGAN-MNIST
4. https://github.com/moono/moo-dl-practice/tree/master/Work-place/DCGAN-celebA

dcgan's Issues

why tanh at last layer in generator

hey,
values in input data ie. mnist digits are between 0-255, then why are you applying tanh at the last layer in generator and making it between -1 and 1. For discriminator, input came from traindata and generated data will be on different scale then how will it learn or for that matter how can we even compare and make it learn.

output of generator

Hi,

Thank you for you code open source.
I am currently doing similar things, but I found that the output of generator is 68*68, not 64*64. It is very strange that I can not get 64 after calculations.