DCGAN in Pytorch

PyTorch implementation of Deep Convolutional Generative Adversarial Networks (DCGAN), which is a stabilize Generative Adversarial Networks. The origin code can be found here.

Network architecture

• Generator

  • input: a vector with z_size.
  • hidden layers: Four 4x4 transposed convolutional layers (1024, 512, 256, and 128 kernels, respectively) with ReLU
  • output layer: 4x4 transposed convolutional layer (channel_size kernels, 4096 nodes = 64x64 size image) with Tanh.
  • BatchNormalization is used except for output layer.

• Discriminator

  • input: a vector with channel_size * image_size * image_size.
  • hidden layers: Four 4x4 convolutional layers (128, 256, 512, and 1024 kernels, respectively) with LeakyReLU (negative slope is 0.2).
  • output layer: 4x4 convolutional layer (1 node) with Sigmoid.
  • BatchNormalization is used except for 1st hidden layer and output layer.

All of the transposed convolutional layer and convolutional layer are initilized by a normal distribution with 0.0 mean and 0.02 std.

MNIST dataset

• For MNIST image, the channel_size is 1 and image_size is 64.

Results

• The learning rate is 0.0002, batch size is 128 and the optimizer is Adam.

Gnerated Images	DCGAN Loss

CelebA dataset

• For CelebA image, the channel_size is 3 and image_size is 180 x 180, which has been aligned and cropped. And then, it will be resized to 64 x 64.

Results

• The learning rate is 0.0002, batch size is 128 and the optimizer is Adam.

Gnerated Images	DCGAN Loss

LSUN dataset

• For LSUN image, the channel_size is 3. And then, it will be resized to 64 x 64.

Results

• The learning rate is 0.0002, batch size is 1024 and the optimizer is Adam.

Gnerated Images	DCGAN Loss

References

1. https://github.com/soumith/dcgan.torch
2. https://github.com/znxlwm/pytorch-MNIST-CelebA-GAN-DCGAN
3. https://github.com/togheppi/DCGAN
4. https://github.com/carpedm20/DCGAN-tensorflow