Conditional DCGAN

PyTorch implementation of Conditional Deep Convolutional Generative Adversarial Networks (cDCGAN)

Generating MNIST dataset

• MNIST image is resized to 32x32 size image

Network architecture

• Generator

  • hidden layers: Three 4x4 strided convolutional layers (512, 256, and 128 kernels, respectively) with ReLU
  • output layer: 4x4 strided convolutional layer (1024 nodes = 32x32 size image) with Tanh
  • Batch normalization is used except for output layer

• Discriminator

  • hidden layers: Three 4x4 convolutional layers (128, 256, and 512 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

Results

• Adam optimizer is used. Learning rate for generator = 0.0002 and 0.0001 for discriminator, batch size = 128, # of epochs = 20:
• Loss values are plotted using Tensorboard in PyTorch.

GAN losses	Generated images

Generating CelebA dataset

• Center region of CelebA image is cropped (108x108 for male/female, 150x150 for black/brown hair) and resized to 64x64 size image

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

Results

• Adam optimizer is used. Learning rate = 0.0002 both for generator and discriminator, batch size = 128, # of epochs = 20:
• Images are generated with fixed noise and varying gender label (1st, 3rd rows: female / 2nd, 4th rows: male)

GAN losses	Generated images

• Images are generated with fixed noise and varying hair color label (1st, 3rd rows: brown hair / 2nd, 4th rows: black hair)
• Only 87,772 images (black hair: 47,336 / brown hair: 40,436) are sampled from original CelebA data for training.

GAN losses	Generated images

• Generated images varying latent variable
  • Odd rows: female / Even rows: male
  • Odd rows: brown hair / Even rows: black hair

References

1. https://wiseodd.github.io/techblog/2016/12/24/conditional-gan-tensorflow/
2. https://github.com/znxlwm/pytorch-MNIST-CelebA-cGAN-cDCGAN
3. https://github.com/moono/moo-dl-practice/tree/master/Work-place/CDCGAN-MNIST
4. https://github.com/moono/moo-dl-practice/tree/master/Work-place/CDCGAN-celebA