DCGAN

Deep Convolutional Generative Adversarial Network (DCGAN) inspired by PyTorch-GAN and DCGAN Tutorial, trained to generate 64x64 anime faces from 20,000 training images over 500 epochs. Mixed precision training was used to speed up training times.

Generator Architecture

• Linear (Latent Dim -> (Img size / 8) ** 2)
• BatchNorm2d
• ConvTranspose2d (256 -> 128)
• BatchNorm2d
• Leaky ReLU
• ConvTranspose2d (128 -> 64)
• BatchNorm2d
• Leaky ReLU
• ConvTranspose2d (64 -> Num Colour Channels)
• Tanh

Discriminator Architecture

• Conv2d(3 -> 32)
• Leaky ReLU
• Dropout(0.25)
• Conv2d (32 -> 64)
• Leaky ReLU
• Dropout(0.25)
• BatchNorm2d
• Conv2d (64 -> 256)
• Leaky ReLU
• Dropout(0.25)
• BatchNorm2d
• Conv2d (256 -> 512)
• Leaky ReLU
• Dropout(0.25)
• BatchNorm2d
• Linear (512 -> 1)
• Sigmoid (unapplied in example as loss function is BCEWithLogitsLoss)