DCGANs

Overview

From-scratch replication of DCGAN (Deep Convolutional GAN). DCGANs introduced convolutional architectures for both generator and discriminator, replacing fully-connected layers and enabling stable GAN training on natural images. Trained on both CelebA (faces) and CIFAR-10 (objects). Based on Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks (Radford et al., 2016).

Architecture

Generator: Random noise → ConvTranspose2d upsampling layers → generated image Discriminator: Image → Conv2d downsampling layers → real/fake logit

• BatchNorm in generator, no pooling
• LeakyReLU in discriminator, ReLU in generator
• Tanh output activation

Training

Pretrained weights available on Google Drive.

Paper

Unsupervised Representation Learning with DCGANs — Radford et al., 2016