Generative Modeling: VAE & GAN Architectures

Generative Modeling: VAE & GAN Architectures

Custom PyTorch implementations investigating latent-space regularization and adversarial training stability.

• Variational Autoencoder (VAE): Designed a custom VAE with a continuous latent bottleneck. Implemented standard reconstruction loss matched with a Kullback-Leibler (KL) Divergence regularization term to enforce a Gaussian distribution on the latent space.
• Generative Adversarial Network (GAN): Implemented Deep Convolutional GANs (DCGAN) and WGAN-GP (Wasserstein GAN with Gradient Penalty) to resolve training instability issues (mode collapse) when generating synthetic images.
• Keywords: PyTorch, Latent Space Interpolation, KL Divergence, Wasserstein Loss, Gradient Penalty.