Generative AI

Implementing and optimizing two generative models from scratch for image generation (ICL):

1. Variational Auto Encoder (VAE)
2. Deep Convolutional Generatve Adversarial Network (DCGAN)

Variational Auto Encoder (VAE)

On MNIST dataset.

Reconstructed images

Generated images

Deep Convolutional Generatve Adversarial Network (DCGAN)

On CIFAR10 dataset

Original images

Generated images

Methods

• DCGAN architecture adapted to the 32x32 size of the images
• Gradient penalty term for discriminator
• 5% of labels are flipped
• One-sided label smooting (Real labels set to 0.9 for stronger gradient signals)
• Batch normalization
• Data augmentation
• Learning rate scheduling
• Leaky ReLU in the discriminator (from literature: Radford et al. Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks)