Tutorial on training generative models (VAEs and GANs)

by Mihaela Rosca

Designed for education purposes. Please do not distribute without permission.

Questions/Correspondence: [email protected]

VAE training tutorial

This is a tutorial training a VAE on MNIST and latent analysis

Structure:

• create encoder and decoder, choose latent and model distributions.
• VAE training
• check likelihood for overfitting.
• latent analysis

Your tasks:

• define the decoder distribution
• define the encoder distribution
• define the terms of the loss
• define the samples and reconstruction tensors
• run the KL analysis
• run the latent traversal task
• run the colab with a different number of latent dimensions (and see how that affects the kl analysis)

GAN training tutorial

This is a tutorial training a GAN on MNIST and doing a latent traversal. We will focus on the original GAN, but there are other GANs out there (such as Wasserstein GAN).

Structure:

• basic GAN training
• latent traversal
• try bigger learning rate to see what happens
• optional: add gradient penalty

Your tasks:

• finish up generator definition
• define the discriminator loss
• define the generator loss
• define the discriminator and generator update operations
• run the latent traversal task
• change the learning rates of the discriminator / generator to see if you can get mode collapse
• change the training to do 5 discriminator updates for generator update
• (optional): implement gradient penalties