by Mihaela Rosca
Designed for education purposes. Please do not distribute without permission.
Questions/Correspondence: [email protected]
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)
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