| Autoencoders and VAE |
Autoencoders, latent variable models, the ELBO and the reparameterization trick, beta-VAEs, perceptual losses, and VQ-VAE |
Open |
| Normalizing Flow Models |
The change of variables formula, coupling layers, NICE, RealNVP and Glow, and flow matching, the current state of the art |
Open |
| Lab 1: VAE, Perceptual Loss and VQ-VAE |
Train a VAE, a VAE with a perceptual loss and a VQ-VAE from scratch on anime faces, and compare what each generates |
Notebook Colab |
| Lab 2: Normalizing Flows on CIFAR-10 |
Planar flows implemented as in Rezende and Mohamed (2015): the change of variables, the log-determinant, and the free-energy bound |
Notebook Colab |
| Optional: Autoregressive Models |
FVSBN, NADE and RNADE, maximum likelihood, RNNs as autoregressive models, PixelRNN, PixelCNN and WaveNet |
Open |
| Optional: VAE for Anime Portraits |
Exercise: train a VAE on anime portraits and study the effect of the latent dimension and of beta |
Notebook Colab |
| Optional: RealNVP on Autoencoder Embeddings (MNIST) |
Train an autoencoder, fit a RealNVP flow to its embeddings, and decode the flow's samples into new digits |
Notebook Colab |
| Optional: RealNVP on Autoencoder Embeddings (Pets) |
The same two-stage pipeline on Oxford Pets images |
Notebook Colab |
