World Models

• two components:

  1. world model (vision model CNN + memory RNN)
  2. controller model

• vision model is a VAE, compress image to small latent vector z

Dataset Generation

You can use the generate_vae_data.py script to generate a dataset for training the VAE model. Before you do this, make sure you have enough space on your disk (~260 GB).

Training

After you have generated your training dataset, you can use it to train your Variational Autoencoder Model.

Todos

• add image transforms for the VAE training