-
two components:
- world model (vision model CNN + memory RNN)
- controller model
-
vision model is a VAE, compress image to small latent vector z
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).
- create a training script for the VAE
- add image transforms for the VAE training
- create the controller model
- create the memory model