- Simplify the NBVAE aspect. Return separate mean, var, mu, etc. Take the class loss out of the NBVAE and handle it separately. Scaled reconstruction loss is very much a win.
- Learn to use PyOT to map between data types, then use contrastive learning.
- Think of contrastive learning generalizations that take into account relative properties.