Potential for use with a deep kernel learning model

[sgbaird]

A first step would be adapting the example in https://botorch.org/v/0.1.0/tutorials/vae_mnist to using DKL based on the excellent dpax repo (https://github.com/ziatdinovmax/gpax) maybe taking from this Pyro example. Once that's in-place, MNIST can be replaced with xtal2png. What remains puzzling to me is what domain knowledge should be incorporated and how.

In the following example:

Related:

• https://github.com/maka89/Deep-Kernel-GP