Explore how data augmentation affects training #587
Description
#581 corrected the likelihood of data augmentation being performed, which means that most training was carried out with the "wrong" likelihood. We should explore how data augmentation affects training, and what a sensible default likelihood is.