Currently train.py is overloaded with many independent features.
Understanding the code and what will be the result of a training
call requires untangling the entire script. It's also an error
prone UX. This is a first step at separating independent parts
into their own scripts.