There's also some confusion over the batch size. If it doesn't work, we probably need to revise our understanding of the batch size in different contexts.
Also, we don't have any code that actually does the training itself either yet.
