Transform Augmentation

[ahundt]

This issue is lower priority and postponed until later, brought over from #294.

We should have a way to augment the input transforms, such as shifting or rotating slightly.

Alternately if there are redundant data channels we can provide offsets to other proposed locations in a non-destructive way so models can better learn.