Document what is being trained in fine-tuning (and multi-class modeling)

[ethanwhite]

See #394. It's a little confusing because all of the layers that we trained are being fine tuned, but (I think) not all of the layers in the underlying Retinanet.