[cmarschner]

It is not necessary to train things from scratch; you take the largest imagenet model available and fine tune it for the task. This way it reuses much of the lower layers the have seen lots of data.