preprocess and save imagenet for faster training

[kutayeroglu]

given enough storage space,
there could be a mirror of imagenet (training set),
instead of applying transformations on the fly we could read from this dir instead.