fine-tune speedup by feature extraction

[Alex Orloff]

Is it possible to reduce network from "image -> all layers -> loss_layer" to "feature extracted -> last(fine-tuned) layers -> loss_layer"?

It could be useful to speedup finetuning in case when first layers are "freezed".

So we can simply make forward pass to all training/vlidation images for those freezed layers only once.

thanks,