601054 (1) [Avatar] Offline
s5.3.1 mentions two approaches to training a new FC layer on top of a pretrained CNN: (1) run samples through the frozen layers once and use this to train the new FC layer only (2) use data augmentation and run the samples through the entire CNN, including the frozen layers, in each epoch.

The book states that approach (1) cannot be combined with data augmentation.

My question is why not? Can you not augment the data then pass it through the frozen CNN once? Is this an issue to do with saving the labels from the augmented data and is there a way around this?

Basically, I'd like to combine the two approaches to use augmented data but only have to run samples through the frozen layers once.