Hello! I read some things about this but I am still confused. So assuming we keep the initial layers frozen, why precompute = True doesn’t do anything with data augmentation. The way i thought it works is like this: you give a new image to the network, the first several layers (which are frozen) will give an output which is used further as an input for the last few layers which are not frozen and the final output (from the last unfrozen layer) tells you what the image is. I don’t understand why a rotated image would give the same result if precompute = True. Don’t you still update the final unfrozen layers? Like, isn’t this augmented image a totally new image, from the point of view of the last layers? Can someone explain to me what precompute = True actually does? Thank you!