In lesson 3 there is a note that the convolutional layers are set to not trainable due to the data being augmented. I have difficulty grasping the arguments for this.
I can imagine that if one has a set of trained convolutional layers with a new set of dense layers, then one does not want to train the conv layers because the new dense layers would influence the conv layers disproportional due to the back propagation.
In a complete new network with no prelearned weights one would like to use augmentation to also create new filters in the conv layers.
These arguments seem different though from the one given. Where am I going amiss?