Hi,
I was wondering if someone has tried to add additional layers to a pre-trained CNN initializing with the identity convolution.
On one hand seems like a “good” idea as final features would come from a trained layer but on it other it seems prejudicial as kernel is almost identical.