So I'm sitting here reexamining the prototext file of the densenet model and i'm confused by what i'm seeing.

I realize its a fancy 'convolutional network' but how do you pretrain a model with what looks like a whollllleee bunch of complex programmatic procedures in between input and output ?

I thought each layer was just an activation code, bias and so many connected weights which act as coefficents to the last layers output ?

how is gradient descent performed where you have imaging procedures being performed ?