Hi, I had the idea to substitute the convolution operator with a FC operator. What I mean is not just to apply a FC layer to the whole image but to a subset of it as the convolution works but instead of just dot product of the weights and the value apply a proper shallow FC layer to catch better relations. How can I start tackling the problem of implementing it?
Thank you a lot.
I sense that you are talking about an idea similar to “Network in Network” paper. One way of implementing it can be found here: https://github.com/yangqiongyongyu/Network-In-Network-Pytorch/blob/master/models/nin.py . They realize the fully connected layer in terms of 1x1 convolution layer.