next_layer = curr_layer + 2 during deconv, but it is + 1 in the paper about deep-image-analogy HOT 1 OPEN

@gxlcliqi
Hi, thanks for your interest on our project.

For deconvolution, the aim is to deconvolve feature map form layer L to the layer L-1, but what we actually do is to deconvolve it to layer L-2 and then forward it to layer L-1. This is because a target feature map in layer L corresponds to many solutions in layer L-1. By incorporating part of the network (form L-2 to L-1 ) into the optimization, we can constrain the feature map we solved is with the same distribution as real network output of layer L-1.

Another way to constrain the solution is to use LBFGS-B and set proper upper and lower bounds. In this way, we can directly deconvolve it to layer L-1.

from deep-image-analogy.