Great writeup! I realize the ConvARC part went over my head in my quick read of the paper. PyTorch and/or your coding style make the code an easier read, too, since I hadn’t seen Theano of the original before. The animations and results made it Facebookable.
I’m looking forward to feeding it diptychs with pics 299x299 for example (training an inception v3 siamese network for the same purpose now). Any ideas how to scale? You can see examples at phobrain.com.