In BinaryConnect ( https://arxiv.org/abs/1511.00363 ) authors always have two sets of weights. Firstly, this is a copy with a high degree of accuracy, on which we apply gradients and update weights, and secondly, binarized (quantized) weights used to calculate output and errors. I can not reproduce this function in Tensorflow. Does anyone have an idea of how I can do network training with two sets of weights that change based on the operation in one Tensorflow graph.
source
share