Trained on large network #39
Description
Hello, I've always been so confused about how could BWN and XNOR-net be trained on large neural network such as vgg-16 or resent-50?
I find it quite difficult to change all the layers into binarized layer at one time, because there is often the gradient explosion or the gradient diminish happens during training time. And I think that change a layer at one time may be able to solve the problem. But is there any approach to deal with it without having to train the binarized layer separately?