about the weight norm

[zimenglan-sysu-512]

hi @happynear,
i read the code of innerproduct, find that, you only norm the weights in the forward pass. so why it does not need corresponding bp?

thanks.

[happynear]

Because we directly change the weights in the forward propagation. We find the performance is similar with calculating the gradients of the normalization, and it looks much clearer.

You may also try a more correct way as in https://github.com/happynear/NormFace/blob/master/prototxt/combine_softmax_contrastive.prototxt .

[zimenglan-sysu-512]

thanks.