Created by: lcy-seso

We are working on NMT project.

Currently, we have the norm_op, however, its name is quite confusing, because it is cross channel norm, not a general purpose norm operator.

ConvS2S need weight normalization, which in essence is the l2 norm along a specified axis. Currently, we do not have any operator to do this.

A better way is to implement an operator not only for weight normalization but also for general purpose normalization.

Such an operator accepts a Tensor with an arbitrary rank and a number specifying along which axis to perform a chosen norm, for example, infinite norm, l1, l2, p-norm and so on.