Normalize embedding and weight for imprinted weights training?

Question

I am trying to use imprinted weights for a lowshot classification convolutional network, from this paper: Low-Shot Learning with Imprinted Weights, where it states that

&