In your paper you write that g is normalized, but you don't describe how. You write that w it is normalized with the l1 norm. That works fine in my code as well. But how did you normalize g?
Because when I use l1 normalization (time axis) for g, my adversary has a hard time increasing the loss. When I don't use normalization, only the train SR increases, but not as much as yours in the paper.
Hi Louis, sorry to bother you again.
In your paper you write that g is normalized, but you don't describe how. You write that w it is normalized with the l1 norm. That works fine in my code as well. But how did you normalize g?
Because when I use l1 normalization (time axis) for g, my adversary has a hard time increasing the loss. When I don't use normalization, only the train SR increases, but not as much as yours in the paper.
Regards