Closed PyoJunCode closed 1 year ago
In paper Eq(4), loss function of $\mathcal{L}_{LAN}$ is $(1-\alpha)||sg(\tilde{x}_1)-\tilde{x}_2||_1$. But in code, it just using l1 for loss.
# alpha = generate_alpha(BNN) # self.loss = self.criteron(BNN.detach() * (1 - alpha), LAN * (1 - alpha)) self.loss = self.criteron(BNN, LAN)
What loss should I use in LAN stage?
Either will be ok.
Thanks for fast reply, closing this issue
In paper Eq(4), loss function of $\mathcal{L}_{LAN}$ is $(1-\alpha)||sg(\tilde{x}_1)-\tilde{x}_2||_1$. But in code, it just using l1 for loss.
What loss should I use in LAN stage?