search

zbh2047 / L_inf-dist-net-v2

[ICLR 2022] Training L_inf-dist-net with faster acceleration and better training strategies
22 stars 4 forks source link

关于模型的两个输出 #1

Closed flyleeee closed 2 years ago

flyleeee commented 2 years ago

请问return x self.scalar, margin / (2 eps)中的后者是什么

zbh2047 commented 2 years ago

你好！你可以看下论文里的loss，前者用于cross entropy loss，后者用于clipped hinge loss，除以2 eps这里的2eps=theta对应hinge threshold超参