TreB1eN / InsightFace_Pytorch

Pytorch0.4.1 codes for InsightFace
MIT License
1.72k stars 418 forks source link

a question about forward propagation in head model #51

Closed lilinwei777 closed 5 years ago

lilinwei777 commented 5 years ago

我理解的论文中的意思是 Additive Angular Margin punishment 是对θyi+m,而不是对 jε1-n θj +m。但很显然你的Arcface 的forward 里是对所有的θj 都加m了。这样做对intra_class 的收紧 就没有什么作用了吧