Values of m for SphereFace loss

[RitiP]

Hi,

Thank you for sharing the code for SphereFace in Pytorch! Very interesting work! I have been using your code for SphereFace loss and was wondering how do you arrive at the value of m. I want to explore the usage of this loss function for different datasets(AWA, SUN etc.) and have currently used only the default values(s=30, m=1.5). It'd be great if you can elaborate how one can arrive at the value of m and tune it for different datasets. Thanks again!