AM-Softmax Loss is also called additive margin softmax. It is an improvement of softmax loss and proposed in paper: Additive Margin Softmax for Face Verification. This loss function often be used in face and speaker verification task.
It is denoted as:
You can find this loss in this page:
https://github.com/happynear/AMSoftmax
The effect of AM-Softmax
From this paper, we can find:
s = 30, the performance become the best when m = 0.35 to m = 0.4. We usually use it when s = 30 and m = 0.35