I have a keras version of amsoftmax, but it not well in face converge is very slow

[FlyEgle]

I do not know where have problem, can you have a look?

[Joker316701882]

@FlyEgle Which layer are u using as face embeddings? Given this code, I assumed you didn't use "nomalized inputs", but original inputs as face embedding right? If so ,this is problem.

[FlyEgle]

A new keras amsoftmax below

This is the net-work cnn struct Can you see if it is right?Beacuse i use m=0.35 and s = 30, i only can train 0.93 acc and val_acc is 0.84 in a small face datasets called FaceScrub which have 529 peoples

[Joker316701882]

@FlyEgle Logic seems no problem here. Evaluation code matters more. And to see whether your code is fine, simply train on larger dataset(CASIA, vggface,vggfacd2 etc...). 529 is too small. Reopen if needed.

[FlyEgle]

ok