how to set the m when the feature without norm? about amsoftmax

[happynear]

When we don't use the feature normalization, it is still necessary to use annealing strategy to set the m like the lambda in SphereFace. So I need to change the code to add the annealing strategy. I will give an example tomorrow.

[happynear]

The example is uploaded. You also need to update the label_specific_add_layer to get the annealing codes.

https://github.com/happynear/AMSoftmax/blob/master/prototxt/face_train_test_wo_fn.prototxt

Note that I am still running the experiment to reproduce the result, so the prototxt may be changed in the next a few hours...

By the way, I can get similar results on LFW BLUFR with normalization by finetuning the network using scale 60 and margin 0.4 after iteration 16000. Maybe you can also try this way.

[Allan9977]

Thanks a lot! I will have a try.