Hello, ReDimNet is a very valuable work, and thank you very much for your open source sharing of the code.
I saw in the paper that using sphereface 2 as the loss function would yield better results. May I ask what specific hyper-parameters you used when adopting sphereface 2? (including margin, t, scale, and lambuda for both pretraining and LMF stages.)
Hello, ReDimNet is a very valuable work, and thank you very much for your open source sharing of the code.
I saw in the paper that using sphereface 2 as the loss function would yield better results. May I ask what specific hyper-parameters you used when adopting sphereface 2? (including margin, t, scale, and lambuda for both pretraining and LMF stages.)
Thanks again