Fully-connected layer in Arcface

[HongtaoYang]

Hi Wei,

Thanks for you awesome work, its really nice to see someone start coding in TF2.0!

I have one question about the last fully connected layer in arcface, which is the NormDense layer defined in recognition/models/models.py. Have you tried training with very large number of identities? The model would be ridiculously big with just ~10k identities. In the ArcFace paper, the author even reported results on some dataset with 94k identities.

Do you find it sufficient to train on just 1000 identities and model would generalise well to other unseen classes? Or it is generally required to fine-tune the model using triplet loss afterwards to make learned feature transferable?

Thanks for your help!