Weiyang Liu

Yes, you can train deeper networks. But you may need to modify the code to support multi-gpu setting for more gpu memory. We have tried ResNet-34 and ResNet-50, and they...

With proper setting, BatchNorm can help SphereFace-64 converge easily on MS-1M.

You may need to change the hyperparameters for ReLU to converge.

Please see Appendix H in SphereFace paper (http://wyliu.com/papers/LiuCVPR17v3.pdf).

the network architectures are the same. The only thing that is different is the cropped patch.

The code will be released soon. Thanks for your interest. :)

Simply set the output feature dimension as 3. It corresponds to the weight matrix size of the layer before entering the classifier layer.

Thanks for the contribution!

Please see https://github.com/kpzhang93/MTCNN_face_detection_alignment

Yes, we have trained A-Softmax loss on much larger dataset (MS dataset) and it definitely can work. You should consider to modify the function of lambda or use the fine-tuning...