A-softmax by PyTorch

This an implementation of SphereFace: Deep Hypersphere Embedding for Face Recognition

Requirements

PyTorch(0.4 or later)
Torchvision
tensorboardX
imageio(2.5.0, must be this version)
matplotlib
numpy

Train

$ cd a-softmax
$ python main.py

After train, you can get the acc about 98.7%. All feature picture is in logs/images. The gif(at logs dir) is shown as below.

The left is feature distribution in train step and the right is in test step.

look train resulst by tensorboard

$ tensorboard --logdir=logs/tblog --port=6006

Then you can go to http://0.0.0.0:6006. The results is shown as below:

Reference

[1].https://github.com/clcarwin/sphereface_pytorch
[2].https://github.com/Joyako/SphereFace-pytorch
[3].https://github.com/woshildh/center_loss
[4]. https://github.com/wujiyang/Face_Pytorch