Various_softmax_by_Keras

Keras implementation of various loss functions in face recognition and visualization on Mnist

Requirements

• tensorflow 1.5
• keras 2.2.0
• some common packages like numpy and so on.

Quick start

• Now we support original softmax, center loss, A-softmax and AM-softmax.
• Just change the loss_name in keras_mnist_loss_compare.py

  loss_name = 'AM-softmax'

Mnist Visualization

original softmax

center-loss

A-softmax

AM-softmax

To do

• [x] original softmax
• [x] center loss
• [x] A-softmax
• [x] AM-softmax
• [ ] AAM-softmax(ArcFace)

Reference

margin-softmax