Skip to content

Latest commit

 

History

History
executable file
·
44 lines (35 loc) · 1.13 KB

File metadata and controls

executable file
·
44 lines (35 loc) · 1.13 KB

Various_softmax_by_Keras

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


Author LCorleone
E-mail [email protected]

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

original-softmax

center-loss

center-loss

A-softmax

A-softmax

AM-softmax

AM-softmax

To do

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

Reference

margin-softmax