Awesome

Face Recognition for Caffe2

Caffe2 is a lightweight, modular, speed, so i want to do face recognition experiment with it, like light cnn, center loss, large margin softmax, Angular softmax and so on.

Reference

• origin_lsoftmax
• mxnet_lsoftmax

Files

• Original Caffe2 library
• Large margin softmax operator
  • caffe2/operators/lsoftmax_with_loss.h
  • caffe2/operators/lsoftmax_with_loss.cc
  • caffe2/operators/lsoftmax_with_loss.cu
  • caffe2/python/operator_test/lsoftmax_with_loss_test.py
• face_example
  • caffe2/python/examples/sphereface_trainer.py
  • caffe2/python/models/sphereface.py
• mnist_example
  • caffe2/python/examples/mnist_trainer.py

Result

• mnist_example for lsoftmax

  • margin = 1
  <p align='left'> <img src='caffe2/python/examples/result/mnist/distance-margin-1.png' style='max-width:600px'></img> </p>
  • margin = 2
  <p align='left'> <img src='caffe2/python/examples/result/mnist/distance-margin-2.png' style='max-width:600px'</img> </p>
  • margin = 3
  <p align='left'> <img src='caffe2/python/examples/result/mnist/distance-margin-3.png' style='max-width:600px'</img> </p>
  • margin = 4
  <p align='left'> <img src='caffe2/python/examples/result/mnist/distance-margin-4.png' style='max-width:600px'</img> </p>

Build

Use my version of caffe2, follow origin caffe2 installation Installation