HVisionSensing/Triplet-Loss

Apply triplet loss to face recognition， adding new loss layer to the latest version of caffe.

Triplet Loss for FaceRecognition

Introduction

Modified from xiaolonw/caffe-video_triplet, update the source code to fit the new verison of BVLC/caffe.

Details:

1. add pair_size parameter for image_data_layer to shuffle training examples in each pair_size. Normally, pair_size = 2.

   in the prototxt:

layer {
  name: "data"
  type: "ImageData"
  top: "data"
  top: "label"
  include { 
    phase: TRAIN
  }
  transform_param {
    scale: 0.0078125
    mirror: true
  }
  image_data_param {
    source: "/media/wujiyang/data/FaceData/train.txt"
    batch_size: 128
    shuffle: true
    pair_size: 2
  }
}   

2. add .hpp files for norm_layer and rank_hard_loss_layer

   rank_hard_loss_layer.hpp
   norm_layer.hpp

Usage

In the prototxt  

layer {
  name: "triplet-loss"
  type: "RankHardLoss" 
  rank_hard_loss_param{
    neg_num: 4
    pair_size: 2
    hard_ratio: 0
    rand_ratio: 1.0
    margin: 0.1
  }
  bottom: "norml2"
  bottom: "label"
} 

Attention:

When training models with the CASIA-WebFace dataset，trying to set margin = 0.1 or 0.2.
In my experiments, when margin = 1, the loss didn't converge at all.