Caffe学习系列(15)：添加新层

如何在Caffe中增加一层新的Layer呢？主要分为四步：

（1）在./src/caffe/proto/caffe.proto 中增加对应layer的paramter message；

（2）在./include/caffe/***layers.hpp中增加该layer的类的声明，***表示有common_layers.hpp,

data_layers.hpp, neuron_layers.hpp, vision_layers.hpp 和loss_layers.hpp等；

（3）在./src/caffe/layers/目录下新建.cpp和.cu（GPU）文件，进行类实现。

（4）在./src/caffe/gtest/中增加layer的测试代码，对所写的layer前传和反传进行测试，测试还包括速度。（可省略，但建议加上）

这位博主添加了一个计算梯度的网络层，简介明了：

http://blog.csdn.net/shuzfan/article/details/51322976

这几位博主增加了自定义的loss层，可供参考：

http://blog.csdn.net/langb2014/article/details/50489305

http://blog.csdn.net/tangwei2014/article/details/46815231

我以添加precision_recall_loss层来学习代码,主要是precision_recall_loss_layer.cpp的实现

#include <algorithm>
#include <cfloat>
#include <cmath>
#include <vector>
#include <opencv2/opencv.hpp>  #include "caffe/layer.hpp"
#include "caffe/util/io.hpp"
#include "caffe/util/math_functions.hpp"
#include "caffe/vision_layers.hpp"  namespace caffe {  //初始化，调用父类进行相应的初始化
template <typename Dtype>
void PrecisionRecallLossLayer<Dtype>::LayerSetUp(
  const vector<Blob<Dtype>*> &bottom, const vector<Blob<Dtype>*> &top) {
  LossLayer<Dtype>::LayerSetUp(bottom, top);
}
//进行维度变换
template <typename Dtype>
void PrecisionRecallLossLayer<Dtype>::Reshape(
  const vector<Blob<Dtype>*> &bottom,
  const vector<Blob<Dtype>*> &top) {
  //同样先调用父类的Reshape,通过成员变量loss_来改变输入维度
  LossLayer<Dtype>::Reshape(bottom, top);
  loss_.Reshape(bottom[]->num(), bottom[]->channels(),
                bottom[]->height(), bottom[]->width());    // Check the shapes of data and label  检查两个输入的维度是否想等
  CHECK_EQ(bottom[]->num(), bottom[]->num())
      << "The number of num of data and label should be same.";
  CHECK_EQ(bottom[]->channels(), bottom[]->channels())
      << "The number of channels of data and label should be same.";
  CHECK_EQ(bottom[]->height(), bottom[]->height())
      << "The heights of data and label should be same.";
  CHECK_EQ(bottom[]->width(), bottom[]->width())
      << "The width of data and label should be same.";
}  
//前向传导
template <typename Dtype>
void PrecisionRecallLossLayer<Dtype>::Forward_cpu(
  const vector<Blob<Dtype>*> &bottom, const vector<Blob<Dtype>*> &top) {
  const Dtype *data = bottom[]->cpu_data();
  const Dtype *label = bottom[]->cpu_data(); 
  const int num = bottom[]->num();  //num和count什么区别
  const int dim = bottom[]->count() / num;
  const int channels = bottom[]->channels();
  const int spatial_dim = bottom[]->height() * bottom[]->width();  
  //存疑？
  const int pnum =
    this->layer_param_.precision_recall_loss_param().point_num();
  top[]->mutable_cpu_data()[] = ;  
  //对于每个通道
  for (int c = ; c < channels; ++c) {
    Dtype breakeven = 0.0;
    Dtype prec_diff = 1.0;
    for (int p = ; p <= pnum; ++p) {
      int true_positive = ;  //统计每类的个数
      int false_positive = ;
      int false_negative = ;
      int true_negative = ;  
 
      for (int i = ; i < num; ++i) {
        const Dtype thresh = 1.0 / pnum * p;  //计算阈值？
        for (int j = ; j < spatial_dim; ++j) {  
          //取得相应的值和标签
          const Dtype data_value = data[i * dim + c * spatial_dim + j];
          const int label_value = (int)label[i * dim + c * spatial_dim + j];  
          //统计
          if (label_value ==  && data_value >= thresh) {
            ++true_positive;
          }
          if (label_value ==  && data_value >= thresh) {
            ++false_positive;
          }
          if (label_value ==  && data_value < thresh) {
            ++false_negative;
          }
          if (label_value ==  && data_value < thresh) {
            ++true_negative;
          }
        }
      }  
      //计算precision和recall
      Dtype precision = 0.0;
      Dtype recall = 0.0;
      if (true_positive + false_positive > ) {
        precision =
          (Dtype)true_positive / (Dtype)(true_positive + false_positive);
      } else if (true_positive == ) {  //都是负类？
        precision = 1.0;
      }
      if (true_positive + false_negative > ) {
        recall =
          (Dtype)true_positive / (Dtype)(true_positive + false_negative);
      } else if (true_positive == ) {
        recall = 1.0;
      }
      if (prec_diff > fabs(precision - recall)  //如果二c者相差小
          && precision >  && precision <
          && recall >  && recall < ) {
        breakeven = precision;  //保留
        prec_diff = fabs(precision - recall);
      }
    }
    top[]->mutable_cpu_data()[] += 1.0 - breakeven;  //计算误差
  }
  top[]->mutable_cpu_data()[] /= channels;  //？？？
}
//反向
template <typename Dtype>
void PrecisionRecallLossLayer<Dtype>::Backward_cpu(
  const vector<Blob<Dtype>*> &top,
  const vector<bool> &propagate_down,
  const vector<Blob<Dtype>*> &bottom) {
  for (int i = ; i < propagate_down.size(); ++i) {
    if (propagate_down[i]) { NOT_IMPLEMENTED; }
  }
}
      #ifdef CPU_ONLY
    STUB_GPU(PrecisionRecallLossLayer);
    #endif  //注册该层
    INSTANTIATE_CLASS(PrecisionRecallLossLayer);
    REGISTER_LAYER_CLASS(PrecisionRecallLoss);      }  // namespace caffe