caffe源码解读(10)－hinge_loss_layer.cpp

公式

参数：两个输入bottom[0]、bottom[1],一个输出top[0]。
bottom[0]: N*C*1*1 预测值；
bottom[1]: N*1*1*1 真实值；
p: 范数，可选$L_{1}、L_{2}$范数；
$\delta \left \{ l_{n}=k \right \}$: 示性函数，如果第$n$个样本的真实标签为$k$,则$\delta \left \{ l_{n}=k \right \}=1$,否则为$-1$；
$t_{nk}$: bottom[0]中,第$n$个样本，第$k$维的预测值；

$$E=\frac{1}{N}\sum_{n=1}^{N}\sum_{k=1}^{K}[max(0,1-\delta \left \{ l_{n}=k \right \}t_{nk})]^{p}$$

代码

(1)Forward

template <typename Dtype>
void HingeLossLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
    const vector<Blob<Dtype>*>& top) {
  const Dtype* bottom_data = bottom[0]->cpu_data();//预测值tnk
  Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
  const Dtype* label = bottom[1]->cpu_data();//真实值（不用反向传播计算）
  int num = bottom[0]->num();
  int count = bottom[0]->count();
  int dim = count / num;
　//bottom_diff=bottom_data
  caffe_copy(count, bottom_data, bottom_diff);
  for (int i = 0; i < num; ++i) {
    bottom_diff[i * dim + static_cast<int>(label[i])] *= -1;//bottom_diff=-1*tnk
  }
  for (int i = 0; i < num; ++i) {
    for (int j = 0; j < dim; ++j) {
      //bottom_diff=max(0,1-tnk)
      bottom_diff[i * dim + j] = std::max(
        Dtype(0), 1 + bottom_diff[i * dim + j]);
    }
  }
  Dtype* loss = top[0]->mutable_cpu_data();
  switch (this->layer_param_.hinge_loss_param().norm()) {
  case HingeLossParameter_Norm_L1://L1范数
  　//caffe_cpu_asum:计算bottom_diff所有元素的绝对值之和
    loss[0] = caffe_cpu_asum(count, bottom_diff) / num;//基于L1范数的hinge_loss计算公式
    break;
  case HingeLossParameter_Norm_L2://L2范数
  　//caffe_cpu_dot:计算bottom_diff*bottom_diff
    loss[0] = caffe_cpu_dot(count, bottom_diff, bottom_diff) / num;//基于L2范数的hinge_loss计算公式
    break;
  default:
    LOG(FATAL) << "Unknown Norm";
  }
}

(2)Backward
bottom[1]是label的groundtruth，不需要进行反向传播运算，只需要对bottom[0]进行反向传播运算，反向传播是损失$E$对$t$的偏导。以$L_{2}$范数为例，求偏导为：

$$\frac{2}{N}*hinge*\frac{\partial hinge}{\partial t_{nk}},记max(0,1-\delta \left \{ l_{n}=k \right \}t_{nk})为hinge$$
其中： $$\frac{\partial hinge}{\partial t_{nk}}=\begin{Bmatrix} 0, &hinge=0 \\ -1,&hinge>0 \end{Bmatrix}$$

template <typename Dtype>
void HingeLossLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
    const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
  if (propagate_down[1]) {//label不需要做反向传播运算
    LOG(FATAL) << this->type()
               << " Layer cannot backpropagate to label inputs.";
  }
  if (propagate_down[0]) {//对bottom[0]进行反向传输运算
    Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
    const Dtype* label = bottom[1]->cpu_data();
    int num = bottom[0]->num();
    int count = bottom[0]->count();
    int dim = count / num;

    for (int i = 0; i < num; ++i) {
      bottom_diff[i * dim + static_cast<int>(label[i])] *= -1;
    }

    const Dtype loss_weight = top[0]->cpu_diff()[0];
    switch (this->layer_param_.hinge_loss_param().norm()) {
    case HingeLossParameter_Norm_L1:
      caffe_cpu_sign(count, bottom_diff, bottom_diff);
      caffe_scal(count, loss_weight / num, bottom_diff);
      break;
    case HingeLossParameter_Norm_L2://L2范数
      caffe_scal(count, loss_weight * 2 / num, bottom_diff);
      break;
    default:
      LOG(FATAL) << "Unknown Norm";
    }
  }
}