Caffe源碼解析7：Pooling_Layer

轉載請注明出處，樓燚(yì)航的blog，http://home.cnblogs.com/louyihang-loves-baiyan/

Pooling 層一般在網絡中是跟在Conv卷積層之后，做采樣操作，其實是為了進一步縮小feature map，同時也能增大神經元的視野。在Caffe中，pooling層屬於vision_layer的一部分，其相關的定義也在vision_layer.hpp的頭文件中。Pooling層的相關操作比較少，在Caffe的自帶模式下只有Max pooling和Average poooling兩種

下圖是一個LeNet的網絡結構圖，全連接之前主要有2個卷基層，2個池化層，其中sub_sampling layer就是pooling的操作。pooling的范圍是給定的一個region。

PoolingLayer

caffe中Pooling的操作相對比較少，結構也簡單，首先看它的Forward_cpu函數，在forward的時候根據相應的Pooling_method選擇相應的pooling方法

forward_cpu

void PoolingLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
      const vector<Blob<Dtype>*>& top) {
  const Dtype* bottom_data = bottom[0]->cpu_data();
  Dtype* top_data = top[0]->mutable_cpu_data();
  const int top_count = top[0]->count();
  //將mask信息輸出到top[1],如果top大於1
  const bool use_top_mask = top.size() > 1;
  int* mask = NULL;  // suppress warnings about uninitalized variables
  Dtype* top_mask = NULL;
  switch (this->layer_param_.pooling_param().pool()) {
  case PoolingParameter_PoolMethod_MAX://這里的case主要是實現max pooling的方法
    // Initialize
    if (use_top_mask) {
      top_mask = top[1]->mutable_cpu_data();
      caffe_set(top_count, Dtype(-1), top_mask);
    } else {
      mask = max_idx_.mutable_cpu_data();
      caffe_set(top_count, -1, mask);
    }
    caffe_set(top_count, Dtype(-FLT_MAX), top_data);
    // The main loop
    for (int n = 0; n < bottom[0]->num(); ++n) {
      for (int c = 0; c < channels_; ++c) {
        for (int ph = 0; ph < pooled_height_; ++ph) {
          for (int pw = 0; pw < pooled_width_; ++pw) {
            int hstart = ph * stride_h_ - pad_h_;//這里的hstart，wstart,hend,wend指的是pooling窗口在特征圖中的坐標，對應左上右下即x1 y1 x2 y2
            int wstart = pw * stride_w_ - pad_w_;
            int hend = min(hstart + kernel_h_, height_);
            int wend = min(wstart + kernel_w_, width_);
            hstart = max(hstart, 0);
            wstart = max(wstart, 0);
            const int pool_index = ph * pooled_width_ + pw;
            for (int h = hstart; h < hend; ++h) {
              for (int w = wstart; w < wend; ++w) {
                const int index = h * width_ + w;//記錄index偏差
                if (bottom_data[index] > top_data[pool_index]) {//不停迭代
                  top_data[pool_index] = bottom_data[index];
                  if (use_top_mask) {
                    top_mask[pool_index] = static_cast<Dtype>(index);//記錄當前最大值的的坐標索引
                  } else {
                    mask[pool_index] = index;
                  }
                }
              }
            }
          }
        }
        // 計算偏移量，進入下一張圖的index起始地址
        bottom_data += bottom[0]->offset(0, 1);
        top_data += top[0]->offset(0, 1);
        if (use_top_mask) {
          top_mask += top[0]->offset(0, 1);
        } else {
          mask += top[0]->offset(0, 1);
        }
      }
    }
    break;
 case PoolingParameter_PoolMethod_AVE://average_pooling
    for (int i = 0; i < top_count; ++i) {
      top_data[i] = 0;
    }
    // The main loop
    for (int n = 0; n < bottom[0]->num(); ++n) {//同樣是主循環
      for (int c = 0; c < channels_; ++c) {
        for (int ph = 0; ph < pooled_height_; ++ph) {
          for (int pw = 0; pw < pooled_width_; ++pw) {
            int hstart = ph * stride_h_ - pad_h_;
            int wstart = pw * stride_w_ - pad_w_;
            int hend = min(hstart + kernel_h_, height_ + pad_h_);
            int wend = min(wstart + kernel_w_, width_ + pad_w_);
            int pool_size = (hend - hstart) * (wend - wstart);
            hstart = max(hstart, 0);
            wstart = max(wstart, 0);
            hend = min(hend, height_);
            wend = min(wend, width_);
            for (int h = hstart; h < hend; ++h) {
              for (int w = wstart; w < wend; ++w) {
                top_data[ph * pooled_width_ + pw] +=
                    bottom_data[h * width_ + w];
              }
            }
            top_data[ph * pooled_width_ + pw] /= pool_size;//獲得相應的平均值
          }
        }
        // compute offset同理計算下一個圖的起始地址
        bottom_data += bottom[0]->offset(0, 1);
        top_data += top[0]->offset(0, 1);
      }
    }
    break;
  case PoolingParameter_PoolMethod_STOCHASTIC:
    NOT_IMPLEMENTED;
    break;
  default:
    LOG(FATAL) << "Unknown pooling method.";
  }

backward_cpu

對於誤差的反向傳導
對於pooling層的誤差傳到，根據下式

\[\delta^l_j=upsample(\delta^{l+1}_{j})\cdot h(a^l_j)' \]

這里的Upsample具體可以根據相應的pooling方法來進行上采樣，upsample的基本思想也是將誤差進行的平攤到各個采樣的對應點上。在這里pooling因為是線性的所以h這一項其實是可以省略的。
具體的計算推導過程請結合http://www.cnblogs.com/tornadomeet/p/3468450.html有詳細的推導過程，結合代碼中主循環中的最里項會更清晰的明白

template <typename Dtype>
void PoolingLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
  if (!propagate_down[0]) {
    return;
  }
  const Dtype* top_diff = top[0]->cpu_diff();//首先獲得上層top_blob的diff
  Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
  caffe_set(bottom[0]->count(), Dtype(0), bottom_diff);
  // We'll output the mask to top[1] if it's of size >1.
  const bool use_top_mask = top.size() > 1;
  const int* mask = NULL;  // suppress warnings about uninitialized variables
  const Dtype* top_mask = NULL;
  switch (this->layer_param_.pooling_param().pool()) {
  case PoolingParameter_PoolMethod_MAX:
    // The main loop
    if (use_top_mask) {
      top_mask = top[1]->cpu_data();
    } else {
      mask = max_idx_.cpu_data();
    }
    for (int n = 0; n < top[0]->num(); ++n) {
      for (int c = 0; c < channels_; ++c) {
        for (int ph = 0; ph < pooled_height_; ++ph) {
          for (int pw = 0; pw < pooled_width_; ++pw) {
            const int index = ph * pooled_width_ + pw;
            const int bottom_index =
                use_top_mask ? top_mask[index] : mask[index];//根據max pooling記錄的mask位置，進行誤差反轉
            bottom_diff[bottom_index] += top_diff[index];
          }
        }
        bottom_diff += bottom[0]->offset(0, 1);
        top_diff += top[0]->offset(0, 1);
        if (use_top_mask) {
          top_mask += top[0]->offset(0, 1);
        } else {
          mask += top[0]->offset(0, 1);
        }
      }
    }
    break;
  case PoolingParameter_PoolMethod_AVE:
    // The main loop
    for (int n = 0; n < top[0]->num(); ++n) {
      for (int c = 0; c < channels_; ++c) {
        for (int ph = 0; ph < pooled_height_; ++ph) {
          for (int pw = 0; pw < pooled_width_; ++pw) {
            int hstart = ph * stride_h_ - pad_h_;
            int wstart = pw * stride_w_ - pad_w_;
            int hend = min(hstart + kernel_h_, height_ + pad_h_);
            int wend = min(wstart + kernel_w_, width_ + pad_w_);
            int pool_size = (hend - hstart) * (wend - wstart);
            hstart = max(hstart, 0);
            wstart = max(wstart, 0);
            hend = min(hend, height_);
            wend = min(wend, width_);
            for (int h = hstart; h < hend; ++h) {
              for (int w = wstart; w < wend; ++w) {
                bottom_diff[h * width_ + w] +=
                  top_diff[ph * pooled_width_ + pw] / pool_size;//mean_pooling中，bottom的誤差值按pooling窗口中的大小計算，從上一層進行填充后，再除窗口大小
              }
            }
          }
        }
        // offset
        bottom_diff += bottom[0]->offset(0, 1);
        top_diff += top[0]->offset(0, 1);
      }
    }
    break;
  case PoolingParameter_PoolMethod_STOCHASTIC:
    NOT_IMPLEMENTED;
    break;
  default:
    LOG(FATAL) << "Unknown pooling method.";
  }
}