Caffe源码中Pooling Layer文件分析

Caffe源码(caffe version commit: 09868ac , date: 2015.08.15)中有一些重要的头文件，这里介绍下include/caffe/vision_layers文件中PoolingLayer类，在最新版caffe中，PoolingLayer类被单独放在了include/caffe/layers/pooling_layer.hpp文件中，这两个文件中PoolingLayer类的内容及实现是完全一致的：

1．  include文件：

(1)、<caffe/blob.hpp>：此文件的介绍可以参考：http://blog.csdn.net/fengbingchun/article/details/59106613

(2)、<caffe/layer.hpp>：此文件的介绍可以参考：http://blog.csdn.net/fengbingchun/article/details/60871052

(3)、<caffe/proto/caffe.pb.h>：此文件的介绍可以参考：http://blog.csdn.net/fengbingchun/article/details/55267162

2．  类PoolingLayer：池化层，Layer类的子类

Pooling layer的主要作用是降维，缩小feature map，图像降采样，方法有：

(1)、均值采样：取区域平均值作为降采样值；

(2)、最大值采样：取区域最大值作为降采样值；

(3)、随机采样：取区域内随机一个像素值。

<caffe/layers/pooling_layer.hpp>文件的详细介绍如下：

#ifndef CAFFE_POOLING_LAYER_HPP_
#define CAFFE_POOLING_LAYER_HPP_#include <vector>#include "caffe/blob.hpp"
#include "caffe/layer.hpp"
#include "caffe/proto/caffe.pb.h"namespace caffe {
/*** @brief Pools the input image by taking the max, average, etc. within regions.** TODO(dox): thorough documentation for Forward, Backward, and proto params.*/
// 池化层，Layer类的子类，图像降采样，有三种Pooling方法：Max、Avx、Stochastic
template <typename Dtype>
class PoolingLayer : public Layer<Dtype> {public:
// 显示构造函数explicit PoolingLayer(const LayerParameter& param) : Layer<Dtype>(param) {}
// 参数初始化，通过类PoolingParameter获取成员变量值，包括：
// global_pooling_、kernel_h_、kernel_w_、pad_h_、pad_w_、stride_h_、stride_w_virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top);
// 调整top blobs的shape，并有可能会reshape rand_idx_或max_idx_；
// 获取成员变量值，包括：channels_、height_、width_、pooled_height_、pooled_width_virtual void Reshape(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top);
// 获得Pooling layer的类型: Poolingvirtual inline const char* type() const { return "Pooling"; }
// 获得Pooling layer所需的bottom blobs的个数: 1virtual inline int ExactNumBottomBlobs() const { return 1; }
// 获得Pooling layer所需的bottom blobs的最少个数: 1virtual inline int MinTopBlobs() const { return 1; }// MAX POOL layers can output an extra top blob for the mask;// others can only output the pooled inputs.
// 获得Pooling layer所需的bottom blobs的最多个数: Max为2，其它(Avg, Stochastic)为1virtual inline int MaxTopBlobs() const {return (this->layer_param_.pooling_param().pool() ==PoolingParameter_PoolMethod_MAX) ? 2 : 1;}protected:
// CPU实现Pooling layer的前向传播，仅有Max和Ave两种方法实现virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top);
// GPU实现Pooling layer的前向传播，Max、Ave、Stochastic三种方法实现virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top);
// CPU实现Pooling layer的反向传播，仅有Max和Ave两种方法实现virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
// GPU实现Pooling layer的反向传播，Max、Ave、Stochastic三种方法实现virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);// Caffe中类的成员变量名都带有后缀"_"，这样就容易区分临时变量和类成员变量int kernel_h_, kernel_w_; // 滤波器(卷积核)大小int stride_h_, stride_w_; // 步长大小int pad_h_, pad_w_; // 图像扩充大小int channels_; // 图像通道数int height_, width_; // 图像高、宽
// 池化后图像高、宽
// pooled_height_ = (height_ + 2 * pad_h_ - kernel_h_) / stride_h_)) + 1
// pooled_width_ = (width_ + 2 * pad_w_ - kernel_w_) / stride_w_)) + 1int pooled_height_, pooled_width_;bool global_pooling_; // 是否全区域池化(将整幅图像降采样为1*1)Blob<Dtype> rand_idx_; // 随机采样索引,Pooling方法为STOCHASTIC时用到并会ReshapeBlob<int> max_idx_; // 最大值采样索引，Pooling方法为MAX时用到并会Reshape
};}  // namespace caffe#endif  // CAFFE_POOLING_LAYER_HPP_

在caffe.proto文件中，有一个message是与pooling layer 相关的，如下：

message PoolingParameter { // Pooling层参数类enum PoolMethod { // 枚举类型，Pooling的方法：Max(最大值采样)、AVE(均值采样)、STOCHASTIC(随机采样)MAX = 0;AVE = 1;STOCHASTIC = 2;}optional PoolMethod pool = 1 [default = MAX]; // The pooling method, pooling方法// Pad, kernel size, and stride are all given as a single value for equal// dimensions in height and width or as Y, X pairs.optional uint32 pad = 4 [default = 0]; // The padding size (equal in Y, X)，图像扩充大小(添加图像边界的像素大小)optional uint32 pad_h = 9 [default = 0]; // The padding height,图像扩充大小，Yoptional uint32 pad_w = 10 [default = 0]; // The padding width，图像扩充大小，Xoptional uint32 kernel_size = 2; // The kernel size (square)，滤波器(卷积核、滑动窗)的大小(高=宽)optional uint32 kernel_h = 5; // The kernel height，滤波器(卷积核、滑动窗)的高optional uint32 kernel_w = 6; // The kernel width，滤波器(卷积核、滑动窗)的宽optional uint32 stride = 3 [default = 1]; // The stride (equal in Y, X),滑动步长(高=宽)，卷积核卷积时平移的步幅optional uint32 stride_h = 7; // The stride height,滑动步长，高optional uint32 stride_w = 8; // The stride width，滑动步长，宽enum Engine {DEFAULT = 0;CAFFE = 1;CUDNN = 2;}optional Engine engine = 11 [default = DEFAULT]; //// If global_pooling then it will pool over the size of the bottom by doing// kernel_h = bottom->height and kernel_w = bottom->widthoptional bool global_pooling = 12 [default = false]; // 是否是全区域池化
}

pooling layer的测试代码如下：

#include "funset.hpp"
#include <string>
#include <vector>
#include "common.hpp"int test_caffe_layer_pooling()
{caffe::Caffe::set_mode(caffe::Caffe::CPU); // set run caffe mode// set layer parametercaffe::LayerParameter layer_param;layer_param.set_phase(caffe::Phase::TRAIN);// cv::Mat -> caffe::Blobstd::string image_name = "E:/GitCode/Caffe_Test/test_data/images/a.jpg";cv::Mat mat1 = cv::imread(image_name, 1);if (!mat1.data) {fprintf(stderr, "read image fail: %s\n", image_name.c_str());return -1;}mat1.convertTo(mat1, CV_32FC3);std::vector<cv::Mat> mat2;cv::split(mat1, mat2);std::vector<int> mat_reshape{ 1, (int)mat2.size(), mat2[0].rows, mat2[0].cols };caffe::Blob<float> blob;blob.Reshape(mat_reshape);size_t size = mat2[0].rows * mat2[0].cols;float* data = new float[mat2.size() * size];memcpy(data, mat2[0].data, size * sizeof(float));memcpy(data + size, mat2[1].data, size * sizeof(float));memcpy(data + 2 * size, mat2[2].data, size * sizeof(float));blob.set_cpu_data(data);for (int method = 0; method < 2; ++method) {// set pooling parametercaffe::PoolingParameter* pooling_param = layer_param.mutable_pooling_param();if (method == 0) pooling_param->set_pool(caffe::PoolingParameter::MAX);else pooling_param->set_pool(caffe::PoolingParameter::AVE);pooling_param->set_kernel_size(3);pooling_param->set_pad(2);pooling_param->set_stride(2);pooling_param->set_global_pooling(false);std::vector<caffe::Blob<float>*> bottom_blob{ &blob }, top_blob{ &caffe::Blob<float>()/*, &caffe::Blob<float>() */ };// test PoolingLayer functioncaffe::PoolingLayer<float> pooling_layer(layer_param);pooling_layer.SetUp(bottom_blob, top_blob);fprintf(stderr, "top blob info: channels: %d, height: %d, width: %d\n",top_blob[0]->channels(), top_blob[0]->height(), top_blob[0]->width());pooling_layer.Forward(bottom_blob, top_blob);int height = top_blob[0]->height();int width = top_blob[0]->width();const float* p = top_blob[0]->cpu_data();std::vector<cv::Mat> mat3{ cv::Mat(height, width, CV_32FC1, (float*)p),cv::Mat(height, width, CV_32FC1, (float*)(p + height * width)),cv::Mat(height, width, CV_32FC1, (float*)(p + height * width * 2)) };cv::Mat mat4;cv::merge(mat3, mat4);mat4.convertTo(mat4, CV_8UC3);if (method == 0) image_name = "E:/GitCode/Caffe_Test/test_data/images/forward0.jpg";else image_name = "E:/GitCode/Caffe_Test/test_data/images/forward1.jpg";cv::imwrite(image_name, mat4);for (int i = 0; i < bottom_blob[0]->count(); ++i)bottom_blob[0]->mutable_cpu_diff()[i] = bottom_blob[0]->cpu_data()[i];for (int i = 0; i < top_blob[0]->count(); ++i)top_blob[0]->mutable_cpu_diff()[i] = top_blob[0]->cpu_data()[i];std::vector<bool> propagate_down{ true };pooling_layer.Backward(top_blob, propagate_down, bottom_blob);height = bottom_blob[0]->height();width = bottom_blob[0]->width();p = bottom_blob[0]->cpu_diff();std::vector<cv::Mat> mat5{ cv::Mat(height, width, CV_32FC1, (float*)p),cv::Mat(height, width, CV_32FC1, (float*)(p + height * width)),cv::Mat(height, width, CV_32FC1, (float*)(p + height * width * 2)) };cv::Mat mat6;cv::merge(mat5, mat6);mat6.convertTo(mat6, CV_8UC3);if (method == 0) image_name = "E:/GitCode/Caffe_Test/test_data/images/backward0.jpg";else image_name = "E:/GitCode/Caffe_Test/test_data/images/backward1.jpg";cv::imwrite(image_name, mat6);}delete[] data;return 0;
}

执行结果如下图：

图像结果(Lena.jpg)如下：前向传播、反向传播，前两幅为Max，后两幅为Avg方法

GitHub：https://github.com/fengbingchun/Caffe_Test