池化层(PoolingLayer)在深度学习的结构中占很重要的一部分,其功能是缩小输入特征图的尺度,增大输出特征图的感受野,在有的网络中,通过全局池化层,将二维特征图变成一维特征向量。池化层有两个特点:(1)输出特征图尺度普遍小于输入特征图尺度;(2)池化层并没有可学习的参量,因此,再反向传播的过程中,只需要求输入特征图的梯度(bottom_diff)。下面,我们来看看在Caffe中,怎么实现PoolingLayer的呢?
首先,我们还是来看看,配置文件,了解PoolingLayer需要哪些用户定义的参数
message PoolingParameter {
//池化方法,平均池化,最大池化,默认为,最大池化
enum PoolMethod {
MAX = 0;
AVE = 1;
STOCHASTIC = 2;
}
optional PoolMethod pool = 1 [default = MAX]; // The pooling method
// Pad, kernel size, and stride are all given as a single value for equal
// dimensions in height and width or as Y, X pairs.
//需要在输入特征图padding的数量,默认为0,
//既可以分别指定(pad_h、pad_w),又可以一起指定(pad)
optional uint32 pad = 4 [default = 0]; // The padding size (equal in Y, X)
optional uint32 pad_h = 9 [default = 0]; // The padding height
optional uint32 pad_w = 10 [default = 0]; // The padding width
//池化核大小,同padding一样两种指定方式
optional uint32 kernel_size = 2; // The kernel size (square)
optional uint32 kernel_h = 5; // The kernel height
optional uint32 kernel_w = 6; // The kernel width
//stride,同padding指定方式一样
optional uint32 stride = 3 [default = 1]; // The stride (equal in Y, X)
optional uint32 stride_h = 7; // The stride height
optional uint32 stride_w = 8; // The stride width
//计算引擎
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->width
//全局池化层,如果设置为全局池化层
//则kernel_h = bottom_h kernel_w = bottom_w
optional bool global_pooling = 12 [default = false];
}
#ifndef CAFFE_POOLING_LAYER_HPP_
#define CAFFE_POOLING_LAYER_HPP_
#include
#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.
*/
template <typename Dtype>
class PoolingLayer : public Layer {
public:
//显式构造函数
explicit PoolingLayer(const LayerParameter& param)
: Layer(param) {}
virtual void LayerSetUp(const vector *>& bottom,
const vector *>& top);
virtual void Reshape(const vector *>& bottom,
const vector *>& top);
virtual inline const char* type() const { return "Pooling"; }
//只能有一个输入Blob
virtual inline int ExactNumBottomBlobs() const { return 1; }
virtual 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.
virtual inline int MaxTopBlobs() const {
return (this->layer_param_.pooling_param().pool() ==
PoolingParameter_PoolMethod_MAX) ? 2 : 1;
}
protected:
//前向传播
virtual void Forward_cpu(const vector *>& bottom,
const vector *>& top);
virtual void Forward_gpu(const vector *>& bottom,
const vector *>& top);
//反向传播
virtual void Backward_cpu(const vector *>& top,
const vector<bool>& propagate_down, const vector *>& bottom);
virtual void Backward_gpu(const vector *>& top,
const vector<bool>& propagate_down, const vector *>& bottom);
//以下参数都和caffe.proto文件中的对应
int kernel_h_, kernel_w_;
int stride_h_, stride_w_;
int pad_h_, pad_w_;
int channels_;
//输入特征图的尺度
int height_, width_;
//Pooling之后特征图的尺度
int pooled_height_, pooled_width_;
bool global_pooling_;
Blob rand_idx_;
//max_idx_要单独解释
//在Pooling的过程中,必然有信息的丢失(空间信息)
//有些算法(SegNet)中,希望保存,Pooling之后(仅针对MaxPooling)的数值,在原特征图的位置,或者索引,有啥用?请看SegNet这篇论文。
//因此,max_idx_就是保存MaxPooling之前,Max值得位置
Blob<int> max_idx_;
};
} // namespace caffe
#endif // CAFFE_POOLING_LAYER_HPP_
#include
#include
#include
#include "caffe/layers/pooling_layer.hpp"
#include "caffe/util/math_functions.hpp"
namespace caffe {
using std::min;
using std::max;
template <typename Dtype>
void PoolingLayer::LayerSetUp(const vector *>& bottom,
const vector *>& top) {
PoolingParameter pool_param = this->layer_param_.pooling_param();
//首先,如果是全局池化层的话,强制赋值;
//如果,不是全局池化层的话,就按用户的来
//以下,padding和stride,一样的套路
//LayerSetUp函数就是来设置kernel、padding、stride的
if (pool_param.global_pooling()) {
CHECK(!(pool_param.has_kernel_size() ||
pool_param.has_kernel_h() || pool_param.has_kernel_w()))
<< "With Global_pooling: true Filter size cannot specified";
} else {
CHECK(!pool_param.has_kernel_size() !=
!(pool_param.has_kernel_h() && pool_param.has_kernel_w()))
<< "Filter size is kernel_size OR kernel_h and kernel_w; not both";
CHECK(pool_param.has_kernel_size() ||
(pool_param.has_kernel_h() && pool_param.has_kernel_w()))
<< "For non-square filters both kernel_h and kernel_w are required.";
}
CHECK((!pool_param.has_pad() && pool_param.has_pad_h()
&& pool_param.has_pad_w())
|| (!pool_param.has_pad_h() && !pool_param.has_pad_w()))
<< "pad is pad OR pad_h and pad_w are required.";
CHECK((!pool_param.has_stride() && pool_param.has_stride_h()
&& pool_param.has_stride_w())
|| (!pool_param.has_stride_h() && !pool_param.has_stride_w()))
<< "Stride is stride OR stride_h and stride_w are required.";
global_pooling_ = pool_param.global_pooling();
if (global_pooling_) {
kernel_h_ = bottom[0]->height();
kernel_w_ = bottom[0]->width();
} else {
if (pool_param.has_kernel_size()) {
kernel_h_ = kernel_w_ = pool_param.kernel_size();
} else {
kernel_h_ = pool_param.kernel_h();
kernel_w_ = pool_param.kernel_w();
}
}
CHECK_GT(kernel_h_, 0) << "Filter dimensions cannot be zero.";
CHECK_GT(kernel_w_, 0) << "Filter dimensions cannot be zero.";
if (!pool_param.has_pad_h()) {
pad_h_ = pad_w_ = pool_param.pad();
} else {
pad_h_ = pool_param.pad_h();
pad_w_ = pool_param.pad_w();
}
if (!pool_param.has_stride_h()) {
stride_h_ = stride_w_ = pool_param.stride();
} else {
stride_h_ = pool_param.stride_h();
stride_w_ = pool_param.stride_w();
}
if (global_pooling_) {
CHECK(pad_h_ == 0 && pad_w_ == 0 && stride_h_ == 1 && stride_w_ == 1)
<< "With Global_pooling: true; only pad = 0 and stride = 1";
}
if (pad_h_ != 0 || pad_w_ != 0) {
CHECK(this->layer_param_.pooling_param().pool()
== PoolingParameter_PoolMethod_AVE
|| this->layer_param_.pooling_param().pool()
== PoolingParameter_PoolMethod_MAX)
<< "Padding implemented only for average and max pooling.";
CHECK_LT(pad_h_, kernel_h_);
CHECK_LT(pad_w_, kernel_w_);
}
}
template <typename Dtype>
void PoolingLayer::Reshape(const vector *>& bottom,
const vector *>& top) {
CHECK_EQ(4, bottom[0]->num_axes()) << "Input must have 4 axes, "
<< "corresponding to (num, channels, height, width)";
channels_ = bottom[0]->channels();//输入(输出)特征图的通道数
//输入特征图的尺度
height_ = bottom[0]->height();
width_ = bottom[0]->width();
if (global_pooling_) {
kernel_h_ = bottom[0]->height();
kernel_w_ = bottom[0]->width();
}
//输出特征图的尺度,这里有个公式,非常重要,在之后卷积层也会涉及到
//输出特征图的尺度 = (输入 + 2 * pad - kernel) / stride + 1
pooled_height_ = static_cast<int>(ceil(static_cast<float>(
height_ + 2 * pad_h_ - kernel_h_) / stride_h_)) + 1;
pooled_width_ = static_cast<int>(ceil(static_cast<float>(
width_ + 2 * pad_w_ - kernel_w_) / stride_w_)) + 1;
if (pad_h_ || pad_w_) {
// If we have padding, ensure that the last pooling starts strictly
// inside the image (instead of at the padding); otherwise clip the last.
//然后 做有个确保,每次滑动窗口的起始位置都在输入特征图的里面
//只有到每行、每列的最后位置时,才可能会超出图像
if ((pooled_height_ - 1) * stride_h_ >= height_ + pad_h_) {
--pooled_height_;
}
if ((pooled_width_ - 1) * stride_w_ >= width_ + pad_w_) {
--pooled_width_;
}
CHECK_LT((pooled_height_ - 1) * stride_h_, height_ + pad_h_);
CHECK_LT((pooled_width_ - 1) * stride_w_, width_ + pad_w_);
}
//分配内存(top)
top[0]->Reshape(bottom[0]->num(), channels_, pooled_height_,
pooled_width_);
if (top.size() > 1) {
top[1]->ReshapeLike(*top[0]);
}
// If max pooling, we will initialize the vector index part.
if (this->layer_param_.pooling_param().pool() ==
PoolingParameter_PoolMethod_MAX && top.size() == 1) {
//最大索引位置
max_idx_.Reshape(bottom[0]->num(), channels_, pooled_height_,
pooled_width_);
}
// If stochastic pooling, we will initialize the random index part.
if (this->layer_param_.pooling_param().pool() ==
PoolingParameter_PoolMethod_STOCHASTIC) {
rand_idx_.Reshape(bottom[0]->num(), channels_, pooled_height_,
pooled_width_);
}
}
// TODO(Yangqing): Is there a faster way to do pooling in the channel-first
// case?
template <typename Dtype>
void PoolingLayer::Forward_cpu(const vector *>& bottom,
const vector *>& top) {
const Dtype* bottom_data = bottom[0]->cpu_data();
Dtype* top_data = top[0]->mutable_cpu_data();
const int top_count = top[0]->count();
// We'll output the mask to top[1] if it's of size >1.
//use_top_mask仅在最大池化层处有用,当use_top_mask = true时,则
//池化层中的最大值索引,被当做top[1]向后面层传输,SegNet有使用
const bool use_top_mask = top.size() > 1;
int* mask = NULL; // suppress warnings about uninitalized variables
Dtype* top_mask = NULL;
// Different pooling methods. We explicitly do the switch outside the for
// loop to save time, although this results in more code.
switch (this->layer_param_.pooling_param().pool()) {
case PoolingParameter_PoolMethod_MAX:
// 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
//前向传播的核心代码
for (int n = 0; n < bottom[0]->num(); ++n) {//大循环,batch_size
for (int c = 0; c < channels_; ++c) {//特征图通道数
for (int ph = 0; ph < pooled_height_; ++ph) {//输出尺度的两层循环
for (int pw = 0; pw < pooled_width_; ++pw) {
//输出特征图位置,对应于输入特征图的一块区域
int hstart = ph * stride_h_ - pad_h_;
int wstart = pw * stride_w_ - pad_w_;
//保证都在输入特征图里面
int hend = min(hstart + kernel_h_, height_);
int wend = min(wstart + kernel_w_, width_);
hstart = max(hstart, 0);
wstart = max(wstart, 0);
//pool_index表示输出特征图的内存位置
//在Caffe或者C++中,数据块是内存的线性存储的,没有分什么维度
//通过解析可以解析成不同维度的
//举个例子有一个连续的12个int类型的数组
//通过不同的解析可以是(2,6)、(3,4)、(2,2,3)不同形状的
//但是所占内存空间都是一样的
const int pool_index = ph * pooled_width_ + pw;
//开始遍历(ph,pw)对于与输入特征图的区域
for (int h = hstart; h < hend; ++h) {
for (int w = wstart; w < wend; ++w) {
//同理输入特征图的索引
const int index = h * width_ + w;
//找最大位置,然后保存最大位置的索引
if (bottom_data[index] > top_data[pool_index]) {
top_data[pool_index] = bottom_data[index];
if (use_top_mask) {
top_mask[pool_index] = static_cast(index);
} else {
mask[pool_index] = index;
}
}
}
}
}
}
//一个特征图完事了,然后就计算下一个特征图,也就是下一个通道
// compute offset
bottom_data += bottom[0]->offset(0, 1);//输入内存偏移
top_data += top[0]->offset(0, 1);//输出内存偏移
//索引偏移
if (use_top_mask) {
top_mask += top[0]->offset(0, 1);
} else {
mask += top[0]->offset(0, 1);
}
}
}
break;
//平均池化代码上和最大池化前向传播差不多
case PoolingParameter_PoolMethod_AVE:
for (int i = 0; i < top_count; ++i) {
top_data[i] = 0;
}
// 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_;
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;
}
}
// compute offset
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.";
}
}
//反向传播
template <typename Dtype>
void PoolingLayer::Backward_cpu(const vector *>& top,
const vector<bool>& propagate_down, const vector *>& bottom) {
if (!propagate_down[0]) {
return;
}
//获取相应的指针
const Dtype* top_diff = top[0]->cpu_diff();
Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
// Different pooling methods. We explicitly do the switch outside the for
// loop to save time, although this results in more codes.
//首先填充0
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
//最大池化和平均池化有所不同
const Dtype* top_mask = NULL;
switch (this->layer_param_.pooling_param().pool()) {
//在最大池化的反向传播中,输入层最大值索引位置的梯度为相应的top_diff的值
//其余位置的梯度均为0
//因此,可以看出最大池化的反向传播时稀疏的!
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;
//获取最大值索引
const int bottom_index =
use_top_mask ? top_mask[index] : mask[index];
//最大索引的位置
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;
//平均池化,就是将top_diff的值平均分配到输入特征图对应的区域
//也就是(ph, pw) -> {(hstart:hend), (wstart:wend)}
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;
}
}
}
}
// 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.";
}
}
#ifdef CPU_ONLY
STUB_GPU(PoolingLayer);
#endif
INSTANTIATE_CLASS(PoolingLayer);
} // namespace caffe