从零开始编写深度学习库（五）PoolingLayer 网络层CPU编写

记录：编写卷积层和池化层，比较需要注意的细节就是边界问题，还有另外一个就是重叠池化的情况，这两个小细节比较重要，边界问题pad在反向求导的时候，由于tensorflow是没有计算的，另外一个比较烦人的是Eigen::Tensor的rowmajor、和colmajor问题，也很烦人。为了跟tensorflow做比较，一些边界处理上的细节，需要特别注意。

一、c++、maxpooling、avgpooling

#pragma once
#include "config.h"
#include <vector>
enum PoolingMethod
{
	max,
	avg
 
};
 
class CPoolingLayer
{
public:
 
	CPoolingLayer(std::vector<int>pooling_shape,PoolingMethod pooling_method, int padding = 0) {
		m_hksize = pooling_shape[0];
		m_wksize = pooling_shape[1];
		m_hstride = pooling_shape[2];
		m_wstride = pooling_shape[3];
 
		m_padding = padding;
		m_pooling_method = pooling_method;
 
	};
	~CPoolingLayer() {
	};
	//返回(bottom[0],bottom[1]/hstride*bottom[2]/wstride,hsize,wsize,bottom[3])
	void CPoolingLayer::extract_image_patches(const Tensor4xf &bottom, Tensor5xf &patches) {
		//这个Eigen tensor类的extract_image_patches函数，由于有数据存储列排列、行排列两种不同的模式。
		//在下面函数中，如果是采用rowmajor，下面的调用方式才是正确的
		//不能采用bottom.extract_image_patches(  m_hksize,m_wksize, m_hstride,m_wstride, 1, 1);
		//switch (m_padding_method)
		//{
		//case valid:
			patches = bottom.extract_image_patches(m_hksize, m_wksize, m_hstride, m_wstride, 1, 1,
				Eigen::PADDING_VALID);
			//break;
		//case same:
			//patches = bottom.extract_image_patches(m_hksize, m_wksize, m_hstride, m_wstride, 1, 1,
				//Eigen::PADDING_SAME );
			//break;
		//default:
			//break;
		//}
 
	}
	//根据stride、size等计算输出数据的维度形状
	Eigen::DSizes<int, 4> CPoolingLayer::get_top_shape(const Tensor4xf&bottom) {
		Eigen::DSizes<int, 4>top_shape;
		top_shape[0] = bottom.dimension(0);
		//switch (m_padding_method)
		//{
		//case valid:
			top_shape[1] = Eigen::divup(float(bottom.dimension(1) - m_hksize + 1), float(m_hstride));
			top_shape[2] = Eigen::divup(float(bottom.dimension(2) - m_wksize + 1), float(m_wstride));
			//break;
		//case same:
			//top_shape[1] = Eigen::divup(float(bottom.dimension(1)), float(m_hstride));
			//top_shape[2] = Eigen::divup(float(bottom.dimension(2)), float(m_wstride));
 
			//break;
		//default:
			//break;
		//}
		top_shape[3] = bottom.dimension(3);
		return top_shape;
	}
 
	//需要特别注意这边的均值池化，与tesorflow，在same模式下处理方式不同，tensorflow的在计算池化的时候，
	//不管有没有padding，padding值在计算池化操作都被忽略
	// bottom(batch_size, input_height, input_width, input_channel);
	void CPoolingLayer::forward(const Tensor4xf&bottom,Tensor4xf&top, const Eigen::ThreadPoolDevice &device) {
		Tensor4xf pad_bottom;
		CBaseFunction::padding_forward(bottom, m_padding, m_padding, pad_bottom);
 
		Eigen::array<int, 2> reduction_dims{1,2};//第二维、第三维的大小等于（hksize、wksize）
		Eigen::DSizes<int, 4>post_reduce_dims=get_top_shape(pad_bottom);
 
		Tensor5xf patches; //(bottom[0],  hsize, wsize,bottom[1] / hstride*bottom[2] / wstride, bottom[3])
		extract_image_patches(pad_bottom, patches);
 
		Tensor3xf pooling(post_reduce_dims[0],post_reduce_dims[1]*post_reduce_dims[2],post_reduce_dims[3]);
		switch (m_pooling_method)
		{
		case avg:
			pooling.device(device) = patches.mean(reduction_dims);//对reduction_dims内对应的维度索引进行统计，比如统计第3、2
			break;
		case max:
			pooling.device(device) = patches.maximum(reduction_dims);//最大池化
			break;
		default:
			break;
		}
		top=pooling.reshape(post_reduce_dims);
 
	}
	//本函数主要用于索引解码，从一维索引到获取多维下标值。主要原因在于：max
	std::vector<int> CPoolingLayer::decode_index(std::vector<int>dim,int index) {
		std::vector<int>result;
		for (int i=0;i<5;i++)
		{
			int accu = 1;
			for (int j=5-1;j>i;j--)
			{
				accu *= dim[j];
 
			}
			result.push_back(std::floor(index / accu));
			index = index%accu;
		}
 
		return result;
 
	}
	//主要是重叠池化的时候，反向求导的时候是微分值累加。
	void CPoolingLayer::maxpooling_backward(const Tensor4xf&bottom,const Tensor4xf&dtop,Tensor4xf&dbottom) {
		Tensor4xf pad_bottom;
		CBaseFunction::padding_forward(bottom, m_padding, m_padding, pad_bottom);
		Tensor5xf patches;
		extract_image_patches(pad_bottom, patches);
		Tensor4xf dpad_bottom(pad_bottom.dimension(0), pad_bottom.dimension(1), pad_bottom.dimension(2), pad_bottom.dimension(3));
		dpad_bottom.setZero();
 
		Eigen::DSizes<int, 4>post_reduce_dims = get_top_shape(pad_bottom);
		Eigen::array<Eigen::DenseIndex, 2> reduce_dims{ 1,2 };
		auto index_tuples = patches.index_tuples();
		Eigen::Tensor<Eigen::Tuple<Eigen::DenseIndex, float>, 3, Eigen::internal::traits<Tensor5xf>::Layout> reduced_by_dims;
		reduced_by_dims = index_tuples.reduce(reduce_dims, Eigen::internal::ArgMaxTupleReducer<Eigen::Tuple<Eigen::DenseIndex, float> >());
 
		int batch = dtop.dimension(0);
		int height = dtop.dimension(1);
		int widht = dtop.dimension(2);
		int channel = dtop.dimension(3);
		bool isColMajor = (Eigen::internal::traits<Tensor4xf>::Layout ==Eigen::ColMajor);
 
		for (int b= 0; b < batch; b++)
		{
			for (int h = 0; h< height; h++)
			{
				for (int w = 0; w < widht; w++)
				{
					for (int c = 0; c <channel ; c++)
					{
						const auto &dmax_element = dtop(b, h, w, c);
 
						int max_inpatch_height;
						int max_inpatch_width;
						if (isColMajor) {//如果是列主元存储，那么维度的序号刚好相反，由(b,h,w,c)变成(c,w,h,b)
							const Eigen::Tuple<Eigen::DenseIndex, float>&v = reduced_by_dims(c*widht*height*batch + w*height*batch + h*batch + b);
							int index_in_patch = v.first % (m_wksize*m_hksize);//最大值在每个块中的索引
							max_inpatch_height = index_in_patch%m_hksize;
							max_inpatch_width = index_in_patch / m_hksize;
 
						}
						else{
							const Eigen::Tuple<Eigen::DenseIndex, float>&v = reduced_by_dims(b*height*widht*channel + h*widht*channel + w*channel + c);
							int index_in_patch = v.first % (m_wksize*m_hksize);//最大值在每个块中的索引
							max_inpatch_height = index_in_patch/m_wksize;
							max_inpatch_width = index_in_patch % m_wksize;
						}
 
						int patch_height = h*m_hstride + max_inpatch_height;
						int patch_width = w*m_wstride + max_inpatch_width;
						dpad_bottom(b, patch_height, patch_width, c) += dmax_element;
						/*if (patch_height < dbottom.dimension(1) && patch_width < dbottom.dimension(2))
						{
 
							dbottom(b, patch_height, patch_width, c) += dmax_element;
						}
						else
						{
							std::cout << "out of range" << std::endl;
						}*/
					}
				}
			}
		}
		CBaseFunction::padding_backward(dpad_bottom, m_padding, m_padding, dbottom);
 
	}
	//均值池化，也可以看成是卷积
	void CPoolingLayer::avgpooling_backward(const Tensor4xf&dtop, Tensor4xf&dbottom) {
 
		Tensor4xf mean_coffe = dtop*(1.f / (m_wksize*m_hksize));//均值池化反向求导要除以均值系数
 
		for (int b=0;b<mean_coffe.dimension(0);b++)
		{
			for (int h=0;h<mean_coffe.dimension(1);h++)
			{
				for (int w=0;w<mean_coffe.dimension(2);w++)
				{
					for (int c=0;c<mean_coffe.dimension(3);c++)
					{
						const auto &mean_element= mean_coffe(b, h, w, c);
						for (int kh=0;kh<m_hksize;kh++)
						{
							for (int kw=0;kw<m_wksize;kw++)
							{
								int patch_height = h*m_hstride + kh - m_padding;
								int patch_width = w*m_wstride + kw - m_padding;
								if (patch_height>=0 &&patch_width>=0&&patch_width<dbottom.dimension(2)&&patch_height<dbottom.dimension(1))
								{
									dbottom(b, patch_height, patch_width,c) += mean_element;
								}
							}
						}
					}
				}
			}
		}
 
		//CBaseFunction::padding_backward(dpad_bottom, m_padding_method, m_padding_method, dbottom);
 
	}
	void CPoolingLayer::backward(const Tensor4xf&bottom,const Tensor4xf&dtop, Tensor4xf&dbottom, const Eigen::ThreadPoolDevice &device) {
		dbottom.setZero();
		//计算第2、3维的降维
		switch (m_pooling_method)
		{
		case max:
			maxpooling_backward(bottom, dtop, dbottom);
			break;
		case avg:
			avgpooling_backward(dtop, dbottom);
			break;
		default:
			break;
		}
 
	}
private:
	int m_hksize;//池化块的长宽
	int m_wksize;
	int m_hstride;//池化步长
	int m_wstride;
 
	int m_padding;//边界处理方法
	PoolingMethod m_pooling_method;//池化方法：均值池化、最大池化等
 
};
 
class CPoolingLayer_test
{
public:
	static void CPoolingLayer_test::test() {
 
		Eigen::ThreadPool *tp = new Eigen::ThreadPool(8);
		Eigen::ThreadPoolDevice device(tp, 8);
 
		int batch_size = 1;
		int input_channel =1;
		int input_height =5;
		int input_width =5;
		int kenel_height = 3;
		int kenel_widht = 2;
		int khstride =2;
		int kwstride = 3;
		int pad = 0;
 
		Tensor4xf bottom(batch_size, input_height, input_width, input_channel);
		int count = 0;
		for (int i=0;i<batch_size;i++)
		{
			for (int j=0;j<input_height;j++)
			{
				for (int k=0;k<input_width;k++)
				{
					for (int h=0;h<input_channel;h++)
					{
						bottom(i, j, k, h) = 0.1f*count;
						count++;
					}
				}
			}
		}
 
		Tensor1xf label_1d(batch_size);
		for (int i = 0; i < batch_size; i++)
		{
			label_1d(i) = i;
		}
 
		//第一层：pooling层
		CPoolingLayer layer({kenel_height,kenel_widht,khstride,kwstride },PoolingMethod::max,pad);
 
		Tensor4xf top;
		layer.forward(bottom, top,device);
 
		Tensor2xf top_flatten;
		CBaseFunction::flatten(top, top_flatten);
 
		//第二层：sotfmax网络层
		Tensor2xf one_hot;
		CBaseFunction::onehot(label_1d, top_flatten.dimension(1), one_hot);
		Tensor2xf dtop_flatten(top_flatten);
		float loss = CBaseFunction::softmax_with_loss(top_flatten, one_hot, dtop_flatten, device);
 
		Tensor4xf dtop;
		CBaseFunction::reshape_like(dtop_flatten, top, dtop);
 
		Tensor4xf dbottom(bottom);
 
		layer.backward(bottom, dtop,dbottom,device);
 
		//Tensor4rf dbottom_swap = dbottom.swap_layout();
		std::cout << "***************forward************" << std::endl;
		//CBaseFunction::print_shape(one_hot);
		CBaseFunction::print_shape(dbottom);
		CBaseFunction::print_element(dbottom);
		//std::cout << "bottom" << bottom<< std::endl;
		//std::cout << "top" << top << std::endl;
		//std::cout << "dbottom" << dbottom << std::endl;
		std::cout << "loss" << loss << std::endl;
 
		//std::cout << "dbottom" << dbottom << std::endl;
		//std::cout << "dtop" << top << std::endl;
 
	}
 
};

二、tensorflow 验证结果：

import  tensorflow as tf
 
batch_size = 1
input_channel = 1
input_height =5
input_width = 5
 
kenel_height =3
kenel_widht =2
khstride =2
kwstride=3
pad=0
bottom=tf.constant([i*0.1 for i in range(batch_size*input_channel*input_height*input_width)],shape=(batch_size,input_height,input_width,input_channel),dtype=tf.float32)
 
pool1=tf.nn.max_pool(tf.pad(bottom,[[0,0],[pad,pad],[pad,pad],[0,0]]),[1,kenel_height,kenel_widht,1],strides=[1,khstride,kwstride,1],padding='VALID')
pool_flatten=tf.reshape(pool1,[batch_size,-1])
 
label=tf.constant([i for i in range(batch_size)])
one_hot=tf.one_hot(label,pool_flatten.get_shape().as_list()[1])
 
predicts=tf.nn.softmax(pool_flatten)
loss =-tf.reduce_mean(one_hot * tf.log(predicts))
 
#打印相关变量，梯度等，验证是否与c++结果相同
dbottom,dpool1=tf.gradients(loss,[bottom,pool1])
 
with tf.Session() as sess:
	sess.run(tf.global_variables_initializer())
 
	print (sess.run([dbottom]))
 
	print (sess.run(loss))
 
	#print ('dbottom_data',dbottom_data)