这是一份充满野指针的只是用来理解的玩具代码
最近用耶鲁人脸库简单实现了人脸表情识别,写几篇博客留个纪念。
分为 BP神经网络 与 实现人脸识别 两篇博客,对于遇到的问题与代码进行总结分析。
第一篇需要理解 BP神经网络的原理,先给出代码再总结。
直接复制粘贴 代码很丑。。。
//DateSet.h
//created by WK
#ifndef DATASET_H
#define DATASET_H
#include <vector>
#include <cfloat>
using namespace std;
//数据集
class DataSet
{
private:
int InputNum;
int OutputNum;
vector<double*> DataMap;
double *Max;
double *Mini;
public:
DataSet(int inputnum, int outputnum)
{
this->InputNum = inputnum;
this->OutputNum = outputnum;
this->Max = NULL;
this->Mini = NULL;
}
void AddRow(double *inputArray, double *outputArray)
{
double *data = new double[this->InputNum + this->OutputNum];
for (int i = 0; i<this->InputNum; ++i)
{
data[i] = inputArray[i];
}
for (int i = 0; i<this->OutputNum; ++i)
{
data[InputNum + i] = outputArray[i];
}
this->DataMap.push_back(data);
}
void Normaliz()
{
if (this->Max)
{
delete this->Max;
this->Max = NULL;
}
if (this->Mini)
{
delete this->Mini;
this->Mini = NULL;
}
this->Max = new double[this->InputNum + this->OutputNum];
this->Mini = new double[this->InputNum + this->OutputNum];
for (int i = 0; i < this->InputNum + this->OutputNum; ++i)
{
double max = INT_MIN;
double mini = DBL_MAX;
for (int j = 0; j < DataMap.size(); ++j)
{
max = (max > DataMap[j][i] ? max : DataMap[j][i]);
mini = (mini < DataMap[j][i] ? mini : DataMap[j][i]);
}
if (max == mini)
continue;
for (int j = 0; j < DataMap.size(); ++j)
{
DataMap[j][i] = (DataMap[j][i] - mini) / (max - mini);
}
Max[i] = max;
Mini[i] = mini;
}
}
double *GetMax()
{
return this->Max;
}
double *GetMini()
{
return this->Mini;
}
int GetInputNum()
{
return this->InputNum;
}
int GetOutputNum()
{
return this->OutputNum;
}
int GetRows()
{
return DataMap.size();
}
vector<double*> GetDataMap()
{
return DataMap;
}
};
#endif // !DATASET_H
这是对数据集的简单封装
private:
int InputNum;
int OutputNum;
vector<double*> DataMap;
double *Max;
double *Mini;
1.保留输入输出的变量的个数,每一行数据,前InputNum 个为输入 后OutputNum 个为输出
2.采用double* 的vector 便于动态添加数据,使用AddRow添加一行数据
3.Max Mini 分别保存了每一列数据的最大最小值,注意是每一列即不同组别的同类数据,用于归一化
4.Normaliz 采用常见的归一化,分别将每一列进行归一化,即不同组别的同类数据进行归一化
/********************************************************************************************************************************************************************/
//TransferFunc.h
//created by WK
#ifndef TRANSFERFUNC_H
#define TRANSFERFUNC_H
#include <cmath>
enum Functypes
{
FUNCTYPE_TANH,
FUNCTYPE_STEP,
FUNCTYPE_LINEAR,
FUNCTYPE_SIGMOID,
FUNCTYPE_SGN,
FUNVTYPE_RAMP
};
class Function
{
private:
double Step(double input)
{
if (input <= 0)
return 0;
else
return 1;
}
double Linear(double input)
{
return input;
}
double Sigmoid(double input)
{
return 1.0 / (1.0 + exp(-1.0*input));
}
double Sgn(double input)
{
if (input < 0)
return -1;
else
return 1;
}
double Ramp(double input)
{
if (input < 0)
return 0;
else if (input >= 0 && input <= 1)
return input;
else
return 1;
}
public:
double GetResult(int funcType, double input)
{
switch (funcType)
{
case FUNCTYPE_TANH:
return tanh(input);
case FUNCTYPE_STEP:
return Step(input);
case FUNCTYPE_LINEAR:
return Linear(input);
case FUNCTYPE_SIGMOID:
return Sigmoid(input);
case FUNCTYPE_SGN:
return Sgn(input);
case FUNVTYPE_RAMP:
return Ramp(input);
default:
return input;
}
}
};
#endif // !TRANSFERFUNC_H
这是对传递函数的简单封装
1.传入函数类型以及输入数据,即可获取相应函数计算所得结果
/****************************************************************************************************************************************************************/
//net.h
//created by WK
#ifndef NET_H
#define NET_H
#include <vector>
#include <iostream>
#include "TransferFunc.h"
#include "DataSet.h"
#include <cstdlib>
using namespace std;
#define WINITVALUE 0.001
#define TINITVALUE 0
//神经元
class Neuron
{
private:
double Input;
double Output;
double Threshold;
double *Last_weight; //神经元维护后向的权重
int LastLayerNeuNum;
int TransferFunctionType;
Function Transferfunction;
public:
Neuron(double threshold, int lastlayerneunum, int funcType)
{
this->Input = 0;
this->Output = 0;
this->Threshold = threshold;
this->LastLayerNeuNum = lastlayerneunum;
this->TransferFunctionType = funcType;
this->Last_weight = new double[lastlayerneunum];
//关键的初始化权值
for (int i = 0; i < lastlayerneunum; ++i)
this->Last_weight[i] = (2.0*(double)rand() / RAND_MAX) - 1;
}
void SetInput(double input)
{
this->Input = input;
}
double GetOutput()
{
this->Output = Transferfunction.GetResult(this->TransferFunctionType, this->Input - this->Threshold);
return this->Output;
}
double* GetThreshold()
{
return &this->Threshold;
}
double *GetWeight()
{
return this->Last_weight;
}
void SetFuncType(int functype)
{
this->TransferFunctionType = functype;
}
};
//多层感知机
class MultiLayerPerceptron
{
private:
int OutTransfetFunctionType;
int HideTransfetFunctionType;
int InTransfetFunctionType;
int InLayerNeuNum;
int HideLayerNeuNum;
int OutLayerNeuNum;
double Speed;
Neuron **InputNeurons;
Neuron **OutputNeurons;
Neuron **HidenNeurons;
public:
MultiLayerPerceptron(int intransferfunctiontype, int inLayerNeuNum, int hidetransferfunctiontype, int hideLayerNeuNum, int outtransferfunctiontype, int outLayerNeuNum, double speed)
{
this->InTransfetFunctionType = intransferfunctiontype;
this->HideTransfetFunctionType = hidetransferfunctiontype;
this->OutTransfetFunctionType = outtransferfunctiontype;
this->InLayerNeuNum = inLayerNeuNum;
this->HideLayerNeuNum = hideLayerNeuNum;
this->OutLayerNeuNum = outLayerNeuNum;
this->Speed = speed;
this->InputNeurons = (Neuron**)new void*[inLayerNeuNum];
for (int i = 0; i < inLayerNeuNum; ++i)
this->InputNeurons[i] = new Neuron(TINITVALUE, 0, intransferfunctiontype);
this->HidenNeurons = (Neuron**)new void*[hideLayerNeuNum];
for (int i = 0; i < hideLayerNeuNum; ++i)
this->HidenNeurons[i] = new Neuron(TINITVALUE, inLayerNeuNum, hidetransferfunctiontype);
this->OutputNeurons = (Neuron**)new void*[outLayerNeuNum];
for (int i = 0; i < outLayerNeuNum; ++i)
this->OutputNeurons[i] = new Neuron(TINITVALUE, hideLayerNeuNum, outtransferfunctiontype);
}
//获取正向的输出
void GetOutput(double *output)
{
double sum;
for (int i = 0; i < this->HideLayerNeuNum; ++i)
{
sum = 0;
for (int j = 0; j < this->InLayerNeuNum; ++j)
sum += this->HidenNeurons[i]->GetWeight()[j] * this->InputNeurons[j]->GetOutput();
this->HidenNeurons[i]->SetInput(sum);
}
for (int i = 0; i < this->OutLayerNeuNum; ++i)
{
sum = 0;
for (int j = 0; j < this->HideLayerNeuNum; ++j)
sum += this->OutputNeurons[i]->GetWeight()[j] * this->HidenNeurons[j]->GetOutput();
this->OutputNeurons[i]->SetInput(sum);
output[i] = this->OutputNeurons[i]->GetOutput();
}
}
//学习所有数据一次
void Learn(DataSet *trainingSet)
{
double *expect;
double *data;
double *output = new double[this->OutLayerNeuNum];
for (int i = 0; i < trainingSet->GetRows(); ++i)
{
data = trainingSet->GetDataMap()[i];
expect = data + trainingSet->GetInputNum();
for (int j = 0; j < trainingSet->GetInputNum(); ++j)
this->InputNeurons[j]->SetInput(data[j]);
this->GetOutput(output);
//更改隐藏层到输出层权重以及阈值
//更新公式详见机器学习
for (int j = 0; j < this->OutLayerNeuNum; ++j)
{
double delta = this->Speed * output[j] * (1 - output[j]) * (expect[j] - output[j]);
for (int k = 0; k < this->HideLayerNeuNum; ++k)
this->OutputNeurons[j]->GetWeight()[k] += (delta * this->HidenNeurons[k]->GetOutput());
*this->OutputNeurons[j]->GetThreshold() -= delta;
}
//更改输入层到隐藏层的权重以及阈值
//更新公式详见机器学习
for (int j = 0; j < this->HideLayerNeuNum; ++j)
{
double t = 0;
for (int k = 0; k < this->OutLayerNeuNum; ++k)
t += (this->OutputNeurons[k]->GetWeight()[j] * output[k] * (1 - output[k])*(expect[k] - output[k]));
double delta = this->HidenNeurons[j]->GetOutput() * (1 - this->HidenNeurons[j]->GetOutput()) * t;
for (int k = 0; k < this->InLayerNeuNum; ++k)
this->HidenNeurons[j]->GetWeight()[k] += (this->Speed * this->InputNeurons[k]->GetOutput() * delta);
*this->HidenNeurons[j]->GetThreshold() -= (this->Speed * delta);
}
}
}
void Test(DataSet *Set)
{
double *output = new double[this->OutLayerNeuNum];
double *expect = new double[this->OutLayerNeuNum];
for (int i = 0; i < Set->GetRows(); ++i)
{
for (int j = 0; j < Set->GetInputNum(); ++j)
this->InputNeurons[j]->SetInput(Set->GetDataMap()[i][j]);
this->GetOutput(output);
for (int j = 0; j < Set->GetOutputNum(); ++j)
{
cout << "output: ";
cout << output[j] << " ";
cout << "expect: ";
cout << Set->GetDataMap()[i][Set->GetInputNum() + j] << " ";
}
cout << endl;
}
cout << endl;
cout << "in to hide Weight:" << endl;
for (int i = 0; i < this->HideLayerNeuNum; ++i)
{
for (int j = 0; j < this->InLayerNeuNum; ++j)
{
cout << this->HidenNeurons[i]->GetWeight()[j] << " ";
}
cout << endl;
}
cout << endl;
cout << "hide to out Weight:" << endl;
for (int i = 0; i < this->OutLayerNeuNum; ++i)
{
for (int j = 0; j < this->HideLayerNeuNum; ++j)
{
cout << this->OutputNeurons[i]->GetWeight()[j] << " ";
}
cout << endl;
}
}
};
#endif // !NET_H
以上为 神经元 以及 BP神经网络的代码 以下分开说
/****************************************************************************************************************************************/
//神经元
class Neuron
{
private:
double Input;
double Output;
double Threshold;
double *Last_weight; //神经元维护后向的权重
int LastLayerNeuNum;
int TransferFunctionType;
Function Transferfunction;
public:
Neuron(double threshold, int lastlayerneunum, int funcType)
{
this->Input = 0;
this->Output = 0;
this->Threshold = threshold;
this->LastLayerNeuNum = lastlayerneunum;
this->TransferFunctionType = funcType;
this->Last_weight = new double[lastlayerneunum];
//关键的初始化权值
for (int i = 0; i < lastlayerneunum; ++i)
this->Last_weight[i] = (2.0*(double)rand() / RAND_MAX) - 1;
}
void SetInput(double input)
{
this->Input = input;
}
double GetOutput()
{
this->Output = Transferfunction.GetResult(this->TransferFunctionType, this->Input - this->Threshold);
return this->Output;
}
double* GetThreshold()
{
return &this->Threshold;
}
double *GetWeight()
{
return this->Last_weight;
}
void SetFuncType(int functype)
{
this->TransferFunctionType = functype;
}
};
以上为神经元
1.输入值,输出值,阈值,以及传输函数类型无需多言,构造必须指定 阈值,上一层神经元个数用于分配内存给权重数组,以及传输函数类型
2.每一个神经元维护后向的权值Last_weight 举例
s前1 代表 前一层的1号神经元输出 s后1 代表后一层1号神经元的输入 w表示权重 有:
s后1 = w1 * s前1 + w2 * s前2....... 暂时忽略阈值与传输函数
即由 s后1 这个神经元维护一个数组,这个数组的元素是 w1 w2 w3
(权重的维护有多种实现方式 随意)
/***********************************************************************************************************************************************************************************/
//多层感知机
class MultiLayerPerceptron
{
private:
int OutTransfetFunctionType;
int HideTransfetFunctionType;
int InTransfetFunctionType;
int InLayerNeuNum;
int HideLayerNeuNum;
int OutLayerNeuNum;
double Speed;
Neuron **InputNeurons;
Neuron **OutputNeurons;
Neuron **HidenNeurons;
public:
MultiLayerPerceptron(int intransferfunctiontype, int inLayerNeuNum, int hidetransferfunctiontype, int hideLayerNeuNum, int outtransferfunctiontype, int outLayerNeuNum, double speed)
{
this->InTransfetFunctionType = intransferfunctiontype;
this->HideTransfetFunctionType = hidetransferfunctiontype;
this->OutTransfetFunctionType = outtransferfunctiontype;
this->InLayerNeuNum = inLayerNeuNum;
this->HideLayerNeuNum = hideLayerNeuNum;
this->OutLayerNeuNum = outLayerNeuNum;
this->Speed = speed;
this->InputNeurons = (Neuron**)new void*[inLayerNeuNum];
for (int i = 0; i < inLayerNeuNum; ++i)
this->InputNeurons[i] = new Neuron(TINITVALUE, 0, intransferfunctiontype);
this->HidenNeurons = (Neuron**)new void*[hideLayerNeuNum];
for (int i = 0; i < hideLayerNeuNum; ++i)
this->HidenNeurons[i] = new Neuron(TINITVALUE, inLayerNeuNum, hidetransferfunctiontype);
this->OutputNeurons = (Neuron**)new void*[outLayerNeuNum];
for (int i = 0; i < outLayerNeuNum; ++i)
this->OutputNeurons[i] = new Neuron(TINITVALUE, hideLayerNeuNum, outtransferfunctiontype);
}
//获取正向的输出
void GetOutput(double *output)
{
double sum;
for (int i = 0; i < this->HideLayerNeuNum; ++i)
{
sum = 0;
for (int j = 0; j < this->InLayerNeuNum; ++j)
sum += this->HidenNeurons[i]->GetWeight()[j] * this->InputNeurons[j]->GetOutput();
this->HidenNeurons[i]->SetInput(sum);
}
for (int i = 0; i < this->OutLayerNeuNum; ++i)
{
sum = 0;
for (int j = 0; j < this->HideLayerNeuNum; ++j)
sum += this->OutputNeurons[i]->GetWeight()[j] * this->HidenNeurons[j]->GetOutput();
this->OutputNeurons[i]->SetInput(sum);
output[i] = this->OutputNeurons[i]->GetOutput();
}
}
//学习所有数据一次
void Learn(DataSet *trainingSet)
{
double *expect;
double *data;
double *output = new double[this->OutLayerNeuNum];
for (int i = 0; i < trainingSet->GetRows(); ++i)
{
data = trainingSet->GetDataMap()[i];
expect = data + trainingSet->GetInputNum();
for (int j = 0; j < trainingSet->GetInputNum(); ++j)
this->InputNeurons[j]->SetInput(data[j]);
this->GetOutput(output);
//更改隐藏层到输出层权重以及阈值
//更新公式详见机器学习
for (int j = 0; j < this->OutLayerNeuNum; ++j)
{
double delta = this->Speed * output[j] * (1 - output[j]) * (expect[j] - output[j]);
for (int k = 0; k < this->HideLayerNeuNum; ++k)
this->OutputNeurons[j]->GetWeight()[k] += (delta * this->HidenNeurons[k]->GetOutput());
*this->OutputNeurons[j]->GetThreshold() -= delta;
}
//更改输入层到隐藏层的权重以及阈值
//更新公式详见机器学习
for (int j = 0; j < this->HideLayerNeuNum; ++j)
{
double t = 0;
for (int k = 0; k < this->OutLayerNeuNum; ++k)
t += (this->OutputNeurons[k]->GetWeight()[j] * output[k] * (1 - output[k])*(expect[k] - output[k]));
double delta = this->HidenNeurons[j]->GetOutput() * (1 - this->HidenNeurons[j]->GetOutput()) * t;
for (int k = 0; k < this->InLayerNeuNum; ++k)
this->HidenNeurons[j]->GetWeight()[k] += (this->Speed * this->InputNeurons[k]->GetOutput() * delta);
*this->HidenNeurons[j]->GetThreshold() -= (this->Speed * delta);
}
}
}
void Test(DataSet *Set)
{
double *output = new double[this->OutLayerNeuNum];
double *expect = new double[this->OutLayerNeuNum];
for (int i = 0; i < Set->GetRows(); ++i)
{
for (int j = 0; j < Set->GetInputNum(); ++j)
this->InputNeurons[j]->SetInput(Set->GetDataMap()[i][j]);
this->GetOutput(output);
for (int j = 0; j < Set->GetOutputNum(); ++j)
{
cout << "output: ";
cout << output[j] << " ";
cout << "expect: ";
cout << Set->GetDataMap()[i][Set->GetInputNum() + j] << " ";
}
cout << endl;
}
cout << endl;
cout << "in to hide Weight:" << endl;
for (int i = 0; i < this->HideLayerNeuNum; ++i)
{
for (int j = 0; j < this->InLayerNeuNum; ++j)
{
cout << this->HidenNeurons[i]->GetWeight()[j] << " ";
}
cout << endl;
}
cout << endl;
cout << "hide to out Weight:" << endl;
for (int i = 0; i < this->OutLayerNeuNum; ++i)
{
for (int j = 0; j < this->HideLayerNeuNum; ++j)
{
cout << this->OutputNeurons[i]->GetWeight()[j] << " ";
}
cout << endl;
}
}
};
以上为bp神经网络
1.构造需指定 输入层神经元个数,输入层神经元传输函数类型,隐藏层(单隐)神经元个数,隐藏层神经元传输函数类型,输出层神经元个数,输出层神经元传输函数类型,学习速率
2.Neuron** 的成员维护一层的神经元,保存指针数组
3.Getoutput 设置输入层各神经元的input 调用Getoutput 获取输出神经元的输出值
4.Learn 根据 周志华所著的机器学习中的权重更新公式,更新权重,这里的代码略为臃肿
5.Test 用于测试
/********************************************************************************************************************************************************************************************/
以下为神经网络测试 解决异或问题
#include "net.h"
int main()
{
DataSet *trainingSet = new DataSet(2, 1);
trainingSet->AddRow(new double[2]{ 1,1 }, new double[1]{ 0 });
trainingSet->AddRow(new double[2]{ 1,0 }, new double[1]{ 1 });
trainingSet->AddRow(new double[2]{ 0,1 }, new double[1]{ 1 });
trainingSet->AddRow(new double[2]{ 0,0 }, new double[1]{ 0 });
//层激励函数类型 神经元个数... 学习速率
MultiLayerPerceptron *m = new MultiLayerPerceptron(FUNCTYPE_LINEAR, 2, FUNCTYPE_SIGMOID, 5, FUNCTYPE_SIGMOID, 1, 0.9);
//学习1000次
for (int i = 0; i < 1000; ++i)
m->Learn(trainingSet);
m->Test(trainingSet);
system("pause");
}