Logistic Regression的OpenCV代码



前言

本文的实现主要是参照了Andrew NG的机器学习课程所讲的内容。理论知识上一篇博文已经介绍。由于刚接触C++，代码写得比较粗糙，望见谅。

实验环境

Visual Studio 2013

OpenCV 2.4

数据

本次实验数据包含了手写数字0和1，每类样本分别有20个。如下图所示：

实验代码

1.定义一个LogisticRegression的类：

头文件 LogisticRegression.h

#ifndef  _LOGISTIC_REGRESSION_H_
#define  _LOGISTIC_REGRESSION_H_
#include <vector>
#include <string>
#include <iostream>
#include <strstream>
#include <fstream>
#include <opencv2\core\core.hpp>
using namespace cv;
using namespace std;
class LogisticRegression
{
public:
	//加载数据
	void loadData(string filePath);//xml
	//数据返回
	Mat& getX();
	Mat& getEX();
	Mat& getTheta();
	Mat& getY();
	//特征归一化
	void featureNormalize(Mat& x);
	//预测
	float predict(Mat& x);
	//假设函数，此处设成静态函数，是方便在其他类中直接调用
	static void calculateHx(Mat& x, Mat& theta, Mat& Hx);
	static void sigmoid(Mat& z, Mat& g);//激活函数
	//计算分类器的正确率
	float classiferAccuracy(Mat& x,Mat& y,Mat& theta);
private:
	Mat originalX;
	Mat extendX;
	Mat y;
	int dim;//数据维数
	Mat theta;//列向量
};
#endif

实现文件 LogisticRegression.cpp

#include "LogisticRegression.h"
//此处的训练样本数据存放在xml中
//包含了数字0和1，每类样本20个
void LogisticRegression::loadData(String filename)
{
	cout << "loading the dataset...";
	FileStorage f;
	if (f.open(filename, FileStorage::READ))//release模式下运行正常，debug模式下打不开
	{
		f["datamat"] >> originalX;
		f["labelsmat"] >> y;
		f.release();
	}
	else
	{
		cerr << "file can not be opened: " << filename << endl;
		return;
	}
	originalX.convertTo(originalX, CV_32F);
	y.convertTo(y, CV_32F);
	cout << "read " << originalX.rows << " rows of data" << endl;
	extendX = Mat::ones(originalX.rows, originalX.cols + 1, CV_32FC1);
	originalX.copyTo(extendX.colRange(1, extendX.cols));
	extendX.col(0) = Mat::ones(extendX.rows, 1, extendX.type());
	theta = Mat::zeros(originalX.cols + 1, 1, originalX.type());
}

//数据返回
Mat& LogisticRegression::getX()
{
	return originalX;
}

Mat& LogisticRegression::getY()
{
	return y;
}

Mat& LogisticRegression::getEX()
{
	return extendX;
}

Mat& LogisticRegression::getTheta()
{
	return theta;
}

//特征归一化
void LogisticRegression::featureNormalize(Mat& ex)
{
	if (ex.cols <= 2) return;//只有一个特征的时候不用归一化
	for (int col = 1; col < ex.cols; col++)
	{
		Mat mean;
		Mat stddev;
		meanStdDev(ex.col(col), mean, stddev);
		//归一化
		ex.col(col) = ex.col(col) - mean;
		ex.col(col) = ex.col(col) / stddev;
	}
}

//预测
//x:一个样本
//theta 训练得到的参数
float LogisticRegression::predict(Mat& ex)
{
	Mat hx;
	calculateHx(ex, this->getTheta(), hx);
	return hx.at<float>(0, 0)>0.5?1:0;
}

//计算假设函数
void LogisticRegression::calculateHx(Mat& x, Mat& theta, Mat& Hx)
{
	Mat thetaX = x*theta;
	sigmoid(thetaX, Hx);
}
//计算激活函数
void LogisticRegression::sigmoid(Mat& z, Mat& g)
{
	Mat expItem;
	exp(-z, expItem);
	g = 1.0 / (1 + expItem);
}

//计算分类器精度
float LogisticRegression::classiferAccuracy(Mat& x, Mat& y, Mat& theta)
{
	Mat hx;
	calculateHx(x, theta,hx);
	Mat predict = (hx > 0.5)/255;
	predict.convertTo(predict, y.type());
	return 100 * (float)countNonZero(y == predict) / predict.rows;
}

2.定义一个梯度下降的类GradientDescent：
头文件 GradientDescent.h

#ifndef _GRADIENTDESCENT_H
#define _GRADIENTDESCENT_H
#include <opencv2\core\core.hpp>
using namespace cv;
class GradientDescent
{
public:
	//批量梯度下降
	void gradientDescent(Mat& x, Mat& y, Mat &theta, int num_iters, float alpha);
	//计算代价
	float computeCost(Mat& x, Mat& y, Mat &theta);
};
#endif

实现文件 GradientDescent.cpp

#include "GradientDescent.h"
#include "LogisticRegression.h"
#include <iostream>
using namespace std;
void GradientDescent::gradientDescent(Mat& x, Mat& y, Mat &theta, int num_iters, float alpha)
{
	int dataSize = x.rows;//数据总数
	int dim = theta.rows;//theta的维数
	for (int i = 0; i < num_iters; i++)
	{
		//方法1：for-loop
		for (int j = 0; j < dim; j++)
		{
			float cost = 0.0;
			for (int k = 0; k < dataSize; k++)
			{
				Mat xRow = x.row(k);
				Mat yRow = y.row(k);
				Mat hx;
				LogisticRegression::calculateHx(xRow, theta, hx);
				Mat tmp = (hx - yRow)*xRow.col(j);
				cost += tmp.at<float>(0, 0);
			}
			theta.at<float>(j, 0) -= alpha*cost / dataSize;
		}
	}
}

float GradientDescent::computeCost(Mat& x, Mat& y, Mat &theta)
{
	float error = .0;
	Mat hx;
	Mat j;
	LogisticRegression::calculateHx(x,theta,hx);
	Mat logHx;
	Mat logHx1;
	log(hx, logHx);
	log(1 - hx, logHx1);
	j = -(y.t()*logHx + (1 - y.t())*logHx1);
	error = j.at<float>(0, 0)/x.rows;
	return error;
}

3.代码测试：

样本集来自mnist数据库，可在官网下载http://yann.lecun.com/exdb/mnist/

为了避免每次都从图片中去读数据，可用下面的语句把已经读好的样本数据和类标号保存到xml文件中

	FileStorage f(".\\ex_data.xml", FileStorage::WRITE);
	f<<"datamat"<< originalX;//originalX:存放原始数据的Mat
	f<<"labelsmat"<< y;//y:存放类标号的Mat
	f.release()

<span style="font-size: 14px;">
</span>

main.cpp内容如下：

#include "LogisticRegression.h"
#include "GradientDescent.h"
#include <iostream>
#include <opencv2\opencv.hpp>
using namespace cv;
using namespace std;

int main()
{
	LogisticRegression regression;
	const String filename = "data/data01.xml";
	regression.loadData(filename);
	GradientDescent gd;
	float cost=gd.computeCost(regression.getEX(), regression.getY(), regression.getTheta());
	cout << "cost:" << cost << endl;
	gd.gradientDescent(regression.getEX(), regression.getY(), regression.getTheta(), 400, 0.01);
	cout<<"第1个样本的分类结果："<<regression.predict(regression.getEX().row(0))<<endl;
	cout << "最后一个样本的分类结果：" << regression.predict(regression.getEX().row(39))<<endl;
	float accuracy=regression.classiferAccuracy(regression.getEX(), regression.getY(), regression.getTheta());
	cout << "分类器对训练集的正确率为：" << accuracy <<"%"<< endl;
	getchar();
}

输出结果：

由于样本量少，同时0和1比较容易区分，因此在训练集上达到了100%的正确率