c++实现LogisticRegression/用c++实现逻辑回归

今天没事学习c++，试着用c++实现了一个简单的逻辑回归。
因为重在学习c++语法和逻辑回归的具体实现细节，因此没有用到向量化矩阵化等操作，而是老老实实用遍历一步一步求解梯度。
代码具体实现了梯度下降、随梯度下降、批梯度下降三种学习算法。
本文仅包含代码实现部分，逻辑回归的理论可以参考：
http://blog.csdn.net/abcjennifer/article/details/7716281
代码包括两个头文件和一个cpp文件：

logisticregression.h

#ifndef _LOGISTICREGRESSION_H_#define _LOGISTICREGRESSION_H_#include <iostream>#include <fstream>#include <string>#include <vector>#include <assert.h>struct DataSet {    int numData;    int numFeature;    std::vector<std::vector<float> > features;    std::vector<int> labels;};class LogisticRegression {public:    LogisticRegression(int max_iter, float learn_rate, float tol);    ~LogisticRegression();    DataSet loadData(std::string filename);    void initWeights(int length);    std::vector<float> oneSampleGrident(std::vector<float> feature, int label);    void train(DataSet* dataset, std::string gdType);    std::vector<float> getWeight();    int predict(std::vector<float> feature);    float predict_proba(std::vector<float> feature);    float score(DataSet* dataset);private:    int maxIter_;    float learnRate_;    float tol_;    std::vector<float> weights;};#endif

common.h

#ifndef _COMMON_H_#define _COMMON_H_#include <iostream>#include <string>#include <vector>#include <cmath>namespace common {std::vector<std::string> Split(std::string line, char flag) {    std::vector<std::string> ret;    size_t start = 0;    size_t index = line.find_first_of(flag, start);    while (index != std::string::npos) {        ret.push_back(line.substr(start, index));        start = index + 1;        index = line.find_first_of(flag, start);    }    ret.push_back(line.substr(start, index));    return ret;}float Sigmoid(float inx) {    return 1 / (1 + exp(-inx));}}#endif

logisticregression.cpp

/* *author: july-zj *usage: g++ -std=c++0x LogisticRegression.cpp -I../include -o lr_test *run: ./lr ../data/iris_train.data ../data/iris_train.data 10 0.1 0.0001 gd *dataset: csvtype file, label is the last column **/#include "LogisticRegression.h"#include "common.h"#include <random>#include <chrono>LogisticRegression::LogisticRegression(int maxIter, float learnRate, float tol = 0.0001) {    this->maxIter_ = maxIter;    this->learnRate_ = learnRate;    this->tol_ = tol;}LogisticRegression::~LogisticRegression() {}DataSet LogisticRegression::loadData(std::string filename) {    std::ifstream ifile(filename);    DataSet dataset;    if (!ifile) {        std::cout << "can not open file: " << filename << std::endl;        return dataset;    }    std::string line;    while (getline(ifile, line)) {        std::vector<std::string> tokens = common::Split(line, ',');        std::vector<float> feature;        for (int i = 0; i < tokens.size(); ++i) {            if (i == tokens.size() - 1) {                dataset.labels.push_back(atoi(tokens[i].c_str()));            } else {                feature.push_back(atof(tokens[i].c_str()));            }        }        dataset.features.push_back(feature);        dataset.numData += 1;    }    dataset.numFeature = dataset.features[0].size();    return dataset; }void LogisticRegression::initWeights(int length) {    this->weights.push_back(1.0);    for (int i = 0; i < length; ++i) {        this->weights.push_back(1.0);       }}std::vector<float> LogisticRegression::oneSampleGrident(std::vector<float> feature, int label) {    std::vector<float> grident(this->weights.size(), 0.0);    float predY = predict_proba(feature);    grident[0] = predY - label;    for (int i = 0; i < feature.size(); ++i) {        grident[i + 1] = (predY - label) * feature[i];    }    return grident; }void LogisticRegression::train(DataSet* dataset, std::string gdType) {    DataSet* traindata = dataset;    int dataNum = traindata->numData;    int featureNum = traindata->numFeature;    initWeights(featureNum);    for (int iter = 0; iter < this->maxIter_; ++iter) {        if (gdType == "gd") {            std::vector<float> overallGrident(this->weights.size(), 0.0);            for (int i = 0; i < dataNum; ++i) {                std::vector<float> tmpGrid = oneSampleGrident(traindata->features[i], traindata->labels[i]);                for (int j = 0; j < tmpGrid.size(); ++j) {                    overallGrident[j] += tmpGrid[j];                }            }            for (int i = 0; i < this->weights.size(); ++i) {                this->weights[i] -= this->learnRate_ * overallGrident[i] / dataNum;            }            float loss = 0.0;            for (int i = 0; i < dataNum; ++i) {                float predY = predict_proba(traindata->features[i]);                if (predY < 1e-5) predY = 1e-5;                if (predY > 1 - 1e-5) predY = 1 - 1e-5;                loss -= (traindata->labels[i] * log(predY) + (1 - traindata->labels[i]) * log(1 - predY));            }            loss /= dataNum;            std::cout << "iter " << iter << " loss is: " << loss << std::endl;            if (loss < this->tol_) {                std::cout << "early stoping for loss is less then tolerance" << std::endl;                return;            }        } else if (gdType == "sgd") {            std::vector<int> indexes;            for (int i = 0; i < dataNum; ++i) {                indexes.push_back(i);            }            unsigned seed = std::chrono::system_clock::now().time_since_epoch().count();            std::shuffle(indexes.begin(), indexes.end(), std::default_random_engine(seed));            for (int i : indexes) {                std::vector<float> grident = oneSampleGrident(traindata->features[i], traindata->labels[i]);                for (int j = 0; j < this->weights.size(); ++j) {                    this->weights[i] -= this->learnRate_ * grident[i];                }                float loss = 0.0;                for (int n = 0; n < dataNum; ++n) {                    float predY = predict_proba(traindata->features[n]);                    if (predY < 1e-5) predY = 1e-5;                    if (predY > 1 - 1e-5) predY = 1 - 1e-5;                    loss -= (traindata->labels[n] * log(predY) + (1 - traindata->labels[n]) * log(1 - predY));                }                loss /= dataNum;                std::cout << "iter " << iter << " loss is: " << loss << std::endl;                if (loss < this->tol_) {                    std::cout << "early stoping for loss is less then tolerance" << std::endl;                    return;                }            }        } else {            std::vector<int> indexes;            for (int i = 0; i < dataNum; ++i) {                indexes.push_back(i);            }            unsigned seed = std::chrono::system_clock::now().time_since_epoch().count();            std::shuffle(indexes.begin(), indexes.end(), std::default_random_engine(seed));            for (int i = 0; i < dataNum; i += 10) {                std::vector<float> batchGrident(this->weights.size(), 0.0);                int batchSize = 0;                for (int j = 0; j < 10; ++j) {                    int idx = i + j;                    if (idx >= dataNum) {                        break;                    }                    batchSize += 1;                    std::vector<float> tmpGrid = oneSampleGrident(traindata->features[idx], traindata->labels[idx]);                    for (int m = 0; m < tmpGrid.size(); ++m) {                        batchGrident[m] += tmpGrid[m];                    }                }                for (int n = 0; n < this->weights.size(); ++n) {                    this->weights[n] -= this->learnRate_ * batchGrident[n] / batchSize;                }                float loss = 0.0;                for (int i = 0; i < dataNum; ++i) {                    float predY = predict_proba(traindata->features[i]);                    if (predY < 1e-5) predY = 1e-5;                    if (predY > 1 - 1e-5) predY = 1 - 1e-5;                    loss -= (traindata->labels[i] * log(predY) + (1 - traindata->labels[i]) * log(1 - predY));                }                loss /= dataNum;                std::cout << "iter " << iter << " loss is: " << loss << std::endl;                if (loss < this->tol_) {                    std::cout << "early stoping for loss is less then tolerance" << std::endl;                    return;                }            }        }    }}std::vector<float> LogisticRegression::getWeight() {    return this->weights;   }int LogisticRegression::predict(std::vector<float> feature) {    float predY = predict_proba(feature);    if (predY > 0.5) {        return 1;    } else {        return 0;    }}float LogisticRegression::predict_proba(std::vector<float> feature) {    float summation = this->weights[0];    for (int i = 0; i < feature.size(); ++i) {        summation += this->weights[i + 1] * feature[i];    }    return common::Sigmoid(summation);}float LogisticRegression::score(DataSet* dataset) {    int rightNum = 0;    for (int i = 0; i < dataset->numData; ++i) {        if (predict(dataset->features[i]) == dataset->labels[i]) {            rightNum++;        }    }    std::cout << "right number: " << rightNum << std::endl;    return rightNum * 1.0 / dataset->numData; }int main(int argc, char* argv[]) {    if (argc < 7) {        std::cout << "usage: trainfile, valfile, maxiter, learnrate, tol, gdtype" << std::endl;        return -1;    }    std::string trainFile = std::string(argv[1]);    std::string valFile = std::string(argv[2]);    int maxIter = atoi(argv[3]);    float learnRate = atof(argv[4]);    float tol = atof(argv[5]);    std::string gdType = std::string(argv[6]);    assert(gdType == "gd" || gdType == "sgd" || gdType == "bachgd");    LogisticRegression lr(maxIter, learnRate);    DataSet trainData = lr.loadData(trainFile);    DataSet valData = lr.loadData(valFile);    lr.train(&trainData, gdType);    std::cout << "precistion is: " << lr.score(&trainData) << std::endl;    return 0;}

直接拷备过来编译即可，输入参数按usage说明来写。需要指定训练数据路径，验证数据路径，迭代次数，学习速率，训练结束的容差，训练方式等六个参数。