参考以上文件，用C语言
神经网络模型，编写代码，预测成熟度。改成一下几种特征firmness hue saturation brightness color_category sound_db weight_g size_cm3 最后成熟度有5种程度
#include <iostream>
#include "rapidcsv.h"
#include <opencv2/opencv.hpp>
using namespace std;
using namespace cv;
using namespace cv::ml;
typedef struct {
double x1,x2,x3,x4; //嘴峰长度、嘴峰深度、脚掌长度、 体重
int x5_0,x5_1; //是否male(1:male,0:female)；是否famale(1:female,0:male)
int label; //类别标签, 0:Adelie,1:Gentoo,2: Chinstrap
} DataPoint;
int ReadCsv(const char* fileName, DataPoint data[]) {
rapidcsv::Document doc(fileName, rapidcsv::LabelParams(),
rapidcsv::SeparatorParams(),rapidcsv::ConverterParams(true));//解析csv文件
int j=0;//样本计数
for (int i=0;i<doc.GetRowCount();i++){ //每次处理一行
double bill_length_mm = doc.GetCell<double>("bill_length_mm", i);//嘴峰长度
double bill_depth_mm = doc.GetCell<double>("bill_depth_mm", i);//嘴峰深度
double flipper_length_mm = doc.GetCell<double>("flipper_length_mm", i);//脚掌长度
double body_mass_g = doc.GetCell<double>("body_mass_g", i);//体重
std::string sex = doc.GetCell<std::string>("sex", i); //性别
std::string type_name = doc.GetCell<std::string>("species", i); //种族
if ( !std::isnan(bill_length_mm) && !std::isnan(bill_depth_mm)
&& !std::isnan(flipper_length_mm) && !std::isnan(body_mass_g) && sex!="NA" ){//处理缺失值
data[j].x1 = bill_length_mm;data[j].x2 = bill_depth_mm;
data[j].x3 = flipper_length_mm;data[j].x4 = body_mass_g;
data[j].x5_0 = sex=="male"?1:0; data[j].x5_1 = sex=="female"?1:0;
if (type_name == "Adelie") data[j].label=0;
else if (type_name == "Gentoo") data[j].label=1;
else data[j].label=2;
j++;//剔除缺失值后的样本数目
}
}
return j;
}
int main() {
DataPoint data[500];
int n = ReadCsv("penguins.csv", data);//从文件读入数据
//将数据拷贝到Mat矩阵中 , 75%训练数据，25%测试数据
int n1=n*3/4, n2=n-n1;
Mat train_samples(n1, 6, CV_32F); //训练数据集
Mat train_labels(n1, 3, CV_32F);//训练数据的标签（独热编码）
Mat test_samples(n2, 6, CV_32F); //测试数据集
Mat test_labels(n2, 1, CV_32S);//测试数据的标签
int i=0, j=0, k=0;
for(i = 0; i<n; i++){
if (i%4==0){//每3个训练数据搭配1个测试数据
test_samples.at<float>(j, 0) = data[i].x1;test_samples.at<float>(j, 1) = data[i].x2;
test_samples.at<float>(j, 2) = data[i].x3;test_samples.at<float>(j, 3) = data[i].x4;
test_samples.at<float>(j, 4) = data[i].x5_0;test_samples.at<float>(j, 5) = data[i].x5_1;
test_labels.at<int>(j, 0) = data[i].label; j++;
}else{
train_samples.at<float>(k, 0) = data[i].x1;train_samples.at<float>(k, 1) = data[i].x2;
train_samples.at<float>(k, 2) = data[i].x3;train_samples.at<float>(k, 3) = data[i].x4;
train_samples.at<float>(k, 4) = data[i].x5_0;train_samples.at<float>(k, 5) = data[i].x5_1;
if (data[i].label==0) {//标签转换为独热编码
train_labels.at<float>(k, 0)=1;train_labels.at<float>(k, 1)=0;train_labels.at<float>(k, 2)=0;
}else if (data[i].label==1) {
train_labels.at<float>(k, 0)=0;train_labels.at<float>(k, 1)=1;train_labels.at<float>(k, 2)=0;
}else{
train_labels.at<float>(k, 0)=0;train_labels.at<float>(k, 1)=0;train_labels.at<float>(k, 2)=1;
}
k++;
}
}
cout<<"n="<<n<<",n1="<<n1<<",n2="<<n2<<endl;
Ptr<ANN_MLP> nn = ANN_MLP::create(); //创建
神经网络模型
nn->setLayerSizes(vector<int>{6, 10, 3});//输入层6个节点，隐藏层10个节点，输出层3个节点（假设3类企鹅）
nn->setActivationFunction(ANN_MLP::SIGMOID_SYM);//激活函数目前只支持Sigmoid
nn->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER, 10000, 1e-5));//设置迭代次数与收敛条件
nn->setTrainMethod(ANN_MLP::
RPROP); //
弹性反向传播算法
nn->train(train_samples, ROW_SAMPLE, train_labels);//训练模型
Mat predictions; //预测结果
nn->predict(test_samples,predictions);//对测试集进行预测
int correct = 0; //正确分类的次数
for (int i = 0; i < predictions.rows; i++){
Point max_loc; // 找到预测值中的最大索引，即预测的类别
minMaxLoc(predictions.row(i), nullptr, nullptr, nullptr, &max_loc);
cout<<"predictions:"<<predictions.row(i)<<", max_loc.x:"<<max_loc.x<<", label:"<<test_labels.at<int>(i, 0)<<endl;
if (max_loc.x == test_labels.at<int>(i, 0)) correct++;
}
double accuracy = (double)correct / predictions.rows; //计算准确率
cout<<"Accuracy: " << accuracy * 100 << "%" << endl;
return 0;
}
本文介绍了弹性反向传播（Rprop）算法，解决了反向传播中的学习率选择难题和梯度弥散问题。Rprop为每个权重分配独立的学习率，权重更新不受误差函数梯度大小影响，仅受梯度符号影响，决定更新方向。同时，学习率会根据梯度符号变化动态调整，以优化搜索最小值的精度。
最低0.47元/天 解锁文章
扫一扫
2136
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
