神经网络(c++)【转】

来自：http://blog.csdn.net/calcular/article/details/50392502

#ifndef DATASER_H#define DATASER_H#include<iostream>#include<math.h>#include<time.h>#include<stdlib.h>#include<algorithm>#include<queue>using namespace std;#define NETL 5//输入维度#define NETX 10//每层多少个单元,大于输入维度/*产生正太分布*/double gaussrand(){static double V1, V2, S;static int phase = 0;double X;if (phase == 0) {do {double U1 = (double)rand() / RAND_MAX;double U2 = (double)rand() / RAND_MAX;V1 = 2 * U1 - 1;V2 = 2 * U2 - 1;S = V1 * V1 + V2 * V2;} while (S >= 1 || S == 0);X = V1 * sqrt(-2 * log(S) / S);}elseX = V2 * sqrt(-2 * log(S) / S);phase = 1 - phase;return X;}/* 找到最小值 */int fminm(double *in,int dim,double &minm)//已绝对化处理{int i,c;c = -1;double minx = 99999.9;for (i = 0;i < dim;i++){if (fabs(in[i]) < minx){minx = fabs(in[i]);c = i;}}minm = minx;return c;}/*找到第二小*/int fmin_d2(double *in, int dim,const double minm)//三个都是输入{int i, c;c = -1;double minx = 99999.9;for (i = 0;i < dim;i++){if (fabs(in[i]) < minx){if (fabs(in[i]) != minm){minx = fabs(in[i]);c = i;}}}return c;}/* 打印一行数据 */void disp(const double *data,const int dim){int i;for (i = 0;i < dim;i++){printf("%-10f ", data[i]);}printf("\n");}/*计算单元:多维降一维in:输入信息w:权重dim:维度*/double calo(const double *in,const double *w,const int dim){double r = 0;int i;for (i = 0;i < dim;i++){r += in[i] * w[i];}r = 1.0 / (1.0 + exp(-1.0*r));return r;}/* 降维得到误差值 */double acalo(const double *in,const double *w,const int dim,const double expect){double r;r = calo(in, w, dim);//线性降维return r - expect;//当前点在期望点上方还是下方,相对位置}/* 单次维度系数调整 用法示例:error[i] = tra(in, w[i], N, ex, offset[i], e[i]);*/double tra(//返回更改权重后的误差const double *in,double *w,//会得到更改的权重系数const int dim,const double expect,double *offset,//权重调整值记录数组double &e//得到误差改进值){double r,err,re,v;int i;r=calo(in, w, dim);err = r - expect;v = 0.8;//学习速率  34,16,9,8,8,9,15,59,x,xfor (i = 0;i < dim;i++){offset[i] = v*err*r*(1.0 - r)*in[i];}for (i = 0;i < dim;i++){w[i] -= offset[i];}re= acalo(in, w, dim,expect);e= fabs(re) - fabs(err);//disp(offset, dim);//printf("err:%-10f e:%-10f\n", re, e);return re;}/* 多层复杂网络求输出 */double calo2(const double in[NETL],//输入const double w1[NETX][NETL],//一层权值,输入层const double w2[NETX][NETX],//二层权值，隐层const double w3[NETX][NETX],//三层权值，隐层const double w4[NETX]//输出层权值){int i, j;double sum;double out[NETX][2];//输出缓冲for (i = 0;i < NETX;i++) {sum = 0;//第一层输出for (j = 0;j < NETL;j++)  sum += w1[i][j] * in[j];out[i][0] = 1.0 / (1.0 + exp(-1.0*sum));}for (i = 0;i < NETX;i++) {//第二层输出sum = 0;for (j = 0;j < NETX;j++)  sum += w2[i][j] * out[j][0];out[i][1] = 1.0 / (1.0 + exp(-1.0*sum));}for (i = 0;i < NETX;i++) {//第三层输出sum = 0;for (j = 0;j < NETX;j++)  sum += w3[i][j] * out[j][1];out[i][0] = 1.0 / (1.0 + exp(-1.0*sum));}sum = 0;for (i = 0;i < NETX;i++)  sum += out[i][0] * w4[i];return(1.0 / (1.0 + exp(-1.0*sum)));}/* 复杂网络单次训练 */double tra2(//训练网络，将会改变权重系数,得到误差const double in[NETL],//输入double w1[NETX][NETL],//一层权值,输入层double w2[NETX][NETX],//二层权值，隐层double w3[NETX][NETX],//三层权值，隐层double w4[NETX],//输出层权值const double expect){int i, j;double sum,res,err;double out[NETX][3];//输出缓冲for (i = 0;i < NETX;i++) {sum = 0;//第一层输出for (j = 0;j < NETL;j++)  sum += w1[i][j] * in[j];out[i][0] = 1.0 / (1.0 + exp(-1.0*sum));}for (i = 0;i < NETX;i++) {//第二层输出sum = 0;for (j = 0;j < NETX;j++)  sum += w2[i][j] * out[j][0];out[i][1] = 1.0 / (1.0 + exp(-1.0*sum));}for (i = 0;i < NETX;i++) {//第三层输出sum = 0;for (j = 0;j < NETX;j++)  sum += w3[i][j] * out[j][1];out[i][2] = 1.0 / (1.0 + exp(-1.0*sum));}sum = 0;for (i = 0;i < NETX;i++)  sum += out[i][2] * w4[i];res=(1.0 / (1.0 + exp(-1.0*sum)));err = res - expect;/* 梯度下降传递误差 */double lost[NETX][3];//误差矩阵for (i = 0;i < NETX;i++)//传递第三层误差{lost[i][2] = err*res*(1.0 - res)*w4[i];}for (i = 0;i < NETX;i++)//传递第二层误差{sum = 0;for (j = 0;j < NETX;j++){sum += lost[j][2] * out[j][2] * (1.0 - out[j][2])*w3[j][i];}lost[i][1] = sum;}for (i = 0;i < NETX;i++)//传递第一层误差{sum = 0;for (j = 0;j < NETX;j++){sum += lost[j][1] * out[j][1] * (1.0 - out[j][1])*w2[j][i];}lost[i][0] = sum;}/* 更新权值 */double v;//学习速率v = 0.1;for (i = 0;i < NETX;i++)//更新第一层权值{for (j = 0;j < NETL;j++){w1[i][j] -= v*lost[i][0] * out[i][0] * (1.0 - out[i][0])*in[j];}}for (i = 0;i < NETX;i++)//更新第二层权值{for (j = 0;j < NETX;j++){w2[i][j] -= v*lost[i][1] * out[i][1] * (1.0 - out[i][1])*out[j][0];}}for (i = 0;i < NETX;i++)//更新第三层权值{for (j = 0;j < NETX;j++){w3[i][j] -= v*lost[i][2] * out[i][2] * (1.0 - out[i][2])*out[j][1];}}for (i = 0;i < NETX;i++)//更新第四层权值{w4[i] -= v*err*res*(1.0 - res)*out[i][2];}double cals;cals=calo2(in, w1, w2, w3, w4);return cals - expect;}/*网络结构体*/typedef struct stu_unit2 {double w1[NETX][NETL];double w2[NETX][NETX];double w3[NETX][NETX];double w4[NETX];}unit2;void ninit(unit2 &inet){int  i, j;for (i = 0;i < NETX;i++)for (j = 0;j < NETL;j++)inet.w1[i][j] = (2.0*rand() / RAND_MAX) - 1.0;for (i = 0;i < NETX;i++)for (j = 0;j < NETX;j++)inet.w2[i][j] = (2.0*rand() / RAND_MAX) - 1.0;for (i = 0;i < NETX;i++)for (j = 0;j < 5;j++)inet.w3[i][j] = (2.0*rand() / RAND_MAX) - 1.0;for (i = 0;i < NETX;i++)inet.w4[i] = (2.0*rand() / RAND_MAX) - 1.0;}/*随机梯度下降训练*/double strain2(unit2 inet,double *in[NETL],double *expect,const int exdim){double los;int co;co = rand() % exdim;los=tra2(in[co], inet.w1, inet.w2, inet.w3, inet.w4, expect[co]);return los;}/*直接训练组*/double dtrain2(unit2 &inet, double in[][NETL], double *expect,  int exdim){double *los = new double[exdim];for (int i = 0;i < exdim;i++)los[i] = tra2(in[i], inet.w1,inet.w2, inet.w3, inet.w4, expect[i]);double sum = 0.0;for (int i = 0;i < exdim;i++){sum += fabs(los[i]);//cout << los[i] << endl;}delete[]los;return sum/(double)exdim;}/*输出网络*/double ncal(unit2 inet, double in[NETL]){return calo2(in, inet.w1, inet.w2, inet.w3, inet.w4);}#define SAVEPATH "net.dat"/*储存网络*/void nsave(unit2 inet){FILE *pFile = fopen(SAVEPATH, "wb");fwrite(&inet, sizeof(unit2), 1, pFile);fclose(pFile);}/*加载网络*/void nload(unit2 &inet){FILE *fp = fopen(SAVEPATH, "rb");fread(&inet, sizeof(unit2), 1, fp);fclose(fp);}#endif