前言:本人旨在交流代码,细节和原理不清楚的可以留言,以后做完了再整理,现在先放一部分,可以直接使用
1.代码可以直接在c上运行,输入为5*5矩阵,比如数字0:
1 1 1 1 1
1 0 0 0 1
1 0 0 0 1
1 0 0 0 1
1 1 1 1 1
因为训练集合太小,决定用matlab生成大量训练样本,放在后文中。
2.c语言实现代码:(BP推导很简单,不会的自己可以留言,代码中有详细注释)
#include <stdio.h>
#include <math.h>
#include <stdlib.h>
#include <time.h>
#define COUNT 50//样本数量
#define IN_NUM 25//输入层神经元数量
#define OUT_NUM 10//输出层神经元数量
#define numHidden 30//隐含层神经元数量
double weight_hidden[IN_NUM][numHidden];//输入层到隐含层的权值
double bias_hidden[numHidden];//隐含层的阈值
double weight_output[numHidden][OUT_NUM];//隐含层到输出层的权值
double bias_output[OUT_NUM];//输出层的阈值
double learnRate = 0.4;//学习数率
double accuracy = 0.001;//最大容许误差0.000001
int maxloopCount = 1000000;//最大学习数率
double fnet(double net)//Sigmoid
{
return 1 / (1 + exp(-net));
}
double dfnet(double net)//Sigmoid导函数 y = s*(s-1)
{
return net * (1 - net);
}
int InitBP()//得到-1到1随机数
{
int i, j;
srand((unsigned)time(NULL));
for (i = 0; i < IN_NUM; i++)
for (j = 0; j < numHidden; j++)
{
weight_hidden[i][j] = rand() / (double)(RAND_MAX)-0.5;
bias_hidden[j] = rand() / (double)(RAND_MAX)-0.5;
}
for (i = 0; i < numHidden; i++)
for (j = 0; j < OUT_NUM; j++)
{
weight_output[i][j] = rand() / (double)(RAND_MAX)-0.5;
bias_output[j] = rand() / (double)(RAND_MAX)-0.5;
}
return 1;
}
int TrainBP(float x[COUNT][IN_NUM], float y[COUNT][OUT_NUM])
{
double delta_hidden[numHidden], delta_output[OUT_NUM];//中间计算量
double output_hidden[numHidden], output_output[OUT_NUM];//隐含层输出 输出层输出
double temp;//中间累加值
double loss;
int i, j, k, n;
for (n = 0; n < maxloopCount; n++)
{
loss = 0;
for (i = 0; i < COUNT; i++)//逐个数据进行计算,整个数据计算一遍后n+1
{
/*前相传播*/
//计算隐含层输出
for (k = 0; k < numHidden; k++)
{
temp = 0;
for (j = 0; j < IN_NUM; j++)
temp += x[i][j] * weight_hidden[j][k];
output_hidden[k] = fnet(temp + bias_hidden[k]);
}
//计算输出层输出
for (k = 0; k < OUT_NUM; k++)
{
temp = 0;
for (j = 0; j < numHidden; j++)
temp += output_hidden[j] * weight_output[j][k];
output_output[k] = fnet(temp + bias_output[k]);
}
//计算误差
for (j = 0; j < OUT_NUM; j++)
loss += 0.5*(y[i][j] - output_output[j])*(y[i][j] - output_output[j]);
/*方向传播阶段*/
//更新输出层的权值
for (j = 0; j < OUT_NUM; j++)
delta_output[j] = (y[i][j] - output_output[j])*dfnet(output_output[j]);
for (j = 0; j < numHidden; j++)
for (k = 0; k < OUT_NUM; k++)
{
weight_output[j][k] += learnRate * delta_output[k] * output_hidden[j];
}
//更新输出层的偏重
for (k = 0; k < OUT_NUM; k++)
bias_output[k] += learnRate * delta_output[k];
//更新隐含层权重
for (j = 0; j < numHidden; j++)
{
temp = 0;
for (k = 0; k < OUT_NUM; k++)
temp += weight_output[j][k] * delta_output[k];
delta_hidden[j] = temp * dfnet(output_hidden[j]);
}
for (j = 0; j<IN_NUM; j++)
for (k = 0; k < numHidden; k++)
weight_hidden[j][k] += learnRate*de