C++ 卷积神经网络
C++ 卷积神经网络
注意：此项目未使用GPU加速
训练时使用文件（可自己制作）：
t10k-images.idx3-ubyte
t10k-labels.idx1-ubyte
读取图片文件时使用opencv
TP_NNW.h 对使用的相关方法进行封装
#pragma once
#includeTP_NNW.cpp实现封装的方法
#include"TP_NNW.h"
void Weight::apply(int H, int W)
{
fz = true;
this->len.height = H;
this->len.width = W;
this->WG = apply2(H, W);//申请内存
for (int i = 0; i < H; i++)
for (int j = 0; j < W; j++)
this->WG[i][j] = Get_rand();//得到随机值
}
void Weight::apply(int H, int W, float(*def)())
{
fz = true;
this->len.height = H;
this->len.width = W;
this->WG = apply2(H, W);
for (int i = 0; i < H; i++)
for (int j = 0; j < W; j++)
this->WG[i][j] = def();
}
Weight::~Weight()
{
this->release();
}
Weight::Weight(int H/*高度*/, int W/*宽度*/)
{
W = W <= 0 ? 1 : W;//防止出现0和负数
H = H <= 0 ? 1 : H;//防止出现0和负数
fz = true;
this->apply(H, W);
}
Weight::Weight(int H/*高度*/, int W/*宽度*/, float(*def)())
{
W = W <= 0 ? 1 : W;
H = H <= 0 ? 1 : H;
fz = true;
this->apply(H,W, def);
}
void Weight::re(float * delta, float * inp, float alpha)
{
for (int i = 0; i < this->len.height; i++)
{
for (int j = 0; j < this->len.width; j++)
this->WG[i][j] += alpha*delta[i] * inp[j];
}
}
void Weight::save(FILE * fp)
{
for (int i = 0; i < this->len.height; i++)
for (int j = 0; j < this->len.width; j++)
fwrite(&this->WG[i][j], sizeof(float), 1, fp);
}
void Weight::load(FILE * fp)
{
for (int i = 0; i < this->len.height; i++)
for (int j = 0; j < this->len.width; j++)
fread(&this->WG[i][j], sizeof(float), 1, fp);
}
void Weight::release()
{
if (this->fz)
{
Free2(this->WG, this->len.height);
//free(this->WG);
}
this->fz = false;
}
void Weight::operator >> (Weight & temp)
{
temp.release();
//free(temp.WG);
temp.apply(this->len.height, this->len.width, zeros);
}
void Weight::operator+=(Weight & temp)
{
for (int i = 0; i < this->len.height; i++)
for (int j = 0; j < this->len.width; j++)
this->WG[i][j] += temp.WG[i][j];
}
//void Weight::operator/=(int & temp)
//{
// for (int i = 0; i < this->len.height; i++)
// for (int j = 0; j < this->len.width; j++)
// this->WG[i][j] /= temp;
//}
void Weight::operator/=(int temp)
{
for (int i = 0; i < this->len.height; i++)
for (int j = 0; j < this->len.width; j++)
this->WG[i][j] /= temp;
}
void Weight::operator<<(Weight & temp)
{
Free2(this->WG, this->len.height);
this->len.height = temp.len.height;
this->len.width = temp.len.width;
this->WG = temp.WG;
}
void WD(Weight * WGS, int H, int W, int len)
{
for (int i = 0; i < len; i++)
{
WGS[i].apply(H, W);
}
}
void WD(Weight * WGS, int H, int W, int len, float(*def)())
{
for (int i = 0; i < len; i++)
{
WGS[i].apply(H, W,def);
}
}
float zeros()
{
return 0;
}
void print(float * y, int y_len)
{
for (int i = 0; i < y_len; i++)
{
printf("%0.2f ", y[i]);
//printf("%d ", y[i]>0?1:0);
}
puts("");
}
void print(float * y, Vector2 & vec)
{
print(y, vec.height);
}
void print(float ** y, Vector2 & vec)
{
for (int i = 0; i < vec.height; i++)
print(y[i], vec.width);
}
void print(char * y, int y_len)
{
for (int i = 0; i < y_len; i++)
{
printf("%d ", y[i]);
}
puts("");
}
void print(char ** y, Vector2 & vec)
{
for (int i = 0; i < vec.height; i++)
print(y[i], vec.width);
}
void print(Weight &w)
{
print(w.WG, w.len);
}
void print(Weight * w, int len)
{
for (int i = 0; i < len; i++)
{
printf("\n第%d层\n",i+1);
print(w[i]);
}
}
float ** apply2(int H, int W)
{
float **temp = (float**)malloc(sizeof(float**)*H);
for(int i=0;i<H;i++)
temp[i]= (float*)malloc(sizeof(float*)*W);
return temp;
}
float *** apply3(int P, int H/*高度*/, int W/*宽度*/)
{
float ***temp = (float***)malloc(sizeof(float***)*P);
for (int i = 0; i < P; i++)
temp[i] = apply2(H, W);
return temp;
}
char ** apply2_char(int H, int W)
{
char **temp = (char**)malloc(sizeof(float**)*H);
for (int i = 0; i<H; i++)
temp[i] = (char*)malloc(sizeof(float*)*W);
return temp;
}
float ones()
{
return 1;
}
float ***Conv(float **X, Vector2 &inp, Vector2 &out, Weight*W, int W_len)
{
out.height = inp.height - W[0].len.height + 1;
out.width = inp.width - W[0].len.width + 1;
float ***temp = (float***)malloc(sizeof(float***)*W_len);
for (int k = 0; k < W_len; k++)
temp[k] = conv2(X, inp, W[k].WG, W[0].len);
return temp;
}
float ***Pool(float ***y, Vector2 &inp, int P, Vector2 &out)
{
int h = inp.height / 2, w = inp.width / 2;
out.height = h;
out.width = w;
float ***temp = apply3(P, h, w);
float **filter = apply2(2, 2);
for (int i = 0; i < 2; i++)
for (int j = 0; j < 2; j++)
filter[i][j] = 0.25;
for (int k = 0; k < P; k++)
{
Vector2 len;
float **img = conv2(y[k], inp, filter, Vector2(2, 2), &len);
for (int i = 0; i < h; i++)
for (int j = 0; j < w; j++)
temp[k][i][j] = img[i * 2][j * 2];
Free2(img, len.height);
}
Free2(filter, 2);
return temp;
}
float * apply1(int H)
{
float *temp = (float*)malloc(sizeof(float*)*H);
return temp;
}
char * apply1_char(int H)
{
char *temp = (char*)malloc(sizeof(char*)*H);
return temp;
}
float Get_rand()
{
float temp = (float)(rand() % 10)/(float)10;
return rand()%2==0?temp:-temp;
}
float Sigmoid(float x)
{
return 1/(1+exp(-x));
}
float * Sigmoid(float * x, Weight &w)
{
return Sigmoid(x, w.len.height);
}
float * Sigmoid(float * x, int height)
{
float *y = (float*)malloc(sizeof(float*)*height);
for (int i = 0; i < height; i++)
y[i] = Sigmoid(x[i]);
return y;
}
float ReLU(float x)
{
return x>0?x:0;
}
float * ReLU(float * x, Weight & w)
{
return ReLU(x, w.len.height);
}
float * ReLU(float * x, int height)
{
float *y = (float*)malloc(sizeof(float*)*height);
for (int i = 0; i < height; i++)
y[i] = ReLU(x[i]);
return y;
}
float * Softmax(float * x, Weight & w)
{
return Softmax(x, w.len.height);
}
float dsigmoid(float x)
{
return x*(1-x);
}
float * Softmax(float * x, int height)
{
float *t = new float[height];
float *ex = new float[height];
float sum = 0;
for (int i = 0; i < height; i++)
{
ex[i] = exp(x[i]);
sum += ex[i];
}
for (int i = 0; i < height; i++)
{
t[i] = ex[i] / sum;
}
delete ex;
return t;
}
float * FXCB_err(Weight & w, float * delta)
{
float *temp = (float*)malloc(sizeof(float*)*w.len.width);
for (int i = 0; i < w.len.width; i++)
temp[i] = 0;
for (int i = 0; i < w.len.width; i++)
for (int j = 0; j < w.len.height; j++)
temp[i] += w.WG[j][i] * delta[j];
return temp;
}
float * Delta1(float * y, float *e, Weight & w)
{
float *temp = (float*)malloc(sizeof(float*)*w.len.height);
for (int i = 0; i < w.len.height; i++)
temp[i] = y[i] * (1 - y[i])*e[i];
return temp;
}
float * Delta2(float * v, float *e, Weight & w)
{
float *temp = (float*)malloc(sizeof(float*)*w.len.height);
for (int i = 0; i < w.len.height; i++)
temp[i] = v[i] > 0 ? e[i] : 0;
return temp;
}
float * dot(Weight & W, float * inp, int *len)
{
float *temp = (float*)malloc(sizeof(float*)*W.len.height);
for (int i = 0; i < W.len.height; i++)
temp[i] = 0;
for (int i = 0; i < W.len.height; i++)
{
for (int j = 0; j < W.len.width; j++)
temp[i] += (W.WG[i][j] * inp[j]);
}
if (len != NULL)
*len = W.len.height;
return temp;
}
char * randperm(int max, int count)
{
char *temp = new char[count] {0};
for (int i = 0; i < count; i++)
{
while (1)
{
char t = rand() % max;
bool nothave = true;
for (int j = 0; j < i; j++)
if (t == temp[j])
{
nothave = false;
break;
}
if (nothave)
{
temp[i] = t;
break;
}
}
}
return temp;
}
void Dropout(float * y, float ratio, Weight & w)
{
float *ym = new float[w.len.height] {0};
float round = w.len.height*(1 - ratio);
int num = (round - (float)(int)round >= 0.5f ? (int)round + 1 : (int)round);
char *idx = randperm(w.len.height, num);
for (int i = 0; i < num; i++)
{
ym[idx[i]] = (1 / (1 - ratio));
}
for (int i = 0; i < w.len.height; i++)
{
y[i] *= ym[i];
}
delete idx;
delete ym;
}
float ** conv2(float ** x, Vector2 & x_len, float ** fiter, Vector2 & fiter_len, Vector2*out_len, int flag, int distance, int fill)
{
switch (flag)
{
case Valid:return VALID(x,x_len.height,x_len.width, fiter, fiter_len.height, fiter_len.width, distance,out_len);
case Same:return SAME(x, x_len.height, x_len.width, fiter, fiter_len.height, fiter_len.width, distance, fill, out_len);
}
return nullptr;
}
float ** VALID(float ** x, int x_h, int x_w, float ** fiter, int fiter_h, int fiter_w, int distance, Vector2*out_len)
{
int h = VALID_out_len(x_h, fiter_h, distance);
int w = VALID_out_len(x_w, fiter_w, distance);
float **temp = apply2(h, w);
float **t = fiter;
if (out_len != NULL)
{
out_len->height = h;
out_len->width = w;
}
for(int i=0;i<x_h+1- fiter_h;i+=distance)
for (int j = 0; j < x_w + 1 - fiter_w; j += distance)
{
float count = 0;
for(int n=i;n<i+fiter_h;n++)
for (int m = j; m < j + fiter_w; m++)
{
if (n >= x_h || m >= x_w)
continue;
count += (x[n][m] * t[n - i][m - j]);
}
temp[(i / distance)][(j / distance)] = count;
}
//free(t);
return temp;
}
float ** SAME(float ** x, int x_h, int x_w, float ** fiter, int fiter_h, int fiter_w, int distance, int fill, Vector2*out_len)
{
return nullptr;
}
int VALID_out_len(int x_len, int fiter_len, int distance)
{
float temp = (float)(x_len - fiter_len) / (float)distance;
int t = temp - (int)((float)temp) >= 0.5 ? (int)temp + 1 : (int)temp;
t++;
return t;
}
void show_Weight(Weight & W)
{
for (int i = 0; i < W.len.height; i++)
{
for (int j = 0; j < W.len.width; j++)
{
printf("%0.3f ", W.WG[i][j]);
}
puts("");
}
}
void rot90(Weight & x)
{
int h = x.len.width, w = x.len.height;
x.WG=rot90(x.WG, x.len,true);
x.len.width = w;
x.len.height = h;
}
float** rot90(float ** x, Vector2 & x_len, bool release)
{
float**temp = apply2(x_len.width,x_len.height);
for(int i=0;i<x_len.height;i++)
for (int j = 0; j < x_len.width; j++)
{
temp[x_len.width-1-j][i]=x[i][j];
}
if (release)
{
Free2(x, x_len.height);
//free(x);
}
return temp;
}
float ** rot180(float ** x, Vector2 & x_len, bool release)
{
float**temp = apply2(x_len.height, x_len.width);
for (int i = 0; i < x_len.height; i++)
{
for (int j = 0; j < x_len.width; j++)
{
temp[x_len.height - 1 - i][x_len.width - 1 - j] = x[i][j];
}
}
if (release)
{
Free2(x, x_len.height);
//free(x);
}
return temp;
}
float ** Get_data_by_Mat(char * filepath, Vector2 & out_len)
{
cv::Mat mat = cv::imread(filepath,0);
cv::resize(mat, mat, cv::Size(28, 28));
/*cv::imshow("tt", mat);
cv::waitKey(0);*/
out_len.height = mat.rows;
out_len.width = mat.cols;
float **temp = apply2(mat.rows, mat.cols);
int c = 0;
for(int i=0;i<mat.rows;i++)
for (int j = 0; j < mat.cols; j++)
{
temp[i][j] = ((float)mat.data[c++] / (float)255);
//temp[i][j] = 1 - temp[i][j];
}
mat.release();
return temp;
}
char ** Get_data_by_Mat_char(char * filepath, Vector2 & out_len, int threshold)
{
cv::Mat mat = cv::imread(filepath, 0);
out_len.height = mat.rows;
out_len.width = mat.cols;
char **temp = apply2_char(mat.rows, mat.cols);
int c = 0;
for (int i = 0; i<mat.rows; i++)
for (int j = 0; j < mat.cols; j++)
{
temp[i][j] =mat.data[c++]>threshold?0:1;
}
mat.release();
return temp;
}
void Get_data_by_Mat(char *filepath, Weight & w)
{
w.WG = Get_data_by_Mat(filepath, w.len);
}
Weight Get_data_by_Mat(char * filepath)
{
Weight temp;
Get_data_by_Mat(filepath, temp);
return temp;
}
Vector2::Vector2()
{
this->height = 0;
this->width = 0;
}
Vector2::Vector2(char height, int width)
{
this->height = height;
this->width = width;
}
XML::XML(FILE * fp, char * name, int layer)
{
this->fp = fp;
this->name = name;
this->layer = layer;
}
void XML::showchild()
{
char reader[500];
while (fgets(reader, 500, this->fp))
{
int len = strlen(reader);
int lay = 0;
for (; lay < len; lay++)
{
if (reader[lay] != '\t')break;
}
if (lay == this->layer)
{
if (reader[lay + 1] == '/')continue;
char show[500];
memset(show, 0, 500);
for (int i = lay + 1; i < len - 2; i++)
{
if (reader[i] == '>')break;
show[i - lay - 1] = reader[i];
}
puts(show);
}
}
fseek(this->fp, 0, 0);
}
void bit::operator=(int x)
{
this->B = x;
}
float * reshape(float ** x, int h, int w)
{
float *temp = (float*)malloc(sizeof(float*)*w*h);
int count = 0;
for (int i = 0; i<h; i++)
for (int j = 0; j < w; j++)
{
temp[count++] = x[i][j];
}
return temp;
}
float * reshape(float ** x, Vector2 & x_len)
{
return reshape(x, x_len.height, x_len.width);
}
float * reshape(float *** x, Vector2 & x_len, int P, bool releace)
{
float *temp = apply1(x_len.height*x_len.width*P);
int c = 0;
for (int i = 0; i < P; i++)
for (int n = 0; n < x_len.height; n++)
for (int m = 0; m < x_len.width; m++)
temp[c++] = x[i][n][m];
if (releace)
Free3(x, P, x_len.height);
//free(x);
return temp;
}
int * bList(int distance, int max, int *out_len)
{
int num = (max%distance!=0);
int t = (int)(max / distance);
t += num;
if (out_len != NULL)
*out_len = t;
int *out = (int*)malloc(sizeof(int*)*t);
for (int i = 0; i < t; i++)
{
out[i] = i*distance;
}
return out;
}
void Free2(float ** x, int h)
{
for (int i = 0; i < h; i++)
free(x[i]);
free(x);
}
void Free3(float *** x, int p, int h)
{
for (int i = 0; i < p; i++)
for (int j = 0; j < h; j++)
free(x[i][j]);
for (int i = 0; i < p; i++)
free(x[i]);
free(x);
}
void kron(float ** out, Vector2 & out_len, float ** inp, Vector2 & inp_len, float ** filter, Vector2 & filter_len)
{
for(int i=0;i<inp_len.height;i++)
for (int j = 0; j < inp_len.width; j++)
{
for(int n=i*2;n<out_len.height&&n<((i*2)+filter_len.height);n++)
for (int m = (j*2); m < ((j*2)+ filter_len.width)&&m<out_len.width; m++)
{
out[n][m] = inp[i][j] * filter[n - (i * 2)][m - (j * 2)]*0.25;
}
}
}
char ** mnist::Toshape2(char * x, int h, int w)
{
char **temp = apply2_char(h, w);
int c = 0;
for (int i = 0; i < h; i++)
for (int j = 0; j < w; j++)
temp[i][j] = x[c++];
return temp;
}
char ** mnist::Toshape2(char * x, Vector2 & x_len)
{
return mnist::Toshape2(x, x_len.height, x_len.width);
}
void mnist::Toshape2(char ** out, char * x, int h, int w)
{
int c = 0;
for (int i = 0; i < h; i++)
for (int j = 0; j < w; j++)
out[i][j] = x[c++];
}
void mnist::Toshape2(char ** out, char * x, Vector2 & x_len)
{
mnist::Toshape2(out,x, x_len.height, x_len.width);
}
float ** mnist::Toshape2_F(char * x, int h, int w)
{
float **temp = apply2(h, w);
int c = 0;
for (int i = 0; i < h; i++)
for (int j = 0; j < w; j++)
temp[i][j] = ((float)x[c++]/(float)255);
return temp;
}
float ** mnist::Toshape2_F(char * x, Vector2 & x_len)
{
return mnist::Toshape2_F(x, x_len.height, x_len.width);
}
void mnist::Toshape2(float ** out, char * x, int h, int w)
{
int c = 0;
for (int i = 0; i < h; i++)
for (int j = 0; j < w; j++)
out[i][j] = ((float)x[c++] / (float)255);
}
void mnist::Toshape2(float ** out, char * x, Vector2 & x_len)
{
mnist::Toshape2(out, x, x_len.height, x_len.width);
}
void mnist::Toshape2(float ** out, unsigned char * x, int h, int w)
{
int c = 0;
for (int i = 0; i < h; i++)
for (int j = 0; j < w; j++)
{
out[i][j] = ((float)x[c++] / (float)255);
}
}
void mnist::Toshape2(float ** out, unsigned char * x, Vector2 & x_len)
{
mnist::Toshape2(out, x, x_len.height, x_len.width);
}
float *** mnist::Toshape3(float * x, int P, Vector2 & x_len)
{
float ***temp = apply3(P, x_len.height, x_len.width);
int c = 0;
for (int i = 0; i < P; i++)
for (int j = 0; j < x_len.height; j++)
for (int n = 0; n < x_len.width; n++)
temp[i][j][n] = x[c++];
return temp;
}
int mnist::ReverseInt(int i)
{
unsigned char ch1, ch2, ch3, ch4;
ch1 = i & 255;
ch2 = (i >> 8) & 255;
ch3 = (i >> 16) & 255;
ch4 = (i >> 24) & 255;
return((int)ch1 << 24) + ((int)ch2 << 16) + ((int)ch3 << 8) + ch4;
}
conv.cpp 主函数在这里
#include"TP_NNW.h"
using namespace std;
using namespace cv;
using namespace mnist;
float *SGD(Weight *W1, Weight &W5, Weight &Wo, float **X)
{
Vector2 ve(28, 28);
float *temp = new float[10];
Vector2 Cout;
float ***y1 = Conv(X, ve, Cout, W1, 20);
for (int i = 0; i < 20; i++)
for (int n = 0; n < Cout.height; n++)
for (int m = 0; m < Cout.width; m++)
y1[i][n][m] = ReLU(y1[i][n][m]);
float ***y2 = y1;
Vector2 Cout2;
float ***y3 = Pool(y1, Cout, 20, Cout2);
float *y4 = reshape(y3, Cout2, 20, true);
float *v5 = dot(W5, y4);
float *y5 = ReLU(v5, W5);
float *v = dot(Wo, y5);
float *y = Softmax(v, Wo);
for (int i = 0; i < Wo.len.height; i++)
temp[i] = y[i];
return temp;
}
void trainSGD(Weight *W1, Weight &W5, Weight &Wo, FILE *fp,FILE *tp)
{
Vector2 ve(28, 28);
unsigned char *reader = new unsigned char[ve.height * ve.width];
float **X = apply2(ve.height, ve.width);
unsigned char hao;
hot_one<char> D(10);
Weight *momentum1 = new Weight[20];//动量
Weight momentum5;
Weight momentumo;
Weight *dW1 = new Weight[20];//动量
Weight dW5;
Weight dWo;
for (int i = 0; i < 20; i++)
W1[0] >> momentum1[i];
W5 >> momentum5;
Wo >> momentumo;
int N = 8000;//训练集取前8000个
int bsize = 100;//100个纠正一次
int b_len;
int *blist = bList(bsize, N, &b_len);
for (int batch = 0; batch < b_len; batch++)
{
for (int i = 0; i < 20; i++)
W1[0] >> dW1[i];
W5 >> dW5;
Wo >> dWo;
int begins = blist[batch];
for (int k = begins; k < begins+bsize&&k<N; k++)
{
fread(reader, sizeof(unsigned char), ve.height * ve.width, fp);//读取图像
Toshape2(X,reader, ve);//组合成二维数组
Vector2 Cout;//储存卷积后数组的尺寸 20
float ***y1 = Conv(X, ve, Cout, W1, 20);//卷积
for (int i = 0; i < 20; i++)
for (int n = 0; n < Cout.height; n++)
{
for (int m = 0; m < Cout.width; m++)
{
y1[i][n][m] = ReLU(y1[i][n][m]);//通过ReLU函数
}
}
float ***y2 = y1;//给变量y2
Vector2 Cout2;//记录池化后的尺寸 10
float ***y3 = Pool(y1, Cout, 20, Cout2);//池化层
float *y4 = reshape(y3, Cout2, 20, true);//作为神经元输入
float *v5 = dot(W5, y4);//矩阵乘法
float *y5 = ReLU(v5, W5);//ReLU函数
float *v = dot(Wo, y5);//举证乘法
float *y = Softmax(v, Wo);//soft分类
fread(&hao, sizeof(unsigned char), 1, tp);//读取标签
D.re(hao);
float *e = new float[10];
for (int i = 0; i < 10; i++)
e[i] = ((float)D.one[i]) - y[i];
float *delta = e;
float *e5 = FXCB_err(Wo, delta);
float *delta5 = Delta2(y5, e5, W5);
float *e4 = FXCB_err(W5, delta5);
float ***e3 = Toshape3(e4, 20, Cout2);
float ***e2 = apply3(20, Cout.height, Cout.width);
Weight one(2, 2, ones);
/*for (int i = 0; i < 20; i++)
{
printf("第%d层\n", i);
for (int n = 0; n < Cout2.height; n++)
{
for (int m = 0; m < Cout2.width; m++)
printf("%0.3f ", e3[i][n][m]);
puts("");
}
}
getchar();*/
for (int i = 0; i < 20; i++)//---------------------------------
kron(e2[i], Cout, e3[i], Cout2, one.WG, one.len);
/*for (int i = 0; i < 20; i++)
{
printf("第%d层\n", i);
for (int n = 0; n < Cout.height; n++)
{
for (int m = 0; m < Cout.width; m++)
printf("%f ", e2[i][n][m]);
puts("");
}
}
getchar();*/
float ***delta2 =apply3(20, Cout.height, Cout.width);
for (int i = 0; i < 20; i++)
for (int n = 0; n < Cout.height; n++)
for (int m = 0; m < Cout.width; m++)
delta2[i][n][m] = (y2[i][n][m] > 0)*e2[i][n][m];
float ***delta_x = (float***)malloc(sizeof(float***) * 20);
Vector2 t1;
for (int i = 0; i < 20; i++)
delta_x[i] = conv2(X, ve, delta2[i], Cout,&t1);
for (int i = 0; i < 20; i++)
for (int n = 0; n < t1.height; n++)
for (int m = 0; m < t1.width; m++)
dW1[i].WG[n][m] += delta_x[i][n][m];
dW5.re(delta5, y4,1);
dWo.re(delta, y5, 1);
Free3(delta_x, 20, t1.height);
Free3(delta2, 20, Cout.height);
one.release();
Free3(e2, 20, Cout.height);
Free3(e3, 20, Cout2.height);
free(e4);
free(delta5);
free(e5);
free(v5);
delete e;
free(y5);
free(v);
free(y);
Free3(y1,20, Cout.height);
free(y4);
}
for (int i = 0; i < 20; i++)
dW1[i] /= (bsize);
dW5/=(bsize);
dWo/=(bsize);
for (int i = 0; i < 20; i++)
for (int n = 0; n < W1[0].len.height; n++)
for (int m = 0; m < W1[0].len.width; m++)
{
momentum1[i].WG[n][m] = ALPHA*dW1[i].WG[n][m] + BETA*momentum1[i].WG[n][m];
W1[i].WG[n][m] += momentum1[i].WG[n][m];
}
for (int n = 0; n < W5.len.height; n++)
for (int m = 0; m < W5.len.width; m++)
momentum5.WG[n][m] = ALPHA*dW5.WG[n][m] + BETA*momentum5.WG[n][m];
W5 += momentum5;
for (int n = 0; n < Wo.len.height; n++)
for (int m = 0; m < Wo.len.width; m++)
momentumo.WG[n][m] = ALPHA*dWo.WG[n][m] + BETA*momentumo.WG[n][m];
Wo += momentumo;
}
for (int i = 0; i < 20; i++)
{
momentum1[i].release();
dW1[i].release();
}
momentum5.release();
momentumo.release();
Free2(X,ve.height);
free(blist);
delete reader;
D.release();
dW5.release();
dWo.release();
return;
}
void trainSGD1(Weight *W1, Weight &W5, Weight &Wo, FILE *fp, FILE *tp)
{
Vector2 ve(28, 28);
unsigned char *reader = new unsigned char[ve.height * ve.width];
float **X = apply2(ve.height, ve.width);
unsigned char hao;
hot_one<char> D(10);
Weight *momentum1 = new Weight[20];//动量
Weight momentum5;
Weight momentumo;
Weight *dW1 = new Weight[20];//动量
Weight dW5;
Weight dWo;
for (int i = 0; i < 20; i++)
W1[0] >> momentum1[i];
W5 >> momentum5;
Wo >> momentumo;
int N = 108;//训练集取前8000个
int bsize = 12;//100个纠正一次
int b_len;
int *blist = bList(bsize, N, &b_len);
for (int batch = 0; batch < b_len; batch++)
{
for (int i = 0; i < 20; i++)
W1[0] >> dW1[i];
W5 >> dW5;
Wo >> dWo;
int begins = blist[batch];
for (int k = begins; k < begins + bsize&&k<N; k++)
{
fread(reader, sizeof(unsigned char), ve.height * ve.width, fp);//读取图像
Toshape2(X, reader, ve);//组合成二维数组
Vector2 Cout;//储存卷积后数组的尺寸 20
float ***y1 = Conv(X, ve, Cout, W1, 20);//卷积
for (int i = 0; i < 20; i++)
for (int n = 0; n < Cout.height; n++)
{
for (int m = 0; m < Cout.width; m++)
{
y1[i][n][m] = ReLU(y1[i][n][m]);//通过ReLU函数
}
}
float ***y2 = y1;//给变量y2
Vector2 Cout2;//记录池化后的尺寸 10
float ***y3 = Pool(y1, Cout, 20, Cout2);//池化层
float *y4 = reshape(y3, Cout2, 20, true);//作为神经元输入
float *v5 = dot(W5, y4);//矩阵乘法
float *y5 = ReLU(v5, W5);//ReLU函数
float *v = dot(Wo, y5);//举证乘法
float *y = Softmax(v, Wo);//soft分类
fread(&hao, sizeof(unsigned char), 1, tp);//读取标签
D.re(hao);
float *e = new float[10];
for (int i = 0; i < 10; i++)
e[i] = ((float)D.one[i]) - y[i];
float *delta = e;
float *e5 = FXCB_err(Wo, delta);
float *delta5 = Delta2(y5, e5, W5);
float *e4 = FXCB_err(W5, delta5);
float ***e3 = Toshape3(e4, 20, Cout2);
float ***e2 = apply3(20, Cout.height, Cout.width);
Weight one(2, 2, ones);
/*for (int i = 0; i < 20; i++)
{
printf("第%d层\n", i);
for (int n = 0; n < Cout2.height; n++)
{
for (int m = 0; m < Cout2.width; m++)
printf("%0.3f ", e3[i][n][m]);
puts("");
}
}
getchar();*/
for (int i = 0; i < 20; i++)//---------------------------------
kron(e2[i], Cout, e3[i], Cout2, one.WG, one.len);
/*for (int i = 0; i < 20; i++)
{
printf("第%d层\n", i);
for (int n = 0; n < Cout.height; n++)
{
for (int m = 0; m < Cout.width; m++)
printf("%f ", e2[i][n][m]);
puts("");
}
}
getchar();*/
float ***delta2 = apply3(20, Cout.height, Cout.width);
for (int i = 0; i < 20; i++)
for (int n = 0; n < Cout.height; n++)
for (int m = 0; m < Cout.width; m++)
delta2[i][n][m] = (y2[i][n][m] > 0)*e2[i][n][m];
float ***delta_x = (float***)malloc(sizeof(float***) * 20);
Vector2 t1;
for (int i = 0; i < 20; i++)
delta_x[i] = conv2(X, ve, delta2[i], Cout, &t1);
for (int i = 0; i < 20; i++)
for (int n = 0; n < t1.height; n++)
for (int m = 0; m < t1.width; m++)
dW1[i].WG[n][m] += delta_x[i][n][m];
dW5.re(delta5, y4, 1);
dWo.re(delta, y5, 1);
Free3(delta_x, 20, t1.height);
Free3(delta2, 20, Cout.height);
one.release();
Free3(e2, 20, Cout.height);
Free3(e3, 20, Cout2.height);
free(e4);
free(delta5);
free(e5);
free(v5);
delete e;
free(y5);
free(v);
free(y);
Free3(y1, 20, Cout.height);
free(y4);
}
for (int i = 0; i < 20; i++)
dW1[i] /= (bsize);
dW5 /= (bsize);
dWo /= (bsize);
for (int i = 0; i < 20; i++)
for (int n = 0; n < W1[0].len.height; n++)
for (int m = 0; m < W1[0].len.width; m++)
{
momentum1[i].WG[n][m] = ALPHA*dW1[i].WG[n][m] + BETA*momentum1[i].WG[n][m];
W1[i].WG[n][m] += momentum1[i].WG[n][m];
}
for (int n = 0; n < W5.len.height; n++)
for (int m = 0; m < W5.len.width; m++)
momentum5.WG[n][m] = ALPHA*dW5.WG[n][m] + BETA*momentum5.WG[n][m];
W5 += momentum5;
for (int n = 0; n < Wo.len.height; n++)
for (int m = 0; m < Wo.len.width; m++)
momentumo.WG[n][m] = ALPHA*dWo.WG[n][m] + BETA*momentumo.WG[n][m];
Wo += momentumo;
}
for (int i = 0; i < 20; i++)
{
momentum1[i].release();
dW1[i].release();
}
momentum5.release();
momentumo.release();
Free2(X, ve.height);
free(blist);
delete reader;
D.release();
dW5.release();
dWo.release();
return;
}
float rand1()
{
float temp = (rand() % 20) / (float)10;
if (temp < 0.0001)
temp = 0.07;
temp *= (rand() % 2 == 0) ? -1 : 1;
return temp*0.01;
}
float rand2()
{
float temp = (rand() % 10) / (float)10;
float ret = (2 * temp - 1)*sqrt(6) / sqrt(360 + 2000);
if (ret < 0.0001&&ret>-0.0001)
ret = 0.07;
return ret;
}
float rand3()
{
float temp = (rand() % 10) / (float)10;
float ret = (2 * temp - 1)*sqrt(6) / sqrt(10 + 100);
if (ret < 0.0001&&ret>-0.0001)
ret = 0.07;
return ret;
}
void train()
{
FILE *fp = fopen("t10k-images.idx3-ubyte", "rb");
FILE *tp = fopen("t10k-labels.idx1-ubyte", "rb");
int rdint;
fread(&rdint, sizeof(int), 1, fp);
printf("训练集幻数：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, fp);
printf("训练集数量：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, fp);
printf("训练集高度：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, fp);
printf("训练集宽度：%d\n", ReverseInt(rdint));
int start1 = ftell(fp);
fread(&rdint, sizeof(int), 1, tp);
printf("标签幻数：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, tp);
printf("标签数量：%d\n", ReverseInt(rdint));
int start2 = ftell(tp);
Weight *W1 = new Weight[20];
WD(W1, 9, 9, 20, rand1);
Weight W5(100, 2000, rand2);
Weight Wo(10, W5.len.height, rand3);
for (int k = 0; k < 3; k++)
{
trainSGD(W1, W5, Wo, fp, tp);
fseek(fp, start1, 0);
fseek(tp, start2, 0);
printf("第%d次训练结束\n", k+1);
}
fclose(fp);
fclose(tp);
fp = fopen("mnist_Weight.acp", "wb");
for (int i = 0; i < 20; i++)
W1[i].save(fp);
W5.save(fp);
Wo.save(fp);
fclose(fp);
printf("训练完成");
getchar();
}
void train1()
{
FILE *fp = fopen("out_img.acp", "rb");
FILE *tp = fopen("out_label.acp", "rb");
int rdint;
fread(&rdint, sizeof(int), 1, fp);
printf("训练集幻数：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, fp);
printf("训练集数量：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, fp);
printf("训练集高度：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, fp);
printf("训练集宽度：%d\n", ReverseInt(rdint));
int start1 = ftell(fp);
fread(&rdint, sizeof(int), 1, tp);
printf("标签幻数：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, tp);
printf("标签数量：%d\n", ReverseInt(rdint));
int start2 = ftell(tp);
Weight *W1 = new Weight[20];
WD(W1, 9, 9, 20, rand1);
Weight W5(100, 2000, rand2);
Weight Wo(10, W5.len.height, rand3);
for (int k = 0; k < 1000; k++)
{
trainSGD1(W1, W5, Wo, fp, tp);
fseek(fp, start1, 0);
fseek(tp, start2, 0);
printf("第%d次训练结束\n", k + 1);
}
fclose(fp);
fclose(tp);
fp = fopen("mnist_Weight.acp", "wb");
for (int i = 0; i < 20; i++)
W1[i].save(fp);
W5.save(fp);
Wo.save(fp);
fclose(fp);
printf("训练完成");
getchar();
}
void test()
{
FILE *fp = fopen("mnist_Weight.acp", "rb");
Weight *W1 = new Weight[20];
WD(W1, 9, 9, 20, rand1);
Weight W5(100, 2000, rand1);
Weight Wo(10, W5.len.height, rand1);
for (int i = 0; i < 20; i++)
W1[i].load(fp);
W5.load(fp);
Wo.load(fp);
fclose(fp);
fp = fopen("t10k-images.idx3-ubyte", "rb");
FILE *tp = fopen("t10k-labels.idx1-ubyte", "rb");
int rdint;
fread(&rdint, sizeof(int), 1, fp);
printf("训练集幻数：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, fp);
printf("训练集数量：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, fp);
printf("训练集高度：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, fp);
printf("训练集宽度：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, tp);
printf("标签幻数：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, tp);
printf("标签数量：%d\n", ReverseInt(rdint));
unsigned char *res = new unsigned char[28 * 28];
float **X = apply2(28, 28);
unsigned char biaoqian;
for (int i = 0; i < 50; i++)
{
fread(res, sizeof(unsigned char), 28 * 28, fp);
Toshape2(X, res, 28, 28);
print(X, Vector2(28, 28));
float *h = SGD(W1, W5, Wo, X);//带入神经网络
int c = -1;
for (int i = 0; i < 10; i++)
{
if (h[i] > 0.85)
{
c = i;
break;
}
}
fread(&biaoqian, sizeof(unsigned char), 1, tp);
printf("正确结果应当为“%d”, 神经网络识别为“%d” \n", biaoqian, c);
}
}
Weight *W1;
Weight W5(100, 2000, rand1);
Weight Wo(10, W5.len.height, rand1);
void sb()
{
//printf("加载权重完毕\n");
Vector2 out;
char path[256];
for (int r = 0; r < 4; r++)
{
sprintf(path, "acp%d.png", r);
float **img = Get_data_by_Mat(path, out);
//print(img, out);
float *h = SGD(W1, W5, Wo, img);//带入神经网络
int c = -1;
float x = 0;
for (int i = 0; i < 10; i++)
{
if (h[i] > 0.85&&h[i]>x)
{
x = h[i];
c = i;
}
}
printf("%d ", c);
Free2(img, out.height);
free(h);
remove(path);
}
puts("");
}
void sb(char *path)
{
FILE *fp = fopen("mnist_Weight.acp", "rb");
//puts("开始加载权重");
WD(W1, 9, 9, 20, rand1);
for (int i = 0; i < 20; i++)
W1[i].load(fp);
W5.load(fp);
Wo.load(fp);
fclose(fp);
//printf("加载权重完毕\n");
Vector2 out;
float **img = Get_data_by_Mat(path, out);
print(img, out);
float *h = SGD(W1, W5, Wo, img);//带入神经网络
int c = -1;
float x = 0;
for (int i = 0; i < 10; i++)
{
printf("%f\n", h[i]);
if (h[i] > 0.65&&h[i]>x)
{
x = h[i];
c = i;
}
}
printf("%d %f", c,x);
Free2(img, out.height);
free(h);
}
bool thank(Vec3b &a, Vec3b &b)
{
int dis = 0;
for (int i = 0; i < 3; i++)
{
int t = (a[i] - b[i]);
dis += t*t;
}
dis = (int)pow((float)dis, 0.5);
if (dis < 100)
return true;
return false;
}
void qg(char *path)
{
printf(path);
printf("识别为：");
Mat img = imread(path);
Vec3b yes = Vec3b(204, 198, 204);
Mat sav = Mat(120, 80, CV_8UC3);
resize(img, img, Size(img.cols * 10, img.rows * 10));
for (int i = 0; i<img.rows; i++)
for (int j = 0; j < img.cols; j++)
{
Vec3b rgb = img.at<Vec3b>(i, j);
if (thank(rgb, yes))
img.at<Vec3b>(i, j) = Vec3b(255, 255, 255);
/*else
img.at(i, j) = Vec3b(0, 0, 0);*/
}
char p[256];
for (int k = 0; k < 4; k++)
{
sprintf(p, "acp%d.png", k);
for (int i = 35 + (k * 80); i < 115 + (k * 80); i++)
for (int j = 30; j < 150; j++)
sav.at<Vec3b>(j-30, i-(35 + (k * 80))) = img.at<Vec3b>(j, i);
imwrite(p, sav);
}
img.release();
sav.release();
sb();
}
void test1()
{
FILE *fp = fopen("mnist_Weight.acp", "rb");
Weight *W1 = new Weight[20];
WD(W1, 9, 9, 20, rand1);
Weight W5(100, 2000, rand1);
Weight Wo(10, W5.len.height, rand1);
for (int i = 0; i < 20; i++)
W1[i].load(fp);
W5.load(fp);
Wo.load(fp);
fclose(fp);
fp = fopen("out_img.acp", "rb");
FILE *tp = fopen("out_label.acp", "rb");
int rdint;
fread(&rdint, sizeof(int), 1, fp);
printf("训练集幻数：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, fp);
printf("训练集数量：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, fp);
printf("训练集高度：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, fp);
printf("训练集宽度：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, tp);
printf("标签幻数：%d\n", ReverseInt(rdint));
fread(&rdint, sizeof(int), 1, tp);
printf("标签数量：%d\n", ReverseInt(rdint));
unsigned char *res = new unsigned char[28 * 28];
float **X = apply2(28, 28);
unsigned char biaoqian;
for (int i = 0; i < 10; i++)
{
fread(res, sizeof(unsigned char), 28 * 28, fp);
Toshape2(X, res, 28, 28);
print(X, Vector2(28, 28));
float *h = SGD(W1, W5, Wo, X);//带入神经网络
int c = -1;
for (int i = 0; i < 10; i++)
{
if (h[i] > 0.85)
{
c = i;
break;
}
}
fread(&biaoqian, sizeof(unsigned char), 1, tp);
printf("正确结果应当为“%d”, 神经网络识别为“%d” \n", biaoqian, c);
}
}
void main(int argc,char **argv)
{
//train1();
//train();
if (argc > 1)
{
FILE *fp = fopen("mnist_Weight.acp", "rb");
//puts("开始加载权重");
W1 = new Weight[20];
WD(W1, 9, 9, 20, rand1);
for (int i = 0; i < 20; i++)
W1[i].load(fp);
W5.load(fp);
Wo.load(fp);
fclose(fp);
for(int i=1;i<argc;i++)
qg(argv[i]);
printf("完毕");
getchar();
}
}