前言：本文不对CNN卷积神经网络做深入探究，CNN卷积神经网络的基本知识请移步本文的相关链接；本文不对LeNet-5神经网络模型做深入探究，该部分的知识可以自行查阅或者查看本文的链接！MNIST数据集请自行在官网下载。此外，如果使用本文的代码，请将该数据集放置于源代码同级目录下。
笔记：从单纯的BP算法，到DNN再到CNN，是一个奇妙的旅程。CNN于DNN的不同之处在于对于局部特征的抽取（个人理解），通过卷积运算，来压缩原有的特征数据，并将压缩后的数据通过池化操作，进一步提取特征。当然可以通过各种不一样的组合来进行这些操作，比如全连接，多次卷积，多次池化，采用不同的激活函数，采用不同的卷积核，控制feature map的数量，控制步长...最后经过训练以及优化完成想要的效果。
1.有关CNN卷积神经网络，请移步：CNN笔记
2.有关LeNet-5 神经网络模型（如下图），请移步：LeNet详解
3.程序说明：
3.1 该程序只针对MNIST数据训练一次，准确率在0.95左右，当然，你可以试着更改一些参数来调试该网络，例如learning rate，epoch，batch size...。多次epoch后准确率更高！一次训练的时间大概在6-10分钟，使用visual studio 2019编译该代码比codeblocks自带的gcc编译后，运行的速度更快一些。（本人实测）。
3.2 该程序提供了多个显示图像的方法，自己debug过程中可以方便调用观察，具体的使用已经在代码中注释。.当然你也可以自己写一些更好的方法加进去！
3.3 log函数只为调试代码使用，运行时请将is_logging参数设置为false。
4.程序代码
#include <iostream>
#include <stdlib.h>
#include <memory.h>
#include <math.h>
#include <ctime>
#ifndef MAX
#define MAX(A,B) (((A) > (B)) ? (A) : (B))
#endif
using namespace std;
typedef struct data_sample
{
double* data;
double* label;
int sample_width;
int sample_height;
int sample_count;
}Sample;
typedef struct convolution_kernel
{
double* weight;
double* delta_weight;
}Kernel;
typedef struct layer_map
{
double* data;
double* error;
double bias;
double delta_bias;
}Map;
typedef struct layer
{
int map_width;
int map_height;
int map_count;
Map* map;
int kernel_width;
int kernel_height;
int kernel_count;
Kernel* kernel;
double* map_common;
}Layer;
const int batch_size = 10;
const int class_count = 10;
const int width = 32;
const int height = 32;
const int train_sample_count = 60000;
const int test_sample_count = 10000;
const int zero_padding = 2;
const int padded_matrix_size = (28 + zero_padding * 2) * (28 + 2 * zero_padding);
Layer input_layer;
Layer output_layer;
Layer c1_convolution_layer;
Layer s2_pooling_layer;
Layer c3_convolution_layer;
Layer s4_pooling_layer;
Layer c5_convolution_layer;
bool is_logging = false;
void log_exec(const char* msg)
{
if (is_logging)
{
cout << "exec part in -->> [ " << msg << " ]" << endl;
}
}
void show_pic(Sample sample, int target)
{
cout << "img:" << endl;
for (int w = 0; w < 1024; w++)
{
cout << sample.data[w] << ",";
if ((w + 1) % 32 == 0)
{
cout << endl;
}
}
cout << endl << "target:" << target << endl;
cout << endl << endl;
}
void show_pic(Sample sample)
{
cout << "img:" << endl;
for (int w = 0; w < 1024; w++)
{
cout << sample.data[w] << ",";
if ((w + 1) % 32 == 0)
{
cout << endl;
}
}
int index = 0;
for (int i = 0; i < class_count; i++)
{
if (sample.label[i] == 1)
{
index = i;
}
}
cout << endl << "target:" << index << endl;
cout << endl << endl;
}
void show_map(Map map, int width, int height)
{
cout << "map info:" << endl;
for (int i = 0; i < height; i++)
{
for (int j = 0; j < width; j++)
{
cout << map.data[i * width + j] << ",";
}
cout << endl;
}
cout << endl << endl;
}
void load_mnist_data(Sample* sample, const char* file_name, int sample_count) //checked 2.0
{
FILE* mnist_file = NULL;
int err = fopen_s(&mnist_file, file_name, "rb");
unsigned char image_buffer[784]; //保存图片信息
int head_info[1000]; //读取出文件的头信息，该信息不参与计算
fread(head_info, 1, 16, mnist_file); //读取16字节头部信息
if (mnist_file == NULL)
{
cout << "load data from your file err..." << endl;
return;
}
else
{
cout << "loading data...[in func -->> load_mnist_data]" << endl;
}
for (int i = 0; i < sample_count; i++)
{
sample[i].data = (double*)malloc(padded_matrix_size * sizeof(double));
memset(sample[i].data, 0, padded_matrix_size * sizeof(double));
fread(image_buffer, 1, 784, mnist_file);
unsigned int value;
int index = 0;
for (int j = 0; j < 28; j++)
{
for (int k = 0; k < 28; k++)
{
int shuffle_index = (j + zero_padding) * width + k + zero_padding;
value = (unsigned int)image_buffer[index++];
if (value < 128)
{
sample[i].data[shuffle_index] = 0;
}
else
{
sample[i].data[shuffle_index] = 1;
}
}
}
}
fclose(mnist_file);
mnist_file = NULL;
}
void load_mnist_label(Sample* sample, const char* file_name, const int sample_count) //checked 2.0
{
FILE* mnist_file = NULL;
int err = fopen_s(&mnist_file, file_name, "rb");
unsigned char label;
if (mnist_file == NULL)
{
cout << "load label from your file err..." << endl;
return;
}
else
{
cout << "loading label...[in func -->> load_mnist_label]" << endl;
}
int head_info[1000]; //读取出文件的头信息，该信息不参与计算
fread(head_info, 1, 8, mnist_file); //读取8字节头部信息
for (int i = 0; i < sample_count; i++)
{
sample[i].label = (double*)malloc(class_count * sizeof(double));
for(int k = 0; k < class_count; k ++)
{
sample[i].label[k] = 0;
}
fread((char*)&label, sizeof(label), 1, mnist_file);
int value_index = (unsigned int)label;
sample[i].label[value_index] = 1;
//show_pic(sample[i],value_index); //test
}
fclose(mnist_file);
mnist_file = NULL;
}
double weight_base_calc(int pre_map_count,int cur_map_count,int kernel_width,int kernel_height,bool is_pooling)
{
const double scale = 6.0;
int fan_in = 0;
int fan_out = 0;
if (is_pooling)
{
fan_in = 4;
fan_out = 1;
}
else
{
fan_in = pre_map_count * kernel_width * kernel_height;
fan_out = cur_map_count * kernel_width * kernel_height;
}
int denominator = fan_in + fan_out;
double weight_base = (denominator != 0) ? sqrt(scale / (double)denominator) : 0.5;
return weight_base;
}
void init_kernel(double* kernel, int size, double weight_base)
{
log_exec("init_kernel");
for (int i = 0; i < size; i++)
{
kernel[i] = ((double)rand() / RAND_MAX - 0.5) * 2 * weight_base;
}
}
void init_layer(Layer* layer, int prelayer_map_count, int map_count, int kernel_width, int kernel_height, int map_width, int map_height, bool is_pooling)
{
log_exec("init_kernel");
int single_kernel_memsize = 0;
double weight_base = weight_base_calc(prelayer_map_count,map_count,kernel_width,kernel_height,is_pooling);
//cout<<"weight base ："<<weight_base<<endl;
layer->kernel_count = prelayer_map_count * map_count;//需要训练的kernel数目
layer->kernel_width = kernel_width;
layer->kernel_height = kernel_height;
layer->kernel = (Kernel*)malloc(layer->kernel_count * sizeof(Kernel)); //为该层的所有卷积核分配内存
single_kernel_memsize = layer->kernel_width * layer->kernel_height * sizeof(double);
for (int i = 0; i < prelayer_map_count; i++) //初始化两层之间每一个卷积核的参数weight delta weight
{
for (int j = 0; j < map_count; j++)
{
int valued_index = (i * map_count) + j;
layer->kernel[valued_index].weight = (double*)malloc(single_kernel_memsize);
init_kernel(layer->kernel[valued_index].weight, kernel_height * kernel_width, weight_base);
layer->kernel[valued_index].delta_weight = (double*)malloc(single_kernel_memsize);
memset(layer->kernel[valued_index].delta_weight, 0, single_kernel_memsize);
}
}
layer->map_count = map_count;
layer->map_height = map_height;
layer->map_width = map_width;
layer->map = (Map*)malloc(layer->map_count * sizeof(Map));
int single_map_size = layer->map_height * layer->map_width * sizeof(double);
for (int i = 0; i < layer->map_count; i++) //初始化每一个map的参数 bias ,delta bias, data, error
{
layer->map[i].bias = 0.0;
layer->map[i].delta_bias = 0.0;
layer->map[i].data = (double*)malloc(single_map_size);
memset(layer->map[i].data, 0, single_map_size);
layer->map[i].error = (double*)malloc(single_map_size);
memset(layer->map[i].error, 0, single_map_size);
}
layer->map_common = (double*)malloc(single_map_size);
memset(layer->map_common, 0, single_map_size);
}
/*
重置某layer层的所有map的delta_bias为0；
重置某layer层所有的kernel的delta_weight为0
*/
void reset_layer_deltabias_kernel_deltaweight(Layer *layer) //checked 2.0
{
log_exec("reset_layer_deltabias_kernel_deltaweight");
int single_kernel_memsize = layer->kernel_height * layer->kernel_width * sizeof(double);
for (int i = 0; i < layer->kernel_count; i++)
{
memset(layer->kernel[i].delta_weight, 0, single_kernel_memsize);
}
for (int i = 0; i < layer->map_count; i++)
{
layer->map[i].delta_bias = 0.0;
}
}
void reset_all_layer() //checked 2.0
{
log_exec("reset_all_layer");
reset_layer_deltabias_kernel_deltaweight(&c1_convolution_layer);
reset_layer_deltabias_kernel_deltaweight(&s2_pooling_layer);
reset_layer_deltabias_kernel_deltaweight(&c3_convolution_layer);
reset_layer_deltabias_kernel_deltaweight(&s4_pooling_layer);
reset_layer_deltabias_kernel_deltaweight(&c5_convolution_layer);
reset_layer_deltabias_kernel_deltaweight(&output_layer);
}
/*
激活函数以及导函数计算
*/
struct activation_function //checked 2.0
{
inline static double tan_h(double x)
{
return ((exp(x) - exp(-x)) / (exp(x) + exp(-x)));
}
inline static double d_tan_h(double x)
{
return (1.0 - (x * x));
}
inline static double relu(double x)
{
return (x > 0.0 ? x : 0.0);
}
inline static double d_relu(double x)
{
return (x > 0.0 ? 1.0 : 0.0);
}
inline static double sigmod(double x)
{
return (1.0 / (1.0 + exp(-x)));
}
inline static double d_sigmod(double x)
{
return (x * (1.0 - x));
}
};
/*
损失函数
*/
struct loss_function //checked 2.0
{
inline static double cal_loss(double real, double target)
{
return (real - target) * (real - target) / 2;
}
inline static double d_cal_loss(double real, double target)
{
return (real - target);
}
};
/*
单次计算map与kernel的卷积
*/
void convolution_calcu(double* input_map_data, int input_map_width, int input_map_height, double* kernel_data, int kernel_width, int kernel_height, double* result_map_data, int result_map_width, int result_map_height)
{
log_exec("convolution_calcu");
double sum = 0.0;
for (int i = 0; i < result_map_height; i++)
{
for (int j = 0; j < result_map_width; j++)
{
sum = 0.0;
for (int n = 0; n < kernel_height; n++)
{
for (int m = 0; m < kernel_width; m++)
{
int index_input_reshuffle = (i + n) * input_map_width + j + m;
int index_kernel_reshuffle = n * kernel_width + m;
sum += input_map_data[index_input_reshuffle] * kernel_data[index_kernel_reshuffle];
}
}
int index_result_reshuffle = i * result_map_width + j;
result_map_data[index_result_reshuffle] += sum;
}
}
}
#define O true
#define X false
//S2到C3不完全连接映射表
bool connection_table[6 * 16] = {
O, X, X, X, O, O, O, X, X, O, O, O, O, X, O, O,
O, O, X, X, X, O, O, O, X, X, O, O, O, O, X, O,
O, O, O, X, X, X, O, O, O, X, X, O, X, O, O, O,
X, O, O, O, X, X, O, O, O, O, X, X, O, X, O, O,
X, X, O, O, O, X, X, O, O, O, O, X, O, O, X, O,
X, X, X, O, O, O, X, X, O, O, O, O, X, O, O, O
};
#undef O
#undef X
void convolution_forward_propagation(Layer* pre_layer, Layer* cur_layer, bool* connection_table) //checked and fixed /checked 2.0
{
log_exec("convolution_forward_propagation");
int index_layer = 0;
int layer_size = cur_layer->map_height * cur_layer->map_width;
for (int i = 0; i < cur_layer->map_count; i++)
{
memset(cur_layer->map_common, 0, layer_size * sizeof(double)); //清空公共map的暂存数据。
for (int j = 0; j < pre_layer->map_count; j++)
{
index_layer = j * cur_layer->map_count + i;
if (connection_table != NULL && !connection_table[index_layer])
{
continue;
}
//fix para 3 map height
convolution_calcu(
pre_layer->map[j].data, pre_layer->map_width, pre_layer->map_height,
cur_layer->kernel[index_layer].weight, cur_layer->kernel_width, cur_layer->kernel_height,
cur_layer->map_common, cur_layer->map_width, cur_layer->map_height
);
}
for (int k = 0; k < layer_size; k++)
{
cur_layer->map[i].data[k] = activation_function::tan_h(cur_layer->map_common[k] + cur_layer->map[i].bias);
}
}
}
void max_pooling_forward_propagation(Layer* pre_layer, Layer* cur_layer) //checked 2.0
{
log_exec("max_pooling_forward_propagation");
int map_width = cur_layer->map_width;
int map_height = cur_layer->map_height;
int pre_map_width = pre_layer->map_width;
for (int k = 0; k < cur_layer->map_count; k++)
{
for (int i = 0; i < map_height; i++)
{
for (int j = 0; j < map_width; j++)
{
double max_value = pre_layer->map[k].data[2 * i * pre_map_width + 2*j];
for (int n = 2 * i; n < 2 * (i + 1); n++)
{
for (int m = 2 * j; m < 2 * (j + 1); m++)
{
max_value = MAX(max_value, pre_layer->map[k].data[n * pre_map_width + m]);
}
}
cur_layer->map[k].data[i * map_width + j] = activation_function::tan_h(max_value);
}
}
}
}
void fully_connection_forward_propagation(Layer* pre_layer, Layer* cur_layer)
{
log_exec("fully_connection_forward_propagation");
for (int i = 0; i < cur_layer->map_count; i++)
{
double sum = 0.0;
for (int j = 0; j < pre_layer->map_count; j++)
{
sum += pre_layer->map[j].data[0] * cur_layer->kernel[j * cur_layer->map_count + i].weight[0];
}
sum += cur_layer->map[i].bias;
cur_layer->map[i].data[0] = activation_function::tan_h(sum);
}
}
void forward_propagation()
{
log_exec("forward_propagation");
convolution_forward_propagation(&input_layer, &c1_convolution_layer, NULL);
//show_map(c1_convolution_layer.map[0],28,28);
max_pooling_forward_propagation(&c1_convolution_layer, &s2_pooling_layer);
convolution_forward_propagation(&s2_pooling_layer, &c3_convolution_layer, connection_table);
max_pooling_forward_propagation(&c3_convolution_layer, &s4_pooling_layer);
convolution_forward_propagation(&s4_pooling_layer, &c5_convolution_layer, NULL);
fully_connection_forward_propagation(&c5_convolution_layer, &output_layer);
}
/*
全连接层反向更新
*/
void fully_connection_backward_propagation(Layer* cur_layer, Layer* pre_layer) //checked
{
log_exec("fully_connection_backward_propagation");
//更新 c5 error值
for (int i = 0; i < pre_layer->map_count; i++)
{
pre_layer->map[i].error[0] = 0.0;
for (int j = 0; j < cur_layer->map_count; j++)
{
pre_layer->map[i].error[0] += cur_layer->map[j].error[0] * cur_layer->kernel[i * cur_layer->map_count + j].weight[0];
}
pre_layer->map[i].error[0] *= activation_function::d_tan_h(pre_layer->map[i].data[0]);
}
//更新c5 --> output kernel delta_weight
for (int i = 0; i < pre_layer->map_count; i++)
{
for (int j = 0; j < cur_layer->map_count; j++)
{
cur_layer->kernel[i * cur_layer->map_count + j].delta_weight[0] += cur_layer->map[j].error[0] * pre_layer->map[i].data[0];
}
}
//更新output delta_bias
for (int i = 0; i < cur_layer->map_count; i++)
{
cur_layer->map[i].delta_bias += cur_layer->map[i].error[0];
}
}
void convolution_backward_propagation(Layer* cur_layer, Layer* pre_layer, bool* connection_table) //checked checked 2.0 fixed
{
log_exec("convolution_backward_propagation");
int connected_index = 0;
int pre_layer_mapsize = pre_layer->map_height * pre_layer->map_width;
//更新S4 error
for (int i = 0; i < pre_layer->map_count; i++)
{
memset(pre_layer->map_common, 0, sizeof(double) * pre_layer_mapsize);
for (int j = 0; j < cur_layer->map_count; j++)
{
connected_index = i * cur_layer->map_count + j;
if (connection_table != NULL && !connection_table[connected_index])
{
continue;
}
for (int n = 0; n < cur_layer->map_height; n++) //fixed cur_layer->kernel_height -->> cur_layer->map_height
{
for (int m = 0; m < cur_layer->map_width; m++)
{
int valued_index = n * cur_layer->map_width + m;
double error = cur_layer->map[j].error[valued_index];
for (int kernel_y = 0; kernel_y < cur_layer->kernel_height; kernel_y++)
{
for (int kernel_x = 0; kernel_x < cur_layer->kernel_width; kernel_x++)
{
int index_convoltion_map = (n + kernel_y) * pre_layer->map_width + m + kernel_x;
int index_kernel = connected_index;
int index_kernel_weight = kernel_y * cur_layer->kernel_width + kernel_x;
pre_layer->map_common[index_convoltion_map] += error * cur_layer->kernel[index_kernel].weight[index_kernel_weight];
}
}
}
}
}
for (int k = 0; k < pre_layer_mapsize; k++)
{
pre_layer->map[i].error[k] = pre_layer->map_common[k] * activation_function::d_tan_h(pre_layer->map[i].data[k]);
//pre_layer->map[i].error[k] = pre_layer->map_common[k] * activation_func::dtan_h(prev_layer->map[i].data[k]); source
}
}
//更新 S_x ->> C_x kernel 的 delta_weight
for (int i = 0; i < pre_layer->map_count; i++)
{
for (int j = 0; j < cur_layer->map_count; j++)
{
connected_index = i * cur_layer->map_count + j;
if (connection_table != NULL && !connection_table[connected_index])
{
continue;
}
//fixed cur_layer->map[i] -->> cur_layer->map[j]
convolution_calcu(
pre_layer->map[i].data, pre_layer->map_width, pre_layer->map_height,
cur_layer->map[j].error, cur_layer->map_width, cur_layer->map_height,
cur_layer->kernel[connected_index].delta_weight, cur_layer->kernel_width, cur_layer->kernel_height
);
}
}
//更新C_x 的delta_bias
int cur_layer_mapsize = cur_layer->map_height * cur_layer->map_width;
for (int i = 0; i < cur_layer->map_count; i++)
{
double delta_sum = 0.0;
for (int j = 0; j < cur_layer_mapsize; j++)
{
delta_sum += cur_layer->map[i].error[j];
}
cur_layer->map[i].delta_bias += delta_sum;
}
}
void max_pooling_backward_propagation(Layer* cur_layer, Layer* pre_layer) //checked
{
log_exec("max_pooling_backward_propagation");
int cur_layer_mapwidth = cur_layer->map_width;
int cur_layer_mapheight = cur_layer->map_height;
int pre_layer_mapwidth = pre_layer->map_width;
for (int k = 0; k < cur_layer->map_count; k++)
{
for (int i = 0; i < cur_layer_mapheight; i++)
{
for (int j = 0; j < cur_layer_mapwidth; j++)
{
int index_row = 2 * i;
int index_col = 2 * j;
double max_value = pre_layer->map[k].data[index_row * pre_layer_mapwidth + index_col];
for (int n = 2 * i; n < 2 * (i + 1); n++)
{
for (int m = 2 * j; m < 2 * (j + 1); m++)
{
if (pre_layer->map[k].data[n * pre_layer_mapwidth + m] > max_value)
{
index_row = n;
index_col = m;
max_value = pre_layer->map[k].data[n * pre_layer_mapwidth + m];
}
else
{
pre_layer->map[k].error[n * pre_layer_mapwidth + m] = 0.0;
}
}
}
pre_layer->map[k].error[index_row * pre_layer_mapwidth + index_col] = cur_layer->map[k].error[i * cur_layer_mapwidth + j] * activation_function::d_tan_h(max_value);
}
}
}
}
void backward_propagation(double* label) //checked
{
log_exec("backward_propagation");
//前面并未初始化输出层error，在此初始化
for (int i = 0; i < output_layer.map_count; i++)
{
output_layer.map[i].error[0] = loss_function::d_cal_loss(output_layer.map[i].data[0], label[i]) * activation_function::d_tan_h(output_layer.map[i].data[0]);
}
fully_connection_backward_propagation(&output_layer, &c5_convolution_layer);
convolution_backward_propagation(&c5_convolution_layer, &s4_pooling_layer, NULL);
max_pooling_backward_propagation(&s4_pooling_layer, &c3_convolution_layer);
convolution_backward_propagation(&c3_convolution_layer, &s2_pooling_layer, connection_table);
max_pooling_backward_propagation(&s2_pooling_layer, &c1_convolution_layer);
convolution_backward_propagation(&c1_convolution_layer, &input_layer, NULL);
}
//梯度下降
inline double gradient_descent(double para, double delta_para, double learning_rate) //checked
{
log_exec("init_kernel");
// cout<<"exec fun -->> [gradient_descent]"<<endl;
return para - learning_rate * delta_para;
}
/*
梯度下降法
更新某layer的 kernel 的weight 和 map的 bias
*/
void update_param(Layer* layer, double learning_rate) //checked and fixed
{
log_exec("update_param");
int kernel_size = layer->kernel_height * layer->kernel_width;
//update weight
for (int i = 0; i < layer->kernel_count; i++)
{
for (int k = 0; k < kernel_size; k++)
{
layer->kernel[i].weight[k] = gradient_descent(layer->kernel[i].weight[k], layer->kernel[i].delta_weight[k] / batch_size, learning_rate);
}
}
for (int i = 0; i < layer->map_count; i++) //fixed add ：/ batch_size
{
layer->map[i].bias = gradient_descent(layer->map[i].bias, layer->map[i].delta_bias / batch_size, learning_rate);
}
}
void update_all_layer_param(double learning_rate)
{
log_exec("update_all_layer_param");
update_param(&c1_convolution_layer, learning_rate);
update_param(&s2_pooling_layer, learning_rate);
update_param(&c3_convolution_layer, learning_rate);
update_param(&s4_pooling_layer, learning_rate);
update_param(&c5_convolution_layer, learning_rate);
update_param(&output_layer, learning_rate);
}
void training(Sample* samples, double learning_rate)
{
log_exec("training");
int batch_count = samples->sample_count / batch_size;
int single_sample_size = samples->sample_height * samples->sample_width * sizeof(double);
cout << "training started!!!" << endl;
cout << "sample count : " << samples->sample_count << "\t batch size :" << batch_size << " \t batch count : " << batch_count << endl;
for (int i = 0; i < batch_count; i++)
{
reset_all_layer(); //重新初始化各层map的delta bias 和 kernel 的 delta weight
for (int j = 0; j < batch_size; j++)
{
int sample_index = i * batch_size + j;
memcpy(input_layer.map[0].data, samples[sample_index].data, single_sample_size);
//show_map(input_layer.map[0],input_layer.map_height,input_layer.map_width); //test
forward_propagation();
backward_propagation(samples[sample_index].label);
}
update_all_layer_param(learning_rate);
if (i % 1500 == 0)
{
// show_pic(samples[i]); //test
cout << "training data process : " << i / (1.0 * batch_count) * 100.0 << " %" << endl;
}
}
cout << "training data process : 100 %" << endl;
cout << "training data over!!!!" << endl;
}
void testing(Sample* samples)
{
log_exec("testing");
int success_count = 0;
int result_test = 0;
int result_label = 0;
int single_sample_size = samples->sample_height * samples->sample_width * sizeof(double);
int* result_test_label_matrix = (int*)malloc(class_count * class_count * sizeof(int));
memset(result_test_label_matrix, 0, sizeof(int) * class_count * class_count);
for (int i = 0; i < samples->sample_count; i++)
{
memcpy(input_layer.map[0].data, samples[i].data, single_sample_size);
forward_propagation();
int index_result = 0;
double max_value = *(output_layer.map[0].data);
for (int k = 0; k < output_layer.map_count; k++)
{
if (*(output_layer.map[k].data) > max_value)
{
max_value = *(output_layer.map[k].data);
index_result = k;
}
}
result_test = index_result;
int index_label = 0;
max_value = samples->label[0];
for (int k = 1; k < class_count; k++)
{
if (samples[i].label[k] > max_value)
{
max_value = samples[i].label[k];
index_label = k;
}
}
result_label = index_label;
if (result_test == result_label)
{
success_count++;
}
//cout << "label value : " << result_label << "\t test value : " << result_test << endl;
result_test_label_matrix[result_test * class_count + result_label] ++;
if (i % 2500 == 0)
{
cout << "testing data process : " << i / (1.0 * samples->sample_count) * 100.0 << " %" << endl;
}
}
//输出结果矩阵
cout << "testing over !!! success rate : " << (1.0) * success_count / samples->sample_count << endl;
cout << "\t ";
for (int i = 0; i < class_count; i++)
{
cout << "\t" << i;
}
cout << endl;
for (int i = 0; i < class_count; i++)
{
cout << "\t" << i;
for (int j = 0; j < class_count; j++)
{
cout << "\t" << result_test_label_matrix[i * class_count + j];
}
cout << endl;
}
int sum = 0;
for (int i = 0; i < class_count * class_count; i++)
{
sum += result_test_label_matrix[i];
}
cout << "total sum : " << sum << endl;
free(result_test_label_matrix);
result_test_label_matrix = NULL;
}
void release_layer(Layer* layer)
{
log_exec("release_layer");
for (int i = 0; i < layer->kernel_count; i++)
{
free(layer->kernel[i].weight);
free(layer->kernel[i].delta_weight);
layer->kernel[i].weight = NULL;
layer->kernel[i].delta_weight = NULL;
}
free(layer->kernel);
layer->kernel = NULL;
for (int i = 0; i < layer->map_count; i++)
{
free(layer->map[i].data);
free(layer->map[i].error);
layer->map[i].data = NULL;
layer->map[i].error = NULL;
}
free(layer->map_common);
layer->map_common = NULL;
free(layer->map);
layer->map = NULL;
}
void free_memory(Sample *test_sample, Sample *train_sample)
{
log_exec("free_memory");
for (int i = 0; i < train_sample_count; i++) {
free(train_sample[i].data);
free(train_sample[i].label);
train_sample[i].data = NULL;
train_sample[i].label = NULL;
}
free(train_sample);
for (int i = 0; i < test_sample_count; i++) {
free(test_sample[i].data);
free(test_sample[i].label);
test_sample[i].data = NULL;
test_sample[i].label = NULL;
}
free(test_sample);
}
int main()
{
log_exec("MAIN");
double learning_rate = 0.018;
//初始化训练集样本
Sample* train_sample = (Sample*)malloc(train_sample_count * sizeof(Sample));
memset(train_sample, 0, train_sample_count * sizeof(Sample));
train_sample->sample_height = height;
train_sample->sample_width = width;
train_sample->sample_count = train_sample_count;
const char* train_sample_path = "./train-images.idx3-ubyte";
const char* train_label_path = "./train-labels.idx1-ubyte";
load_mnist_data(train_sample, train_sample_path, 60000);
load_mnist_label(train_sample, train_label_path, 60000);
//初始化测试集样本
Sample* test_sample = (Sample*)malloc(test_sample_count * sizeof(Sample));
memset(test_sample, 0, test_sample_count * sizeof(Sample));
test_sample->sample_height = height;
test_sample->sample_width = width;
test_sample->sample_count = test_sample_count;
const char* test_sample_path = "./t10k-images.idx3-ubyte";
const char* test_label_path = "./t10k-labels.idx1-ubyte";
load_mnist_data(test_sample, test_sample_path, 10000);
load_mnist_label(test_sample, test_label_path, 10000);
//初始化各层
/*
init_layer(*layer, prelayer_map_count, map_count, kernel_width, kernel_height, map_width, map_height, is_pooling)
*/
init_layer(&input_layer, 0, 1, 0, 0, 32, 32, false);
init_layer(&c1_convolution_layer, 1, 6, 5, 5, 28, 28, false);
init_layer(&s2_pooling_layer, 1, 6, 1, 1, 14, 14, true);
init_layer(&c3_convolution_layer, 6, 16, 5, 5, 10, 10, false);
init_layer(&s4_pooling_layer, 1, 16, 1, 1, 5, 5, true);
init_layer(&c5_convolution_layer, 16, 120, 5, 5, 1, 1, false);
init_layer(&output_layer, 120, 10, 1, 1, 1, 1, false);
//开始训练和测试
training(train_sample, learning_rate);
testing(test_sample);
cout << "run over 1" << endl;
//清理内存
free_memory(test_sample,train_sample);
release_layer(&input_layer);
release_layer(&c1_convolution_layer);
release_layer(&s2_pooling_layer);
release_layer(&c3_convolution_layer);
release_layer(&s4_pooling_layer);
release_layer(&c5_convolution_layer);
release_layer(&output_layer);
return 0;
}
扫一扫
2089
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
