该程序是模拟tensflow游乐场写的,实现了基本的神经网络效果并验证通过,不多废话,上代码。
核心代码在nn.c中,包含激活函数和损失函数,前向传播,反向传播以及更新权重与偏执的函数。
#include <stdint.h> #include <stdlib.h> #include <math.h> #include "config.h" #include "dataset.h" #include "nn.h" int networkShape[] = {2, 8, 8, 8, 8, 8, 8, 1}; NODE **network; double getOutPut() { return network[sizeof(networkShape) / sizeof(int) - 1][0].output; } double square(double output, double target) { double r = output - target; return r * r / 2; } double squareder(double output, double target) { return output - target; } double activation(double x) { #if ACTIVATIONFUNCTION == RELU if (x > 0) { return x; } else { return 0; } #elif ACTIVATIONFUNCTION == TANH return tanh(x); #endif } double activationder(double x) { #if ACTIVATIONFUNCTION == RELU if (x > 0) { return 1; } else { return 0; } #elif ACTIVATIONFUNCTION == TANH // tanh的倒数 double y = tanh(x); return 1 - y * y; #endif } double outlayeractivation(double x) { #if OUTLAYERACTIVATIONFUNCTION == TANH return tanh(x); #endif } double outlayeractivationder(double x) { #if OUTLAYERACTIVATIONFUNCTION == TANH // tanh的倒数 double y = tanh(x); return 1 - y * y; #endif } void buildNetwork() { network = (PPNODE)malloc((sizeof(networkShape) / sizeof(int)) * sizeof(PNODE)); // 输入层 network[0] = (PNODE)malloc(networkShape[0] * sizeof(NODE)); // 隐藏层与输出层 for (int i = 1, leni = sizeof(networkShape) / sizeof(int); i < leni; i++) { network[i] = (PNODE)malloc(networkShape[i] * sizeof(NODE)); int prenodeNum = networkShape[i - 1]; for (int j = 0, lenj = networkShape[i]; j < lenj; j++) { network[i][j].link = (PLINK)malloc(prenodeNum * sizeof(LINK)); } } // 输入层 for (int i = 0; i < networkShape[0]; i++) { network[0][i].bias = 0.1; } // 隐藏层与输出层 for (int i = 1, leni = sizeof(networkShape) / sizeof(int); i < leni; i++) { for (int j = 0, lenj = networkShape[i]; j < lenj; j++) { network[i][j].bias = 0.1; network[i][j].inputDer = 0; network[i][j].outputDer = 0; network[i][j].accInputDer = 0; network[i][j].numAccumulatedDers = 0; for (int k = 0, lenk = networkShape[i - 1]; k < lenk; k++) { network[i][j].link[k].weight = (double)rand() / RAND_MAX - 0.5; network[i][j].link[k].errorDer = 0; network[i][j].link[k].accErrorDer = 0; network[i][j].link[k].numAccumulatedDers = 0; } } } } void forwardProp(POINT point) { int outlayerNum = sizeof(networkShape) / sizeof(int) - 1; // 输出层所在层 // 输入层 network[0][0].output = point.x; network[0][1].output = point.y; // 隐藏层 for (int i = 1, leni = outlayerNum; i < leni; i++) { for (int j = 0, lenj = networkShape[i]; j < lenj; j++) { network[i][j].totalInput = network[i][j].bias; for (int k = 0, lenk = networkShape[i - 1]; k < lenk; k++) { network[i][j].totalInput += network[i][j].link[k].weight * network[i - 1][k].output; } network[i][j].output = activation(network[i][j].totalInput); } } // 输出层 for (int i = 0, leni = networkShape[outlayerNum]; i < leni; i++) { network[outlayerNum][i].totalInput = network[outlayerNum][i].bias; for (int j = 0, lenj = networkShape[outlayerNum - 1]; j < lenj; j++) { network[outlayerNum][i].totalInput += network[outlayerNum][i].link[j].weight * network[outlayerNum - 1][j].output; } network[outlayerNum][i].output = outlayeractivation(network[outlayerNum][i].totalInput); } } void backProp(POINT point) { // 清空所有节点的outputDer for (int i = 0, leni = sizeof(networkShape) / sizeof(int); i < leni; i++) { for (int j = 0; j < networkShape[i]; j++) { network[i][j].outputDer = 0; } } int outlayerNum = sizeof(networkShape) / sizeof(int) - 1; // 输出层所在层 // 输出层 for (int i = 0, leni = networkShape[outlayerNum]; i < leni; i++) { network[outlayerNum][i].outputDer = squareder(network[outlayerNum][i].output, point.label); // 目标和结果的差距 network[outlayerNum][i].inputDer = network[outlayerNum][i].outputDer * outlayeractivationder(network[outlayerNum][i].totalInput); network[outlayerNum][i].accInputDer += network[outlayerNum][i].inputDer; network[outlayerNum][i].numAccumulatedDers++; for (int j = 0, lenj = networkShape[outlayerNum]; j < lenj; j++) { network[outlayerNum][i].link[i].errorDer = network[outlayerNum][i].inputDer * network[outlayerNum - 1][i].output; network[outlayerNum][i].link[i].accErrorDer += network[outlayerNum][i].link[i].errorDer; network[outlayerNum][i].link[i].numAccumulatedDers++; network[outlayerNum - 1][i].outputDer += network[outlayerNum][i].link[i].weight * network[outlayerNum][i].inputDer; } } // 隐藏层 for (int i = outlayerNum; i > 0; i--) { for (int j = 0; j < networkShape[i]; j++) { network[i][j].inputDer = network[i][j].outputDer * activationder(network[i][j].totalInput); network[i][j].accInputDer += network[i][j].inputDer; network[i][j].numAccumulatedDers++; for (int k = 0; k < networkShape[i - 1]; k++) { network[i][j].link[k].errorDer = network[i][j].inputDer * network[i - 1][k].output; network[i][j].link[k].accErrorDer += network[i][j].link[k].errorDer; network[i][j].link[k].numAccumulatedDers++; network[i - 1][k].outputDer += network[i][j].link[k].weight * network[i][j].inputDer; } } } } void updateWeights() { // 隐藏层与输出层 for (int i = 1; i < sizeof(networkShape) / sizeof(int); i++) { for (int j = 0; j < networkShape[i]; j++) { if (network[i][j].numAccumulatedDers > 0) { network[i][j].bias -= LEARNINGRATE * network[i][j].accInputDer / network[i][j].numAccumulatedDers; network[i][j].accInputDer = 0; network[i][j].numAccumulatedDers = 0; } for (int k = 0; k < networkShape[i - 1]; k++) { if (network[i][j].link[k].numAccumulatedDers > 0) { network[i][j].link[k].weight -= LEARNINGRATE * network[i][j].link[k].accErrorDer / network[i][j].link[k].numAccumulatedDers; network[i][j].link[k].accErrorDer = 0; network[i][j].link[k].numAccumulatedDers = 0; } } } } }
对应头文件为nn.h
#ifndef __NN_H__ #define __NN_H__ #include "config.h" typedef struct LINK { double weight; double errorDer; double accErrorDer; int numAccumulatedDers; } LINK; typedef LINK *PLINK; typedef struct NODE { double bias; PLINK link; double output; double inputDer; double outputDer; double accInputDer; int numAccumulatedDers; double totalInput; } NODE; typedef NODE *PNODE; typedef PNODE *PPNODE; double getOutPut(); double square(double output, double target); double squareder(double output, double target); double tanhder(double x); // tanh的倒数 double activation(double x); double activationder(double x); double outlayeractivation(double x); double outlayeractivationder(double x); void buildNetwork(); void forwardProp(POINT point); void backProp(POINT point); void updateWeights(); #endif
自动创建与生成训练集与测试集的程序,这里就创建了一个基于半径为5的圆型,圆中间是一部分数据,圆外围是一部分数据。
#include <stdlib.h> #include <math.h> #include "config.h" #include "dataset.h" POINT points[NUMSAMPLES]; void shuffle() { for (int i = 0; i < NUMSAMPLES; i++) { int index = i * ((double)rand() / RAND_MAX); POINT point = points[i]; points[i] = points[index]; points[index] = point; } } // 创建NUMSAMPLES个参数,按照原型来创建 void classifyCircleData() { double radius = 5; // 创建内部圆上的点 for (int i = 0; i < NUMSAMPLES / 2; i++) { double r = 0.5 * radius * rand() / RAND_MAX; // 生成随机的半径 double angle = 2.0 * M_PI * rand() / RAND_MAX; // 生成随机的角度 points[i].x = r * cos(angle); points[i].y = r * sin(angle); points[i].label = 1; } // 创建外部圆上的点 for (int i = NUMSAMPLES / 2; i < NUMSAMPLES; i++) { double r = 0.7 * radius + 0.3 * radius * rand() / RAND_MAX; // 生成随机的半径 double angle = 2.0 * M_PI * rand() / RAND_MAX; // 生成随机的角度 points[i].x = r * cos(angle); points[i].y = r * sin(angle); points[i].label = -1; } shuffle(); }
对应头文件为dataset.h
#ifndef __DATASET_H__ #define __DATASET_H__ typedef struct { double x; double y; double label; } POINT; void classifyCircleData(); #endif
程序配置部分为config.h,定义了数据集大小,学习率以及batchsize大小,还有激活函数,损失函数等应该选什么。
#ifndef __CONFIG_H__ #define __CONFIG_H__ #define NUMSAMPLES 500 // 创建测试点的数量,其中前一半作为训练集,后一半作为测试集 #define LEARNINGRATE 0.03 #define BATCHSIZE 10 #define ACTIVATIONFUNCTION RELU #define OUTLAYERACTIVATIONFUNCTION TANH #endif
main.c主要是调用上述函数,初始化网络以及数据集,以及训练。
#include <stdio.h> #include <stdlib.h> #include <time.h> #include "config.h" #include "dataset.h" #include "nn.h" extern POINT points[NUMSAMPLES]; double getLoss(int mode) // 0代表训练集,1代表测试集 { double loss = 0; if (mode) { for (int i = NUMSAMPLES / 2; i < NUMSAMPLES; i++) { forwardProp(points[i]); loss += square(getOutPut(), points[i].label); } } else { for (int i = 0; i < NUMSAMPLES / 2; i++) { forwardProp(points[i]); loss += square(getOutPut(), points[i].label); } } return loss / (NUMSAMPLES / 2); } void training() { for (int i = 0; i < NUMSAMPLES / 2; i++) { forwardProp(points[i]); backProp(points[i]); if ((i + 1) % BATCHSIZE == 0) { updateWeights(); } } double lossTrain = getLoss(0); double lossTest = getLoss(1); printf("lossTrain:%f,lossTest:%f\n", lossTrain, lossTest); } int main(int argc, char **argv) { srand((unsigned)time(NULL)); classifyCircleData(); buildNetwork(); double lossTrain = getLoss(0); double lossTest = getLoss(1); printf("lossTrain:%f,lossTest:%f\n", lossTrain, lossTest); for (int i = 0; i < 100; i++) { training(); } return 0; }
代码完整地址为:传送门
后期可能会根据我学习的深入继续更新这份代码,就不另行通知了。
文章作者:沃航科技
优秀产品推荐:可编程网络IO控制器