Merge pull request #1 from CrystSw/dev

Merge dev-branch.

README.md

@@ -1,4 +1,4 @@
-# Simple-Neural-Network
+# Simple-Neural-Network
 ## 概要
 Sample of 2-layer categorization-neural-network(CNN).  
 This CNN consists of 2-layer(input-layer, output-layer). not exist hidden-layer.  
@@ -16,7 +16,9 @@ This CNN consists of 2-layer(input-layer, output-layer). not exist hidden-layer.
 
 の4つのファイルをダウンロードおよび解凍し，src/mnistの中へ入れてください．  
 
-Windows10 64bit(gcc 6.3.0)とCentOS6.10(gcc 4.4.7)で動作確認を行っています．  
+Windows10 64bit(gcc 6.3.0)とCentOS6.10(gcc 4.4.7)で動作確認を行っています． 
+
+**◎ver1.1以前(旧バージョン)**  
 コンパイルを行うには次のコマンドを実行します．  
 ```gcc -lm snnet.c main.c -o snnet```  
 最適化オプションを付けたほうが，動作は早くなります．  
@@ -25,7 +27,18 @@ Windows10 64bit(gcc 6.3.0)とCentOS6.10(gcc 4.4.7)で動作確認を行ってい
 ```./snnet learn```
 
 テストデータを基に識別を行うには次のコマンドを実行します．  
-```./snnet test```
+```./snnet test```  
+
+**◎ver1.2以降(新バージョン)**  
+コンパイルを行うには次のコマンドを実行します．  
+```gcc -lm nnet.c util.c mnist.c main.c -o snnet```  
+最適化オプションを付けたほうが，動作は早くなります．  
+
+教師データを基に学習を行うには次のコマンドを実行します．  
+```./snnet -l```
+
+テストデータを基に識別を行うには次のコマンドを実行します．  
+```./snnet -t```
 
 ---
 ## 注意事項

src/header/mnist.h

@@ -0,0 +1,10 @@
+#ifndef __SNN_MNIST_H_INCLUDED__
+#define __SNN_MNIST_H_INCLUDED__
+
+#include "nnet_types.h"
+
+extern FVALUE read_mnistimg(char *filename);
+extern LABELOHV read_mnistlbl_ohv(char *filename);
+extern LABEL read_mnistlbl(char *filename);
+
+#endif

src/snnet.h → src/header/nnet.h

@@ -1,17 +1,13 @@
-#ifndef __SIMPLE_NNET_H__
-#define __SIMPLE_NNET_H__
+#ifndef __SNN_NNET_H_INCLUDED__
+#define __SNN_NNET_H_INCLUDED__
 
-#include "snnet-types.h"
+#include "nnet_types.h"
 
-extern FVALUE read_mnistimg(char *filename);
-extern LABEL read_mnistlbl(char *filename);
-extern LABELOHV read_mnistlbl_ohv(char *filename);
 extern double softmax(const double *output, const int id, const int size);
 extern double cross_entropy(const double *rpp, const double *wpp, const int size);
 extern void nncout_train(const TRAINDATA td, double **weight, const double *bias, const int id, double *catout);
 extern void nncout_test(const TESTDATA td, double **weight, const double *bias, const int id, double *catout);
 extern void nncpp(const double *catout, const int size, double *catpp);
 extern double xentrloss(const TRAINDATA td, double **weight, const double *bias);
-extern int getmax(const double *p, int size);
 
 #endif

src/snnet-types.h → src/header/nnet_types.h

@@ -1,3 +1,6 @@
+#ifndef __SNN_NNET_TYPES_H_INCLUDED__
+#define __SNN_NNET_TYPES_H_INCLUDED__
+
 typedef unsigned char u_char;
 
 typedef struct {
@@ -29,3 +32,5 @@ typedef struct{
 	FVALUE fval;
 	LABEL label;
 } TESTDATA;
+
+#endif

src/header/util.h

@@ -0,0 +1,16 @@
+#ifndef __SNN_UTIL_H_INCLUDED__
+#define __SNN_UTIL_H_INCLUDED__
+
+extern double *malloc_d1(const int size1);
+extern double *calloc_d1(const int size1);
+extern double **malloc_d2(const int size1, const int size2);
+extern double **calloc_d2(const int size1, const int size2);
+extern void free_d1(double *array);
+extern void free_d2(double **array, const int size1);
+extern int getmax(const double *array, int size1);
+extern void writefile_d1(const char *filename, double *array_d1, const int size1);
+extern void readfile_d1(const char *filename, double *array_d1, const int size1);
+extern void writefile_d2(const char *filename, double **array_d2, const int size1, const int size2);
+extern void readfile_d2(const char *filename, double **array_d2, const int size1, const int size2);
+
+#endif

src/main.c

@@ -2,19 +2,20 @@
 #include <string.h>
 #include <stdlib.h>
 
-#include "snnet.h"
+#include "header/nnet.h"
+#include "header/util.h"
+#include "header/mnist.h"
 
 #define LOSS_CALC_RATE	5000
-//#define OUTPUT_IMAGE
 
 void learn(void);
 void test(void);
 
 int main(int argc, char *argv[]){
 	if(argc > 1){
-		if(!strcmp(argv[1],"learn")){
+		if(!strcmp(argv[1],"-l")){
 			learn();
-		}else if(!strcmp(argv[1],"test")){
+		}else if(!strcmp(argv[1],"-t")){
 			test();
 		}
 	}
@@ -39,20 +40,15 @@ void learn(void){
 	double alpha = 0.0000001;
 
 	//重みとバイアスの宣言(callocで確保します)
-	double **weight, *bias;
-	weight = (double**)calloc(td.label.size, sizeof(double*));
-	for(j = 0; j < td.label.size; ++j){
-		weight[j] = (double*)calloc(td.fval.size, sizeof(double));
-	}
-	bias = (double*)calloc(td.label.size, sizeof(double));
+	double **weight = calloc_d2(td.label.size, td.fval.size);
+	double *bias = calloc_d1(td.label.size);
 
 	//学習前のクロスエントロピーの総和
 	printf("Data: %6d, Loss: %.8f\n", 0, xentrloss(td, weight, bias));
 
 	//カテゴリ出力とカテゴリ事後確率の宣言
-	double *catout, *catpp;
-	catout = (double*)malloc(sizeof(double)*td.label.size);
-	catpp = (double*)malloc(sizeof(double)*td.label.size);
+	double *catout = malloc_d1(td.label.size);
+	double *catpp = malloc_d1(td.label.size);
 
 	//学習
 	for(i = 0; i < td.fval.num; ++i){
@@ -69,43 +65,25 @@ void learn(void){
 			//バイアスの更新
 			bias[j] += alpha*(td.label.data[i][j]-catpp[j]);
 		}
-		//クロスエントロピーの総和
+		//クロスエントロピー誤差(計算負荷の高い処理なので，実行回数は少なめ)
 		if((i+1) % LOSS_CALC_RATE == 0) printf("Data: %6d, Loss: %.8f\n", i+1, xentrloss(td, weight, bias));
 	}
-	free(catout);
-	free(catpp);
+	free_d1(catout);
+	free_d1(catpp);
 
 	//重みの書き出し
-	FILE *wwp;
-	if((wwp = fopen("weight-value","w")) != NULL){
-		for(i = 0; i < td.label.size; ++i){
-			for(j = 0; j < td.fval.size; ++j){
-				fprintf(wwp, "%.20f\n", weight[i][j]);
-			}
-		}
-	}else{
-		fprintf(stderr, "Error: Cannot open file.");
-		exit(1);
-	}
-	fclose(wwp);
-
+	writefile_d2("./result/weight-value", weight, td.label.size, td.fval.size);	
 	//バイアスの書き出し
-	FILE *wbp;
-	if((wbp = fopen("bias-value","w")) != NULL){
-		for(i = 0; i < td.label.size; ++i){
-			fprintf(wbp, "%.20f\n", bias[i]);
-		}
-	}else{
-		fprintf(stderr, "Error: Cannot open file.");
-		exit(1);
-	}
-	fclose(wbp);
+	writefile_d1("./result/bias-value", bias, td.label.size);
 
-	for(i = 0; i < td.label.size; ++i) free(weight[i]);
-	free(weight);
-	free(bias);
+	free_d2(weight, td.label.size);
+	free_d1(bias);
 }
 
+/**
+ * 学習したニューラルネットワークを用いてテストデータの分類を行う．
+ * 結果は混同行列として出力される．
+ */
 void test(void){
 	int i, j, k;
 
@@ -117,50 +95,20 @@ void test(void){
 	puts("Label data has loaded.");
 
 	//重みとバイアスの宣言(callocで確保します)
-	double **weight, *bias;
-	weight = (double**)calloc(td.label.size, sizeof(double*));
-	for(j = 0; j < td.label.size; ++j){
-		weight[j] = (double*)calloc(td.fval.size, sizeof(double));
-	}
-	bias = (double*)calloc(td.label.size, sizeof(double));
+	double **weight = calloc_d2(td.label.size, td.fval.size);
+	double *bias = calloc_d1(td.label.size);
 
 	//重みの読み込み
-	FILE *rwp;
-	if((rwp = fopen("weight-value","r")) != NULL){
-		for(i = 0; i < td.label.size; ++i){
-			for(j = 0; j < td.fval.size; ++j){
-				fscanf(rwp, "%lf", &weight[i][j]);
-			}
-		}
-	}else{
-		fprintf(stderr, "Error: Cannot open file.");
-		exit(1);
-	}
-	fclose(rwp);
-
+	readfile_d2("./result/weight-value", weight, td.label.size, td.fval.size);
 	//バイアスの読み込み
-	FILE *rbp;
-	if((rbp = fopen("bias-value","r")) != NULL){
-		for(i = 0; i < td.label.size; ++i){
-			fscanf(rbp, "%lf", &bias[i]);
-		}
-	}else{
-		fprintf(stderr, "Error: Cannot open file.");
-		exit(1);
-	}
-	fclose(rbp);
+	readfile_d1("./result/bias-value", bias, td.label.size);
 
 	//混同行列
-	double **confmat;
-	confmat = (double**)calloc(td.label.size, sizeof(double*));
-	for(j = 0; j < td.label.size; ++j){
-		confmat[j] = (double*)calloc(td.fval.size, sizeof(double));
-	}
+	double **confmat = calloc_d2(td.label.size, td.fval.size);
 
 	//カテゴリ出力とカテゴリ事後確率の宣言
-	double *catout, *catpp;
-	catout = (double*)malloc(sizeof(double)*td.label.size);
-	catpp = (double*)malloc(sizeof(double)*td.label.size);
+	double *catout = malloc_d1(td.label.size);
+	double *catpp = malloc_d1(td.label.size);
 
 	//カテゴリ分類
 	int m_category;
@@ -171,21 +119,13 @@ void test(void){
 		nncpp(catout, td.label.size, catpp);
 		//事後確率が最大であるカテゴリを導出
 		m_category = getmax(catpp, td.label.size);
+		//分類結果を格納(混同行列の計算に利用)
 		confmat[td.label.data[i]][m_category] += 1;
-		printf("[Data:%d]right:%d predict:%d(pp:%.1f)\n", i+1, td.label.data[i], m_category, catpp[m_category]*100);
-#ifdef OUTPUT_IMAGE
-		//誤った画像の表示
-		if(td.label.data[i] != m_category){
-			puts("-----Image-----");
-			for(j = 0; j < td.fval.size; ++j){
-				putchar(td.fval.data[i][j] < 128 ? '-' : '#');
-				if((j+1) % td.fval.c_size == 0) putchar('\n');
-			}
-		}
-#endif
+		//分類結果を出力
+		printf("[Data:%d]truth:%d predict:%d(pp:%.1f)\n", i+1, td.label.data[i], m_category, catpp[m_category]*100);
 	}
-	free(catout);
-	free(catpp);
+	free_d1(catout);
+	free_d1(catpp);
 
 	//混同行列の計算
 	int cat_num;
@@ -194,7 +134,6 @@ void test(void){
 		for(j = 0; j < td.label.size; ++j) cat_num += confmat[i][j];
 		for(j = 0; j < td.label.size; ++j) confmat[i][j] /= cat_num;
 	}
-
 	//混同行列の出力
 	puts("-----Confusion Matrix-----");
 	printf("   ");
@@ -208,10 +147,16 @@ void test(void){
 		putchar('\n');
 	}
 
-	for(i = 0; i < td.label.size; ++i) free(weight[i]);
-	free(weight);
-	free(bias);
+	//正解率の出力
+	puts("-----Accuracy Rate-----");
+	double arate = 0;
+	for(i = 0; i < td.label.size; ++i){
+		arate += confmat[i][i];
+	}
+	arate /= td.label.size;
+	printf("%3.1f\n", arate*100);
 
-	for(i = 0; i < td.label.size; ++i) free(confmat[i]);
-	free(confmat);
-}
+	free_d2(weight, td.label.size);
+	free_d1(bias);
+	free_d2(confmat, td.label.size);
+}