simplenn/src/main/java/Back.java

   1 import java.applet.Applet;\r
   2 import java.awt.*;\r
   3 import java.awt.event.*;\r
   4 \r
   5 public class Back extends Applet implements MouseListener,MouseMotionListener,ActionListener{\r
   6 \r
   7    Button button1,button2,button3,button4;\r
   8 \r
   9    int X0=10,X1=125;\r
  10    int Y0=55,Y1=70,Y2=160,Y3=240,Y4=305;\r
  11 \r
  12    int RX0=30,RX1=60,RX2=210,RX3=260;\r
  13    int RY0=225,RY1=240;\r
  14 \r
  15    int WIDTH=7;              //入力データの幅\r
  16    int HEIGHT=11;            //入力データの高さ\r
  17    int INPUT=WIDTH*HEIGHT;   //入力層の数（入力データ数）\r
  18    int HIDDEN=16;            //隠れ層の数\r
  19    int PATTERN=10;           //パターンの種類\r
  20    int OUTPUT=PATTERN;       //出力層の数（出力データ数）\r
  21    int OUTER_CYCLES=200;     //外部サイクル（一連のパターンの繰返し学習）の回数\r
  22    int INNER_CYCLES=200;     //内部サイクル（同一パターンの繰返し学習）の回数\r
  23    float ALPHA=1.2f;         //学習の加速係数\r
  24    float BETA=1.2f;          //シグモイド曲線の傾斜\r
  25 \r
  26    int[] sample_in=new int[INPUT];                  //学習用入力\r
  27    int[] written_in=new int[INPUT];                 //認識用手書き入力\r
  28 \r
  29    float[][] weight_ih=new float[INPUT][HIDDEN];    //入力層と隠れ層の間の重み係数\r
  30    float[] thresh_h=new float[HIDDEN];              //隠れ層の閾値\r
  31    float[] hidden_out=new float[HIDDEN];            //隠れ層の出力\r
  32 \r
  33    float[][] weight_ho=new float[HIDDEN][OUTPUT];   //隠れ層と出力層の間の重み係数\r
  34    float[] thresh_o=new float[OUTPUT];              //出力層の閾値\r
  35    float[] recog_out=new float[OUTPUT];             //認識出力（出力層の出力）\r
  36 \r
  37    int[] teach=new int[PATTERN];                    //教師信号\r
  38 \r
  39 \r
  40 \r
  41 \r
  42 \r
  43    boolean learning_flag;  //「学習モード」フラグ\r
  44 \r
  45    //学習用入力データの基となるパターン\r
  46    int[][] sample_array={{0,0,1,1,1,0,0,  //'0'\r
  47                          0,1,0,0,0,1,0,\r
  48                          1,0,0,0,0,0,1,\r
  49                          1,0,0,0,0,0,1,\r
  50                          1,0,0,0,0,0,1,\r
  51                          1,0,0,0,0,0,1,\r
  52                          1,0,0,0,0,0,1,\r
  53                          1,0,0,0,0,0,1,\r
  54                          1,0,0,0,0,0,1,\r
  55                          0,1,0,0,0,1,0,\r
  56                          0,0,1,1,1,0,0},\r
  57   \r
  58                         {0,0,0,1,0,0,0,  //'1'\r
  59                          0,0,0,1,0,0,0,\r
  60                          0,0,0,1,0,0,0,\r
  61                          0,0,0,1,0,0,0,\r
  62                          0,0,0,1,0,0,0,\r
  63                          0,0,0,1,0,0,0,\r
  64                          0,0,0,1,0,0,0,\r
  65                          0,0,0,1,0,0,0,\r
  66                          0,0,0,1,0,0,0,\r
  67                          0,0,0,1,0,0,0,\r
  68                          0,0,0,1,0,0,0},\r
  69  \r
  70                         {0,0,1,1,1,0,0,  //'2'\r
  71                          0,1,0,0,0,1,0,\r
  72                          1,0,0,0,0,0,1,\r
  73                          0,0,0,0,0,0,1,\r
  74                          0,0,0,0,0,0,1,\r
  75                          0,0,0,0,0,1,0,\r
  76                          0,0,0,0,1,0,0,\r
  77                          0,0,0,1,0,0,0,\r
  78                          0,0,1,0,0,0,0,\r
  79                          0,1,0,0,0,0,0,\r
  80                          1,1,1,1,1,1,1},\r
  81   \r
  82                         {0,0,1,1,1,0,0,  //'3'\r
  83                          0,1,0,0,0,1,0,\r
  84                          1,0,0,0,0,0,1,\r
  85                          0,0,0,0,0,1,0,\r
  86                          0,0,0,0,1,0,0,\r
  87                          0,0,0,0,0,1,0,\r
  88                          0,0,0,0,0,0,1,\r
  89                          0,0,0,0,0,0,1,\r
  90                          1,0,0,0,0,0,1,\r
  91                          0,1,0,0,0,1,0,\r
  92                          0,0,1,1,1,0,0},\r
  93  \r
  94                         {0,0,0,0,1,0,0,  //'4'\r
  95                          0,0,0,1,1,0,0,\r
  96                          0,0,1,0,1,0,0,\r
  97                          0,0,1,0,1,0,0,\r
  98                          0,1,0,0,1,0,0,\r
  99                          0,1,0,0,1,0,0,\r
 100                          1,0,0,0,1,0,0,\r
 101                          1,1,1,1,1,1,1,\r
 102                          0,0,0,0,1,0,0,\r
 103                          0,0,0,0,1,0,0,\r
 104                          0,0,0,0,1,0,0},\r
 105  \r
 106                         {1,1,1,1,1,1,1,  //'5'\r
 107                          1,0,0,0,0,0,0,\r
 108                          1,0,0,0,0,0,0,\r
 109                          1,0,0,0,0,0,0,\r
 110                          1,1,1,1,1,0,0,\r
 111                          0,0,0,0,0,1,0,\r
 112                          0,0,0,0,0,0,1,\r
 113                          0,0,0,0,0,0,1,\r
 114                          0,0,0,0,0,0,1,\r
 115                          1,0,0,0,0,1,0,\r
 116                          0,1,1,1,1,1,0},\r
 117  \r
 118                         {0,0,0,0,1,1,0,  //'6'\r
 119                          0,0,0,1,0,0,0,\r
 120                          0,0,1,0,0,0,0,\r
 121                          0,1,0,0,0,0,0,\r
 122                          0,1,0,0,0,0,0,\r
 123                          1,0,0,0,0,0,0,\r
 124                          1,0,1,1,1,0,0,\r
 125                          1,1,0,0,0,1,0,\r
 126                          1,0,0,0,0,0,1,\r
 127                          0,1,0,0,0,1,0,\r
 128                          0,0,1,1,1,0,0},\r
 129  \r
 130                         {1,1,1,1,1,1,1,  //'7'\r
 131                          0,0,0,0,0,0,1,\r
 132                          0,0,0,0,0,0,1,\r
 133                          0,0,0,0,0,1,0,\r
 134                          0,0,0,0,0,1,0,\r
 135                          0,0,0,0,1,0,0,\r
 136                          0,0,0,0,1,0,0,\r
 137                          0,0,0,1,0,0,0,\r
 138                          0,0,0,1,0,0,0,\r
 139                          0,0,1,0,0,0,0,\r
 140                          0,0,1,0,0,0,0},\r
 141  \r
 142                         {0,0,1,1,1,0,0,  //'8'\r
 143                          0,1,0,0,0,1,0,\r
 144                          1,0,0,0,0,0,1,\r
 145                          1,0,0,0,0,0,1,\r
 146                          0,1,0,0,0,1,0,\r
 147                          0,0,1,1,1,0,0,\r
 148                          0,1,0,0,0,1,0,\r
 149                          1,0,0,0,0,0,1,\r
 150                          1,0,0,0,0,0,1,\r
 151                          1,0,0,0,0,0,1,\r
 152                          0,1,1,1,1,1,0},\r
 153                  \r
 154                         {0,1,1,1,1,1,0,  //'9'\r
 155                          1,0,0,0,0,0,1,\r
 156                          1,0,0,0,0,0,1,\r
 157                          1,0,0,0,0,0,1,\r
 158                          0,1,1,1,1,1,1,\r
 159                          0,0,0,0,0,0,1,\r
 160                          0,0,0,0,0,0,1,\r
 161                          0,0,0,0,0,0,1,\r
 162                          0,0,0,0,0,0,1,\r
 163                          1,0,0,0,0,0,1,\r
 164                          0,1,1,1,1,1,0}};\r
 165 \r
 166    int[][] teach_array=new int[PATTERN][OUTPUT];  //パターンと出力すべき教師信号の比較表\r
 167 \r
 168    int x_new,y_new,x_old,y_old;           //手書き文字入力用座標\r
 169 \r
 170 \r
 171    public void init(){\r
 172 \r
 173       setBackground(Color.gray);\r
 174 \r
 175       //ボタンの設定\r
 176       add(button1=new Button("  再学習  "));\r
 177       add(button2=new Button(" 学習終了 "));\r
 178       add(button3=new Button("入力クリヤ"));\r
 179       add(button4=new Button("  認  識  "));\r
 180       button1.addActionListener(this);\r
 181       button2.addActionListener(this);\r
 182       button3.addActionListener(this);\r
 183       button4.addActionListener(this);\r
 184 \r
 185       //マウスの設定\r
 186       addMouseListener(this);\r
 187       addMouseMotionListener(this);\r
 188 \r
 189       //教師信号の設定\r
 190       for(int q=0;q<PATTERN;q++)\r
 191          for(int k=0;k<OUTPUT;k++){\r
 192             if(q==k) teach_array[q][k]=1;\r
 193             else     teach_array[q][k]=0;\r
 194          }\r
 195 \r
 196       //モードの初期設定\r
 197       learning_flag=true;\r
 198 \r
 199    }\r
 200 \r
 201    //------------------- ボタン関係のメソッド ------------------\r
 202 \r
 203    public void actionPerformed(ActionEvent ae){\r
 204 \r
 205       if(ae.getSource()==button1){      //「再学習」\r
 206          learning_flag=true;\r
 207          repaint();\r
 208       }\r
 209       if(ae.getSource()==button2){      //「学習終了」\r
 210          learning_flag=false;\r
 211          repaint();\r
 212       }\r
 213       if(ae.getSource()==button3){      //「入力クリヤ」\r
 214          if(!learning_flag)\r
 215             repaint();\r
 216       }\r
 217       if(ae.getSource()==button4){      //「認識」\r
 218          if(!learning_flag)\r
 219             recognizeCharacter();\r
 220       }\r
 221 \r
 222    }\r
 223 \r
 224    //---------- マウス関係のメソッド（手書き文字入力）----------\r
 225 \r
 226    public void mousePressed(MouseEvent me){\r
 227       int x=me.getX();\r
 228       int y=me.getY();\r
 229       if(!learning_flag && x>=RX1 && x<=RX1+WIDTH*10 && y>=RY1 && y<=RY1+HEIGHT*10){\r
 230          x_old=me.getX();\r
 231          y_old=me.getY();\r
 232          written_in[(y_old-RY1)/10*WIDTH+(x_old-RX1)/10]=1;\r
 233       }\r
 234    }\r
 235 \r
 236    public void mouseClicked(MouseEvent me){}\r
 237    public void mouseEntered(MouseEvent me){}\r
 238    public void mouseExited(MouseEvent me){}\r
 239    public void mouseReleased(MouseEvent me){}\r
 240 \r
 241    public void mouseDragged(MouseEvent me){\r
 242       int x=me.getX();\r
 243       int y=me.getY();\r
 244       if(!learning_flag && x>=RX1 && x<=RX1+WIDTH*10 && y>=RY1 && y<=RY1+HEIGHT*10){\r
 245          Graphics g=getGraphics(); \r
 246          x_new=me.getX();\r
 247          y_new=me.getY();\r
 248          g.drawLine(x_old,y_old,x_new,y_new);\r
 249          x_old=x_new;\r
 250          y_old=y_new;\r
 251          written_in[(y_old-RY1)/10*WIDTH+(x_old-RX1)/10]=1;\r
 252       }\r
 253  \r
 254    }\r
 255 \r
 256    public void mouseMoved(MouseEvent me){}\r
 257  \r
 258 \r
 259 \r
 260    //---------- 起動時およびrepaint()で呼び出されるメソッド ----------\r
 261  \r
 262    public void paint(Graphics g){\r
 263 \r
 264       int i,j,k,p,q,r,x;\r
 265 \r
 266       String string;\r
 267 \r
 268       float outer_error;          //外部サイクルエラー累計\r
 269       float inner_error;          //内部サイクルエラー累計\r
 270       float temp_error;           //隠れ層の誤差の累計 \r
 271 \r
 272       //学習モードの背景\r
 273       if(learning_flag){\r
 274          g.setColor(new Color(255,255,192));\r
 275          g.fillRect(5,35,590,460);\r
 276          g.setColor(Color.black);\r
 277          g.drawString("学習モード",500,55);\r
 278       }\r
 279 \r
 280       //認識モードの背景\r
 281       else{\r
 282          g.setColor(new Color(192,255,255));\r
 283          g.fillRect(5,35,590,460);\r
 284          g.setColor(Color.black);\r
 285          g.drawString("認識モード",500,55);\r
 286       }\r
 287 \r
 288       //学習用パターンの表示\r
 289       g.drawString("使用している学習用パターン",X0,Y0);\r
 290       for(q=0;q<PATTERN;q++){\r
 291          x=56*q;\r
 292          for(j=0;j<HEIGHT;j++)\r
 293             for(i=0;i<WIDTH;i++){\r
 294                if(sample_array[q][WIDTH*j+i]==1)     g.setColor(Color.red);\r
 295                else                                  g.setColor(Color.cyan);\r
 296                g.fillRect(X0+x+6*i,Y1+6*j,5,5);\r
 297             }\r
 298       }\r
 299       g.setColor(Color.black);\r
 300 \r
 301       //-------------------------------------------------------------------\r
 302       //--------------------------- 学習モード ----------------------------\r
 303       //-------------------------------------------------------------------\r
 304       if(learning_flag){\r
 305 \r
 306          //閾値と重みの乱数設定\r
 307          for(j=0;j<HIDDEN;j++){\r
 308             thresh_h[j]=(float)Math.random()-0.5f;\r
 309             for(i=0;i<INPUT;i++)\r
 310                weight_ih[i][j]=(float)Math.random()-0.5f;\r
 311          }\r
 312          for(k=0;k<OUTPUT;k++){\r
 313             thresh_o[k]=(float)Math.random()-0.5f;\r
 314             for(j=0;j<HIDDEN;j++)\r
 315                weight_ho[j][k]=(float)Math.random()-0.5f;\r
 316          }\r
 317 \r
 318          //-------------------------- 学習 --------------------------\r
 319 \r
 320          for(p=0;p<OUTER_CYCLES;p++){     //外部サイクル\r
 321 \r
 322             outer_error=0.0f;         //外部二乗誤差のクリヤー\r
 323 \r
 324             for(q=0;q<PATTERN;q++){   //パターンの切り替え\r
 325 \r
 326                //パターンに対応した入力と教師信号の設定\r
 327                sample_in=sample_array[q];\r
 328                teach=teach_array[q];\r
 329 \r
 330                for(r=0;r<INNER_CYCLES;r++){   //内部サイクル\r
 331 \r
 332                   //順方向演算\r
 333                   forwardNeuralNet(sample_in,recog_out);       \r
 334 \r
 335                   //逆方向演算（バックプロパゲーション）\r
 336                   backwardNeuralNet();\r
 337 \r
 338                }\r
 339 \r
 340                //内部二乗誤差の計算\r
 341                inner_error=0.0f;   //内部二乗誤差のクリヤー\r
 342                for(k=0;k<OUTPUT;k++)\r
 343                   inner_error+=(teach[k]-recog_out[k])*(teach[k]-recog_out[k]);\r
 344 \r
 345                outer_error+=inner_error;   //外部二乗誤差への累加算\r
 346 \r
 347             }\r
 348 \r
 349             //外部サイクルの回数と外部二乗誤差の表示\r
 350             g.drawString("実行中の外部サイクルの回数と二乗誤差",X0,Y2);\r
 351             g.setColor(new Color(255,255,192));\r
 352             g.fillRect(X0+5,Y2+10,200,50);   //以前の表示を消去\r
 353             g.setColor(Color.black);\r
 354             g.drawString("OuterCycles="+String.valueOf(p),X0+10,Y2+25);\r
 355             g.drawString("TotalSquaredError="+String.valueOf(outer_error),X0+10,Y2+45);\r
 356 \r
 357          } \r
 358 \r
 359 \r
 360          //--------------------- 学習結果の確認 ---------------------\r
 361 \r
 362          g.drawString("学習結果の確認",X0,Y3);\r
 363          for(k=0;k<OUTPUT;k++){\r
 364             g.drawString("Output",X1+45*k,Y3+25);\r
 365             g.drawString("  ["+String.valueOf(k)+"]",X1+5+45*k,Y3+40);\r
 366          }      \r
 367 \r
 368          for(q=0;q<PATTERN;q++){\r
 369 \r
 370             //入力パターンの設定\r
 371             sample_in=sample_array[q];\r
 372 \r
 373             //順方向演算\r
 374             forwardNeuralNet(sample_in,recog_out);\r
 375 \r
 376             //結果の表示\r
 377             g.setColor(Color.black);\r
 378             g.drawString("TestPattern["+String.valueOf(q)+"]",X0+10,Y4+20*q);\r
 379             for(k=0;k<OUTPUT;k++){\r
 380                if(recog_out[k]>0.99){        //99% より大は、赤で YES と表示\r
 381                   g.setColor(Color.red);\r
 382                   string="YES";\r
 383                }\r
 384                else if(recog_out[k]<0.01){   // 1% より小は、青で NO と表示\r
 385                   g.setColor(Color.blue);\r
 386                   string="NO ";\r
 387                }\r
 388                else{                         // 1% 以上 99% 以下は、黒で ? と表示\r
 389                   g.setColor(Color.black);\r
 390                   string=" ? ";\r
 391                }\r
 392                g.drawString(string,X1+10+45*k,Y4+20*q);\r
 393             }\r
 394 \r
 395          }\r
 396       }\r
 397 \r
 398       //-------------------------------------------------------------------\r
 399       //--------------------------- 認識モード ----------------------------\r
 400       //-------------------------------------------------------------------\r
 401       else{\r
 402          g.setColor(Color.black);\r
 403          g.drawString("マウスで数字を描いて下さい",RX0,RY0);\r
 404          g.drawRect(RX1-1,RY1-1,WIDTH*10+2,HEIGHT*10+2);     //外枠\r
 405          g.setColor(Color.gray);\r
 406          for(j=1;j<HEIGHT;j++)\r
 407             g.drawLine(RX1,RY1+10*j,RX1+WIDTH*10,RY1+10*j);  //横方向区切り\r
 408          for(i=1;i<WIDTH;i++)\r
 409             g.drawLine(RX1+10*i,RY1,RX1+10*i,RY1+HEIGHT*10);  //縦方向区切り\r
 410          for(i=0;i<INPUT;i++)\r
 411             written_in[i]=0;     //手書き入力データのクリヤ\r
 412       }\r
 413 \r
 414    }\r
 415 \r
 416    //順方向演算のメソッド\r
 417    public void forwardNeuralNet(int[] input,float[] output){\r
 418 \r
 419       float[] out=new float[OUTPUT];\r
 420       float[] hidden=new float[HIDDEN];\r
 421 \r
 422       //隠れ層出力の計算\r
 423       for(int j=0;j<HIDDEN;j++){\r
 424          hidden[j]=-thresh_h[j];\r
 425          for(int i=0;i<INPUT;i++)\r
 426             hidden[j]+=input[i]*weight_ih[i][j];\r
 427          hidden_out[j]=sigmoid(hidden[j]);\r
 428       }\r
 429 \r
 430       //出力層出力の計算\r
 431       for(int k=0;k<OUTPUT;k++){\r
 432          out[k]=-thresh_o[k];\r
 433          for(int j=0;j<HIDDEN;j++)\r
 434             out[k]+=hidden_out[j]*weight_ho[j][k];\r
 435          output[k]=sigmoid(out[k]);\r
 436       }\r
 437 \r
 438    }\r
 439 \r
 440    //逆方向演算のメソッド\r
 441    public void backwardNeuralNet(){\r
 442 \r
 443       int i,j,k;\r
 444 \r
 445       float[] output_error=new float[OUTPUT];       //出力層の誤差\r
 446       float[] hidden_error=new float[HIDDEN];       //隠れ層の誤差\r
 447 \r
 448       float temp_error;\r
 449 \r
 450       //出力層の誤差の計算\r
 451       for(k=0;k<OUTPUT;k++)\r
 452          output_error[k]=(teach[k]-recog_out[k])*recog_out[k]*(1.0f-recog_out[k]);\r
 453 \r
 454       //隠れ層の誤差の計算\r
 455       for(j=0;j<HIDDEN;j++){\r
 456          temp_error=0.0f;\r
 457          for(k=0;k<OUTPUT;k++)\r
 458             temp_error+=output_error[k]*weight_ho[j][k];\r
 459          hidden_error[j]=hidden_out[j]*(1.0f-hidden_out[j])*temp_error;\r
 460       }\r
 461 \r
 462       //重みの補正\r
 463       for(k=0;k<OUTPUT;k++)\r
 464          for(j=0;j<HIDDEN;j++)\r
 465             weight_ho[j][k]+=ALPHA*output_error[k]*hidden_out[j];\r
 466       for(j=0;j<HIDDEN;j++)\r
 467          for(i=0;i<INPUT;i++)\r
 468             weight_ih[i][j]+=ALPHA*hidden_error[j]*sample_in[i];\r
 469 \r
 470       //閾値の補正\r
 471       for(k=0;k<OUTPUT;k++)\r
 472          thresh_o[k]-=ALPHA*output_error[k];\r
 473       for(j=0;j<HIDDEN;j++)\r
 474          thresh_h[j]-=ALPHA*hidden_error[j];\r
 475 \r
 476    }\r
 477   \r
 478    //Sigmoid関数を計算するメソッド\r
 479    public float sigmoid(float x){\r
 480 \r
 481       return 1.0f/(1.0f+(float)Math.exp(-BETA*x));\r
 482 \r
 483    }\r
 484 \r
 485    //入力文字を認識するメソッド\r
 486    public void recognizeCharacter(){\r
 487 \r
 488       Graphics g=getGraphics();\r
 489       String string;\r
 490 \r
 491       //順方向演算\r
 492       forwardNeuralNet(written_in,recog_out);\r
 493 \r
 494       //結果の表示\r
 495       for(int k=0;k<OUTPUT;k++){\r
 496           g.setColor(Color.black);\r
 497           g.drawString(String.valueOf(k)+"である",RX2,RY1+20*k);\r
 498           if(recog_out[k]>0.8f)  g.setColor(Color.red);\r
 499           else                   g.setColor(Color.black);\r
 500 \r
 501           g.fillRect(RX3,RY1-10+20*k,(int)(200*recog_out[k]),10);\r
 502           g.drawString(String.valueOf((int)(100*recog_out[k]+0.5f))+"%",RX3+(int)(200*recog_out[k])+10,RY1+20*k);\r
 503        }\r
 504 \r
 505    }\r
 506 \r
 507 }\r
 508 \r
 509 \r
 510 \r