som.class.cpp

人工神经网络BP算法 1、动态改变学习速率 2、加入动量项 3、运用了Matcom4.5的矩阵运算库（可免费下载,头文件matlib.h)

#include <vector>
#include <iostream>
#include <math.h>
#include <stdlib.h>
#include "som.class.const.cpp"

#include "Unit1.h"

using namespace std;
// -----------------------------------------------------------------------
class som_Node
{
    private:
        //
    public:
     vector <double> l_weights;
     som_Node(int,int,int);
     float distance (const vector <double> &);
     void  update(const vector <double> &, const double, const double);
     double color;

     int X;
     int Y;
};

som_Node::som_Node(int numWeights,int n_col,int n_row)
{
  for (int i = 0; i < numWeights; ++i)
     l_weights.push_back(RandFloat());

      X = n_col;
      Y = n_row;
  
}

float som_Node::distance(const vector <double> &Input)
{
    double dist = 0;
    
    for (unsigned int i = 0; i < l_weights.size(); ++i)
     {
          dist += (Input[i] - l_weights[i]) * (Input[i] - l_weights[i]);
     }    
    return dist;
}

void  som_Node::update(const vector <double> &Input, const double LearningRate, const double Influence)
{
    for (unsigned int i = 0; i < l_weights.size(); ++i)
    {
       l_weights[i] += LearningRate * Influence * (Input[i] - l_weights[i]);
    }
}
// ------------------------------------------------------------
class SOM
{
  public:
   SOM(int,int,int,int);
   int Epoch(vector <vector <double> > );
   som_Node* SOM::FindWinner(const vector <double> &);
   int complete();
   double getErrorValue(int,int);
   som_Node **Latice;


   int     s_cols;
   int     s_rows;
   int     Iteration;

  private:
     int     doneTraining;
     int     maxIter;
     int     numIterations;
     double  Radius;
     double  timeConst;
     double  mapRadius;
     double  influence;
     double  learningRate;

     som_Node *winner;
};

SOM::SOM(int rows, int cols,int inputs,int maxI)
{
   Latice = (som_Node **)calloc(rows,sizeof(som_Node *));
   for (int row = 0; row < rows; row++)
   {
     Latice[row] = (som_Node *)calloc(cols,sizeof(som_Node));
      for (int col = 0; col < cols; col++)
         {

            Latice[row][col] = som_Node(
                                    inputs,
                                    row,
                                    col
                             );
         }
    }

   doneTraining = FALSE;
   maxIter = maxI;
   learningRate = constStartLearningRate;
   numIterations = constNumIterations;
   Iteration = 0;
   mapRadius = max(cols, rows)/2;
   timeConst = (double) maxI/log(mapRadius);
   s_cols = cols;
   s_rows = rows;
}

int SOM::Epoch(vector <vector<double> > Input)
{
    if (doneTraining) return 1;

    if (maxIter--)
    {
        int selectPattern = RandInt(0, Input.size()-1);
        winner = FindWinner(Input[selectPattern]);
        Radius = mapRadius * exp(-(double)Iteration/timeConst);

        for (int j = 0; j < s_rows; j++)
         for (int i = 0; i < s_cols; i++)
           {
              double dist = (winner->X - Latice[j][i].X)*(winner->X - Latice[j][i].X) +
                              (winner->Y - Latice[j][i].Y)*(winner->Y - Latice[j][i].Y);

                if (dist < Radius*Radius)
                {
                   influence = exp(-(dist) / (2*Radius*Radius));
                   Latice[j][i].update(Input[selectPattern],learningRate,influence);
                }
           }


       Iteration++;  
       learningRate = constStartLearningRate * exp(-(double)Iteration/numIterations);  
    }  
    else
    {
      doneTraining = 1;
    }

    return 1;
}

double SOM::getErrorValue(int y,int x)     
 {
    double max = 0;

        for (int j = -1; j <= 1; j++)
                for (int i = -1; i <=1; i++)
                   {
                      if  ( (x+i>=0) && (y+j >= 0) && (x + i < s_cols) && (y+j < s_cols) )
                      {
                   
                        if ( (abs(Latice[y][x].X - Latice[y+j][x+i].X) <= 1) &&
                              (abs(Latice[y][x].Y - Latice[y+j][x+i].Y) <= 1) )
                           {
                             if (Latice[y][x].distance(Latice[y+j][x+i].l_weights) > max)
                                     max =  Latice[y][x].distance(Latice[y+j][x+i].l_weights); 
                           } 
                     }

                } 
    return  max;
 }

som_Node* SOM::FindWinner(const vector <double> &pattern)
{
  som_Node *temp_Winner;
  double min = 9999999;
  double dist = 0;


     for (int j = 0; j < s_rows; j++)
      for (int i = 0; i < s_cols; i++)
       {
         dist = Latice[j][i].distance(pattern);
          if (dist < min)
           {
                      min = dist;
                      temp_Winner = &Latice[j][i];
           }
       }
       
    return temp_Winner;
}

int SOM::complete()
{
    return doneTraining;
}