bkneuro.h

類神經網路的BK演算法

		hid.m[h]=1.0/(1.0+exp(-(sum-theta_h.m[h])));
	}
	for(j=0;j<nb_out;j++)
	{
		sum=0.0;
		for(h=0;h<nb_hid;h++)sum=sum+hid.m[h]*poid_hy.m[h][j];
		y.m[j]=1.0/(1.0+exp(-(sum-theta_y.m[j])));
	}
	for(j=0;j<nb_out;j++)delta_y.m[j]=y.m[j]*(1.0-y.m[j])*(desire_y.m[j]-y.m[j]);
	for(h=0;h<nb_hid;h++)
	{
		sum=0.0;
		for(j=0;j<nb_out;j++)sum=sum+poid_hy.m[h][j]*delta_y.m[j];
		delta_h.m[h]=hid.m[h]*(1.0-hid.m[h])*sum;
	}
	for(j=0;j<nb_out;j++)for(h=0;h<nb_hid;h++)dPoid_hy.m[h][j]=0.0;
	for(h=0;h<nb_hid;h++)for(i=0;i<nb_dim;i++)dPoid_xh.m[i][h]=0.0;
	for(j=0;j<nb_out;j++)dTheta_y.m[j]=0.0;
	for(h=0;h<nb_hid;h++)dTheta_h.m[h]=0.0;
	for(j=0;j<nb_out;j++){for(h=0;h<nb_hid;h++)
		{dPoid_hy.m[h][j]=eta*delta_y.m[j]*hid.m[h]+alpha*dPoid_hy.m[h][j];}}
	for(j=0;j<nb_out;j++)dTheta_y.m[j]=-eta*delta_y.m[j]+alpha*dTheta_y.m[j];
	for(h=0;h<nb_hid;h++){for(i=0;i<nb_dim;i++)
		{dPoid_xh.m[i][h]=eta*delta_h.m[h]*x.m[i]+alpha*dPoid_xh.m[i][h];}}
	for(h=0;h<nb_hid;h++)dTheta_h.m[h]=-eta*delta_h.m[h]+alpha*dTheta_h.m[h];
	for(j=0;j<nb_out;j++){for(h=0;h<nb_hid;h++)
		{poid_hy.m[h][j]=poid_hy.m[h][j]+dPoid_hy.m[h][j];}}
	for(j=0;j<nb_out;j++)theta_y.m[j]=theta_y.m[j]+dTheta_y.m[j];
	for(h=0;h<nb_hid;h++){for(i=0;i<nb_dim;i++)
		{poid_xh.m[i][h]=poid_xh.m[i][h]+dPoid_xh.m[i][h];}}
	for(h=0;h<nb_hid;h++)theta_h.m[h]=theta_h.m[h]+dTheta_h.m[h];
	ModifyPara();
	float err=0.0;
    for(j=0;j<nb_out;j++)err+=fabs(desire_y.m[j]-y.m[j]);
	return err/nb_out;
}


void BKNeuroNet::ModifyPara()
{
	alpha=alpha*alpha_rate;
	if(alpha<alpha_min)alpha=alpha_min;
	eta=eta*eta_rate;
	if(eta<eta_min)eta=eta_min;
}
//	Fonction membre BKNeuroNet::Using
f1D& BKNeuroNet::Using()
{
	int h,i,j;
	for(i=0;i<nb_dim;i++)x.m[i]=(x.m[i]-min.m[i])*scaling.m[i];
	for(h=0; h<nb_hid;h++)
	{
		sum=0.0;
		for(i=0;i<nb_dim;i++)sum=sum+x.m[i]*poid_xh.m[i][h];
		hid.m[h]=1.0/(1.0+exp(-(sum-theta_h.m[h])));
	}
	for(j=0;j<nb_out;j++)
	{
		sum=0.0;
		for(h=0;h<nb_hid;h++)sum=sum+hid.m[h]*poid_hy.m[h][j];
		y.m[j]=1.0/(1.0+exp(-(sum-theta_y.m[j])));
		y.m[j]/=scaling[nb_dim+j];
      y.m[j]+=min[nb_dim+j];
	}
	return y;
}

//
f1D& BKNeuroNet::Using(f1D &d)
{
	int h,i,j;
	for(i=0;i<nb_dim;i++)d.m[i]=(d.m[i]-min.m[i])*scaling.m[i];
	for(h=0; h<nb_hid;h++)
	{
		sum=0.0;
		for(i=0;i<nb_dim;i++)sum=sum+d.m[i]*poid_xh.m[i][h];
		hid.m[h]=1.0/(1.0+exp(-(sum-theta_h.m[h])));
	}
	for(j=0;j<nb_out;j++)
	{
		sum=0.0;
		for(h=0;h<nb_hid;h++)sum=sum+hid.m[h]*poid_hy.m[h][j];
		y.m[j]=1.0/(1.0+exp(-(sum-theta_y.m[j])));
		y.m[j]/=scaling[nb_dim+j];
      y.m[j]+=min[nb_dim+j];
	}
	return y;
}

// Transcription des poids de connexion en tableau de d閜endance
f2D BKNeuroNet::TransDependantTable()
{
	f2D dtab(nb_dim,nb_out);
	int i,j,k;
	float acc;
	for(j=0;j<nb_out;j++)for(i=0;i<nb_dim;i++)
	{
		dtab.m[i][j]=0.0;
		for(k=0;k<nb_hid;k++)
		{
			dtab.m[i][j]+=poid_xh.m[i][k]*poid_hy.m[k][j];
		}
	}
	return dtab;
}

void BKNeuroNet::LoadTrainingData(ifstream& in)
{
	 int i,j,n1,n2,n3;
    in>>n1>>n2>>n3;

// Network Initialization
	nb_dim=n1;
	nb_out=n2;
	if(nb_hid==0)nb_hid=(nb_dim+nb_out)*2;
	MemoryAlloc();
	SetRandomPoid();
	eta=5.0;eta_rate=0.99999;eta_min=0.00001;
	alpha=1.0;alpha_rate=0.99998;alpha_min=0.00001;
   input.Initialize(n3,n1);
   output.Initialize(n3,n2);
   nb_record=n3;

// Data loading
	for(j=0;j<nb_dim+nb_out;j++){max[j]=-1e20;min[j]=1e20;scaling[j]=1.0;}

    for(i=0;i<nb_record;i++)
    {
    	for(j=0;j<nb_dim;j++)
    	{
			in>>input.m[i][j];
			if(max[j]<input.m[i][j])max[j]=input.m[i][j];
			if(min[j]>input.m[i][j])min[j]=input.m[i][j];
    	}
		for(j=0;j<nb_out;j++)
		{
        	in>>output.m[i][j];
			if(max[j+nb_dim]<output.m[i][j])max[j+nb_dim]=output.m[i][j];
			if(min[j+nb_dim]>output.m[i][j])min[j+nb_dim]=output.m[i][j];
		}
		for(j=0;j<nb_dim+nb_out;j++)
		{
			if((min[j]==0)&&(max[j]==0))scaling[j]=1.0;
        	else if(max[j]==min[j])scaling[j]=1.0/max[j];
        	else scaling[j]=1./(max[j]-min[j]);
		}
    }
// Nomalization to [0,1]
    for(i=0;i<nb_record;i++)
    {
    	for(j=0;j<nb_dim;j++)input.m[i][j]=(input.m[i][j]-min[j])*scaling[j];
      for(j=0;j<nb_out;j++)output.m[i][j]=(output.m[i][j]-min[j+nb_dim])*scaling[j+nb_dim];
    }
}

void BKNeuroNet::LoadTrainingData(int n1,int n2,int n3, f2D& XV, f2D& YV)
{
	int i,j;
	if(!vide)MemoryFree();
// Network Initialization
	nb_dim=n1;
	nb_out=n2;
	nb_hid=(nb_dim+nb_out)/2;

	MemoryAlloc();
//	SetRandomPoid();
	eta=5.0;eta_rate=0.99999;eta_min=0.00001;
	alpha=1.0;alpha_rate=0.99998;alpha_min=0.00001;
   input.nr=XV.nr;input.nc=XV.nc;input.m=XV.m;
   output.nr=YV.nr;output.nc=YV.nc;output.m=YV.m;
   nb_record=n3;

// Data loading
	for(j=0;j<nb_dim+nb_out;j++){max[j]=-1e20;min[j]=1e20;scaling[j]=1.0;}
	for(i=0;i<nb_record;i++)
	{
		for(j=0;j<nb_dim;j++)
    	{
			if(max[j]<input.m[i][j])max[j]=input.m[i][j];
			if(min[j]>input.m[i][j])min[j]=input.m[i][j];
    	}
		for(j=0;j<nb_out;j++)
		{
			if(max[j+nb_dim]<output.m[i][j])max[j+nb_dim]=output.m[i][j];
			if(min[j+nb_dim]>output.m[i][j])min[j+nb_dim]=output.m[i][j];
		}
		for(j=0;j<nb_dim+nb_out;j++)
		{
			if((min[j]==0)&&(max[j]==0))scaling[j]=1.0;
        	else if(max[j]==min[j])scaling[j]=1.0/max[j];
        	else scaling[j]=1./(max[j]-min[j]);
		}
    }
// Nomalization to [0,1]
    for(i=0;i<nb_record;i++)
    {
    	for(j=0;j<nb_dim;j++)input.m[i][j]=(input.m[i][j]-min[j])*scaling[j];
      for(j=0;j<nb_out;j++)output.m[i][j]=(output.m[i][j]-min[j+nb_dim])*scaling[j+nb_dim];
    }
}

float BKNeuroNet::LearningFromBatchData()
{
  int h,i,j;
  float err=0.0;
  for(int k=0;k<nb_record;k++)
  {
	for(i=0;i<nb_dim;i++)x.m[i]=input.m[k][i];
	for(j=0;j<nb_out;j++)desire_y.m[j]=output.m[k][j];
	for(h=0; h<nb_hid;h++)
	{
		sum=0.0;
		for(i=0;i<nb_dim;i++)sum=sum+x.m[i]*poid_xh.m[i][h];
		hid.m[h]=1.0/(1.0+exp(-(sum-theta_h.m[h])));
	}
	for(j=0;j<nb_out;j++)
	{
		sum=0.0;
		for(h=0;h<nb_hid;h++)sum=sum+hid.m[h]*poid_hy.m[h][j];
		y.m[j]=1.0/(1.0+exp(-(sum-theta_y.m[j])));
	}
	for(j=0;j<nb_out;j++)delta_y.m[j]=y.m[j]*(1.0-y.m[j])*(desire_y.m[j]-y.m[j]);
	for(h=0;h<nb_hid;h++)
	{
		sum=0.0;
		for(j=0;j<nb_out;j++)sum=sum+poid_hy.m[h][j]*delta_y.m[j];
		delta_h.m[h]=hid.m[h]*(1.0-hid.m[h])*sum;
	}
	for(j=0;j<nb_out;j++)for(h=0;h<nb_hid;h++)dPoid_hy.m[h][j]=0.0;
	for(h=0;h<nb_hid;h++)for(i=0;i<nb_dim;i++)dPoid_xh.m[i][h]=0.0;
	for(j=0;j<nb_out;j++)dTheta_y.m[j]=0.0;
	for(h=0;h<nb_hid;h++)dTheta_h.m[h]=0.0;
	for(j=0;j<nb_out;j++){for(h=0;h<nb_hid;h++)
		{dPoid_hy.m[h][j]=eta*delta_y.m[j]*hid.m[h]+alpha*dPoid_hy.m[h][j];}}
	for(j=0;j<nb_out;j++)dTheta_y.m[j]=-eta*delta_y.m[j]+alpha*dTheta_y.m[j];
	for(h=0;h<nb_hid;h++){for(i=0;i<nb_dim;i++)
		{dPoid_xh.m[i][h]=eta*delta_h.m[h]*x.m[i]+alpha*dPoid_xh.m[i][h];}}
	for(h=0;h<nb_hid;h++)dTheta_h.m[h]=-eta*delta_h.m[h]+alpha*dTheta_h.m[h];
	for(j=0;j<nb_out;j++){for(h=0;h<nb_hid;h++)
		{poid_hy.m[h][j]=poid_hy.m[h][j]+dPoid_hy.m[h][j];}}
	for(j=0;j<nb_out;j++)theta_y.m[j]=theta_y.m[j]+dTheta_y.m[j];
	for(h=0;h<nb_hid;h++){for(i=0;i<nb_dim;i++)
		{poid_xh.m[i][h]=poid_xh.m[i][h]+dPoid_xh.m[i][h];}}
	for(h=0;h<nb_hid;h++)theta_h.m[h]=theta_h.m[h]+dTheta_h.m[h];
	ModifyPara();
    for(j=0;j<nb_out;j++)err+=fabs(desire_y.m[j]-y.m[j]);
  }
  err=err/(nb_record*nb_out);
  return err;
}

void BKNeuroNet::ApplyToTestingData(ifstream& in, ofstream& out)
{
	int i,j,n1,n2,n3;
	in>>n1>>n2>>n3;
	if((n1!=nb_dim)||(n2!=nb_out))return;
	out<<"observed"<<"\t"<<"infered\n";
	for(i=0;i<n3;i++)
	{
		for(j=0;j<nb_dim;j++)in>>x.m[j];
		for(j=0;j<nb_out;j++)in>>desire_y.m[j];
		Using();
		for(j=0;j<nb_out;j++){out<<desire_y.m[j]<<"\t"<<y.m[j]<<"\t";}
		out<<endl;
   }
}

void BKNeuroNet::ApplyToTestingData(ifstream& in, f2D& observed, f2D& infered)
{
	int i,j,n1,n2,n3;
	in>>n1>>n2>>n3;
	if((n1!=nb_dim)||(n2!=nb_out))return;
    observed.Initialize(n3,n2);
    infered.Initialize(n3,n2);
	for(i=0;i<n3;i++)
	{
		for(j=0;j<nb_dim;j++)in>>x.m[j];
		for(j=0;j<nb_out;j++)in>>desire_y.m[j];
		Using();
		for(j=0;j<nb_out;j++)
        {
        	observed.m[i][j]=desire_y.m[j];
            infered.m[i][j]=y.m[j];
        }
   }
}

void BKNeuroNet::ApplyToData(f2D& Input, f2D& Output)
{
	int i,j;
	if((Input.nc!=nb_dim)||(Output.nc!=nb_out))return;
	for(i=0;i<Input.nr;i++)
	{
		for(j=0;j<nb_dim;j++)x.m[j]=Input.m[i][j];
		Using();
		for(j=0;j<nb_out;j++)Output.m[i][j]=y.m[j];
   }
}
#endif