public FPA NeighborhoodRatio(FPA t) //Sigma(t)
{
FPA exponent = -t / m_Theta;
FPA Sigmat = m_SigmaInitial * PA.Pow2(PA.Log2(LogBase) * exponent);
return(Sigmat);
}
private void run()
{
FPA t1 = PA.Add(PA.InnerProduct(dinput, diwt), theta); // summation and theta
FPA ohidden = PA.Reciprocal(PA.Add(PA.Pow2(PA.Negate(t1)), 1.0f)); //applying sigmoid function
FPA t2 = PA.Add(PA.InnerProduct(ohidden, dowt), tau); // summation and tau
FPA ooutput = PA.Reciprocal(PA.Add(PA.Pow2(PA.Negate(t2)), 1.0f));
FPA oerror = PA.Subtract(doutput, ooutput); //numpat no
FPA herror = PA.Transpose(PA.InnerProduct(dowt, PA.Transpose(oerror, new int[] { 1, 0 })), new int[] { 1, 0 }); // doubtful transpose
herror = PA.Multiply(PA.Multiply(PA.Subtract(1.0f, t1), t1), herror);
FPA _owt = PA.Add(dowt, PA.Multiply(PA.InnerProduct(PA.Transpose(t1, new int[] { 1, 0 }), oerror), betao)); // orig no transpose
FPA _iwt = PA.Multiply(PA.InnerProduct(PA.Transpose(dinput, new int[] { 1, 0 }), herror), betah); //original dinput herror and no transpose
dtau = PA.Add(PA.Multiply(betao, oerror), dtau);
dtheta = PA.Add(PA.Multiply(betah, herror), dtheta); //orig oerror
PA.ToArray(_owt, out owt);
PA.ToArray(_iwt, out iwt);
cleanup();
diwt = new DFPA(iwt);
dowt = new DFPA(owt);
}
private void run()
{
FPA t1 = PA.Add(PA.InnerProduct(dinput, diwt), theta);
FPA ohidden = PA.Reciprocal(PA.Add(PA.Pow2(PA.Negate(t1)), 1.0f));
FPA t2 = PA.Add(PA.InnerProduct(ohidden, dowt), tau);
FPA ooutput = PA.Reciprocal(PA.Add(PA.Pow2(PA.Negate(t2)), 1.0f));
FPA oerror = PA.Subtract(doutput, ooutput);
FPA herror = PA.InnerProduct(dowt, PA.Transpose(oerror, new int[] { 1, 0 }));
herror = PA.InnerProduct(PA.Multiply(PA.Subtract(1.0f, t1), t1), herror);
FPA _owt = PA.Add(dowt, PA.Multiply(PA.InnerProduct(t1, oerror), betao));
FPA _iwt = PA.Multiply(PA.InnerProduct(herror, dinput), betah); //original dinput herror
dtau = PA.Add(PA.Multiply(betao, oerror), dtau);
dtheta = PA.Add(PA.Multiply(betah, herror), dtheta); //orig herror
PA.ToArray(_owt, out owt);
PA.ToArray(_iwt, out iwt);
diwt = new DFPA(owt);
dowt = new DFPA(iwt);
}
public FPA LearningRate(FPA t) //Epsilon(t)
{
FPA exponent = -t / m_Theta;
FPA Epsilont = m_EpsilonInitial * PA.Pow2(PA.Log2(LogBase) * exponent);
return(Epsilont);
}
/*
* Function which performs all the GPU operations
*/
private void run()
{
/* Note : Inner product --- Matrix multiplication
* Multiply -- Element by element multiplication */
FPA t1 = PA.Add(PA.InnerProduct(dinput, diwt), dtheta);
/* ohidden is the output of hidden layer
* Only Sigmoid function is used for timebeing */
FPA ohidden = PA.Reciprocal(PA.Add(PA.Pow(new FPA(2.71828f, new int[] { numpat, nh }), PA.Negate(t1)), 1.0f));
FPA t2 = PA.Add(PA.InnerProduct(ohidden, dowt), dtau);
/* ooutput is the "actual" output of hidden layer
* Only Sigmoid function is used for timebeing */
FPA ooutput = PA.Reciprocal(PA.Add(PA.Pow(new FPA(2.71828f, new int[] { numpat, no }), PA.Negate(t2)), 1.0f));
/* Error between expected and actual */
FPA oerror = PA.Subtract(doutput, ooutput);
/* Checking if error has fallen below 1% if so terminatinf further cycles */
BoolParallelArray b = PA.All(PA.CompareGreater(derror, PA.Abs(oerror)), 1);
b = PA.All(b);
bool[] bt;
PA.ToArray(b, out bt);
if (bt[0] == true)
{
traincycles = 0;
}
/* herror is the error in the hidden layer */
FPA herror = PA.Transpose(PA.InnerProduct(dowt, PA.Transpose(oerror, new int[] { 1, 0 })), new int[] { 1, 0 });
herror = PA.Multiply(PA.Multiply(PA.Subtract(1.0f, ohidden), ohidden), herror);
/* Weights between hidden and output layer being updated */
FPA _owt = PA.Add(PA.Multiply(PA.InnerProduct(PA.Transpose(ohidden, new int[] { 1, 0 }), oerror), betao), dowt);
/* Weights between input and hidden layer being updated */
FPA _iwt = PA.Add(PA.Multiply(PA.InnerProduct(PA.Transpose(dinput, new int[] { 1, 0 }), herror), betah), diwt);
/*Updating threshold for output layer */
dtau = PA.Add(PA.Multiply(betao, oerror), dtau);
/*Updating threshold for hidden layer */
dtheta = PA.Add(PA.Multiply(betah, herror), dtheta);
/* Casting the Parallel arrays to normal arrays */
PA.ToArray(_owt, out owt);
PA.ToArray(_iwt, out iwt);
/* Rebuilding the disposable arrays from newly formed arrays */
diwt = new DFPA(iwt);
dowt = new DFPA(owt);
}
}//Init
public void FindBMU()
{
//Useful locals
int alen = m_Height * m_Width;
//Compute the distances from pattern to code vectors
FPA a = PA.AddDimension(m_CurrentPatternGPU, 1);
FPA x = PA.Stretch(a, 1, alen);
FPA pol = PA.Subtract(m_Weights, x);
FPA pol2 = PA.Multiply(pol, pol);
FPA pol3 = PA.Sum(pol2, 0);
m_Distances = PA.Sqrt(pol3);
//Find the minimal distance
FPA dist2 = PA.AddDimension(m_Distances, 1);
FPA minval = PA.MinVal(dist2, 0);
FPA xxx = PA.Stretch(minval, alen);
//Prepare trigger array
BPA trigger = PA.CompareEqual(xxx, m_Distances);
//BMU Coord ZEROS
BPA trigger2 = PA.AddDimension(trigger, 0);
BPA trigger3 = PA.Stretch(trigger2, 2, 1);
//Extract BMU Coord
FPA lol = PA.Cond(trigger3, m_Shape, zeros);
m_BMUCoord = PA.Sum(lol, 1);
//m_BMUCoord = PA.Evaluate(m_BMUCoord);
//BMU Code Vector ZEROS
BPA triggercv2 = PA.AddDimension(trigger, 0);
BPA triggercv3 = PA.Stretch(triggercv2, m_PatternLength, 1);
//Extract BMU code vector
FPA mdr = PA.Cond(triggercv3, m_Weights, zeroscv);
m_BMUCodeVector = PA.Sum(mdr, 1);
//m_BMUCodeVector = PA.Evaluate(m_BMUCodeVector);
//Begin computation ! ^^ AND OUTPUT
PA.Evaluate(m_BMUCodeVector);
PA.Evaluate(m_BMUCoord);
}
public float[] Test(float[] iinput)
{
float[,] tinput = new float[1, ni];
for (int i = 0; i < ni; i++)
{
tinput[0, i] = iinput[i];
}
dinput = new DFPA(tinput);
diwt = new DFPA(iwt);
dowt = new DFPA(owt);
dtheta = PA.Section(dtheta, new Slice(0, 1), new Slice(0, nh));
dtau = PA.Section(dtau, new Slice(0, 1), new Slice(0, no));
FPA t1 = PA.Add(PA.InnerProduct(dinput, diwt), dtheta);
FPA ohidden = PA.Reciprocal(PA.Add(PA.Pow2(PA.Negate(t1)), 1.0f));
FPA t2 = PA.Add(PA.InnerProduct(ohidden, dowt), dtau);
FPA ooutput = PA.Reciprocal(PA.Add(PA.Pow2(PA.Negate(t2)), 1.0f));
float[,] output;
float[] routput = new float[no];
PA.ToArray(ooutput, out output);
for (int i = 0; i < no; i++)
{
routput[i] = output[0, i];
}
/*Disposable Floating arrays need to be explicitly "disposed" */
dinput.Dispose();
diwt.Dispose();
dowt.Dispose();
doutput.Dispose();
/*Releasing all GPU Resources*/
PA.UnInit();
return(routput);
}
public void DoEpoch(int nb)
{
int alen = m_Width * m_Height;
for (int i = 0; i < nb; ++i)
{
//Neighborhood Function
FPA sbmuc = PA.AddDimension(m_BMUCoord, 1);
sbmuc = PA.Stretch(sbmuc, 1, alen);
sbmuc = sbmuc - m_Shape;
sbmuc = PA.Pow2(sbmuc);
FPA sqdist = PA.Sum(sbmuc, 0);
sqdist = PA.AddDimension(sqdist, 0);
sqdist = PA.Stretch(sqdist, m_PatternLength, 1);
//PA.Evaluate(sqdist);
//
//Learning Rate
FPA lrate = new FPA((float)_CPU_LearningRate(m_time_val), m_PatternLength, alen);
/*FPA sLearningRate = PA.AddDimension(LearningRate(m_Time), 1);
* sLearningRate = PA.Stretch(sLearningRate, m_PatternLength, alen);*/
//Difference between units and current pattern
FPA a = PA.AddDimension(m_CurrentPatternGPU, 1);
FPA x = PA.Stretch(a, 1, alen);
FPA pol = x - m_Weights;
//Calcul des deltas
FPA deltaW = lrate * pol;
//Mise à jour des poids
m_Weights = m_Weights + deltaW;
//Incrémente le compteur de temps
//m_Time = PA.Add(m_Time, 1.0f);
m_time_val += 1;
}
}
public Network(int ini, int inh, int ino, int inumpat, int itraincycles)
{
ni = ini;
no = ino;
nh = inh;
numpat = inumpat;
traincycles = itraincycles;
input = new float[numpat, ni];
output = new float[numpat, no];
iwt = new float[ni, nh];
owt = new float[nh, no];
dtheta = new FPA(theta, new int[] { numpat, nh });
dtau = new FPA(tau, new int[] { numpat, no });
rand = new Random();
theta = tau = 0.35f;
betao = 0.2f;
betah = 0.15f;
}
public void DoEpoch(int nb)
{
int alen = m_Width * m_Height;
for (int i = 0; i < nb; ++i)
{
//Neighborhood Function
FPA sbmuc = PA.AddDimension(m_BMUCoord, 1);
sbmuc = PA.Stretch(sbmuc, 1, alen);
sbmuc = sbmuc - m_Shape;
sbmuc = PA.Pow2(sbmuc);
FPA sqdist = PA.Sum(sbmuc, 0);
sqdist = PA.AddDimension(sqdist, 0);
sqdist = PA.Stretch(sqdist, m_PatternLength, 1);
//PA.Evaluate(sqdist);
//
//Learning Rate
FPA lrate = new FPA((float)_CPU_LearningRate(m_time_val), m_PatternLength, alen);
/*FPA sLearningRate = PA.AddDimension(LearningRate(m_Time), 1);
sLearningRate = PA.Stretch(sLearningRate, m_PatternLength, alen);*/
//Difference between units and current pattern
FPA a = PA.AddDimension(m_CurrentPatternGPU, 1);
FPA x = PA.Stretch(a, 1, alen);
FPA pol = x - m_Weights;
//Calcul des deltas
FPA deltaW = lrate * pol;
//Mise à jour des poids
m_Weights = m_Weights + deltaW;
//Incrémente le compteur de temps
//m_Time = PA.Add(m_Time, 1.0f);
m_time_val += 1;
}
}
public void SetCurrentPattern(float[] Pattern)
{
m_CurrentPattern = Pattern;
m_CurrentPatternGPU = new DisposableFloatParallelArray(Pattern);
}
private void InitMap()
{
// MAP VALUES AND SHAPE
Random RandomNumber = new Random();
weight_vals = new float[m_PatternLength,m_Width * m_Height];
for (int i = 0; i < m_Width * m_Height; ++i)
{
for(int k = 0; k < m_PatternLength; ++k )
weight_vals[k, i] = (float)(RandomNumber.NextDouble()*255.0f);
}
shape_vals = new float[2,m_Width*m_Height];
for( int i = 0; i < m_Height; i++ )
for (int j = 0; j < m_Width; j++)
{
shape_vals[0,m_Width*i + j] = j+1;
shape_vals[1,m_Width*i + j] = i+1;
}
m_Weights = new DisposableFloatParallelArray(weight_vals);
m_Shape = new DFPA(shape_vals);
//
//BMU INIT
m_bmucoord_vals = new float[2];
m_bmucodevector_vals = new float[m_PatternLength];
//ZEROS FOR FINDBMU
zero_vals = new float[2, m_Width * m_Height];
for (int i = 0; i < m_Width * m_Height; ++i)
{
zero_vals[0, i] = 0;
zero_vals[1, i] = 0;
}
zerocv_vals = new float[m_PatternLength, m_Width * m_Height];
for (int i = 0; i < m_Width * m_Height; ++i)
for (int k = 0; k < m_PatternLength; ++k)
{
zerocv_vals[k, i] = 0;
zerocv_vals[k, i] = 0;
}
zeroscv = new DFPA(zerocv_vals);
zeros = new DFPA(zero_vals);
//
//INIT TIME
//Don't use this
m_Time = new DFPA(new float[] { 1 });
//NEIGHBORHOOD VARIABLE INIT
m_maxtime_val = 100.0f;
m_theta_val = 100.0f / (float)Math.Log(25);
m_sigmainitial_val = 25;
m_epsiloninitial_val = 0.1f;
m_Theta = new DFPA(new float[] { m_theta_val });
m_SigmaInitial = new DFPA(new float[] { m_sigmainitial_val });
m_EpsilonInitial = new DFPA(new float[] { m_epsiloninitial_val });
m_MaxTime = new DFPA(new float[] { m_maxtime_val });
//LOG BASE
LogBase = new DFPA(new float[] { (float)(Math.E) });
}
//Sigma(t)
public FPA NeighborhoodRatio(FPA t)
{
FPA exponent = -t / m_Theta;
FPA Sigmat = m_SigmaInitial * PA.Pow2(PA.Log2(LogBase) * exponent);
return Sigmat;
}
//Epsilon(t)
public FPA LearningRate(FPA t)
{
FPA exponent = -t / m_Theta;
FPA Epsilont = m_EpsilonInitial * PA.Pow2(PA.Log2(LogBase) * exponent);
return Epsilont;
}
public void FindBMU()
{
//Useful locals
int alen = m_Height * m_Width;
//Compute the distances from pattern to code vectors
FPA a = PA.AddDimension(m_CurrentPatternGPU, 1);
FPA x = PA.Stretch(a, 1, alen);
FPA pol = PA.Subtract(m_Weights, x);
FPA pol2 = PA.Multiply(pol, pol);
FPA pol3 = PA.Sum(pol2, 0);
m_Distances = PA.Sqrt(pol3);
//Find the minimal distance
FPA dist2 = PA.AddDimension(m_Distances, 1);
FPA minval = PA.MinVal(dist2, 0);
FPA xxx = PA.Stretch(minval, alen);
//Prepare trigger array
BPA trigger = PA.CompareEqual(xxx, m_Distances);
//BMU Coord ZEROS
BPA trigger2 = PA.AddDimension(trigger, 0);
BPA trigger3 = PA.Stretch(trigger2, 2, 1);
//Extract BMU Coord
FPA lol = PA.Cond(trigger3, m_Shape, zeros);
m_BMUCoord = PA.Sum(lol, 1);
//m_BMUCoord = PA.Evaluate(m_BMUCoord);
//BMU Code Vector ZEROS
BPA triggercv2 = PA.AddDimension(trigger, 0);
BPA triggercv3 = PA.Stretch(triggercv2, m_PatternLength, 1);
//Extract BMU code vector
FPA mdr = PA.Cond(triggercv3, m_Weights, zeroscv);
m_BMUCodeVector = PA.Sum(mdr, 1);
//m_BMUCodeVector = PA.Evaluate(m_BMUCodeVector);
//Begin computation ! ^^ AND OUTPUT
PA.Evaluate(m_BMUCodeVector);
PA.Evaluate(m_BMUCoord);
}
}//_CPU_FindBMU
public void SetCurrentPattern(float[] Pattern)
{
m_CurrentPattern = Pattern;
m_CurrentPatternGPU = new DisposableFloatParallelArray(Pattern);
}
private void InitMap()
{
// MAP VALUES AND SHAPE
Random RandomNumber = new Random();
weight_vals = new float[m_PatternLength, m_Width *m_Height];
for (int i = 0; i < m_Width * m_Height; ++i)
{
for (int k = 0; k < m_PatternLength; ++k)
{
weight_vals[k, i] = (float)(RandomNumber.NextDouble() * 255.0f);
}
}
shape_vals = new float[2, m_Width *m_Height];
for (int i = 0; i < m_Height; i++)
{
for (int j = 0; j < m_Width; j++)
{
shape_vals[0, m_Width *i + j] = j + 1;
shape_vals[1, m_Width *i + j] = i + 1;
}
}
m_Weights = new DisposableFloatParallelArray(weight_vals);
m_Shape = new DFPA(shape_vals);
//
//BMU INIT
m_bmucoord_vals = new float[2];
m_bmucodevector_vals = new float[m_PatternLength];
//ZEROS FOR FINDBMU
zero_vals = new float[2, m_Width *m_Height];
for (int i = 0; i < m_Width * m_Height; ++i)
{
zero_vals[0, i] = 0;
zero_vals[1, i] = 0;
}
zerocv_vals = new float[m_PatternLength, m_Width *m_Height];
for (int i = 0; i < m_Width * m_Height; ++i)
{
for (int k = 0; k < m_PatternLength; ++k)
{
zerocv_vals[k, i] = 0;
zerocv_vals[k, i] = 0;
}
}
zeroscv = new DFPA(zerocv_vals);
zeros = new DFPA(zero_vals);
//
//INIT TIME
//Don't use this
m_Time = new DFPA(new float[] { 1 });
//NEIGHBORHOOD VARIABLE INIT
m_maxtime_val = 100.0f;
m_theta_val = 100.0f / (float)Math.Log(25);
m_sigmainitial_val = 25;
m_epsiloninitial_val = 0.1f;
m_Theta = new DFPA(new float[] { m_theta_val });
m_SigmaInitial = new DFPA(new float[] { m_sigmainitial_val });
m_EpsilonInitial = new DFPA(new float[] { m_epsiloninitial_val });
m_MaxTime = new DFPA(new float[] { m_maxtime_val });
//LOG BASE
LogBase = new DFPA(new float[] { (float)(Math.E) });
}//Init
