public void Train(DenseMatrix X, DenseVector d, DenseVector Kd)
{
int R = X.RowCount;
int N = X.ColumnCount;
int U = 0; //the number of neurons in the structure
var c = new DenseMatrix(R, 1);
var sigma = new DenseMatrix(R, 1);
var Q = new DenseMatrix((R + 1), (R + 1));
var O = new DenseMatrix(1, (R + 1));
var pT_n = new DenseMatrix((R + 1), 1);
double maxPhi = 0;
int maxIndex;
var Psi = new DenseMatrix(N, 1);
Console.WriteLine("Running...");
//for each observation n in X
for (int i = 0; i < N; i++)
{
Console.WriteLine(100 * (i / (double)N) + "%");
var x = new DenseVector(R);
X.Column(i, x);
//if there are neurons in structure,
//update structure recursively.
if (U == 0)
{
c = (DenseMatrix)x.ToColumnMatrix();
sigma = new DenseMatrix(R, 1, SigmaZero);
U = 1;
Psi = CalculatePsi(X, c, sigma);
UpdateStructure(X, Psi, d, ref Q, ref O);
pT_n =
(DenseMatrix)
(CalculateGreatPsi((DenseMatrix)x.ToColumnMatrix(), (DenseMatrix)Psi.Row(i).ToRowMatrix()))
.Transpose();
}
else
{
StructureRecurse(X, Psi, d, i, ref Q, ref O, ref pT_n);
}
bool KeepSpinning = true;
while (KeepSpinning)
{
//Calculate the error and if-part criteria
double ee = pT_n.Multiply(O)[0, 0];
double approximationError = Math.Abs(d[i] - ee);
DenseVector Phi;
double SumPhi;
CalculatePhi(x, c, sigma, out Phi, out SumPhi);
maxPhi = Phi.Maximum();
maxIndex = Phi.MaximumIndex();
if (approximationError > delta)
{
if (maxPhi < threshold)
{
var tempSigma = new DenseVector(R);
sigma.Column(maxIndex, tempSigma);
double minSigma = tempSigma.Minimum();
int minIndex = tempSigma.MinimumIndex();
sigma[minIndex, maxIndex] = k_sigma * minSigma;
Psi = CalculatePsi(X, c, sigma);
UpdateStructure(X, Psi, d, ref Q, ref O);
var psi = new DenseVector(Psi.ColumnCount);
Psi.Row(i, psi);
pT_n =
(DenseMatrix)
CalculateGreatPsi((DenseMatrix)x.ToColumnMatrix(), (DenseMatrix)psi.ToRowMatrix())
.Transpose();
}
else
{
//add a new neuron and update strucutre
double distance = 0;
var cTemp = new DenseVector(R);
var sigmaTemp = new DenseVector(R);
//foreach input variable
for (int j = 0; j < R; j++)
{
distance = Math.Abs(x[j] - c[j, 0]);
int distanceIndex = 0;
//foreach neuron past 1
for (int k = 1; k < U; k++)
{
if ((Math.Abs(x[j] - c[j, k])) < distance)
{
distanceIndex = k;
distance = Math.Abs(x[j] - c[j, k]);
}
}
if (distance < Kd[j])
{
cTemp[j] = c[j, distanceIndex];
sigmaTemp[j] = sigma[j, distanceIndex];
}
else
{
cTemp[j] = x[j];
sigmaTemp[j] = distance;
}
}
//end foreach
c = (DenseMatrix)c.InsertColumn(c.ColumnCount - 1, cTemp);
sigma = (DenseMatrix)sigma.InsertColumn(sigma.ColumnCount - 1, sigmaTemp);
Psi = CalculatePsi(X, c, sigma);
UpdateStructure(X, Psi, d, ref Q, ref O);
U++;
KeepSpinning = false;
}
}
else
{
if (maxPhi < threshold)
{
var tempSigma = new DenseVector(R);
sigma.Column(maxIndex, tempSigma);
double minSigma = tempSigma.Minimum();
int minIndex = tempSigma.MinimumIndex();
sigma[minIndex, maxIndex] = k_sigma * minSigma;
Psi = CalculatePsi(X, c, sigma);
UpdateStructure(X, Psi, d, ref Q, ref O);
var psi = new DenseVector(Psi.ColumnCount);
Psi.Row(i, psi);
pT_n =
(DenseMatrix)
CalculateGreatPsi((DenseMatrix)x.ToColumnMatrix(), (DenseMatrix)psi.ToRowMatrix())
.Transpose();
}
else
{
KeepSpinning = false;
}
}
}
}
out_C = c;
out_O = O;
out_Sigma = sigma;
Console.WriteLine("Done.");
}
/// <summary>
/// Gets the local minimums distribution.
/// </summary>
/// <param name="dmFieldData">The dm field data.</param>
/// <param name="dimensionNumber">The dimension number:
/// 1 - rows (angle)
/// 2 - columns (distance)
/// </param>
/// <returns>DenseMatrix.</returns>
//public static DenseMatrix GetLocalMinimumsDistribution(DenseMatrix dmFieldData, PointD sunCenterPoint, PointD imageCenterPoint, double imageRadius, int imageHeight, double imageCircleCropFactor = 0.9d, int dimensionNumber = 1)
public static List <Point3D> GetLocalMinimumsDistribution(DenseMatrix dmFieldData, RoundData sunDiskData, RoundData imageRoundData, int imageHeight, double imageCircleCropFactor = 0.9d)
{
// DenseMatrix dmFieldminimumsData = DenseMatrix.Create(dmFieldData.RowCount, dmFieldData.ColumnCount, 0.0d);
List <Point3D> lRetPoints = new List <Point3D>();
double imageRadius = imageRoundData.DRadius;
PointD imageCenterPoint = imageRoundData.pointDCircleCenter();
PointPolar imageCenterPointRelatedToSunCenter = new PointPolar(imageCenterPoint - sunDiskData.pointDCircleCenter(), true);
double distanceSunCenterToImageCenter = PointD.Distance(imageCenterPoint, sunDiskData.pointDCircleCenter());
#region // obsolete
//if (dimensionNumber == 1)
//{
#endregion // obsolete
for (int i = 0; i < dmFieldData.RowCount; i++)
{
bool itsTheCropCase = false;
//если направлени на кроп кадра - то не берем в расмотрение
double currentAngle = ((double)i / (double)(dmFieldData.RowCount - 1)) * 2.0d * Math.PI;
LineDescription2D line, lineMargin;
if (currentAngle < Math.PI)
{
//верхняя половина, смотрим направление на y=0.0d
line = new LineDescription2D(sunDiskData.pointDCircleCenter(),
new Vector2D(Math.Cos(currentAngle), -Math.Sin(currentAngle)));
lineMargin = new LineDescription2D(new PointD(0.0d, 0.0d), new Vector2D(1.0d, 0.0d));
}
else
{
line = new LineDescription2D(sunDiskData.pointDCircleCenter(),
new Vector2D(Math.Cos(currentAngle), Math.Sin(currentAngle)));
lineMargin = new LineDescription2D(new PointD(0.0d, imageHeight), new Vector2D(1.0d, 0.0d));
}
PointD crossPointD = LineDescription2D.CrossPoint(line, lineMargin);
if (crossPointD.Distance(imageCenterPoint) < imageRadius)
{
itsTheCropCase = true;
}
#region // obsolete
//double yMargin = 0.0d;
//double xMargin = sunCenterPoint.X + (yMargin - sunCenterPoint.Y) / Math.Tan(currentAngle);
//double dx = xMargin - imageCenterPoint.X;
//double dy = yMargin - imageCenterPoint.Y;
//if (Math.Sqrt(dx * dx + dy * dy) < imageRadius) itsTheCropCase = true;
#endregion // obsolete
#region //obsolete
//else
//{
// //нижняя половина, смотрим направление на y=imageHeight
// double yMargin = (double)imageHeight;
// double xMargin = sunCenterPoint.X + (yMargin - sunCenterPoint.Y) / Math.Tan(currentAngle);
// double dx = xMargin - imageCenterPoint.X;
// double dy = yMargin - imageCenterPoint.Y;
// if (Math.Sqrt(dx * dx + dy * dy) < imageRadius) itsTheCropCase = true;
//}
#endregion //obsolete
//Если слишком близко к краю изображения - тоже исключаем. Минимум должен лежать не ближе, например, 1/15
//DenseMatrix dmSlicedDataMatrix = (DenseMatrix)dmFieldData.SubMatrix(i, 1, 0, dmFieldData.ColumnCount);
DenseVector dvRowDataVector = (DenseVector)dmFieldData.EnumerateRows().ElementAt(i);
#region // debug plotting
//dvRowDataVector.SaveVectorDataAsImagePlot(
// "D:\\_gulevlab\\SkyImagesAnalysis_appData\\patent-samples\\result.2015-03-24\\img-2014-09-20T16-03-58devID1\\dvRowDataVector-plot-image-" +
// i.ToString("D03") + "-step1.png");
#endregion // debug plotting
dvRowDataVector.MapIndexedInplace((idx, x) => ((x == 0.0d) || (idx < sunDiskData.DRadius * 1.5d)) ? (1.0d) : (x));
#region // debug plotting
//dvRowDataVector.SaveVectorDataAsImagePlot(
// "D:\\_gulevlab\\SkyImagesAnalysis_appData\\patent-samples\\result.2015-03-24\\img-2014-09-20T16-03-58devID1\\dvRowDataVector-plot-image-" +
// i.ToString("D03") + "-step2.png");
#endregion // debug plotting
double phiFromImageCenterToDirection = imageCenterPointRelatedToSunCenter.Phi - currentAngle;
double distanceToImageMargin = distanceSunCenterToImageCenter * Math.Cos(phiFromImageCenterToDirection) +
Math.Sqrt(imageRadius * imageRadius -
distanceSunCenterToImageCenter * distanceSunCenterToImageCenter *
Math.Sin(phiFromImageCenterToDirection) *
Math.Sin(phiFromImageCenterToDirection));
dvRowDataVector.MapIndexedInplace(
(idx, x) => ((double)idx / distanceToImageMargin >= imageCircleCropFactor) ? (1.0d) : (x));
#region // debug plotting
//dvRowDataVector.SaveVectorDataAsImagePlot(
// "D:\\_gulevlab\\SkyImagesAnalysis_appData\\patent-samples\\result.2015-03-24\\img-2014-09-20T16-03-58devID1\\dvRowDataVector-plot-image-" +
// i.ToString("D03") + "-step3.png");
#endregion // debug plotting
double minValue = dvRowDataVector.Minimum();
int minValueIndex = dvRowDataVector.MinimumIndex();
//if (!itsTheCropCase) dmFieldminimumsData[i, minValueIndex] = minValue;
if ((!itsTheCropCase) && ((double)minValueIndex > sunDiskData.DRadius))
{
lRetPoints.Add(new Point3D(currentAngle, minValueIndex, minValue));
}
else
{
continue;
}
}
#region // obsolete
//}
//else if (dimensionNumber == 2)
//{
// for (int i = 0; i < dmFieldData.ColumnCount; i++)
// {
// DenseMatrix dmSlicedDataMatrix = (DenseMatrix)dmFieldData.SubMatrix(0, dmFieldData.RowCount, i, 1);
// DenseVector dvSlicedDataVector = DenseVector.OfEnumerable(dmSlicedDataMatrix.Values);
// dvSlicedDataVector.MapInplace(new Func<double, double>(x => (x == 0.0d) ? (1.0d) : (x)));
// double minValue = dvSlicedDataVector.Minimum();
// int minValueIndex = dvSlicedDataVector.MinimumIndex();
// dmFieldminimumsData[minValueIndex, i] = minValue;
// }
//}
#endregion // obsolete
//return dmFieldminimumsData;
return(lRetPoints);
}
