public void TestRandomizeSeemsToHaveMean0()
{
int count = 100000;
Vector <double> vec = new DenseVector(count);
for (int i = 0; i < count; i++)
{
vec[i] = MathHelper.RandomExceptZero(1, 0.2);
}
Assert.AreEqual(0, vec.Average(), 0.004);
}
/// <summary>
/// Tests the network over a given test dataset. Returns the error ( sum(|a - y(x)|^2)/n ). The out param will count the data that was correctly categorized using a given function.
/// </summary>
private static double TestNetwork(DeepNeuralNetwork nn, int testingDataSetSize, GetNextDataDelegate getNextData, out int correctCount, CheckIfCorrectDelegate checkCorrect)
{
correctCount = 0;
Vector <double> error = new DenseVector(nn.OutputLayer.GetNeuronCount());
for (int i = 0; i < testingDataSetSize; ++i)
{
LabeledData labeledData = getNextData(i);
Vector <double> result = nn.ProcessInput(labeledData.InputValues);
if (checkCorrect(result.AsArray(), labeledData.OutputValues.AsArray()))
{
++correctCount;
}
Vector <double> diff = labeledData.OutputValues - result;
error += diff.PointwiseMultiply(diff);
}
error = error.Divide(testingDataSetSize);
return(error.Average());
}
} //обучение
public double getRegressionError(DenseVector Y, double error)
{
double averageError = Y.Subtract(Y.Average()).PointwisePower(2).Sum();
return(1 - (error / averageError));
}
public static RBFNetSolution SolveOLS(List <Vec> xs, List <double> ys, double tolerance, double spread, Func <bool> cancelled = null)
{
Func <Vec, Vec, double> phi = (x, c) => Phi(x, c, spread);
var maxCenters = (int)Math.Min(xs.Count, Math.Max(1, 4 * Math.Sqrt(xs.Count)));
var centers = new List <Vec>(xs);
var n = xs.Count;
var m = centers.Count;
Mat P = new DenseMatrix(n, m);
for (int i = 0; i < n; i++)
{
for (int j = 0; j < m; j++)
{
P[i, j] = phi(xs[i], centers[j]);
}
}
Vec dOrig = new DenseVector(ys.ToArray());
Vec d = dOrig.Subtract(dOrig.Average());
Vec dNorm = d.Normalize(2);
var candidateBases = centers.Zip(P.Columns(), (ci, pi) => new { Center = ci, Basis = pi }).ToList();
var selectedBases = candidateBases.Select(cb => new { Center = cb.Center, Basis = cb.Basis, OrthoBasis = cb.Basis }).Take(0).ToList();
double qualityTotal = 0;
var context = QMCreateOrthoContext(n, m);
var iters = m;
for (int k = 0; k < iters; k++)
{
int stride = n;
double[] selectedOrthonormalBases = new double[selectedBases.Count * stride];
for (int ptr = 0, sb = 0; sb < selectedBases.Count; ptr += stride, sb++)
{
Array.Copy(selectedBases[sb].OrthoBasis.ToArray(), 0, selectedOrthonormalBases, ptr, stride);
}
var best = candidateBases.Select(cb => {
var pi = cb.Basis;
Vec w;
if (!selectedBases.Any())
{
w = pi.Normalize(2);
}
else
{
double[] result = new double[pi.Count];
double[] pis = pi.ToList().ToArray();
double[] sobs = selectedOrthonormalBases.ToList().ToArray();
QMOrthogonalize(context, pis, selectedBases.Count, sobs);
w = new DenseVector(pis.ToList().ToArray());
}
//Vec wCheck = Orthogonalize(pi, selectedBases.Select(b => b.OrthoBasis).ToList()).Normalize(2);
//var w = wCheck;
//Debug.Assert(w.Subtract(wCheck).Norm(2) < 0.0001);
var err = Math.Pow(w.DotProduct(dNorm), 2);
return(new { Center = cb.Center, OrthoBasis = w, Quality = err, Candidate = cb });
}).OrderByDescending(q => q.Quality).First();
candidateBases.Remove(best.Candidate);
selectedBases.Add(new { Center = best.Center, Basis = best.Candidate.Basis, OrthoBasis = best.OrthoBasis });
qualityTotal += best.Quality;
if (1 - qualityTotal < tolerance || (cancelled != null && cancelled()))
{
break;
}
}
QMDestroyOrthoContext(context);
//Trace.WriteLine(string.Format("Centers: {0}, Spread: {1}, Tolerance: {2}", selectedBases.Count, spread, tolerance));
var weights = SolveLS(
new Vec[] { DenseVector.Create(n, _ => 1) }.Concat(selectedBases.Select(sb => sb.Basis)).ColumnsToMatrix(),
new DenseVector(ys.ToArray()));
bool isDegenerate = false;
if (weights.Any(w => double.IsNaN(w)))
{
Trace.WriteLine("! Degenerate RBF network !");
isDegenerate = true;
}
var allBases = selectedBases.ToList();
allBases.Insert(0, null);
var resultBases = allBases.Zip(weights, (b, w) => new RadialBasis(b != null ? b.Center : null, w)).ToList();
return(new RBFNetSolution(resultBases, spread, isDegenerate));
}
