private static JobResult RunJob(JobDescription job)
{
Matrix <double> weights = GenerateRandomWeights(job.k, -1, 1);
Matrix <double> radial = new Matrix <double>(1, job.k);
Random rng = new Random();
// kohonen learning phase
for (int t = 0; t < job.iters_kohonen; t++)
{
int ostIndex = rng.Next(job.ost.Rows);
RadialNumerators(radial, job.ost, weights, ostIndex);
int j_0 = radial.IndexOfMax();
weights[j_0, 0] += job.eta_kohonen * (job.ost[ostIndex, 0] - weights[j_0, 0]);
weights[j_0, 1] += job.eta_kohonen * (job.ost[ostIndex, 1] - weights[j_0, 1]);
}
Lab1.NeuralNetwork nn = new Lab1.NeuralNetwork(job.beta, new int[] { job.k, 1 });
nn.RandomizeBiases(-1, 1);
nn.RandomizeWeights(-1, 1);
Matrix <double>[] ostRadials = new Matrix <double> [job.ost.Rows];
for (int i = 0; i < job.ost.Rows; i++)
{
RadialNumerators(ostRadials[i] = new Matrix <double>(1, job.k), job.ost, weights, i);
Radial(ostRadials[i]);
}
var radialIndices = Enumerable.Range(0, job.ost.Rows).ToList();
radialIndices.Shuffle();
Matrix <double>[] trainingRadials = new Matrix <double> [(int)Math.Round(job.ost.Rows * 0.7)];
int[][] trainingClasses = new int[trainingRadials.Length][];
for (int i = 0; i < trainingRadials.Length; i++)
{
trainingRadials[i] = ostRadials[radialIndices[i]];
trainingClasses[i] = job.ostClasses[radialIndices[i]];
}
Matrix <double>[] validationRadials = new Matrix <double> [job.ost.Rows - trainingRadials.Length];
int[][] validationClasses = new int[validationRadials.Length][];
for (int i = 0; i < validationRadials.Length; i++)
{
validationRadials[i] = ostRadials[radialIndices[trainingRadials.Length + i]];
validationClasses[i] = job.ostClasses[radialIndices[trainingRadials.Length + i]];
}
// perceptron learning phase
for (int t = 0; t < job.iters_perceptron; t++)
{
int ostIndex = rng.Next(trainingRadials.Length);
nn.Train(trainingRadials[ostIndex], trainingClasses[ostIndex], job.eta_perceptron);
}
double trainingErrorRate = Lab1.Task4a.ErrorRate(trainingRadials, trainingClasses, nn);
double validationErrorRate = Lab1.Task4a.ErrorRate(validationRadials, validationClasses, nn);
var result = new JobResult
{
k = job.k,
weights = weights,
nn = nn,
radial = radial,
traniningErrorRate = trainingErrorRate,
validationErrorRate = validationErrorRate
};
Interlocked.Increment(ref JobsCompleted);
return(result);
}
private static void BoundaryDump(string fileStr, Matrix <double> ost, Matrix <double> weights, Matrix <double> radial, Lab1.NeuralNetwork nn)
{
//double x_min = weights.Col(0).Min(), x_max = weights.Col(0).Max();
//double y_min = weights.Col(1).Min(), y_max = weights.Col(1).Max();
double x_min = ost.Col(0).Min(), x_max = ost.Col(0).Max();
double y_min = ost.Col(1).Min(), y_max = ost.Col(1).Max();
double x_stepSize = (x_max - x_min) / 999;
double classification_precision = 1E-6;
Matrix <double> zeroOst = new Matrix <double>(1, 2);
List <Tuple <double, double> > boundary = new List <Tuple <double, double> >(1000);
for (double x = x_min; x <= x_max; x += x_stepSize)
{
zeroOst[0, 0] = x;
zeroOst[0, 1] = y_max;
RadialNumerators(radial, zeroOst, weights, 0);
Radial(radial);
nn.FeedPattern(radial);
double max_classification = nn.Output[0, 0];
zeroOst[0, 1] = y_min;
RadialNumerators(radial, zeroOst, weights, 0);
Radial(radial);
nn.FeedPattern(radial);
double min_classificiaton = nn.Output[0, 0];
if (max_classification * min_classificiaton > 0)
{
continue;
}
double classification = max_classification;
double range_max = y_max, range_min = y_min;
while (Math.Abs(classification) > classification_precision)
{
zeroOst[0, 1] = (range_max - range_min) / 2 + range_min;
RadialNumerators(radial, zeroOst, weights, 0);
Radial(radial);
nn.FeedPattern(radial);
classification = nn.Output[0, 0];
if (classification * max_classification > 0)
{
range_max = zeroOst[0, 1];
}
else
{
range_min = zeroOst[0, 1];
}
}
boundary.Add(Tuple.Create(zeroOst[0, 0], zeroOst[0, 1]));
}
string filename = "lab2task3_boundary_" + fileStr + ".txt";
Console.WriteLine("Writing to " + filename + "...");
using (StreamWriter sw = new StreamWriter(new FileStream(filename, FileMode.CreateNew)))
{
foreach (var xy in boundary)
{
sw.WriteLine(xy.Item1.ToString(CultureInfo.InvariantCulture) + "," + xy.Item2.ToString(CultureInfo.InvariantCulture));
}
}
}
