private List<Tuple<Person, double>> RecogniseSpeaker(List<double[]> mfccs)
{
var results = new List<Tuple<Person, double>>();
var mfccArray = mfccs.ToArray();
double[] input = new double[GetNetworksInputSize()];
int totalSamples = 0;
Dictionary<Person, double> resultsPerNetwork = new Dictionary<Person, double>();
Dictionary<Person, int> samplesPerNetwork = new Dictionary<Person, int>();
Dictionary<Person, GnuPlot> plots = new Dictionary<Person, GnuPlot>();
foreach (var person in people) {
resultsPerNetwork[person] = 0.0f;
samplesPerNetwork[person] = 0;
plots[person] = new GnuPlot(person.FullName, GetRecognitionPlotFileName(person));
plots[person].BeginPlot();
}
for (int start = 0; start < mfccArray.Length - algorithmParams.SignalFramesCount; start++) {
CopyInputData(input, mfccArray, start);
var result = from person in people
let network = neuralNetworks[person]
select new KeyValuePair<Person, double>(
person,
network.Compute(new[] { input })[0][0]
);
foreach (var r in result) {
plots[r.Key].AddDataPoint(start, r.Value);
if (r.Value > THRESHOLD || r.Value < (1.0 - THRESHOLD)) {
//samplesPerNetwork[r.Key]++;
resultsPerNetwork[r.Key] += (r.Value > THRESHOLD) ? r.Value : (r.Value - 1.0);// NormalizeResult(r.Value);
if(r.Value > THRESHOLD)
totalSamples++;
}
}
}
foreach (var p in people) {
plots[p].SetYRange(0, 1);
plots[p].EndPlot();
}
var finalResults = from person in people
let res = Math.Max(0, resultsPerNetwork[person])
let norm = res / (double)totalSamples
select new Tuple<Person, double>(person, norm);
//var finalResults = from person in people
// let samples = samplesPerNetwork[person]
// let res = resultsPerNetwork[person]
// let norm = (samples == 0) ? 0 : (res / (double)samples)
// select new Tuple<Person, double>(person, norm);
return finalResults.ToList();
}
private void TrainPersonNeuralNetwork(
Dictionary<Person, List<double[]>[]> trainingMfccs,
Dictionary<Person, List<double[]>[]> testMfccs,
double[] input, double[][] answer, Person person,
PerceptronWrapper network)
{
GnuPlot plot = new GnuPlot(
person.FullName,
@"D:\GIT\nnetwork_mini_pw\charts\plot_" + person.FullName.Replace(' ', '_') + ".png"
);
plot.BeginPlot();
List<double[]>[] personMfccs = null;
int iterations = 0;
while (iterations < MaxIterationCounts) {
double expectedAnswer;
if (Random.NextDouble() < algorithmParams.TCoef) {
expectedAnswer = 0.0d;
Person otherPerson = SelectOtherPerson(person);
personMfccs = trainingMfccs[otherPerson];
}
else {
expectedAnswer = 1.0d;
personMfccs = trainingMfccs[person];
}
DrawInputData(input, personMfccs);
double scale = Math.Sqrt(1.0 + iterations);
double newLearningRate = algorithmParams.LearningRate / scale;
double newMomentum = algorithmParams.Momentum / scale;
answer[0][0] = expectedAnswer;
network.Train(newLearningRate, 1, newMomentum,
new[] { input },
answer);
iterations++;
if (iterations % 1000 == 0) {
double testResult = TestPersonNeuralNetwork(person, testMfccs, input);
double learnResult = TestPersonNeuralNetwork(person, trainingMfccs, input);
AddPlotPoint(plot, iterations, learnResult, testResult);
if (testResult <= TestSetMaxError)
{
break;
}
}
}
plot.AddAlgorithmParamsToTitle(algorithmParams);
plot.End2Plot();
}
private static void AddPlotPoint(GnuPlot plot, int iterations, double learnResult, double testResult)
{
plot.AddDataPoint(iterations, learnResult, testResult);
string line = String.Format(CultureInfo.InvariantCulture, "{0}\t{1}", iterations, testResult);
Debug.WriteLine(line);
}