public NetworkHelper(IList<Minutia> features, SOMParams somParams)
{
var trainingFeaturesCount = features.Count;
var inputs = new double[trainingFeaturesCount][];
var featureCharacteristicsCount = somParams.CharacteristicsCount;
for (var i = 0; i < trainingFeaturesCount; i++)
{
var feature = features[i];
inputs[i] = new double[featureCharacteristicsCount];
inputs[i][0] = feature.Angle * somParams.AngleMultiplier;
inputs[i][1] = (double)feature.MinutiaType * somParams.TypeMultiplier;
inputs[i][2] = feature.X * somParams.CoordsMultiplier;
inputs[i][3] = feature.Y * somParams.CoordsMultiplier;
inputs[i][4] = feature.GetHashCode()*somParams.HashCodeMultiplier;
}
_networkSize = CreateRectangleFromFeatures(trainingFeaturesCount);
_som = new DistanceNetwork(featureCharacteristicsCount, trainingFeaturesCount);
_teacher = new SOMLearning(_som, _networkSize[0], _networkSize[1]);
_inputs = inputs;
_learningRadius = somParams.LearningRadius;
_fixedLearningRate = somParams.LearningRate / 10;
_driftingLearningRate = _fixedLearningRate * 9;
}
public NetworkHelper(double[][] features, SOMParams somParams)
{
var trainingFeaturesCount = features.GetLength(0);
var featureCharacteristicsCount = somParams.CharacteristicsCount;
_networkSize = CreateRectangleFromFeatures(trainingFeaturesCount);
_som = new DistanceNetwork(featureCharacteristicsCount, trainingFeaturesCount);
_teacher = new SOMLearning(_som, _networkSize[0], _networkSize[1]);
_inputs = features;
_learningRadius = somParams.LearningRadius;
_fixedLearningRate = somParams.LearningRate / 10;
_driftingLearningRate = _fixedLearningRate * 9;
}
static void Main()
{
network = new DistanceNetwork(2, task.NetworkWidth * task.NetworkHeight);
for (int x=0;x<task.NetworkWidth;x++)
for (int y = 0; y < task.NetworkHeight; y++)
{
var n = network.Layers[0].Neurons[x * task.NetworkHeight + y];
n.Weights[0] = rnd.NextDouble() * 0.2 + 0.4;
n.Weights[1] = rnd.NextDouble() * 0.2 + 0.4;
}
learning = new SOMLearning(network, task.NetworkWidth, task.NetworkHeight);
learning.LearningRadius = task.LearningRadius;
learning.LearningRate = task.LearningRate;
Inputs = task.GenerateInputs().ToArray();
pointsPanel = new MyUserControl() { Dock= DockStyle.Fill};
pointsPanel.Paint += DrawPoints;
networkPanel = new MyUserControl() { Dock = DockStyle.Fill };
networkPanel.Paint += DrawNetwork;
networkGraphControl = new MyUserControl { Dock = DockStyle.Fill };
networkGraphControl.Paint += DrawGraph;
var pauseButton = new Button { Text = "Pause/Resume" };
pauseButton.Click+=PauseResume;
var table = new TableLayoutPanel() { Dock = DockStyle.Fill, RowCount=2, ColumnCount=2 };
table.Controls.Add(pointsPanel, 0, 0);
table.Controls.Add(networkPanel, 0, 1);
table.Controls.Add(networkGraphControl, 1, 0);
table.Controls.Add(pauseButton,1,1);
// table.Controls.Add(pause, 1, 1);
table.RowStyles.Add(new RowStyle(SizeType.Percent, 50f));
table.RowStyles.Add(new RowStyle(SizeType.Percent, 50f));
table.ColumnStyles.Add(new ColumnStyle(SizeType.Percent, 50f));
table.ColumnStyles.Add(new ColumnStyle(SizeType.Percent, 50f));
form = new Form()
{
ClientSize = new Size(600, 600),
Controls =
{
table
}
};
timer = new System.Windows.Forms.Timer();
timer.Tick += (sender, args) => { Learning(); form.Invalidate(true); };
timer.Interval = 100;
timer.Start();
Application.Run(form);
}
// Worker thread
void SearchSolution( )
{
// create learning algorithm
SOMLearning trainer = new SOMLearning( network );
// input
double[] input = new double[3];
double fixedLearningRate = learningRate / 10;
double driftingLearningRate = fixedLearningRate * 9;
// iterations
int i = 0;
// loop
while ( !needToStop )
{
trainer.LearningRate = driftingLearningRate * ( iterations - i ) / iterations + fixedLearningRate;
trainer.LearningRadius = (double) radius * ( iterations - i ) / iterations;
input[0] = rand.Next( 256 );
input[1] = rand.Next( 256 );
input[2] = rand.Next( 256 );
trainer.Run( input );
// update map once per 50 iterations
if ( ( i % 10 ) == 9 )
{
UpdateMap( );
}
// increase current iteration
i++;
// set current iteration's info
this.BeginInvoke ( new MethodInvoker (() => currentIterationBox.Text = i.ToString()));
// stop ?
if ( i >= iterations )
break;
}
// enable settings controls
EnableControls( true );
}
/// <summary>
/// Enseña patrones a la red en un hilo secundario.
/// </summary>
protected override void teachingWorker_DoWork(object sender, DoWorkEventArgs e)
{
// Extrae los kanjis
var allKanjis = (List<Kanji>)e.Argument;
// Genera una lista de patrones, una por cada kanji
var allPatterns = new List<Pattern>(allKanjis.Count);
foreach (var kanji in allKanjis)
{
//Dependiendo del modo obj de aprendizaje genera el patron
string imageHash = string.Empty;
Pattern pattern = null;
switch (Method)
{
case GenerationMethod.Normal:
case GenerationMethod.Heightmap:
generatePattern_Normal(kanji.sourceImage, out pattern, out imageHash);
break;
case GenerationMethod.Hashing:
generatePattern_Hashing(kanji.sourceImage, out pattern, out imageHash);
break;
}
// Lo agrega a la lista de patrones a aprender
allPatterns.Add(pattern);
}
// Crea un entrenador
var sqrtOutputSize = (int)Math.Sqrt(OutputSize);
SOMLearning trainer = new SOMLearning(somNN, sqrtOutputSize, sqrtOutputSize);
// Crea el set de datos a entrenar
var trainingSet = generateInputSet(allPatterns);
// Iteraciones de aprendizaje
for (int i = 0; i < LearningIterations; i++)
{
// Establece los valores de aprendizaje y radio para la corrida actual
var completedRatio = (float)i / (LearningIterations - 1);
trainer.LearningRate = completedRatio * LearningEndingRate + (1 - completedRatio) * LearningInitialRate;
trainer.LearningRadius = completedRatio * LearningEndingRadius + (1 - completedRatio) * LearningInitialRadius;
// Ejecuta la corrida
trainer.RunEpoch(trainingSet);
// Informa progreso
((BackgroundWorker)sender).ReportProgress((int)(completedRatio * 100 * .99f));
// Revisa la solicitud de cancelar
if (needToStop)
{
e.Cancel = true;
break;
}
}
// Genera la lista de clases para cada patron aprendido
var classes = new List<int>(allPatterns.Count);
foreach (var pattern in allPatterns) classes.Add(classifyPattern(pattern));
if (classes.GroupBy(i => i).Where(g => g.Count() > 1).ToList().Count > 0)
throw new Exception("No se pudo distinguir la diferencia de clase entre dos kanjis.");
// Para cada patron busca la neurona representante
for (int i = 0; i < allKanjis.Count; i++)
{
// Guarda el kanji en el diccionario
learnedKanjis.Add(classes[i].ToString(), allKanjis[i]);
}
}
// Worker thread
void SearchSolution( )
{
// set random generators range
Neuron.RandRange = new DoubleRange( 0, Math.Max( pointsPanel.ClientRectangle.Width, pointsPanel.ClientRectangle.Height ) );
// create network
DistanceNetwork network = new DistanceNetwork( 2, networkSize * networkSize );
// create learning algorithm
SOMLearning trainer = new SOMLearning( network, networkSize, networkSize );
// create map
map = new int[networkSize, networkSize, 3];
double fixedLearningRate = learningRate / 10;
double driftingLearningRate = fixedLearningRate * 9;
// iterations
int i = 0;
RefreshDelegate refreshIterationsBox = delegate () {
currentIterationBox.Text = i.ToString( );
};
RefreshDelegate enableControls = delegate () {
EnableControls( true );
};
// loop
while ( !needToStop )
{
trainer.LearningRate = driftingLearningRate * ( iterations - i ) / iterations + fixedLearningRate;
trainer.LearningRadius = (double) learningRadius * ( iterations - i ) / iterations;
// run training epoch
trainer.RunEpoch( trainingSet );
// update map
UpdateMap( network );
// increase current iteration
i++;
// set current iteration's info
this.Invoke(refreshIterationsBox);
// stop ?
if ( i >= iterations )
break;
}
// enable settings controls
this.Invoke(enableControls);
}
void KohonenTeach()
{
Neuron.RandRange = new Range(0.0f, 1.0f);
kohonenNetwork = new DistanceNetwork(liczba_wejsc, liczba_neuronow_pion * liczba_neuronow_poziom);
kohonenNetwork.Randomize();
SOMLearning teacher = new SOMLearning(kohonenNetwork);
double driftingLearningRate = this.wps_zmiany_wsp;
double fixedLearningRate = this.pocz_wart_wsp_nauki;
double learningRadius = this.pocz_rozmiar_sasiedz;
double driftingLearningRadius = this.wsp_zmian_rozm_sasiedz;
int iteration = 1;
while (!needToStop)
{
teacher.LearningRate = driftingLearningRate * (numberOfCycles - iteration) / numberOfCycles + fixedLearningRate;
teacher.LearningRadius = (double)learningRadius * (numberOfCycles - iteration) / numberOfCycles;
teacher.RunEpoch(trainingKData);
SetIterationsCount(iteration++);
if (iteration > numberOfCycles) break;
}
teachingDone = true;
needToStop = false;
UpdateLabels();
}
// On "Start" button click
private void startButton_Click(object sender, System.EventArgs e)
{
// get network size
try
{
networkSize = Math.Max( 5, Math.Min( 50, int.Parse( sizeBox.Text ) ) );
}
catch
{
networkSize = 15;
}
// get iterations count
try
{
iterations = Math.Max( 10, Math.Min( 1000000, int.Parse( iterationsBox.Text ) ) );
}
catch
{
iterations = 500;
}
// get learning rate
try
{
learningRate = Math.Max( 0.00001, Math.Min( 1.0, double.Parse( rateBox.Text ) ) );
}
catch
{
learningRate = 0.3;
}
// get radius
try
{
learningRadius = Math.Max( 1, Math.Min( 30, int.Parse( radiusBox.Text ) ) );
}
catch
{
learningRadius = 3;
}
// update settings controls
UpdateSettings( );
// disable all settings controls except "Stop" button
EnableControls( false );
// generate training set
for ( int i = 0; i < pointsCount; i++ )
{
// create new training sample
trainingSet[i] = new double[2] { points[i, 0], points[i, 1] };
}
//// run worker thread
//needToStop = false;
//workerThread = new Thread( new ThreadStart( SearchSolution ) );
//workerThread.Start( );
Neuron.RandRange = new DoubleRange(0, Math.Max(pointsPanel.ClientRectangle.Width, pointsPanel.ClientRectangle.Height));
// create network
network = new DistanceNetwork(2, networkSize * networkSize);
// create learning algorithm
trainer = new SOMLearning(network, networkSize, networkSize);
// create map
map = new int[networkSize, networkSize, 3];
fixedLearningRate = learningRate / 10;
driftingLearningRate = fixedLearningRate * 9;
// iterations
_i = 0;
}
