public SemiSupervisedClassificationStatisticsAggregator(int nClasses, double a, double b)
{
nClasses_ = nClasses;
a_ = a;
b_ = b;
GaussianAggregator2d = new GaussianAggregator2d(a, b);
HistogramAggregator = new HistogramAggregator(nClasses);
}
public double ComputeInformationGain(HistogramAggregator allStatistics, HistogramAggregator leftStatistics, HistogramAggregator rightStatistics)
{
double entropyBefore = allStatistics.Entropy();
int nTotalSamples = leftStatistics.SampleCount + rightStatistics.SampleCount;
if (nTotalSamples <= 1)
{
return(0.0);
}
double entropyAfter = (leftStatistics.SampleCount * leftStatistics.Entropy() + rightStatistics.SampleCount * rightStatistics.Entropy()) / nTotalSamples;
return(entropyBefore - entropyAfter);
}
public double ComputeInformationGain(SemiSupervisedClassificationStatisticsAggregator allStatistics, SemiSupervisedClassificationStatisticsAggregator leftStatistics, SemiSupervisedClassificationStatisticsAggregator rightStatistics)
{
double informationGainLabelled;
{
double entropyBefore = allStatistics.HistogramAggregator.Entropy();
HistogramAggregator leftHistogram = leftStatistics.HistogramAggregator;
HistogramAggregator rightHistogram = rightStatistics.HistogramAggregator;
int nTotalSamples = leftHistogram.SampleCount + rightHistogram.SampleCount;
if (nTotalSamples <= 1)
{
informationGainLabelled = 0;
}
else
{
double entropyAfter = (leftHistogram.SampleCount * leftHistogram.Entropy() + rightHistogram.SampleCount * rightHistogram.Entropy()) / nTotalSamples;
informationGainLabelled = entropyBefore - entropyAfter;
}
}
double informationGainUnlabelled;
{
double entropyBefore = ((SemiSupervisedClassificationStatisticsAggregator)(allStatistics)).GaussianAggregator2d.GetPdf().Entropy();
GaussianAggregator2d leftGaussian = leftStatistics.GaussianAggregator2d;
GaussianAggregator2d rightGaussian = rightStatistics.GaussianAggregator2d;
int nTotalSamples = leftGaussian.SampleCount + rightGaussian.SampleCount;
double entropyAfter = (leftGaussian.SampleCount * leftGaussian.GetPdf().Entropy() + rightGaussian.SampleCount * rightGaussian.GetPdf().Entropy()) / nTotalSamples;
informationGainUnlabelled = entropyBefore - entropyAfter;
}
double gain =
informationGainLabelled
+ alpha_ * informationGainUnlabelled;
return(gain);
}
/// <summary>
/// Apply a trained forest to some test data.
/// </summary>
/// <typeparam name="F">Type of split function</typeparam>
/// <param name="forest">Trained forest</param>
/// <param name="testData">Test data</param>
/// <returns>An array of class distributions, one per test data point</returns>
public static HistogramAggregator[] Test <F>(Forest <F, HistogramAggregator> forest, DataPointCollection testData) where F : IFeatureResponse
{
int nClasses = forest.GetTree(0).GetNode(0).TrainingDataStatistics.BinCount;
int[][] leafIndicesPerTree = forest.Apply(testData);
HistogramAggregator[] result = new HistogramAggregator[testData.Count()];
for (int i = 0; i < testData.Count(); i++)
{
// Aggregate statistics for this sample over all leaf nodes reached
result[i] = new HistogramAggregator(nClasses);
for (int t = 0; t < forest.TreeCount; t++)
{
int leafIndex = leafIndicesPerTree[t][i];
result[i].Aggregate(forest.GetTree(t).GetNode(leafIndex).TrainingDataStatistics);
}
}
return(result);
}
public bool ShouldTerminate(HistogramAggregator parent, HistogramAggregator leftChild, HistogramAggregator rightChild, double gain)
{
return(gain < 0.01);
}
public static Bitmap Visualize <F>(
Forest <F, HistogramAggregator> forest,
DataPointCollection trainingData,
Size PlotSize,
PointF PlotDilation) where F : IFeatureResponse
{
// Size PlotSize = new Size(300, 300), PointF PlotDilation = new PointF(0.1f, 0.1f)
// Generate some test samples in a grid pattern (a useful basis for creating visualization images)
PlotCanvas plotCanvas = new PlotCanvas(trainingData.GetRange(0), trainingData.GetRange(1), PlotSize, PlotDilation);
DataPointCollection testData = DataPointCollection.Generate2dGrid(plotCanvas.plotRangeX, PlotSize.Width, plotCanvas.plotRangeY, PlotSize.Height);
Console.WriteLine("\nApplying the forest to test data...");
int[][] leafNodeIndices = forest.Apply(testData);
// Form a palette of random colors, one per class
Color[] colors = new Color[Math.Max(trainingData.CountClasses(), 4)];
// First few colours are same as those in the book, remainder are random.
colors[0] = Color.FromArgb(183, 170, 8);
colors[1] = Color.FromArgb(194, 32, 14);
colors[2] = Color.FromArgb(4, 154, 10);
colors[3] = Color.FromArgb(13, 26, 188);
Color grey = Color.FromArgb(255, 127, 127, 127);
System.Random r = new Random(0); // same seed every time so colours will be consistent
for (int c = 4; c < colors.Length; c++)
{
colors[c] = Color.FromArgb(255, r.Next(0, 255), r.Next(0, 255), r.Next(0, 255));
}
// Create a visualization image
Bitmap result = new Bitmap(PlotSize.Width, PlotSize.Height);
// For each pixel...
int index = 0;
for (int j = 0; j < PlotSize.Height; j++)
{
for (int i = 0; i < PlotSize.Width; i++)
{
// Aggregate statistics for this sample over all leaf nodes reached
HistogramAggregator h = new HistogramAggregator(trainingData.CountClasses());
for (int t = 0; t < forest.TreeCount; t++)
{
int leafIndex = leafNodeIndices[t][index];
h.Aggregate(forest.GetTree(t).GetNode(leafIndex).TrainingDataStatistics);
}
// Let's muddy the colors with grey where the entropy is high.
float mudiness = 0.5f * (float)(h.Entropy());
float R = 0.0f, G = 0.0f, B = 0.0f;
for (int b = 0; b < trainingData.CountClasses(); b++)
{
float p = (1.0f - mudiness) * h.GetProbability(b); // NB probabilities sum to 1.0 over the classes
R += colors[b].R * p;
G += colors[b].G * p;
B += colors[b].B * p;
}
R += grey.R * mudiness;
G += grey.G * mudiness;
B += grey.B * mudiness;
Color c = Color.FromArgb(255, (byte)(R), (byte)(G), (byte)(B));
result.SetPixel(i, j, c); // painfully slow but safe
index++;
}
}
// Also draw the original training data
using (Graphics g = Graphics.FromImage(result))
{
g.InterpolationMode = System.Drawing.Drawing2D.InterpolationMode.HighQualityBicubic;
g.SmoothingMode = System.Drawing.Drawing2D.SmoothingMode.HighQuality;
for (int s = 0; s < trainingData.Count(); s++)
{
PointF x = new PointF(
(trainingData.GetDataPoint(s)[0] - plotCanvas.plotRangeX.Item1) / plotCanvas.stepX,
(trainingData.GetDataPoint(s)[1] - plotCanvas.plotRangeY.Item1) / plotCanvas.stepY);
RectangleF rectangle = new RectangleF(x.X - 3.0f, x.Y - 3.0f, 6.0f, 6.0f);
g.FillRectangle(new SolidBrush(colors[trainingData.GetIntegerLabel(s)]), rectangle);
g.DrawRectangle(new Pen(Color.Black), rectangle.X, rectangle.Y, rectangle.Width, rectangle.Height);
}
}
return(result);
}
public static Bitmap VisualizeLabels(Forest <LinearFeatureResponse2d, SemiSupervisedClassificationStatisticsAggregator> forest, DataPointCollection trainingData, Size PlotSize, PointF PlotDilation)
{
// Generate some test samples in a grid pattern (a useful basis for creating visualization images)
PlotCanvas plotCanvas = new PlotCanvas(trainingData.GetRange(0), trainingData.GetRange(1), PlotSize, PlotDilation);
// Apply the trained forest to the test data
Console.WriteLine("\nApplying the forest to test data...");
DataPointCollection testData = DataPointCollection.Generate2dGrid(plotCanvas.plotRangeX, PlotSize.Width, plotCanvas.plotRangeY, PlotSize.Height);
int[][] leafNodeIndices = forest.Apply(testData);
Bitmap result = new Bitmap(PlotSize.Width, PlotSize.Height);
// Paint the test data
GaussianPdf2d[][] leafDistributions = new GaussianPdf2d[forest.TreeCount][];
for (int t = 0; t < forest.TreeCount; t++)
{
leafDistributions[t] = new GaussianPdf2d[forest.GetTree(t).NodeCount];
for (int i = 0; i < forest.GetTree(t).NodeCount; i++)
{
Node <LinearFeatureResponse2d, SemiSupervisedClassificationStatisticsAggregator> nodeCopy = forest.GetTree(t).GetNode(i);
if (nodeCopy.IsLeaf)
{
leafDistributions[t][i] = nodeCopy.TrainingDataStatistics.GaussianAggregator2d.GetPdf();
}
}
}
// Form a palette of random colors, one per class
Color[] colors = new Color[Math.Max(trainingData.CountClasses(), 4)];
// First few colours are same as those in the book, remainder are random.
colors[0] = Color.FromArgb(183, 170, 8);
colors[1] = Color.FromArgb(194, 32, 14);
colors[2] = Color.FromArgb(4, 154, 10);
colors[3] = Color.FromArgb(13, 26, 188);
Color grey = Color.FromArgb(255, 127, 127, 127);
System.Random r = new Random(0); // same seed every time so colours will be consistent
for (int c = 4; c < colors.Length; c++)
{
colors[c] = Color.FromArgb(255, r.Next(0, 255), r.Next(0, 255), r.Next(0, 255));
}
int index = 0;
for (int j = 0; j < PlotSize.Height; j++)
{
for (int i = 0; i < PlotSize.Width; i++)
{
// Aggregate statistics for this sample over all leaf nodes reached
HistogramAggregator h = new HistogramAggregator(trainingData.CountClasses());
for (int t = 0; t < forest.TreeCount; t++)
{
int leafIndex = leafNodeIndices[t][index];
SemiSupervisedClassificationStatisticsAggregator a = forest.GetTree(t).GetNode(leafIndex).TrainingDataStatistics;
h.Aggregate(a.HistogramAggregator);
}
// Let's muddy the colors with a little grey where entropy is high.
float mudiness = 0.5f * (float)(h.Entropy());
float R = 0.0f, G = 0.0f, B = 0.0f;
for (int b = 0; b < trainingData.CountClasses(); b++)
{
float p = (1.0f - mudiness) * h.GetProbability(b); // NB probabilities sum to 1.0 over the classes
R += colors[b].R * p;
G += colors[b].G * p;
B += colors[b].B * p;
}
R += grey.R * mudiness;
G += grey.G * mudiness;
B += grey.B * mudiness;
Color c = Color.FromArgb(255, (byte)(R), (byte)(G), (byte)(B));
result.SetPixel(i, j, c);
index++;
}
}
PaintTrainingData(trainingData, plotCanvas, result);
return(result);
}