static ScoreStats GetScoreStats(int Rows, BlockedImage[] images, Kernel kernel, int InLabel, BlockTracker tracker = null)
{
var stats = new ScoreStats();
for (int i = 0; i < Rows; i++)
{
var image = images[i];
var action = tracker == null ? null : tracker.GetTrackInOut(image, InLabel);
float error = GetScore(image, kernel, action: action);
if (image.Label == InLabel)
{
stats.InCount++;
if (error < kernel.ClassificationCutoff)
{
stats.InHit++;
}
}
else
{
stats.OutCount++;
if (error < kernel.ClassificationCutoff)
{
stats.OutHit++;
}
}
}
Console.WriteLine($" - in label {stats.InHit} / {stats.InCount} - out label {stats.OutHit} / {stats.OutCount} - {kernel.ShortDescription()}");
return(stats);
}
static ScoreStats TestThreshold(int Rows, Kernel kernel, float[] source_score, byte[] source_label, float threshold, int InLabel)
{
var stats = new ScoreStats();
for (int i = 0; i < Rows; i++)
{
var label = source_label[i];
var error = source_score[i];
if (label == InLabel)
{
stats.InCount++;
if (error < threshold)
{
stats.InHit++;
}
}
else
{
stats.OutCount++;
if (error < threshold)
{
stats.OutHit++;
}
}
}
//Console.WriteLine($" - in label {stats.InHit} / {stats.InCount} - out label {stats.OutHit} / {stats.OutCount} - error threshold {threshold:0.0000} - {kernel.ShortDescription()}");
return(stats);
}
public ScoreStats[] StatsForClasses()
{
var stats = new ScoreStats[NumLabels];
for (int label = 0; label < NumLabels; label++)
{
stats[label] = StatsForClass(label);
}
return(stats);
}
static ScoreStats OptimizeKernelCutoff(int Rows, BlockedImage[] blocks, Kernel kernel, int InLabel,
float InLabelMinRatio = 0.05f)
{
float[] source_score = GetScores(Rows, blocks, kernel);
var labels = get_labels(blocks);
ScoreStats best = null;
for (int i = 0; i < 1000; i++)
{
float threshold = .001f + .5f * (float)rnd.NextDouble();
var result = TestThreshold(Rows, kernel, source_score, labels, threshold, InLabel);
if (best == null || result.InHitRatio > InLabelMinRatio && result.InOutHitRatio > best.InOutHitRatio)
{
best = result;
kernel.ClassificationCutoff = threshold;
}
}
return(best);
}
public ScoreStats StatsForClass(int label)
{
var stats = new ScoreStats();
for (int l = 0; l <= NumLabels; l++)
{
stats.InCount += Hits(PredictedLabel: l, ActualLabel: label);
}
stats.OutCount = Total - stats.InCount;
for (int l = 0; l <= NumLabels; l++)
{
if (l == label)
{
stats.InHit += Hits(PredictedLabel: label, ActualLabel: l);
}
else
{
stats.OutHit += Hits(PredictedLabel: label, ActualLabel: l);
}
}
return(stats);
}