private void loadToolStripMenuItem_Click(object sender, EventArgs e)
{
imgClassifier = new BowImageClassifier();
imgClassifier.Init(
dataPath + String.Format(@"\train-{0}.bow", bowSize),
dataPath + String.Format(@"\train-{0}.net", bowSize),
mask
);
logger.logStr("Loaded");
}
private bool Train(dynamic bow, SceneFeatureData scenefeatureData)
{
Bitmap mask = Utils.CreateMaskBitmap(new Size(1280, 720), new Rectangle[] { scenefeatureData.feature.area });
mask.Save(scenefeatureData.feature.name + "-mask.png");
var trainData = scenefeatureData.trainData;
int[] labelIndexs = trainData.GetLabelIndexs();
String[] labels = trainData.GetLabels();
double[][] features = trainData.GetFeature(bow, mask);
int numOutput = trainData.GetNumOutput();
var function = new SigmoidFunction();
bool flgFound = false;
int count = 0;
while ((flgFound == false) && (count < 100))
{
count++;
var network = new ActivationNetwork(function, bow.NumberOfOutputs, 20, numOutput);
new NguyenWidrow(network).Randomize();
var teacher = new ParallelResilientBackpropagationLearning(network);
BowImageClassifier trainImgClassifier = new BowImageClassifier();
trainImgClassifier.Init(bow, network, mask);
//creat output
double[][] outputs = trainData.GetOutputs(numOutput);
double avgError = 10000.0;
double prevError = avgError;
double bestError = avgError;
int errorCount = 0;
while ((errorCount < 3) && (avgError > 0.00001))
{
//Application.DoEvents();
double[] errors = new double[10];
for (int i = 0; i < 10; i++)
{
errors[i] = teacher.RunEpoch(features, outputs);
}
avgError = errors.Average();
if (prevError > avgError)
{
int trainError = trainImgClassifier.Evaluate(trainData);
//int testError = trainImgClassifier.Evaluate(testData);
//int testSetError = trainImgClassifier.Evaluate(testDataSet);
logger.logStr(String.Format("{0} {1} {2}", avgError, prevError, trainError));
prevError = avgError;
//save best error
if (bestError > avgError)
{
bestError = avgError;
//Accord.IO.Serializer.Save(network, dataPath + String.Format(@"\train-{0}.net", bow.NumberOfOutputs));
}
if (trainError /*+ testError + testSetError*/ == 0)
{
Accord.IO.Serializer.Save(network, path + @"\" + scenefeatureData.feature.name + String.Format(@"\train-{0}.net", bow.NumberOfOutputs));
logger.logStr("Done " + bestError + " " + trainError);
return(true);
}
}
else
{
logger.logStr(String.Format("{0}", avgError));
prevError = 10000.0;
errorCount++;
}
//Application.DoEvents();
}
logger.logStr("Done " + bestError + " " + count);
}
return(false);
}
private void trainToolStripMenuItem_Click_1(object sender, EventArgs e)
{
ImageTrainDataSet testDataSet = FeatureDetector.GetAllTrainImageData(testDataPath, configure.trainFolders);
testDataSet.flgCache = true;
int[] labelIndexs = trainData.GetLabelIndexs();
String[] labels = trainData.GetLabels();
var bow = Accord.IO.Serializer.Load <BagOfVisualWords>(dataPath + String.Format(@"\train-{0}.bow", bowSize));
double[][] features = trainData.GetFeature(bow, mask);
int numOutput = trainData.GetNumOutput();
var function = new SigmoidFunction();
logger.logStr("Start Training");
bool flgFound = false;
int count = 0;
while ((flgFound == false) && (count < 100))
{
count++;
var network = new ActivationNetwork(function, bow.NumberOfOutputs, 20, numOutput);
new NguyenWidrow(network).Randomize();
var teacher = new ParallelResilientBackpropagationLearning(network);
BowImageClassifier trainImgClassifier = new BowImageClassifier();
trainImgClassifier.Init(bow, network, mask);
//creat output
double[][] outputs = trainData.GetOutputs(numOutput);
double avgError = 10000.0;
double prevError = avgError;
double bestError = avgError;
int errorCount = 0;
while ((errorCount < 3) && (avgError > 0.00001))
{
//Application.DoEvents();
double[] errors = new double[10];
for (int i = 0; i < 10; i++)
{
errors[i] = teacher.RunEpoch(features, outputs);
}
avgError = errors.Average();
if (prevError > avgError)
{
int trainError = trainImgClassifier.Evaluate(trainData);
int testError = trainImgClassifier.Evaluate(testData);
int testSetError = trainImgClassifier.Evaluate(testDataSet);
logger.logStr(String.Format("{0} {1} {2} {3} {4} #{5}", avgError, prevError, trainError, testError, testSetError, errorCount));
prevError = avgError;
//save best error
if (bestError > avgError)
{
bestError = avgError;
Accord.IO.Serializer.Save(network, dataPath + String.Format(@"\train-{0}.net", bow.NumberOfOutputs));
}
if (trainError + testError + testSetError == 0)
{
flgFound = true;
Accord.IO.Serializer.Save(network, dataPath + String.Format(@"\train-{0}.net", bow.NumberOfOutputs));
break;
}
}
else
{
logger.logStr(String.Format("{0}", avgError));
prevError = 10000.0;
errorCount++;
}
Application.DoEvents();
}
logger.logStr("Done " + bestError + " " + count);
}
}
