public ImageTrainDataSet SplitTestData(int num)
{
Random rand = new Random();
//List<ImageTrainData> ret = _self.GetRange(0, num);
//_self.RemoveRange(0, num);
int numOutput = this.GetNumOutput();
ImageTrainDataSet ret = new ImageTrainDataSet();
while (ret.Count() < num)
{
for (int i = 0; i < numOutput; i++)
{
while (true)
{
int index = rand.Next(this.Count);
if (this[index].labelIndex == i)
{
ret.Add(this[index]);
this.RemoveAt(index);
break;
}
}
}
}
return(ret);
}
public ImageTrainDataSet BalanceData()
{
Random rand = new Random();
ImageTrainDataSet ret = new ImageTrainDataSet();
ret.AddRange(this);
TrainDataInfo[] infos = this.GetInfo(this.GetNumOutput());
int maxCount = infos.Max(x => x.count);
for (int i = 0; i < infos.Length; i++)
{
for (int j = infos[i].count; j < maxCount; j++)
{
//Random _self for that index
while (true)
{
int index = rand.Next(this.Count);
if (this[index].labelIndex == i)
{
ret.Add(this[index].Clone());
break;
}
}
}
}
return(ret);
}
private void testWithTestSetToolStripMenuItem_Click(object sender, EventArgs e)
{
ImageTrainDataSet testDataSet = FeatureDetector.GetAllTrainImageData(testDataPath, configure.trainFolders);
testDataSet.flgCache = true;
byte[] hash = imgClassifier.ComputeHash(testDataSet);
logger.logStr(hash.ToHex());
int testErrCount = imgClassifier.Evaluate(testDataSet);
logger.logStr(String.Format("test {0}", testErrCount));
}
private void loadTrainDataToolStripMenuItem_Click_1(object sender, EventArgs e)
{
trainData = FeatureDetector.GetAllTrainImageData(dataPath, configure.trainFolders);
trainData = trainData.BalanceData();
trainData.Shuffle();
testData = trainData.SplitTestData((int)(0.1 * trainData.Count()));
TrainDataInfo[] infos = trainData.GetInfo(configure.trainFolders.Count());
TrainDataInfo[] testinfos = testData.GetInfo(configure.trainFolders.Count());
logger.logStr(Utils.ToJsonString(infos, true));
logger.logStr(Utils.ToJsonString(testinfos, true));
}
public int Evaluate(ImageTrainDataSet trainData)
{
double[][] features = trainData.GetFeature(bow, mask);
int[] labelIndexs = trainData.GetLabelIndexs();
String[] labels = trainData.GetLabels();
int errorCount = 0;
for (int i = 0; i < trainData.Count(); i++)
{
//double[] feature = bow.Transform(images[i]);
double[] answer = network.Compute(features[i]);
int expected = labelIndexs[i];
int actual; answer.Max(out actual);
if (actual != expected)
{
errorCount++;
}
}
return(errorCount);
}
public static ImageTrainDataSet GetAllTrainImageData(String folder, String[] subFolders, String imageFilter = "*.png", String imageFilterOut = "mask.png")
{
ImageTrainDataSet ret = new ImageTrainDataSet();
for (int i = 0; i < subFolders.Length; i++)
{
String folderName = folder + @"\" + subFolders[i];
ImageFileData[] images = GetImagesFromDir(folderName, imageFilter, imageFilterOut);
foreach (ImageFileData img in images)
{
ret.Add(
new ImageTrainData()
{
label = subFolders[i],
labelIndex = i,
fileName = img.fileName,
}
);
}
}
return(ret);
}
public byte[] ComputeHash(ImageTrainDataSet trainData)
{
return(trainData.MD5Feature(bow, mask));
}
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);
}
}
