private void button5_Click(object sender, EventArgs e)
{
//Test loading a MNIST image, distorting it, blurring it, converting it back to a DigitImage and then saving it to disk
ImageProcessing.DigitImageCollection sourceImages = ImageProcessing.DigitImageCollection.LoadMnistDataSet("TrainingImages.dat", "TrainingLabels.dat", 60000);
string imageFolder = Path.Combine(Environment.GetFolderPath(Environment.SpecialFolder.MyDocuments), "ConvertedImages");
if (Directory.Exists(imageFolder))
{
Directory.Delete(imageFolder, true);
}
Directory.CreateDirectory(imageFolder);
DigitImage originalDigitImage = sourceImages.DigitImages[0];
originalDigitImage.SaveToFile(Path.Combine(imageFolder, "Original.jpg"));
Bitmap distortedImage = sourceImages.DigitImages[0].ToBitmap().DistortionBlurFilter(25);
DigitImage distortedDigitImage = new DigitImage(distortedImage, sourceImages.DigitImages[0].Label);
distortedDigitImage.SaveToFile(Path.Combine(imageFolder, "Distorted.jpg"));
Bitmap blurredImage = sourceImages.DigitImages[0].ToBitmap().ImageBlurFilter(BlurringBitmapExtensions.BlurType.GaussianBlur5x5);
DigitImage blurredDigitImage = new DigitImage(blurredImage, sourceImages.DigitImages[0].Label);
blurredDigitImage.SaveToFile(Path.Combine(imageFolder, "Blurred.jpg"));
}
public CallbackModelWrapper PopulateImages(CallbackModelWrapper model)
{
Init(model);
var pixels = model.GenerateNewPixelMatrices();
int amount = 0;
try
{
while (true)
{
for (int i = 0; i < 28; ++i)
{
for (int j = 0; j < 28; ++j)
{
byte b = brImages.ReadByte();
pixels[i][j] = b;
}
}
byte lbl = brLabels.ReadByte();
var image = new DigitImage(pixels, lbl, ++amount);
model.Images.Add(image);
}
}
catch (Exception e)
{
//will get exception here
//BinaryReader cannot read past end of file error
Console.Write(e.ToString());
}
brImages.Close();
brLabels.Close();
return(model);
}
public void Next()
{
DigitImage img = Program.RandomTestingData;
TrainingItem ti = MnistReader.ConvertSingleImage(img);
Image realImg = img.ToBitMap();
input.Image = new Bitmap(realImg, input.Size);
byte b = 0;
for (int i = 0; i < 4; i++)
{
b += (byte)(ti.outputs[i] > 0.5 ? Math.Pow(2, i) : 0);
}
byte o = Program.GetOutput(ti.inputs);
output.Text = (timerRunning ? "Test in progress : \n" : "") + "Found : " + o + "\nExpected : " + b;
if (timerRunning)
{
loop++;
if (o == b)
{
worked++;
}
if (loop >= 100)
{
timerRunning = false;
timer.Stop();
output.Text = "Test finished\nSuccess : " + worked + " %";
}
}
}
private void SetMainText(DigitImage image)
{
pcbMain.Image = image.NumberImage;
lblExpected.Text = image.Label.ToString();
lblActual.Text = network.GetNumber(image).ToString();
lblActual.ForeColor = (lblActual.Text == lblExpected.Text) ? Color.Green : Color.Red;
}
private void button6_Click(object sender, EventArgs e)
{
//Load the original Mnist training images and create a new blurred data set from them
ImageProcessing.DigitImageCollection sourceImages = ImageProcessing.DigitImageCollection.LoadMnistDataSet("TrainingImages.dat", "TrainingLabels.dat", 60000);
DigitImageCollection blurredImages = new DigitImageCollection();
blurredImages.DigitImages = new DigitImage[sourceImages.DigitImages.Length];
Parallel.For(0, sourceImages.DigitImages.Length, i =>
{
int imageIndex = i;
DigitImage blurredDigitImage = new DigitImage(sourceImages.DigitImages[imageIndex].ToBitmap().ImageBlurFilter(BlurringBitmapExtensions.BlurType.GaussianBlur5x5), sourceImages.DigitImages[imageIndex].Label);
blurredImages.DigitImages[imageIndex] = blurredDigitImage;
});
string imageFolder = Path.Combine(Environment.GetFolderPath(Environment.SpecialFolder.MyDocuments), "BlurredTrainingDataset");
if (Directory.Exists(imageFolder))
{
Directory.Delete(imageFolder, true);
}
Directory.CreateDirectory(imageFolder);
blurredImages.SaveToFile(Path.Combine(imageFolder, "BlurredTrainingImages.dat"), Path.Combine(imageFolder, "BlurredTrainingLabels.dat"));
MessageBox.Show("Done");
}
public void StartCycle(DigitImage digitImage)
{
currentLabel = digitImage.label;
ForwardPropagation(digitImage);
BackwardPropagation(digitImage);
}
private void pictureBox1_Paint(object sender, PaintEventArgs e)
{
var mag = DigitImage.MakeBitmap(DigitImage.FromBitmap(new Bitmap(GetInputImage()), 1), 10);
pictureBox2.Image = mag;
pictureBox2.Invalidate();
}
private DigitImage addToListView(DigitImage image)
{
var guid = Guid.NewGuid().ToString();
imageList.Images.Add(guid, image.NumberImage);
lvImages.Items.Add(image.Label.ToString(), imageList.Images.IndexOfKey(guid));
lvImages.Items[lvImages.Items.Count - 1].Tag = image;
lvImages.Refresh();
return(image);
}
public void SetImage(DigitImage image)
{
MainGrid.Children.Clear();
for (int x = 0; x < 28; x++)
{
for (int y = 0; y < 28; y++)
{
SetInputShade(y, x, image.GetValueAt(y, x));
}
}
}
public int DetectNumberInImage(Bitmap inputImage)
{
var di = DigitImage.FromBitmap(inputImage, 1);
var response = network.Query(di.pixels.SelectMany(su => su.Select(DigitImage.ConvertGrayScaleByteToDouble)).ToArray());
var max = response.Max(x => x);
var f = response.ToList().IndexOf(max);
return(f);
}
public static DigitImage[] LoadImage(string pixelFile, string labelFile)
{
int numImages = 60000;
DigitImage[] digitImages = new DigitImage[numImages];
byte[][] pixels = new byte[28][];
for (int i = 0; i < pixels.Length; i++)
{
pixels[i] = new byte[28];
}
FileStream ifsPixels = new FileStream(pixelFile, FileMode.Open);
FileStream ifsLabels = new FileStream(labelFile, FileMode.Open);
BinaryReader brImages = new BinaryReader(ifsPixels);
BinaryReader brLabels = new BinaryReader(ifsLabels);
int magic1 = brImages.ReadInt32(); // обратный порядок байтов
magic1 = ReverseBytes(magic1); // преобразуем в формат Intel
int imageCount = brImages.ReadInt32();
imageCount = ReverseBytes(imageCount);
int numRows = brImages.ReadInt32();
numRows = ReverseBytes(numRows);
int numCols = brImages.ReadInt32();
numCols = ReverseBytes(numCols);
int magic2 = brLabels.ReadInt32();
magic2 = ReverseBytes(magic2);
int numLabels = brLabels.ReadInt32();
numLabels = ReverseBytes(numLabels);
for (int i = 0; i < numImages; i++)
{
for (int j = 0; j < 28; j++)
{
for (int k = 0; k < 28; k++)
{
byte b = brImages.ReadByte();
pixels[j][k] = b;
}
}
byte lb = brLabels.ReadByte();
DigitImage digitImage = new DigitImage(28, 28, pixels, lb);
digitImages[i] = digitImage;
}
ifsPixels.Close();
brImages.Close();
ifsLabels.Close();
brLabels.Close();
return(digitImages);
}
public void PaintFromDatabase(DigitImage image)
{
screen.GiveCaller(this);
for (int i = 0; i < 28; i++)
{
for (int j = 0; j < 28; j++)
{
float temp = image.pixels[i][j];
temp = (255 - temp) / 255f;
screenData[i][j].GetComponent <SpriteRenderer>().color = new Color(temp, temp, temp);
}
}
}
private void ReadMNISTDataset()
{
try
{
var directories = GetDirectories();
var ifsLabels = new FileStream(directories.LabelsFileDir, FileMode.Open);
var ifsImages = new FileStream(directories.ImagesFileDir, FileMode.Open);
ifsLabels.Seek(_labelPosition, SeekOrigin.Begin);
ifsImages.Seek(_datasetPosition, SeekOrigin.Begin);
var brLabels = new BinaryReader(ifsLabels);
var brImages = new BinaryReader(ifsImages);
int magic1 = brImages.ReadInt32();
int numImages = brImages.ReadInt32();
int numRows = brImages.ReadInt32();
int numCols = brImages.ReadInt32();
int magic2 = brLabels.ReadInt32();
int numLabels = brLabels.ReadInt32();
var pixels = new byte[28][];
for (int i = 0; i < pixels.Length; ++i)
{
pixels[i] = new byte[28];
}
for (int i = 0; i < 28; ++i)
{
for (int j = 0; j < 28; ++j)
{
byte b = brImages.ReadByte();
pixels[i][j] = b;
}
}
byte lbl = brLabels.ReadByte();
var dImage = new DigitImage(pixels, lbl);
CorrespondingLabel = lbl;
PixelArray = dImage.GetImageArray();
var formatEmptyLines = Regex.Replace(dImage.ToString(), @"^\s+$[\r\n]*", string.Empty, RegexOptions.Multiline);
DatasetInformation += string.Format("{0}\n\r", formatEmptyLines);
ifsImages.Close();
brImages.Close();
ifsLabels.Close();
brLabels.Close();
}
catch (Exception ex)
{
_exceptionLogging.LogException(ex.ToString());
}
}
public void ForwardPropagation(DigitImage digitImage)
{
for (int i = 0; i < firstLayerSize; i++)
{
for (int j = 0; j < pixelNum; j++)
{
tempValue += (double)(digitImage.arrayPixels[j] / 255f) * input_FirstLayerWeights[j][i];
}
tempValue += firstLayerBias[i];
if (tempValue > gradientClippingMax)
{
tempValue = gradientClippingMax;
}
else if (tempValue < gradientClippingMin)
{
tempValue = gradientClippingMin;
}
firstLayerZ[i] = tempValue;
firstLayerActivations[i] = Sigmoid(tempValue);
tempValue = 0;
}
for (int i = 0; i < secondLayerSize; i++)
{
for (int j = 0; j < firstLayerSize; j++)
{
tempValue += firstLayerActivations[j] * firstLayer_SecondLayerWeights[j][i];
}
tempValue += secondLayerBias[i];
secondLayerZ[i] = tempValue;
secondLayerActivations[i] = Sigmoid(tempValue);
tempValue = 0;
}
for (int i = 0; i < outputLayerSize; i++)
{
for (int j = 0; j < secondLayerSize; j++)
{
tempValue += secondLayerActivations[j] * secondLayer_OutputWeights[j][i];
}
tempValue += outputLayerBias[i];
outputLayerZ[i] = tempValue;
outputLayerActivations[i] = Sigmoid(tempValue);
tempValue = 0;
}
SeeIfCorrect();
}
private async Task <DigitImage> ReadOneImageAsync(CallbackModelWrapper model, int imageSize, int offset_image, int offset_label, int orderInCollection)
{
int bufferSize = imageSize * imageSize;
byte[] image_bytes = new byte[bufferSize];
byte[] label_bytes = new byte[1];
using (FileStream fsi = new FileStream(@"..\..\Files\t10k-images.idx3-ubyte",
FileMode.Open, FileAccess.Read, FileShare.Read,
bufferSize: 4096, useAsync: true))
using (FileStream fsl = new FileStream(@"..\..\Files\t10k-labels.idx1-ubyte",
FileMode.Open, FileAccess.Read, FileShare.Read,
bufferSize: 4096, useAsync: true))
{
//lock (fsi)
//{
// fsi.copy
//}
fsi.Seek(offset_image, SeekOrigin.Begin);
fsl.Seek(offset_label, SeekOrigin.Begin);
var tmp_label = fsl.Read(label_bytes, 0, 1);
var tmp_image = await fsi.ReadAsync(image_bytes, 0, bufferSize);
//var tmp_image = fsi.Read(image_bytes, 0, bufferSize);
}
var pixels = model.GenerateNewPixelMatrices();
for (int i = 0; i < imageSize; ++i)
{
for (int j = 0; j < imageSize; ++j)
{
pixels[i][j] = image_bytes[i * imageSize + j];
}
}
DigitImage image = new DigitImage(pixels, label_bytes[0], orderInCollection);
return(image);
}
public static Bitmap GetBitmap(DigitImage digitImage, int mag)
{
int width = digitImage.Width * mag;
int height = digitImage.Height * mag;
Bitmap bitmap = new Bitmap(width, height);
Graphics graphics = Graphics.FromImage(bitmap);
for (int i = 0; i < digitImage.Height; i++)
{
for (int j = 0; j < digitImage.Width; j++)
{
int pixelColor = 255 - digitImage.Pixels[i][j];
System.Drawing.Color color = System.Drawing.Color.FromArgb(pixelColor, pixelColor, pixelColor);
SolidBrush solidBrush = new SolidBrush(color);
graphics.FillRectangle(solidBrush, j * mag, i * mag, mag, mag);
}
}
return(bitmap);
}
private static System.Drawing.Image GetImage(Layer layer)
{
byte[][] pixels = new byte[layer.NumOfNeurons][];
for (int i = 0; i < layer.NumOfNeurons; i++)
{
pixels[i] = new byte[layer.NumOfNeurons];
}
for (int i = 0; i < layer.NumOfNeurons; i++)
{
for (int j = 0; j < layer.NumOfNeurons; j++)
{
pixels[i][j] = (byte)layer.Neurons.Find(x => x.MapX == i && x.MapY == j).Output;
}
}
DigitImage digitImage = new DigitImage(layer.NumOfNeurons, layer.NumOfNeurons, pixels, 4);
Bitmap bitmap = GetBitmap(digitImage, 1);
System.Drawing.Image image = bitmap;
return(image);
}
public void LoadData(Action <int> progressUpdater)
{
if (File.Exists(jsonFilePath))
{
var deserialized = JsonConvert.DeserializeObject <NeuralNetwork>(File.ReadAllText(jsonFilePath));
network = deserialized;
progressUpdater.Invoke(100);
return;
}
network = new NeuralNetwork(inputLength, 100, 10, 0.3);
var database = MnistDataLoader.LoadData(pixelFile, labelFile);
var toDoubles = database.Select(i => i.pixels.SelectMany(s => s)
.Select(
byt => DigitImage.ConvertGrayScaleByteToDouble(byt))).ToArray();
//Test2(toDoubles, database);
var targets = database.Select(i => Convert.ToInt32(i.label)).ToArray();
var percent = 0;
for (var index = 0; index < toDoubles.Length; index++)
{
var doubleArray = toDoubles[index];
var correctAnswer = targets[index];
var target = Enumerable.Range(0, 10).Select(x => 0.01).ToArray();
target[correctAnswer] = 0.99;
percent = NotifyDataLoadingProgress(index, toDoubles.Length, percent, progressUpdater);
network.Train(doubleArray.ToArray(), target);
}
var networkJson = JsonConvert.SerializeObject(network);
File.WriteAllText(jsonFilePath, networkJson);
}
private CallbackModelWrapper PopulateImages(CallbackModelWrapper model, int amount)
{
var pixels = model.GenerateNewPixelMatrices();
for (int di = 0; di < amount; ++di)
{
for (int i = 0; i < 28; ++i)
{
for (int j = 0; j < 28; ++j)
{
byte b = brImages.ReadByte();
pixels[i][j] = b;
}
}
byte lbl = brLabels.ReadByte();
var image = new DigitImage(pixels, lbl, di + 1);
model.Images.Add(image);
}
brImages.Close();
brLabels.Close();
return(model);
}
public static DigitImage[] ReadTraining(string path)
{
try
{
FileStream ifsLabels =
new FileStream(@path + "train-labels.idx1-ubyte",
FileMode.Open); // test labels
FileStream ifsImages =
new FileStream(@path + "train-images.idx3-ubyte",
FileMode.Open); // test images
BinaryReader brLabels =
new BinaryReader(ifsLabels);
BinaryReader brImages =
new BinaryReader(ifsImages);
int magic1 = brImages.ReadInt32(); // discard
int numImages = brImages.ReadInt32();
int numRows = brImages.ReadInt32();
int numCols = brImages.ReadInt32();
DigitImage[] TestImages = new DigitImage[60000];
int magic2 = brLabels.ReadInt32();
int numLabels = brLabels.ReadInt32();
byte[][] pixels = new byte[28][];
for (int i = 0; i < pixels.Length; ++i)
{
pixels[i] = new byte[28];
}
// each test image
for (int di = 0; di < 60000; ++di)
{
for (int i = 0; i < 28; ++i)
{
for (int j = 0; j < 28; ++j)
{
byte b = brImages.ReadByte();
pixels[i][j] = b;
}
}
byte lbl = brLabels.ReadByte();
DigitImage dImage =
new DigitImage(pixels, lbl);
TestImages[di] = dImage;
} // each image
ifsImages.Close();
brImages.Close();
ifsLabels.Close();
brLabels.Close();
TestImages = TestImages.Take(50000).ToArray();
return(TestImages);
}
catch (Exception ex)
{
Console.WriteLine(ex.Message);
Console.ReadLine();
return(null);
}
}
private async Task Populate_AddImageAsync(CallbackModelWrapper model, CancellationTokenSource cts)
{
var index = 0;
lock (model)
{
index = model.index++;
}
int bufferSize = model.ImageSize * model.ImageSize;
var offset_image = model.offset_origin_image + index * bufferSize;
var offset_label = model.offset_origin_label + index;
var orderInCollection = index + 1;
byte[] image_bytes = new byte[bufferSize];
byte[] label_bytes = new byte[1];
//
using (FileStream fsi = new FileStream(model.imagePath,
FileMode.Open, FileAccess.Read, FileShare.Read,
bufferSize: 4096, useAsync: true))
using (FileStream fsl = new FileStream(model.labelPath,
FileMode.Open, FileAccess.Read, FileShare.Read,
bufferSize: 4096, useAsync: true))
{
var tmp_label = 0;
var tmp_image = 0;
try
{
fsi.Seek(offset_image, SeekOrigin.Begin);
fsl.Seek(offset_label, SeekOrigin.Begin);
tmp_label = fsl.Read(label_bytes, 0, 1);
tmp_image = await fsi.ReadAsync(image_bytes, 0, bufferSize, cts.Token);
//var tmp_image = fsi.Read(image_bytes, 0, bufferSize);
}
catch (TaskCanceledException e)
{
throw e;
}
//throw exception here, no byte can be read
if (tmp_label == 0 || tmp_image == 0)
{
cts.Cancel();
return;
//cancellation token only yields exception for later threads
//to prevent the current thread continue, either throw an exception or return
//the behavior observed is that the taskcancellation exception can't be caught if
//I throw an exception here, so just do a return
//throw new Exception("reach the end of stream, use token to cancel all tasks");
}
}
//var t = System.Threading.Thread.CurrentThread.ManagedThreadId;
var pixels = model.GenerateNewPixelMatrices();
for (int i = 0; i < model.ImageSize; ++i)
{
for (int j = 0; j < model.ImageSize; ++j)
{
pixels[i][j] = image_bytes[i * model.ImageSize + j];
}
}
DigitImage image = new DigitImage(pixels, label_bytes[0], orderInCollection);
model.Images.Add(image);
}
/// <summary>
/// Return number of calculated number
/// </summary>
/// <param name="image">Input image</param>
/// <returns>Calculated number</returns>
public byte GetNumber(DigitImage image)
{
var data = Calculate(image.ToDoubleArray()).ToList();
return((byte)data.IndexOf(data.Max()));
}
public static List <DigitImage> ReadMNIST(bool Visualize)
{
var trainingData = new List <DigitImage>();
try
{
Console.Write("\nLoading MNIST...");
FileStream ifsLabels =
new FileStream(@"MNIST/train-labels.idx1-ubyte",
FileMode.Open); // test labels
FileStream ifsImages =
new FileStream(@"MNIST/train-images.idx3-ubyte",
FileMode.Open); // test images
BinaryReader brLabels =
new BinaryReader(ifsLabels);
BinaryReader brImages =
new BinaryReader(ifsImages);
int magic1 = brImages.ReadInt32(); // discard
int numImages = brImages.ReadInt32();
int numRows = brImages.ReadInt32();
int numCols = brImages.ReadInt32();
int magic2 = brLabels.ReadInt32();
int numLabels = brLabels.ReadInt32();
byte[][] pixels = new byte[28][];
for (int i = 0; i < pixels.Length; ++i)
{
pixels[i] = new byte[28];
}
// each test image
for (int di = 0; di < 60000; ++di)
{
if (di % 1000 == 0)
{
Console.Write(".");
}
for (int i = 0; i < 28; ++i)
{
for (int j = 0; j < 28; ++j)
{
byte b = brImages.ReadByte();
pixels[i][j] = b;
}
}
byte lbl = brLabels.ReadByte();
DigitImage dImage =
new DigitImage(pixels, lbl);
if (Visualize)
{
Console.WriteLine(dImage.ToString());
}
trainingData.Add(dImage);
// Console.ReadLine();
} // each image
ifsImages.Close();
brImages.Close();
ifsLabels.Close();
brLabels.Close();
Console.WriteLine("\nLoaded " + trainingData.Count + " Elements");
Console.WriteLine("\nCompleted\n");
return(trainingData);
}
catch (Exception ex)
{
Console.WriteLine(ex.Message);
Console.ReadLine();
return(null);
}
} // Main
public static void Run(int durchlaufe, int ausgebenPro)
{
double Fehler = 0;
SimpleNet Chantall = new SimpleNet(0.1, 784, 30, 20, 10);
Chantall.SetAfNames(AF.Sigmoid, "n", "n", "n");
Chantall.SetAfT(1, 1);
Chantall.SetAfT(2, 1);
Chantall.SetAfT(3, 1);
FileStream ifsLabels = new FileStream(@"C:\Users\Gregor\DATENBSNKEBN\train-labels.idx1-ubyte", FileMode.Open); // test labels
FileStream ifsImages = new FileStream(@"C:\Users\Gregor\DATENBSNKEBN\train-images.idx3-ubyte", FileMode.Open); // test images
BinaryReader brLabels = new BinaryReader(ifsLabels);
BinaryReader brImages = new BinaryReader(ifsImages);
int magic1 = brImages.ReadInt32(); // discard
int numImages = brImages.ReadInt32();
int numRows = brImages.ReadInt32();
int numCols = brImages.ReadInt32();
int magic2 = brLabels.ReadInt32();
int numLabels = brLabels.ReadInt32();
byte[][] pixels = new byte[28][];
for (int i = 0; i < pixels.Length; ++i)
{
pixels[i] = new byte[28];
}
for (int i = 1; i < durchlaufe; ++i)
{
if (i % 60000 == 0)
{
ifsImages.Close();
brImages.Close();
ifsLabels.Close();
brLabels.Close();
ifsLabels = new FileStream(@"C:\Users\Gregor\memBrain\train-labels.idx1-ubyte", FileMode.Open); // test labels
ifsImages = new FileStream(@"C:\Users\Gregor\memBrain\train-images.idx3-ubyte", FileMode.Open); // test images
brLabels = new BinaryReader(ifsLabels);
brImages = new BinaryReader(ifsImages);
magic1 = brImages.ReadInt32(); // discard
numImages = brImages.ReadInt32();
numRows = brImages.ReadInt32();
numCols = brImages.ReadInt32();
magic2 = brLabels.ReadInt32();
numLabels = brLabels.ReadInt32();
pixels = new byte[28][];
for (int j = 0; j < pixels.Length; ++j)
{
pixels[j] = new byte[28];
}
}
for (int i1 = 0; i1 < 28; ++i1)
{
for (int j = 0; j < 28; ++j)
{
byte b = brImages.ReadByte();
pixels[i1][j] = b;
}
}
byte lbl = brLabels.ReadByte();
double[] erwrteterOutput = new double[10];
erwrteterOutput[lbl] = 1;
Chantall.SetInput(ConvertD(pixels));
Chantall.OutputBerechnen();
Chantall.DeltawertBerechen(erwrteterOutput);
Chantall.Backpropagation();
Fehler += Chantall.GetFehlerDif(erwrteterOutput);
if (i % ausgebenPro == 0)
{
Fehler /= ausgebenPro;
DigitImage dImage = new DigitImage(pixels, lbl);
Console.WriteLine(dImage.ToString());
ZweiArrayAusgeben(erwrteterOutput, Chantall.GetOutput());
Console.WriteLine("Treffer Quote: {0}", 100 - (Fehler * 10));
Console.WriteLine(Chantall.ToString());
Chantall.Etha = Fehler;
Fehler = 0;
}
}
}
public void BackwardPropagation(DigitImage digitImage)
{
for (int i = 0; i < outputLayerSize; i++)
{
if (i == digitImage.label)
{
expectedValues[i] = 1;
}
else
{
expectedValues[i] = 0;
}
}
for (int i = 0; i < outputLayerSize; i++)
{
outputLayerBiasCosts[i] = SigmoidDerivative(outputLayerZ[i]) * 2 * (outputLayerActivations[i] - expectedValues[i]);
for (int j = 0; j < secondLayerSize; j++)
{
secondLayer_OutputWeightCosts[j][i] = outputLayerBiasCosts[i] * secondLayerActivations[j];
}
}
for (int j = 0; j < secondLayerSize; j++)
{
for (int i = 0; i < outputLayerSize; i++)
{
secondLayerActivationCosts[j] += secondLayer_OutputWeights[j][i] * outputLayerBiasCosts[i];
}
}
//L
for (int i = 0; i < secondLayerSize; i++)
{
secondLayerBiasCosts[i] = SigmoidDerivative(secondLayerZ[i]) * secondLayerActivationCosts[i];
for (int j = 0; j < firstLayerSize; j++)
{
firstLayer_SecondLayerWeightCosts[j][i] = secondLayerBiasCosts[i] * firstLayerActivations[j];
}
}
for (int j = 0; j < firstLayerSize; j++)
{
for (int i = 0; i < secondLayerSize; i++)
{
firstLayerActivationCosts[j] += firstLayer_SecondLayerWeights[j][i] * secondLayerBiasCosts[i];
}
}
//L-1
for (int i = 0; i < firstLayerSize; i++)
{
firstLayerBiasCosts[i] = SigmoidDerivative(firstLayerZ[i]) * firstLayerActivationCosts[i];
for (int j = 0; j < pixelNum; j++)
{
input_FirstLayerWeightCosts[j][i] = firstLayerBiasCosts[i] * digitImage.arrayPixels[j];
}
}
//L-2
AddUpToMiniBatch();
miniBatchCounter++;
if (miniBatchCounter >= miniBatchSize)
{
ApplyCostFunction();
miniBatchCounter = 0;
}
ResetData();//reset activation costs, as they are added up in iterations
}
public static void Run(int durchlaufe, int ausgebenPro)
{
double Fehler = 0;
NeuronalNet Chantall = new NeuronalNet(0, 784, 20, 20, 10);
Chantall.StatischeInitialisierung(true);
Chantall.Etha = 4;
FileStream ifsLabels = new FileStream(@"C:\Users\Gregor\memBrain\train-labels.idx1-ubyte", FileMode.Open); // test labels
FileStream ifsImages = new FileStream(@"C:\Users\Gregor\memBrain\train-images.idx3-ubyte", FileMode.Open); // test images
BinaryReader brLabels = new BinaryReader(ifsLabels);
BinaryReader brImages = new BinaryReader(ifsImages);
int magic1 = brImages.ReadInt32(); // discard
int numImages = brImages.ReadInt32();
int numRows = brImages.ReadInt32();
int numCols = brImages.ReadInt32();
int magic2 = brLabels.ReadInt32();
int numLabels = brLabels.ReadInt32();
byte[][] pixels = new byte[28][];
for (int i = 0; i < pixels.Length; ++i)
{
pixels[i] = new byte[28];
}
for (int i = 1; i < durchlaufe; ++i)
{
if (i % 60000 == 0)
{
ifsImages.Close();
brImages.Close();
ifsLabels.Close();
brLabels.Close();
ifsLabels = new FileStream(@"C:\Users\Gregor\memBrain\train-labels.idx1-ubyte", FileMode.Open); // test labels
ifsImages = new FileStream(@"C:\Users\Gregor\memBrain\train-images.idx3-ubyte", FileMode.Open); // test images
brLabels = new BinaryReader(ifsLabels);
brImages = new BinaryReader(ifsImages);
magic1 = brImages.ReadInt32(); // discard
numImages = brImages.ReadInt32();
numRows = brImages.ReadInt32();
numCols = brImages.ReadInt32();
magic2 = brLabels.ReadInt32();
numLabels = brLabels.ReadInt32();
pixels = new byte[28][];
for (int j = 0; j < pixels.Length; ++j)
{
pixels[j] = new byte[28];
}
}
for (int i1 = 0; i1 < 28; ++i1)
{
for (int j = 0; j < 28; ++j)
{
byte b = brImages.ReadByte();
pixels[i1][j] = b;
}
}
byte lbl = brLabels.ReadByte();
double[] erwrteterOutput = new double[10];
erwrteterOutput[lbl] = 1;
Chantall.InputGeben(ConvertD(pixels));
Chantall.AllesBerechnen();
Chantall.DeepLearning(erwrteterOutput);
Fehler += Chantall.GetFehlerDif(erwrteterOutput);
if (i % ausgebenPro == 0)
{
Fehler /= ausgebenPro;
DigitImage dImage = new DigitImage(pixels, lbl);
Console.WriteLine(dImage.ToString());
ZweiArrayAusgeben(erwrteterOutput, Chantall.GetOutput());
Console.WriteLine("Treffer Quote: {0}", 100 - (Fehler * 10));
Console.WriteLine(Chantall.ToString());
Chantall.Etha = Fehler * 2;
Fehler = 0;
}
}
}
