public static void Teach(object A)
{
Between B = (Between)A;
do
{
int curAnswer = MathAI.GetRandom(0, nElements - 1);
//Console.WriteLine(curAnswer + " " + B.nFiles[curAnswer]);
int curDigitVar = MathAI.GetRandom(0, B.nFiles[curAnswer] - 1);
Digit curDigit = B.Digits[curAnswer, curDigitVar];
FirstLayer Layer1 = new FirstLayer(curDigit, dimHiddenLayer, B.Weights1);
SecondLayer Layer2 = new SecondLayer(Layer1, nElements, B.Weights2, curAnswer);
lock (locker)
{
iteration++;
Epoch = iteration / B.nFilesFull + 1;
if ((Convert.ToDouble(iteration) - Convert.ToDouble((Epoch - 1) * B.nFilesFull)) != 0)
{
testdel = (Convert.ToDouble(iteration) - Convert.ToDouble((Epoch - 1) * B.nFilesFull));
}
if (iteration % 2000 == 0)
{
Console.WriteLine("Epoch #" + Epoch + " | Theard #" + Thread.CurrentThread.Name);
Console.WriteLine("Answer = " + (curAnswer));
Console.WriteLine("Full Error = " + (Weight.fError));
Console.WriteLine("Accuration = " + Layer2.ResultAccuracy * 100 + "%");
Console.WriteLine("Epoch Accuration = " + averageEpochAccuracy / (testdel) * 100 + "%");
}
Weight.Correct(Layer1, Layer2, B.Weights1, B.Weights2, curAnswer, nElements, dimHiddenLayer, LearningRate);
if (iteration % B.nFilesFull == 0)
{
averageEpochAccuracy = 0;
}
averageEpochAccuracy += Layer2.ResultAccuracy;
}
} while (averageEpochAccuracy / (testdel) < E || (Convert.ToDouble(iteration) - Convert.ToDouble((Epoch - 1) * B.nFilesFull)) < B.nFilesFull * 0.9);
lock (locker)
{
if (!ThreadSave)
{
ThreadSave = true;
B.Weights1.Save(MainFolder + @"\weights1.txt");
B.Weights2.Save(MainFolder + @"\weights2.txt");
}
}
}
public SecondLayer(FirstLayer Layer, int nElements, Weight Weights)
{
Result = new double[nElements];
ResultLength = nElements;
for (int x = 0; x < nElements; x++)
{
if (Result[x] != 0)
{
Result[x] = 0;
}
for (int y = 0; y < Layer.HiddenLength; y++)
{
Result[x] += Layer.Hidden[y] * Weights.Body[y, x];
}
Result[x] = MathAI.Sigmoid(Result[x]);
}
}
public SecondLayer(FirstLayer Layer, int nElements, Weight Weights, int Answer)
{
Result = new double[nElements];
ResultLength = nElements;
for (int x = 0; x < nElements; x++)
{
if (Result[x] != 0)
{
Result[x] = 0;
}
for (int y = 0; y < Layer.HiddenLength; y++)
{
Result[x] += Layer.Hidden[y] * Weights.Body[y, x];
}
Result[x] = MathAI.Sigmoid(Result[x]);
}
double sumResult = 0;
for (int i = 0; i < nElements; i++)
{
sumResult += Result[i];
}
ResultAccuracy = Result[Answer] / sumResult;
}
public static void Main(string[] args)
{
string firstquestion;
string TrainingFolder; //Основная папка с изображениями для обучения
if (Directory.Exists(@"\mnist\"))
{
MainFolder = @"";
}
else
{
MainFolder = @"D:";
}
Console.WriteLine("DataSet[1] Or Check[2]?");
firstquestion = Console.ReadLine();
if (firstquestion[0] == '1')
{
do
{
Console.WriteLine(@"Address of images (Folder): " + @MainFolder + @"\mnist\mnist_png\training");
TrainingFolder = @MainFolder + @"\mnist\mnist_png\training";
Console.WriteLine(@"Starting...");
int[] nFiles = new int[nElements];
List <string>[] filesDigit = new List <string> [nElements];
int nFilesFull = 0;
for (int i = 0; i < nElements; i++)
{
nFiles[i] = HowManyFiles(@TrainingFolder + @"\" + i + @"\");
Console.WriteLine(@"In " + @TrainingFolder + @"\" + i + @"\ " + nFiles[i] + " files.");
}
for (int i = 0; i < 10; i++)
{
nFilesFull += nFiles[i];
}
Console.WriteLine(@"Total: " + nFilesFull + " files.");
for (int i = 0; i < nElements; i++)
{
filesDigit[i] = Directory.GetFiles(@TrainingFolder + @"\" + i + @"\", "*.png").ToList <string>();
}
//Заносим все файлы в массив
Digit[,] Digits = new Digit[nElements, MathAI.Max(nFiles)];
for (int i = 0; i < Digits.GetLength(0); i++)
{
for (int k = 0; k < nFiles[i]; k++)
{
Digits[i, k] = new Digit(filesDigit[i].ElementAt(k));
}
}
filesDigit = null;
Console.WriteLine("Digits downloaded.");
Weight Weights1 = new Weight(sourcesize * sourcesize, dimHiddenLayer);
Weight Weights2 = new Weight(dimHiddenLayer, nElements);
//StreamWriter graf = new StreamWriter(@"D:\graf.txt", false);
Console.WriteLine("Weights created.");
Between A = new Between(Digits, nFilesFull, nFiles, Weights1, Weights2);
Thread t1 = new Thread(new ParameterizedThreadStart(Teach));
Thread t2 = new Thread(new ParameterizedThreadStart(Teach));
Thread t3 = new Thread(new ParameterizedThreadStart(Teach));
Thread t4 = new Thread(new ParameterizedThreadStart(Teach));
t1.Name = "1";
t2.Name = "2";
t3.Name = "3";
t4.Name = "4";
Console.WriteLine("Threads starting...");
t1.Start(A);
t2.Start(A);
t3.Start(A);
t4.Start(A);
t1.Join();
t2.Join();
t3.Join();
t4.Join();
Console.WriteLine("Teaching end.");
Console.WriteLine();
Digits = null;
Console.WriteLine(@"Address of test images (Folder):" + MainFolder + @"\mnist\mnist_png\testing");
string TestingFolder = MainFolder + @"\mnist\mnist_png\testing";
Console.WriteLine(@"Starting testing...");
int[] nFilesT = new int[nElements];
for (int i = 0; i < nElements; i++)
{
nFilesT[i] = HowManyFiles(@TestingFolder + @"\" + i + @"\");
Console.WriteLine(@"In " + @TestingFolder + @"\" + i + @"\ " + nFilesT[i] + " files.");
}
int nFilesTFull = 0;
for (int i = 0; i < 10; i++)
{
nFilesTFull += nFilesT[i];
}
Console.WriteLine(@"Total: " + nFilesTFull + " files.");
List <string>[] filesTest = new List <string> [nElements];
for (int i = 0; i < nElements; i++)
{
filesTest[i] = Directory.GetFiles(@TestingFolder + @"\" + i + @"\", "*.png").ToList <string>();
}
Digit[,] DigitsTest = new Digit[nElements, MathAI.Max(nFilesT)];
for (int i = 0; i < DigitsTest.GetLength(0); i++)
{
for (int k = 0; k < nFilesT[i]; k++)
{
DigitsTest[i, k] = new Digit(filesTest[i].ElementAt(k));
}
}
//Bitmap imageT;
double[] aResult = new double[nFilesTFull];
double averageResult = 0;
for (int iteration2 = 1; iteration2 <= nFilesTFull; iteration2++)
{
int curAnswer = MathAI.GetRandom(0, nElements - 1);
int curDigitVar = MathAI.GetRandom(0, nFilesT[curAnswer] - 1);
Digit curDigit = DigitsTest[curAnswer, curDigitVar];
FirstLayer Layer1 = new FirstLayer(curDigit, dimHiddenLayer, Weights1);
SecondLayer Layer2 = new SecondLayer(Layer1, nElements, Weights2, curAnswer);
bool GoodAnswer;
if (MathAI.MaxI(Layer2.Result) == curAnswer)
{
GoodAnswer = true;
}
else
{
GoodAnswer = false;
}
Console.WriteLine("Iteration #" + iteration2 + " | " + GoodAnswer);
Console.WriteLine("Answer = " + curAnswer);
Console.WriteLine("Output = " + MathAI.MaxI(Layer2.Result));
Console.WriteLine("Average = " + Layer2.Result[curAnswer]);
for (int i = 0; i < Layer2.ResultLength; i++)
{
Console.WriteLine(i + ": " + (Layer2.Result[i]));
}
//for (int i = 0; i < Layer2.ResultLength; i++) fullResultWork += Result[i];
aResult[iteration2 - 1] = Layer2.Result[curAnswer];
}
for (int i = 0; i < nFilesTFull; i++)
{
averageResult += aResult[i];
}
averageResult /= nFilesTFull;
Console.WriteLine("-------------\nTotal Average = " + averageResult);
Console.ReadKey();
DigitsTest = null;
break;
} while (TrainingFolder != @"\q");
}
else
{
string FileAddress;
do
{
Console.WriteLine("Enter address of image: ");
FileAddress = Console.ReadLine();
Weight Weights1 = new Weight(@"D:\weights1.txt");
Weight Weights2 = new Weight(@"D:\weights2.txt");
Digit curDigit = new Digit(@FileAddress);
FirstLayer Layer1 = new FirstLayer(curDigit, dimHiddenLayer, Weights1);
SecondLayer Layer2 = new SecondLayer(Layer1, nElements, Weights2);
Console.WriteLine("Answer: " + MathAI.MaxI(Layer2.Result));
Console.WriteLine("Detail: ");
for (int i = 0; i < Layer2.ResultLength; i++)
{
Console.WriteLine(i + ": " + (Layer2.Result[i]));
}
Console.ReadKey();
} while (FileAddress != @"\q");
}
}
public static void Correct(FirstLayer Layer1, SecondLayer Layer2, Weight Weights1, Weight Weights2, int curAnswer, int nElements, int dimHiddenLayer, double LearningRate)
{
double[] lError = new double[nElements];
double[] deltaWeights = new double[nElements];
double fullResultWork = 0;
for (int i = 0; i < Layer2.ResultLength; i++)
{
fullResultWork += Layer2.Result[i];
}
for (int i = 0; i < nElements; i++)
{
if (i != curAnswer)
{
lError[i] = Layer2.Result[i];
}
else
{
lError[i] = (Layer2.Result[i] - 1);
}
}
for (int i = 0; i < nElements; i++)
{
deltaWeights[i] = lError[i] * MathAI.derSigmoid2(Layer2.Result[i]);
}
double[] lError2 = new double[dimHiddenLayer];
double[] deltaWeights2 = new double[dimHiddenLayer];
for (int i = 0; i < dimHiddenLayer; i++)
{
lError2[i] = 0;
for (int j = 0; j < nElements; j++)
{
lError2[i] += Weights2.Body[i, j] * deltaWeights[j];
}
}
for (int i = 0; i < dimHiddenLayer; i++)
{
deltaWeights2[i] = lError2[i] * MathAI.derSigmoid2(Layer1.Hidden[i]);
}
for (int x = 0; x < Weights2.Width; x++)
{
for (int y = 0; y < Weights2.Height; y++)
{
//Weights2.Body[x, y] -= Layer1.Hidden[x] * deltaWeights[y] * LearningRate;
Weights2.Body[x, y] -= Layer1.Hidden[x] * deltaWeights[y] * LearningRate;
}
}
for (int x = 0; x < Weights1.Width; x++)
{
for (int y = 0; y < Weights1.Height; y++)
{
Weights1.Body[x, y] -= Layer1.curLayer[x] * deltaWeights2[y] * LearningRate;
}
}
//aResult[(k - 1) * nButch + Epoch - 1] = Result[answer] / fullResultWork;
fError = 0;
for (int i = 0; i < lError.GetLength(0); i++)
{
fError += lError[i] * lError[i];
}
fError /= 2;
}