private List <InputOutputPair> ReadFromDirectory(string directory)
{
List <InputOutputPair> result = new List <InputOutputPair>();
string[] directories = Directory.GetDirectories(directory);
Task[] tasks = new Task[directories.Length];
for (int i = 0; i < directories.Length; i++)
{
int newI = 0 + i;
tasks[i] = Task.Run(() =>
{
foreach (string input in Directory.GetFiles(directories[newI]))
{
InputOutputPair pair = new InputOutputPair(input, directories[newI]);
Monitor.Enter(result);
result.Add(pair);
Monitor.Exit(result);
}
});
}
Task.WaitAll(tasks);
return(result);
}
public void TrainMnist(InputOutputPair pair, double learningRate)
{
FilteredImage input;
if (ColorScheme == "rgb")
{
input = ImageProcessing.GetNormalizedFilteredImage(new Bitmap(pair.Input));
}
else
{
//input = ImageProcessing.GetNormalizedGrayscaleFilteredImage(new Bitmap(pair.Input));
input = ImageProcessing.GetNormalizedMnist(new Bitmap(pair.Input));
}
double[] actualOutput = RecogniseImage(input);
Backpropagate(actualOutput, pair.Output, pair.OutputChar, learningRate);
if (Math.Abs(((ConvolutionalLayer)NetworkLayers[0]).Filters[0].Kernels[0].ElementSum) > 1000)
{
throw new Exception("Weights over 1000");
}
if (Double.IsNaN(((ConvolutionalLayer)NetworkLayers[0]).Filters[0].Kernels[0].Values[0, 0]))
{
throw new Exception("Nan error");
}
}
private List <InputOutputPair> Shuffle(List <InputOutputPair> set)
{
for (int i = 0; i < set.Count; i++)
{
int x = GlobalRandom.GetRandomInt(i, set.Count);
InputOutputPair aux = set[x];
set[x] = set[i];
set[i] = aux;
}
return(set);
}
public EvaluationMetrics EvaluateParallel(List <InputOutputPair> set)
{
int[,] matrix = new int[26, 26];
int N = 26;
int total = set.Count;
Task[] tasks = new Task[N];
int chunkSize = (N + total - 1) / N;
for (int t = 0; t < N; t++)
{
int start = t * chunkSize;
int end = Math.Min(start + chunkSize, total);
tasks[t] = Task.Run(() =>
{
for (int i = start; i < end; i++)
{
InputOutputPair pair = set[i];
FilteredImage input;
if (ColorScheme == "rgb")
{
input = ImageProcessing.GetNormalizedFilteredImage(new Bitmap(pair.Input));
}
else
{
input = ImageProcessing.GetNormalizedGrayscaleFilteredImage(new Bitmap(pair.Input));
}
double[] actualOutput = RecogniseImage(input);
int expected = pair.OutputChar - 65;
int obtained = GetReconisedChar(actualOutput) - 65;
Monitor.Enter(matrix);
matrix[expected, obtained]++;
Monitor.Exit(matrix);
}
});
}
Task.WaitAll(tasks);
return(new EvaluationMetrics(matrix, total));
}
public EvaluationMetrics EvaluateMetrics(List <InputOutputPair> set)
{
//double error = 0;
//int correct = 0, total = 0;
int[,] matrix = new int[26, 26];
int total = 0;
int N = set.Count;
for (int i = 0; i < N; i++)
{
InputOutputPair pair = set[i];
FilteredImage input;
if (ColorScheme == "rgb")
{
input = ImageProcessing.GetNormalizedFilteredImage(new Bitmap(pair.Input));
}
else
{
input = ImageProcessing.GetNormalizedGrayscaleFilteredImage(new Bitmap(pair.Input));
}
double[] actualOutput = RecogniseImage(input);
//error += LeastSquaredError(actualOutput, pair.Output);
//total++;
//if (GetReconisedChar(actualOutput) == pair.OutputChar)
//{
// correct++;
//}
int expected = pair.OutputChar - 65;
int obtained = GetReconisedChar(actualOutput) - 65;
matrix[expected, obtained]++;
total++;
}
return(new EvaluationMetrics(matrix, total));
}
public double Evaluate(List <InputOutputPair> set)
{
double error = 0;
int correct = 0, total = 0;
int N = set.Count;
for (int i = 0; i < N; i++)
{
InputOutputPair pair = set[i];
FilteredImage input;
if (ColorScheme == "rgb")
{
input = ImageProcessing.GetNormalizedFilteredImage(new Bitmap(pair.Input));
}
else
{
input = ImageProcessing.GetNormalizedGrayscaleFilteredImage(new Bitmap(pair.Input));
}
double[] actualOutput = RecogniseImage(input);
error += LeastSquaredError(actualOutput, pair.Output);
total++;
if (GetReconisedChar(actualOutput) == pair.OutputChar)
{
correct++;
}
int expected = pair.OutputChar - 65;
int obtained = GetReconisedChar(actualOutput) - 65;
total++;
}
return((double)correct / total);
}