static DataGroup LoadTrainingData()
{
var directoryPath = Directory.GetCurrentDirectory();
var file = new FileInfo($"{directoryPath}\\Resources\\number-image-train.csv");
if (!file.Exists)
{
throw new Exception($"File Doesn't Exist: {file.FullName}");
}
var rows = new List <double[]>();
var labels = new List <double>();
using (var reader = new StreamReader(file.FullName))
{
var count = 0;
while (!reader.EndOfStream)
{
count++;
var line = reader.ReadLine();
if (count == 1)
{
continue; // Skip first line that has column names
}
var data = line.Split(",");
var rowData = new double[data.Length];
string label = data[0];
double convertedLabel;
if (!double.TryParse(label, out convertedLabel))
{
convertedLabel = 0.0d;
}
labels.Add(convertedLabel);
for (int i = 1; i < data.Length; i++)
{
string dataPoint = data[i];
double convertedDataPoint;
if (!double.TryParse(dataPoint, out convertedDataPoint))
{
convertedDataPoint = 0.0d;
}
rowData[i] = convertedDataPoint;
}
rows.Add(rowData);
count++;
}
}
var labelList = labels.Distinct().OrderBy(l => l).ToArray();
//var rowLabels = new List<double[]>(labelList.Length);
//foreach (var label in labels)
//{
// var expectedOutput = Enumerable.Repeat(-1.0d, labelList.Length).ToArray();
// for (int i = 0; i < labelList.Length; i++)
// {
// if (labelList[i] == label)
// {
// expectedOutput[i] = label;
// }
// }
// rowLabels.Add(expectedOutput);
//}
//var possibleNeuronOutputs = new List<double[]>(labelList.Length);
//for (int i = 0; i < labelList.Count(); i++)
//{
// var result = Enumerable.Repeat(-1.0d, labelList.Count()).ToArray();
// for (int j = 0; j < labelList.Count(); j++)
// {
// if (j == i)
// {
// result[j] = labelList[j];
// }
// }
// possibleNeuronOutputs.Add(result);
//}
var dataGroup = new DataGroup(labelList, rows.ToArray(), labels.ToArray());
return(dataGroup);
}
public void Train(DataGroup trainGroup)
{
var results = new List <DetectionResult>();
// Iterate through epochs and train network, keeping track of results
for (int e = 0; e < Epochs; e++)
{
for (int i = 0; i < trainGroup.Values.Count; i++)
{
FeedForward(trainGroup.Values[i]);
BackPropagate(trainGroup.LabelList, trainGroup.ValueLabels[i]);
// var label = trainGroup.ValueLabels[i];
var outputLayer = Layers.Last();
var outputResults = outputLayer.Neurons.SelectMany(n => n.OutputDendrites.Select(d => d.Signal.Value)).ToArray();
// var isCorrect = label.Equals(outputResult);
results.Add(new DetectionResult {
Label = trainGroup.ValueLabels[i], OutputValues = outputResults
});
if (i % 100 == 0)
{
var displayList = new List <string>();
for (int l = 0; l < trainGroup.LabelList.Length; l++)
{
var label = trainGroup.LabelList[l];
var resultsList = results.Where(r => r.Label == label).ToList();
if (resultsList.Count > 0)
{
int correctCount = 0;
foreach (var result in resultsList)
{
var maxValue = result.OutputValues.Max();
var maxValueIndex = Array.IndexOf(result.OutputValues, maxValue);
if (l == maxValueIndex)
{
correctCount++;
}
}
var line = $"Label {label} Accuracy: {(correctCount / resultsList.Count) * 100}%";
displayList.Add(line);
}
else
{
displayList.Add($"Label {label} Accuracy: NOT YET CALCULATED");
}
}
Console.WriteLine();
Console.WriteLine(String.Join("\n", displayList));
}
//if (i % 100 == 0)
//{
// Console.WriteLine();
// foreach (var result in accuracyResults.OrderBy(r => r.RelevantKey).ToList())
// {
// Console.WriteLine($"{result.RelevantKey}\tAttempts: {result.Count}\tCorrect: {result.Correct}\tIncorrect: {result.Incorrect}\tAccuracy: {(result.Count == 0 ? 0 : (result.Correct / result.Count))}");
// }
//}
}
}
}