public static void Main()
{
const int numInHiddenLayer = 500;
const int numOfOutputs = 10;
const double normalisation = 255.0d;
var inputFileReader = new InputFileReader();
IList<Tuple<int, IEnumerable<double>>> csvInputs = inputFileReader.ReadTrainingInputFile(@"C:\Users\Pavlos\Desktop\training.csv", normalisation);
int validationFraction = csvInputs.Count / 10; // use all but ten percent for training, hold the rest back for validation
var trainingInputs = csvInputs.Skip(validationFraction).ToList();
var validationInputs = csvInputs.Take(validationFraction).ToList();
// create inputs and the three layers
List<SensoryInput> sensoryInputs = trainingInputs[0].Item2.Select(i => new SensoryInput()).ToList();
List<INeuron> inputLayer = CreateLayer(sensoryInputs.Count, sensoryInputs.Cast<IInput>().ToList());
List<INeuron> hiddenLayer = CreateLayer(numInHiddenLayer, inputLayer.Cast<IInput>().ToList());
List<INeuron> outputLayer = CreateLayer(numOfOutputs, hiddenLayer.Cast<IInput>().ToList());
double previousGlobalError = double.MaxValue;
double globalErrorDelta;
// training:
do
{
foreach (var specimen in trainingInputs)
{
UpdateNetwork(specimen.Item2.ToList(), sensoryInputs, inputLayer, hiddenLayer, outputLayer);
// train output layer
for (int k = 0; k < outputLayer.Count; k++)
{
double desired = k == specimen.Item1 ? 1.0d : 0.0d;
double output = outputLayer[k].GetValue();
double error = desired - output;
outputLayer[k].Train(error);
}
// train hidden layer, then train input layer
BackPropagate(hiddenLayer, outputLayer);
BackPropagate(inputLayer, hiddenLayer);
}
// calculate global error using the validation set that was excluded from training:
double globalError = 0.0d;
foreach (var specimen in validationInputs)
{
UpdateNetwork(specimen.Item2.ToList(), sensoryInputs, inputLayer, hiddenLayer, outputLayer);
for (int i = 0; i < outputLayer.Count; i++)
{
double desired = i == specimen.Item1 ? 1.0d : 0.0d;
globalError += Math.Abs(desired - outputLayer[i].GetValue());
}
}
globalErrorDelta = Math.Abs(previousGlobalError - globalError);
previousGlobalError = globalError;
Console.WriteLine("Global error: {0}", globalError);
} while (globalErrorDelta > 1000.0d); // train until global error begins to level off
// Run on real testing data and write output to console:
var testingInputs = inputFileReader.ReadTestingInputFile(@"C:\Users\Pavlos\Desktop\testing.csv", normalisation);
foreach (var specimen in testingInputs)
{
UpdateNetwork(specimen.ToList(), sensoryInputs, inputLayer, hiddenLayer, outputLayer);
int mostLikelyAnswer = GetMostLikelyAnswer(outputLayer);
Console.WriteLine(mostLikelyAnswer);
}
}
public static void Main()
{
const int numInHiddenLayer = 500;
const int numOfOutputs = 10;
const double normalisation = 255.0d;
var inputFileReader = new InputFileReader();
IList <Tuple <int, IEnumerable <double> > > csvInputs = inputFileReader.ReadTrainingInputFile(@"C:\Users\Pavlos\Desktop\training.csv", normalisation);
int validationFraction = csvInputs.Count / 10; // use all but ten percent for training, hold the rest back for validation
var trainingInputs = csvInputs.Skip(validationFraction).ToList();
var validationInputs = csvInputs.Take(validationFraction).ToList();
// create inputs and the three layers
List <SensoryInput> sensoryInputs = trainingInputs[0].Item2.Select(i => new SensoryInput()).ToList();
List <INeuron> inputLayer = CreateLayer(sensoryInputs.Count, sensoryInputs.Cast <IInput>().ToList());
List <INeuron> hiddenLayer = CreateLayer(numInHiddenLayer, inputLayer.Cast <IInput>().ToList());
List <INeuron> outputLayer = CreateLayer(numOfOutputs, hiddenLayer.Cast <IInput>().ToList());
double previousGlobalError = double.MaxValue;
double globalErrorDelta;
// training:
do
{
foreach (var specimen in trainingInputs)
{
UpdateNetwork(specimen.Item2.ToList(), sensoryInputs, inputLayer, hiddenLayer, outputLayer);
// train output layer
for (int k = 0; k < outputLayer.Count; k++)
{
double desired = k == specimen.Item1 ? 1.0d : 0.0d;
double output = outputLayer[k].GetValue();
double error = desired - output;
outputLayer[k].Train(error);
}
// train hidden layer, then train input layer
BackPropagate(hiddenLayer, outputLayer);
BackPropagate(inputLayer, hiddenLayer);
}
// calculate global error using the validation set that was excluded from training:
double globalError = 0.0d;
foreach (var specimen in validationInputs)
{
UpdateNetwork(specimen.Item2.ToList(), sensoryInputs, inputLayer, hiddenLayer, outputLayer);
for (int i = 0; i < outputLayer.Count; i++)
{
double desired = i == specimen.Item1 ? 1.0d : 0.0d;
globalError += Math.Abs(desired - outputLayer[i].GetValue());
}
}
globalErrorDelta = Math.Abs(previousGlobalError - globalError);
previousGlobalError = globalError;
Console.WriteLine("Global error: {0}", globalError);
} while (globalErrorDelta > 1000.0d); // train until global error begins to level off
// Run on real testing data and write output to console:
var testingInputs = inputFileReader.ReadTestingInputFile(@"C:\Users\Pavlos\Desktop\testing.csv", normalisation);
foreach (var specimen in testingInputs)
{
UpdateNetwork(specimen.ToList(), sensoryInputs, inputLayer, hiddenLayer, outputLayer);
int mostLikelyAnswer = GetMostLikelyAnswer(outputLayer);
Console.WriteLine(mostLikelyAnswer);
}
}
