public bool StopIterations(NeuralNetwork neuralNetwork, ProblemDescriptionModel neuralNetworkTrainModel)
{
var trainSetCount = Convert.ToInt32(Math.Floor(neuralNetworkTrainModel.ValuesCount * .66));
var validationSetCount = Convert.ToInt32(Math.Floor((neuralNetworkTrainModel.ValuesCount - trainSetCount) * .66));
var innerLastOutputDeviation = 0d;
for (var i = trainSetCount; i < trainSetCount + validationSetCount; i++)
{
_feedForward.Compute(neuralNetwork, neuralNetworkTrainModel.GetInputValues(i));
innerLastOutputDeviation = Math.Max(innerLastOutputDeviation, _ouputDeviation.Compute(neuralNetwork, neuralNetworkTrainModel.GetOutputValues(i)));
}
//innerLastOutputDeviation /= validationSetCount;
//check deviation to break training
neuralNetwork.NeuralNetworkError = innerLastOutputDeviation;
//check to stop cycles with this setup
if (Math.Round(stopIterations_lasMaxtOutputDeviation, neuralNetwork.Divisor.ToString().Length)
<= Math.Round(innerLastOutputDeviation, neuralNetwork.Divisor.ToString().Length) || //if important digits stopped correcting, stop iterations
innerLastOutputDeviation < neuralNetworkTrainModel.AcceptedError || //if we are in the accepted error range, stop iterations
Math.Abs(Math.Abs(stopIterations_lasMaxtOutputDeviation) - Math.Abs(innerLastOutputDeviation)) < 1 / (neuralNetwork.Divisor * 1000)) //if the correction is too small stop iterations
{
stopIterations_lasMaxtOutputDeviation = double.MaxValue;
return(true);
}
stopIterations_lasMaxtOutputDeviation = innerLastOutputDeviation;
return(false);
}
public void Test(NeuralNetwork neuralNetwork, ProblemDescriptionModel neuralNetworkTrainModel)
{
var trainSetCount = Convert.ToInt32(Math.Floor(neuralNetworkTrainModel.ValuesCount * .66));
var validationSetCount = Convert.ToInt32(Math.Floor((neuralNetworkTrainModel.ValuesCount - trainSetCount) * .66));
var testSet = Convert.ToInt32(neuralNetworkTrainModel.ValuesCount - trainSetCount - validationSetCount);
var testError = 0d;
for (var i = trainSetCount + validationSetCount; i < trainSetCount + validationSetCount + testSet; i++)
{
_feedForward.Compute(neuralNetwork, neuralNetworkTrainModel.GetInputValues(i));
testError = Math.Max(testError, _ouputDeviation.Compute(neuralNetwork, neuralNetworkTrainModel.GetOutputValues(i)));
}
neuralNetwork.NeuralNetworkError = testError;// / testSet;//update with test error
}
public bool StopTraining(NeuralNetwork neuralNetwork, ProblemDescriptionModel neuralNetworkTrainModel)
{
//iteration < neuralNetwork.HiddenNeurons.Count() ||
if (neuralNetworkTrainModel.AutoAdjuctHiddenLayer && neuralNetwork.NeuralNetworkError > neuralNetworkTrainModel.AcceptedError)
{
//reconfigure for up to ten times the sum of input/output neurons
foreach (var layer in neuralNetwork.HiddenLayers)
{
if (layer.Count() >= (neuralNetwork.InputNeurons.Count() + neuralNetwork.OutputNeurons.Count()) * 10)
{
return(true);
}
}
return(false);
}
return(true);
}
public void Train(NeuralNetwork neuralNetwork, ProblemDescriptionModel neuralNetworkTrainModel)
{
var trainSetCount = Convert.ToInt32(Math.Floor(neuralNetworkTrainModel.ValuesCount * .66));
if (trainSetCount < neuralNetwork.InputNeurons.Count() + neuralNetwork.HiddenLayers.Sum(x => x.Count()) + neuralNetwork.OutputNeurons.Count())
{
trainSetCount = neuralNetworkTrainModel.ValuesCount;
}
//var suffle = Suffle(0, trainSetCount);
//foreach (var i in suffle)
for (var i = 0; i < trainSetCount; i++)
{
_feedForward.Compute(neuralNetwork, neuralNetworkTrainModel.GetInputValues(i));
_backPropagate.Compute(neuralNetwork, neuralNetworkTrainModel.GetOutputValues(i));
}
}
public ProblemDescriptionCreator()
{
neuralNetworkTrainModel = new ProblemDescriptionModel();
}
