private void Start()
{
_process.Start();
int count = 0;
while (_isRunning)
{
using (var reader = _process.StandardOutput)
{
var files = Directory.GetFiles($@"{AppDomain.CurrentDomain.BaseDirectory}script\output");
while (!reader.EndOfStream)
{
ConsolePrint?.Invoke(reader.ReadLine(), null);
}
if (count != files.Length)
{
count = files.Length;
EpochResult?.Invoke(this, null);
}
}
Thread.Sleep(TimeSpan.FromMinutes(1));
}
}
public void OnEpochFinished(EpochResult result)
{
_loss = 0;
_accuracy = 0;
_elapsed = 0;
}
private EpochResult Backpropagate(IReadOnlyCollection <IDataWindow> trainingData)
{
var output = new EpochResult();
var correctClassifications = 0;
var costs = new List <float>();
var n1Costs = new List <float>();
var n2Costs = new List <float>();
var n1S = new List <float>();
var n2S = new List <float>();
//create adjustment neurons
for (var layer = 1; layer < Options.LayerStructure.Length; layer++)
{
for (var neuron = 0; neuron < Options.LayerStructure[layer]; neuron++)
{
_proposedNeurons[layer, neuron] =
new ProposedNeuron(NeuronLayers[layer, neuron].Weights.Length, trainingData.Count);
}
}
// for window
foreach (var window in trainingData)
{
//get activation
CreateActivationArray(window);
FillActivationArray();
var n1 = _exampleActivations[Options.LayerStructure.Length - 1, 0];
var diff = n1 - window.OutputLayer[0];
if (Math.Abs(diff) > 0.5F)
{
correctClassifications++;
}
n1S.Add(_exampleActivations[Options.LayerStructure.Length - 1, 0]);
n2S.Add(_exampleActivations[Options.LayerStructure.Length - 1, 1]);
//calculate offset
// var offset = _exampleActivations[_exampleActivations.GetLength(0) - 1, 0] - window.OutputLayer[0];
// offset *= offset < 0 ? -1 : 1; //if less than 0 times with -1
// successes.Add(offset < 0.5F);
//get cost
ResetExampleCostArray();
FillExampleCostArray(window.OutputLayer);
costs.Add(_exampleCosts[Options.LayerStructure.Length - 1, 0] + _exampleCosts[Options.LayerStructure.Length - 1, 1]);
n1Costs.Add(_exampleCosts[Options.LayerStructure.Length - 1, 0]);
n2Costs.Add(_exampleCosts[Options.LayerStructure.Length - 1, 1]);
// n1S.Add(_exampleActivations[Options.LayerStructure.Length-1, 0]);
// n2S.Add(_exampleActivations[Options.LayerStructure.Length-1, 1]);
//for each neuron
for (var layer = 1; layer < Options.LayerStructure.Length; layer++)
{
for (var neuron = 0; neuron < Options.LayerStructure[layer]; neuron++)
{
//get adjustments
var neuronCost = _exampleCosts[layer, neuron];
var biasAdjustment = NeuronLayers[layer, neuron]
.GetBiasAdjust(_exampleActivations, neuronCost);
var weightAdjustment = NeuronLayers[layer, neuron]
.GetWeightAdjust(_exampleActivations, neuronCost);
//set adjustment
_proposedNeurons[layer, neuron].AddBiasProposal(biasAdjustment);
_proposedNeurons[layer, neuron].AddWeightProposal(weightAdjustment);
}
}
}
//after for
//make adjustment
for (var layer = 1; layer < Options.LayerStructure.Length; layer++)
{
for (var neuron = 0; neuron < Options.LayerStructure[layer]; neuron++)
{
NeuronLayers[layer, neuron].Adjust(_proposedNeurons[layer, neuron], Options.LearningRate);
}
}
//var accuracy = (float)successes.Count(x => x) / trainingData.Count;
//output.AverageOffset = accuracy;
//output.AverageOffset = _exampleCosts[Options.LayerStructure.Length - 1, 0];
var totalCost = costs.Sum();
output.AverageOffset = totalCost / costs.Count;
output.AverageCost = new NeuronPairing(n1Costs.Average(), n2Costs.Average());
output.Accuracy = (float)correctClassifications / trainingData.Count;
output.AverageOutputLayerActivation = new NeuronPairing(n1S.Average(), n2S.Average());
return(output);
}
protected void RaiseEpochFinishedEvent(EpochResult result)
{
OnEpochFinished?.Invoke(result);
}
