/// <summary> Test function to iterate through all learning rates and momentums with different numbers of neurons and layers. </summary>
private void RateTestBtn_Click(object sender, EventArgs e)
{
//Setup paths and lists.
string dataFile = "";
string path = "";
//Normalise file.
List <string> outputTitles = GetFile(false, ref path, ref dataFile, (int)(outputsUpDown.Value));
if (outputTitles == null)
{
return;
}
//False when percentage split, true when cross validation.
bool validation = true;
//Save network to file.
if (!Directory.Exists(path + "tests"))
{
Directory.CreateDirectory(path + "tests");
}
string header = path + "tests" + @"\Results" + dataFile;
using (StreamWriter sw = new StreamWriter(header))
{
sw.WriteLine(String.Format("Error,Network Type,Layers,Neurons,Learning Rate,Momentum"));
}
// initialize input and output values.
Data info = new Data(); info = info.ReturnInfo(path + @"normal\" + dataFile.Replace(".csv", "Normal.csv"), outputTitles, sampleBar.Value, validation);
if (info == null)
{
return;
}
//Setup network.
List <NeuralNetwork> networkList = new List <NeuralNetwork>
{
new EncogNeuralNetwork(),
new EncogDeepNeuralNetwork(),
new AccordNeuralNetwork(),
new AccordDeepNeuralNetwork()
};
List <BestNetwork> bestList = new List <BestNetwork>();
foreach (NeuralNetwork network in networkList)
{
bestList.Add(network.CalculateBest(info, header));
}
BestNetwork best = bestList[0];
int index = 0;
int count = 0;
foreach (BestNetwork b in bestList)
{
double error = b.Error;
if (error < best.Error)
{
best = b;
index = count;
}
count++;
}
output.Text += "The best netork is: " + networkList[index].GetType().ToString().Replace("ENP1.", "")
+ "\nLayers: " + best.Layers + "\nNeurons: " + best.Neurons + "\nLearning Rate: " + best.LearningRate
+ "\nMomentum: " + best.Momentum + "\nInaccuracy: " + best.Error;
}
private BestNetwork BestLearningRate(Data info, string path, int layers, int neurons)
{
float lr = 0.1F;
bool complete = false;
bool firstPass = true;
bool increase = true;
BestNetwork best = BestMomentum(info, path, layers, neurons, lr);
while (!complete)
{
if (firstPass)
{
lr = 0.5F;
}
else if (increase)
{
lr += 0.1F;
}
else
{
lr -= 0.1F;
}
if (Math.Round(lr, 1) == 0.1F || Math.Round(lr, 1) == 1.0F)
{
complete = true;
}
BestNetwork newest = BestMomentum(info, path, layers, neurons, lr);
if (firstPass)
{
if (newest.Error < best.Error)
{
best = newest;
increase = true;
}
else if (newest.Error == best.Error)
{
complete = true;
}
else
{
increase = false;
}
firstPass = false;
}
else
{
if (newest.Error < best.Error)
{
best = newest;
}
else
{
complete = true;
}
}
}
return(best);
}
private BestNetwork BestMomentum(Data info, string path, int layers, int neurons, float lr)
{
bool complete = false;
bool firstPass = true;
bool increase = true;
BestNetwork best = new BestNetwork
{
Momentum = 0.1F,
LearningRate = lr,
Neurons = neurons,
Layers = layers
};
best.Error = Train(info, lr, best.Momentum);
while (!complete)
{
if (firstPass)
{
best.Momentum = 0.5F;
}
else if (increase)
{
best.Momentum += 0.1F;
}
else
{
best.Momentum -= 0.1F;
}
if (Math.Round(best.Momentum, 1) == 0.1F || Math.Round(best.Momentum, 1) == 1.0F)
{
complete = true;
}
double newest = Train(info, lr, best.Momentum);
if (firstPass)
{
if (newest < best.Error)
{
best.Error = newest;
increase = true;
}
else if (newest == best.Error)
{
complete = true;
}
else
{
increase = false;
}
firstPass = false;
}
else
{
if (newest < best.Error)
{
best.Error = newest;
}
else
{
complete = true;
}
}
}
using (StreamWriter sw = new StreamWriter(path, true))
{
sw.WriteLine(String.Format(
"{0},{1},{2},{3},{4},{5}", best.Error, GetType().ToString().Replace("ENP1.", ""), layers, neurons, Math.Round(lr, 1), Math.Round(best.Momentum, 1))); //"ENP1" must change to reflect solution name (name.) if ever changed.
}
return(best);
}
public BestNetwork CalculateBest(Data info, string path)
{
int layers = 1;
bool complete = false;
bool firstPass = true;
bool increase = true;
BestNetwork best = BestNeurons(info, path, layers);;
if (GetType().ToString().Replace("ENP1.", "") == "EncogDeepNeuralNetwork")
{
return(best);
}
while (!complete)
{
if (firstPass)
{
layers = 2;
}
else if (increase)
{
layers += 1;
}
else
{
layers -= 1;
}
if (layers == 1 || layers == 5)
{
complete = true;
}
BestNetwork newest = BestNeurons(info, path, layers);
if (firstPass)
{
if (newest.Error < best.Error)
{
best = newest;
increase = true;
}
else if (newest.Error == best.Error)
{
complete = true;
}
else
{
increase = false;
}
firstPass = false;
}
else
{
if (newest.Error < best.Error)
{
best = newest;
}
else
{
complete = true;
}
}
}
return(best);
}
private BestNetwork BestNeurons(Data info, string path, int layers)
{
int neurons = 5;
bool complete = false;
bool firstPass = true;
bool increase = true;
Create(info.InputNumber, layers, neurons, info.OutputNumber);
BestNetwork best = BestLearningRate(info, path, layers, neurons);
while (!complete)
{
if (firstPass)
{
neurons = 25;
}
else if (increase)
{
neurons += 5;
}
else
{
neurons -= 5;
}
if (neurons == 5 || neurons == 50)
{
complete = true;
}
Create(info.InputNumber, layers, neurons, info.OutputNumber);
BestNetwork newest = BestLearningRate(info, path, layers, neurons);
if (firstPass)
{
if (newest.Error < best.Error)
{
best = newest;
increase = true;
}
else if (newest.Error == best.Error)
{
complete = true;
}
else
{
increase = false;
}
firstPass = false;
}
else
{
if (newest.Error < best.Error)
{
best = newest;
}
else
{
complete = true;
}
}
}
return(best);
}
