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);
}
/// <summary> Returns string to display current neural network (used when correct output is know). </summary>
public string Display(double[][] answers, EncogAnalyst analyst, Data info, List <string> titles, string path)
{
string item = "";
List <int> count = new List <int>();
for (int c = 0; c < titles.Count; c++)
{
count.Add(0);
}
//Get length of CSV, Inputs and Outputs.
using (var reader = new StreamReader(path))
{
var line = reader.ReadLine();
var values = line.Split(',');
for (int v = 0; v < values.Length; v++)
{
for (int k = 0; k < titles.Count; k++)
{
if (values[v].Contains(titles[k]))
{
count[k]++;
}
}
}
}
for (int i = 0; i < info.OutputDataSample.Length; i++)
{
/*for (int j = 0; j < info.InputDataSample[0].Length; j++)
* {
* double input = Math.Round(analyst.Script.Normalize.NormalizedFields[count].DeNormalize(info.InputDataSample[i][j]), 2);
*
* if (input > 0)
* {
* item += String.Format(
* "Input {0}: [{1}] ", j + 1,
* input
* );
*
* count++;
* }
* }*/
item += String.Format(
"Prediction Number: {0}\n\n",
i + 1
);
for (int j = 0; j < count.Count; j++)
{
string prediction = "";
string outpt = "";
if (count[j] == 1)
{
outpt = Math.Round(analyst.Script.Normalize.NormalizedFields[analyst.Script.Normalize.NormalizedFields.Count + j - count.Count].DeNormalize(info.OutputDataSample[i][j]), 1).ToString();
prediction = Math.Round(analyst.Script.Normalize.NormalizedFields[analyst.Script.Normalize.NormalizedFields.Count + j - count.Count].DeNormalize(answers[i][j]), 1).ToString();
}
else
{
double[] temp = new double[count[j]];
for (int c = 0; c < count[j]; c++)
{
temp[c] = info.OutputDataSample[i][c];
}
outpt = analyst.Script.Normalize.NormalizedFields[analyst.Script.Normalize.NormalizedFields.Count + j - count.Count].DetermineClass(temp).Name;
for (int c = 0; c < count[j]; c++)
{
temp[c] = answers[i][c];
}
prediction = analyst.Script.Normalize.NormalizedFields[analyst.Script.Normalize.NormalizedFields.Count + j - count.Count].DetermineClass(temp).Name;
}
item += String.Format(
"Predicted Output: {0} Correct Output: {1}\n\n",
prediction, outpt
);
}
}
return(item);
}
/// <summary> Trains current neural network. </summary> public abstract double Train(Data info, float lr, float mom);
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 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);
}
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);
}