public static int TrainSlow(this NeuralNet net, List <List <double> > inputs, List <List <double> > expectedResults, double performanceThreshold)
{
Contract.Equals(inputs.Count, expectedResults.Count);
bool trained;
int iteration = 0;
do
{
trained = true;
for (int i = 0; i < inputs.Count; i++)
{
net.ChangeData(inputs[i], expectedResults[i]);
trained &= net.CalculatePerformance() >= performanceThreshold;
net.UpdateAdjustment();
}
net.AdjustCumulatedWeights();
net.ResetValues();
net.ResetDerrivates();
iteration++;
if (iteration % 100 == 0)
{
Debug.WriteLine("Performance: " + net.CalculatePerformance());
}
}while (!trained);
Debug.WriteLine("Iteration: " + iteration);
return(iteration);
}
public static int TrainAdaptiveWithConfidenceAndPerformance(this NeuralNet net, List <List <double> > inputs, List <List <double> > expectedResults, double performanceThreshold, int iterationLimit, Reading testData)
{
Contract.Equals(inputs.Count, expectedResults.Count);
bool trained;
int iteration = 0;
double performanceOnOneBefore = -1;
double performance = -1;
List <double> performances;
do
{
performances = new List <double>(inputs.Count);
var setsTrained = inputs.Count;
trained = true;
performanceOnOneBefore = net.CalculatePerformance();
for (int i = 0; i < inputs.Count; i++)
{
performances.Add(net.CalculatePerformance());
net.ChangeData(inputs[i], expectedResults[i]);
var setTrained = performances[i] >= performanceThreshold;
trained &= setTrained;
if (setTrained)
{
setsTrained--;
}
net.UpdateAdjustment();
}
net.AdjustCumulatedWeights();
net.ResetValues();
net.ResetDerrivates();
performance = net.CalculatePerformance();
var performanceIncreased = performance > performanceOnOneBefore;
var performanceIncreaseUnderThreshold = performance - performanceOnOneBefore < 0.0000000001;
if (performanceIncreased && performanceIncreaseUnderThreshold) //+2
{
Link.RenewalFactor = 0.000001;
}
else
{
Link.RenewalFactor = 0;
}
iteration++;
if (iteration % 1000 == 0)
{
Debug.WriteLine("Iteration: " + iteration++ + " performance:" + net.CalculatePerformance() + "Performance average" + performances.Sum() / performances.Count + " Sets untrained:" + setsTrained);
ComputeResultsAndConfidence(iteration, net, testData);
}
if (iteration % iterationLimit == 0)
{
Link.RenewalFactor = 1;
}
}while (!trained);
Debug.WriteLine("Iteration: " + iteration);
return(iteration);
}
public static int TrainAdaptive(this NeuralNet net, List <List <double> > inputs, List <List <double> > expectedResults, double performanceThreshold)
{
Contract.Equals(inputs.Count, expectedResults.Count);
bool trained;
int iteration = 0;
double performanceBefore = -1;
double performance = -1;
do
{
var setsTrained = inputs.Count;
trained = true;
performanceBefore = performance;
for (int i = 0; i < inputs.Count; i++)
{
net.ChangeData(inputs[i], expectedResults[i]);
var setTrained = net.CalculatePerformance() >= performanceThreshold;
trained &= setTrained;
if (setTrained)
{
setsTrained--;
}
net.UpdateAdjustment();
}
net.AdjustCumulatedWeights();
net.ResetValues();
net.ResetDerrivates();
performance = net.CalculatePerformance();
if (performance - performanceBefore < 0.0000001)
{
Link.RenewalFactor = 0.00003;
}
else
{
Link.RenewalFactor = 0.00000000003;
}
if (iteration % 1000 == 0)
{
Debug.WriteLine("Iteration: " + iteration++ + " performance:" + net.CalculatePerformance() + "Sets trained:" + setsTrained);
}
iteration++;
}while (!trained);
Debug.WriteLine("Iteration: " + iteration);
return(iteration);
}