/// <summary>
/// Perform validation procedure, train on trainProblem and test on testProblem
/// </summary>
/// <typeparam name="TProblemElement">Problem element</typeparam>
/// <param name="TrainingProblem">Train <see cref="Problem{TProblemElement}">Problem</see> </param>
/// <param name="TestProblem">Problem to test</param>
/// <param name="Kernel">SVM kernel</param>
/// <param name="C">Penalty parameter fo SVM solver</param>
/// <returns>Accuracy</returns>
public double TrainAndTestValidation(
Problem <TProblemElement> TrainingProblem, Problem <TProblemElement> TestProblem)
{
CSVM <TProblemElement> svm = new CSVM <TProblemElement>(TrainingProblem, Kernel, C, Evaluator);
Stopwatch t = Stopwatch.StartNew();
svm.Init();
//todo: change in in production code
Console.WriteLine("SVM init time {0}", t.Elapsed);
// Debug.WriteLine("SVM init time {0}", t.Elapsed);
svm.Train();
Console.WriteLine("Svm train takes {0}", svm.model.ModelTimeMs);
//svm.model.WriteToFile("modelFile.txt");
int correct = 0;
Debug.WriteLine("Start Predict");
t.Restart();
//for (int i = 0; i < TestProblem.ElementsCount; i++)
//{
// float predictedLabel = svm.Predict(TestProblem.Elements[i]);
// if (predictedLabel == TestProblem.Labels[i])
// ++correct;
//}
var predictions = svm.Predict(TestProblem.Elements);
t.Stop();
for (int i = 0; i < TestProblem.ElementsCount; i++)
{
float predictedLabel = predictions[i];
if (predictedLabel == TestProblem.Y[i])
{
++correct;
}
}
double accuracy = (float)correct / TestProblem.ElementsCount;
Console.WriteLine(string.Format("init, dispose and prediction on {0} elements takes {1}, correct={2}", TestProblem.ElementsCount, t.Elapsed, correct));
return(accuracy);
}
/// <summary>
/// Do cross validation procedure on specified folds.
/// </summary>
/// <typeparam name="TProblemElement">The type of the problem element.</typeparam>
/// <param name="probSize">Size of the problem.</param>
/// <param name="foldsElements">Array of folds, each list contains elements which belonds to fold .</param>
/// <param name="foldsLabels">Arrat of fodls, each list contains elements labels.</param>
/// <param name="Kernel">The kernel.</param>
/// <param name="C">Penalty parameter C.</param>
/// <returns>Accuracy</returns>
public double CrossValidateOnFolds(
int probSize,
List <TProblemElement>[] foldsElements,
List <float>[] foldsLabels)
{
Debug.Assert(foldsElements.Length == foldsLabels.Length, "array lenght should have the same lenght");
int nrFolds = foldsElements.Length;
long correct = 0L;
for (int i = 0; i < nrFolds; i++)
{
//array of sub problem elements (comes form permutation)
int subProbSize = 0;
var trainFoldIndexes = from t in Enumerable.Range(0, nrFolds)
where t != i
select t;
foreach (var s in trainFoldIndexes)
{
subProbSize += foldsElements[s].Count;
}
float[] subLabels;
TProblemElement[] subProbElem = CreateSubProblem(foldsElements, foldsLabels, trainFoldIndexes, subProbSize, out subLabels);
//create sub problem based on previous subProblemElements and subProbLabels
Problem <TProblemElement> trainSubprob = new Problem <TProblemElement>(subProbElem, subLabels);
CSVM <TProblemElement> svm = new CSVM <TProblemElement>(trainSubprob, Kernel, C, Evaluator);
svm.Init();
svm.Train();
for (int j = 0; j < foldsElements[i].Count; j++)
{
var element = foldsElements[i][j];
var prediction = svm.Predict(element);
if (prediction == foldsLabels[i][j])
{
++correct;
}
}
}
//todo: accuracy count for cross validation
return((float)correct / probSize);
}
