public DAGSVM(ClassificationProblem problem)
{
this.m_problem = problem;
this.m_n = problem.CategoryCount;
this.m_k = m_n * (m_n - 1) / 2;
this.m_dag = new Node(problem, 0, m_n - 1);
}
public LinearLeraningMachine(ClassificationProblem problem)
{
this.m_problem = problem;
this.m_t_set = problem.TrainingSet;
//this.m_problem.RetrieveVocabulary(out this.m_voc);
this.m_l = m_t_set.Examples.Count;
//this.m_weight = new SparseVector(m_voc.Count);
}
/// <summary>
/// foamliu, 2009/12/21, please make sure you've uncompressed "2_newsgroups.7z" in the "data" folder.
/// </summary>
/// <returns></returns>
private static ClassificationProblem CreateText()
{
const string DataFolder = @"..\data\2_newsgroups";
ClassificationProblem problem = new ClassificationProblem();
ExampleSet t_set = new ExampleSet();
ExampleSet v_set = new ExampleSet();
CategoryCollection collect = new CategoryCollection();
collect.Add(new Category(+1, "+1"));
collect.Add(new Category(-1, "-1"));
problem.Dimension = 2;
problem.CategoryCollection = collect;
DirectoryInfo dataFolder = new DirectoryInfo(DataFolder);
DirectoryInfo[] subfolders = dataFolder.GetDirectories();
int count = 0;
for (int i = 0; i < subfolders.Count(); i++)
{
DirectoryInfo categoryFolder = subfolders[i];
int cat = i * 2 - 1;
// for all the text files in each category
FileInfo[] files = categoryFolder.GetFiles();
count = 0;
int trainSetCount = Convert.ToInt32(Constants.TrainingSetRatio * files.Count());
for (int j = 0; j < files.Count(); j++)
{
FileInfo textFile = files[j];
Example e = new Example();
if (++count < trainSetCount)
{
t_set.AddExample(e);
}
else
{
v_set.AddExample(e);
}
}
}
problem.TrainingSet = t_set;
problem.ValidationSet = v_set;
return problem;
}
private static ClassificationProblem CreateChessBoard()
{
ClassificationProblem problem = new ClassificationProblem();
CategoryCollection collect = new CategoryCollection();
collect.Add(new Category(+1, "+1"));
collect.Add(new Category(-1, "-1"));
problem.Dimension = 2;
problem.CategoryCollection = collect;
problem.TrainingSet = GetExamples(collect);
problem.ValidationSet = GetExamples(collect);
return problem;
}
//private bool m_isleaf;
public Node(ClassificationProblem problem, int first, int second)
{
this.m_problem = problem;
this.m_first = first;
this.m_second = second;
this.m_llm = new LinearLeraningMachine(problem);
this.m_llm.Train();
if (second > first + 1)
{
this.m_leftChild = new Node(problem, first + 1, second);
this.m_rightChild = new Node(problem, first, second - 1);
//this.m_isleaf = false;
}
else
{
//this.m_isleaf = true;
}
}
private SparseVector m_weight; // weight vector
#endregion Fields
#region Constructors
public Binary_SVM_SMO(ClassificationProblem problem)
{
this.m_problem = problem;
this.m_t_set = this.m_problem.TrainingSet;
// this.m_problem.RetrieveVocabulary(out this.m_voc);
this.m_l = m_t_set.Examples.Count;
this.m_alpha = new double[m_l];
this.m_error = new double[m_l];
this.m_kernel = new LinearKernel();
this.m_NonBound = new List<int>();
this.m_rand = new Random();
this.m_weight = new SparseVector(problem.Dimension);
// foamliu, 2009/01/12, default values
this.m_c = Constants.SVM_C;
this.m_eta = Constants.SVM_Eta;
this.m_tolerance = Constants.SVM_Tolerance;
this.m_epsilon = Constants.SVM_Epsilon;
}
public double CrossValidate(ClassificationProblem problem)
{
ExampleSet t_Set;
ExampleSet v_Set; // validation set
//Logging.Info("Retrieving training set");
t_Set = problem.TrainingSet;
//Logging.Info("Retrieving validation set");
v_Set = problem.ValidationSet;
int numExample = v_Set.Examples.Count;
int numCorrect = 0;
//Logging.Info("Cross Validating on validation set");
foreach (Example example in v_Set.Examples)
{
ClassificationResult result = new ClassificationResult();
this.PredictText(t_Set, example, ref result);
if (result.ResultCategoryId == example.Label.Id)
{
numCorrect++;
}
}
double correctRatio = 1.0 * numCorrect / numExample;
Logger.Info(string.Format("Correct ratio: {0}", correctRatio));
return correctRatio;
}
//private void BuildExample(TextExample example, Vocabulary voc, int exampleCount)
//{
// int dimension = voc.Count;
// SparseVector vector = new SparseVector(dimension);
// foreach (string word in example.Tokens.Keys)
// {
// int pos = voc.GetWordPosition(word);
// if (pos == Constants.KEY_NOT_FOUND)
// continue;
// // phi i(x) = tfi log(idfi) /k
// // tfi: number of occurences of the term i in the document x
// // idfi: the ratio between the total number of documents and the
// // number of documents containing the term
// // k: normalisation constant ensuring that ||phi|| = 1
// double phi = example.Tokens[word] * Math.Log(exampleCount / voc.WordExampleOccurMap[word]);
// vector.Components.Add(pos, phi);
// }
// vector.Normalize();
// example.X = vector;
//}
//private void Preprocess(ClassificationProblem problem)
//{
// Vocabulary voc;
// problem.RetrieveVocabulary(out voc);
// foreach (Category c in problem.CategoryCollection.Collection)
// {
// foreach (TextExample e in c.Examples)
// {
// BuildExample(e, voc, problem.ExampleCount);
// }
// }
// m_weight = new SparseVector(voc.Count);
//}
/// <summary>
/// simple on-line algorithm for the 1-norm soft margin:
/// training SVMs in the non-bias case.
/// </summary>
/// <param name="problem"></param>
public void Train(ClassificationProblem problem)
{
ExampleSet t_Set; // training set
//Logging.Info("Retrieving training set");
t_Set = problem.TrainingSet;
l = t_Set.Examples.Count;
//Logging.Info("Preprocessing all the examples");
//this.Preprocess(problem);
m_Alpha = new double[l];
m_newalpha = new double[l];
for (int i=0;i<m_Alpha.Length;i++)
{
m_Alpha[i] = 0.0;
}
//Logging.Info("Gradient descent");
while (true)
{
for (int i = 0; i < l; i++)
{
double temp = 0.0;
for (int j=0;j<l;j++)
{
temp += m_Alpha[j] * t_Set.Examples[j].Label.Id * m_kernel.Compute(t_Set.Examples[i].X, t_Set.Examples[j].X);
}
m_newalpha[i] = m_Alpha[i] + Constants.SVM_Eta * (1.0 - t_Set.Examples[i].Label.Id * temp);
if (m_newalpha[i] < 0.0)
{
m_newalpha[i] = 0.0;
}
else if (m_newalpha[i] > Constants.SVM_C)
{
m_newalpha[i] = Constants.SVM_C;
}
}
this.CopyAlphas();
W = this.CalculateSVM_W(t_Set);
if (Math.Abs((W - old_W) / W) < Constants.SVM_Tolerance)
{
break;
}
Logger.Info(string.Format("SVM W = {0}", W));
old_W = W;
}
this.CalculateWeight(t_Set);
//this.CalculateB(t_Set);
}