www.pudn.com > libpmk.rar > experiment.h, change:2007-05-27,size:3557b


// Copyright 2007, Massachusetts Institute of Technology.
// The use of this code is permitted for research only. There is
// absolutely no warranty for this software.
//
// Author: John Lee (jjl@mit.edu)
//

#ifndef EXPERIMENT_EXPERIMENT_H
#define EXPERIMENT_EXPERIMENT_H

#include <vector>
#include "kernel/matrix.h"
#include "kernel/kernel-matrix.h"
#include "util/labeled-index.h"

using namespace std;

using namespace libpmk;
namespace libpmk_util {

/// Encapsulates training/testing of any method involving a kernel.
/**
 * All that you need to implement are Train() and Test(). When
 * implementing Test(), the way you report a new prediction is via
 * SetPrediction(). Make sure to call SetPrediction() on each
 * testing example.
 */
class Experiment {
 public:
   /**
    * <kernel> includes pairwise kernel values for all data (both
    * training and testing). The LabeledIndices in <training> and
    * <testing> specify which row of the kernel to look at.
    */
   Experiment(vector<LabeledIndex> training,
              vector<LabeledIndex> testing,
              const KernelMatrix& kernel);

   /**
    * <training_matrix> is a kernel matrix for training examples
    * only. Let N be the number of training examples. Then
    * <testing_matrix> is a NxM Matrix where M is the number of test
    * examples, and the testing[i][j] is the kernel value between the
    * i'th training example and the j'th test example.  <training>
    * must be N-dimensional and <testing> must be M-dimensional.
    */
   Experiment(vector<LabeledIndex> training,
              const KernelMatrix& training_matrix,
              vector<LabeledIndex> testing, const Matrix& testing_matrix);

   virtual ~Experiment() { }

   /// Train a model (if applicable)
   virtual void Train() = 0;

   /// Make predictions for each testing example.
   /**
    * Returns the number of test examples that were correct.
    * When you implement Test(), you must report results via
    * SetPrediction().
    */
   virtual int Test() = 0;


   /// Returns the predicted value of the <test_index>th test example.
   /**
    * Can only call this after Test() is called. 
    */
   int GetPrediction(int test_index) const;

   /// \brief Get the total number of testing examples that were
   /// correctly classified.
   int GetNumCorrect() const;

   /// \brief Get the number of testing examples that had label
   /// <label> that were also correctly classified.
   int GetNumCorrect(int label) const;

   /// Get the total number of test examples.
   int GetNumTestExamples() const;

   /// Get the number of text examples with label <label>.
   int GetNumTestExamples(int label) const;

   /// Same as GetNumCorrect() / GetNumTestExamples().
   double GetAccuracy() const;

   /// Same as GetNumCorrect(label) / GetNumTestExamples(label).
   double GetAccuracy(int label) const;

 protected:
   vector<LabeledIndex> training_;
   vector<LabeledIndex> testing_;

   /// Get a kernel value (wrapper for KernelMatrix)
   double GetKernelValue(int row, int col) const;

   /// Get the kernel value corresponding to the given LabeledIndices.
   double GetKernelValue(const LabeledIndex& row,
                         const LabeledIndex& col) const;

   /** 
    * Call this to tell Experiment's internals that the test example
    * at <test_index> was classified as <prediction>.
    */
   void SetPrediction(int test_index, int prediction);

 private:
   KernelMatrix kernel_;
   vector<int> predictions_;
};
}  // namespace libpmk_util
#endif  // EXPERIMENT_EXPERIMENT_H