source: trunk/c++_tools/classifier/KernelLookup.h @ 628

// $Id$

#ifndef _theplu_classifier_kernel_lookup_
#define _theplu_classifier_kernel_lookup_

#include <c++_tools/classifier/Kernel.h>
#include <c++_tools/classifier/DataLookup2D.h>
#include <c++_tools/classifier/MatrixLookup.h>
#include <vector>

namespace theplu {
namespace classifier {

  class KernelFunction;

  ///
  /// @brief Lookup into Kernel
  ///
  /// This is the KernelLookup class to be used together with kernel
  /// methods such as Support Vector Machines (SVM). The class does
  /// not contain any data or values, but rather is a lookup into a
  /// Kernel object. Each row and each column corresponds to a row and
  /// a column in the Kernel, respectively. This design allow for fast
  /// creation of sub-kernels, which is a common operation in most
  /// traning/validation procedures.
  ///
  /// A KernelLookup can be created directly from a Kernel or from an
  /// other KernelLookup. In the latter case, the resulting
  /// KernelLookup is looking directly into the underlying Kernel to
  /// avoid multiple lookups.
  ///
  /// There is a possibility to set the KernelLookup as owner of the
  /// underlying Kernel. In that case the underlying Kernel will be
  /// destroyed in the destructor. Consequently, the underlying Kernel
  /// must have been dynamically allocated and no other KernelLookup
  /// can own the Kernel.
  ///
  class KernelLookup : public DataLookup2D 
  {

  public:
    
    ///
    /// @brief Constructor a Lookup into a Kernel
    ///
    /// Constructs a KernelLookup corresponding to the Kernel @a
    /// kernel. By default @a owner is set to false, which means
    /// KernelLookup does not own the underlying Kernel. If
    /// KernelLookup owns the Kernel then the Kernel will be deleted
    /// in the destructor.
    ///
    /// @note If underlying Kernel goes out of scope or is deleted, the
    /// KernelLookup becomes invalid and the result of further use is
    /// undefined.
    ///
    KernelLookup(const Kernel& kernel, const bool owner=false);

    /// 
    /// @brief Constructing a Lookup into a subKernel
    ///
    /// Creating a Lookup into parts of the Kernel. In the created
    /// Lookup the element in the \f$ i \f$ th row in the \f$ j \f$ th
    /// column is identical to the element in row row[i] and columns
    /// column[j] in the underlying @a kernel. If @a owner is set to
    /// true yhe underlying @a kernel is destroyed in the destructor.
    ///
    /// @note If @a kernel goes out of scope or is deleted, the
    /// returned pointer becomes invalid and the result of further use is
    /// undefined.
    ///
    /// @note For training usage row index shall always be equal to
    /// column index.
    ///
    KernelLookup(const Kernel& kernel, const std::vector<size_t>& row, 
                 const std::vector<size_t>& column, const bool owner=false);
    
    /// 
    /// @brief Copy constructor. 
    ///
    /// A Lookup is created looking into the
    /// same underlying Kernel as @a kl is looking into. The newly
    /// created KernelLookup does not own the underlying Kernel.
    ///
    KernelLookup(const KernelLookup& kl);


    /// 
    /// @brief Contructing a sub-KernelLookup. 
    ///
    /// Contructor building a sub-KernelLookup from a KernelLookup
    /// defined by row index vector and column index vector. In the
    /// created Lookup the element in the \f$ i \f$ th row in the
    /// \f$ j \f$ th column is identical to the element in row row[i] and
    /// columns column[j] in the copied @a kl. The resulting
    /// KernelLookup is independent of the old KernelLookup, but is
    /// undefined in case underlying Kernel is destroyed.
    ///
    /// @note For training usage row index shall always be equal to
    /// column index.
    ///
    KernelLookup(const KernelLookup& kl, const std::vector<size_t>& row, 
                 const std::vector<size_t>& column);
    
    ///
    /// Constructor taking the column (default) or row index vector as
    /// input. If @a row is false the created KernelLookup will have
    /// equally many rows as @a kernel.
    ///
    /// @note If underlying kernel goes out of scope or is deleted, the
    /// KernelLookup becomes invalid and the result of further use is
    /// undefined.
    ///
    KernelLookup(const KernelLookup& kernel, const std::vector<size_t>&, 
                 const bool row=false);

    ///
    /// @brief Destructor
    ///
    /// Deletes underlying Kernel if KernelLookup owns it. 
    ///
    virtual ~KernelLookup(void);


    ///
    /// Creates a sub-Kernel identical to the one created using
    /// KernelLookup(*this, train, train).
    ///
    /// @return pointer to dynamically allocated sub-Lookup of the KernelLookup
    ///
    /// @Note Returns a dynamically allocated DataLookup2D, which has
    /// to be deleted by the caller to avoid memory leaks.
    ///
    const KernelLookup* training_data(const std::vector<size_t>& train) const;


    ///
    /// @retun a sub-kernel of kernel calculated using data defined by
    /// @a features. Each row and each columns corresponds to a traing
    /// sample defined by @a train.
    ///
    /// @return pointer to dynamically allocated sub-Lookup of the KernelLookup
    ///
    /// @Note Returns a dynamically allocated DataLookup2D, which has
    /// to be deleted by the caller to avoid memory leaks.
    ///
    const KernelLookup* training_data(const std::vector<size_t>& features,
                                      const std::vector<size_t>& train) const;


    ///
    /// In returned kernel each row corresponds to a training sample
    /// and each column corresponds to a validation sample. The
    /// created sub-KernelLookup is equivalent to using
    /// KernelLooup(*this, train, validation).
    ///
    /// @return sub-Lookup of the DataLookup2D
    ///
    /// @Note Returns a dynamically allocated DataLookup2D, which has
    /// to be deleted by the caller to avoid memory leaks.
    ///
    const KernelLookup* 
    validation_data(const std::vector<size_t>& train, 
                    const std::vector<size_t>& validation) const;


    ///
    /// In returned kernel each row corresponds to a training sample
    /// and each column corresponds to a validation sample. The kernel
    /// is based on the features defined by @a features.
    ///
    /// @Note Returns a dynamically allocated DataLookup2D, which has
    /// to be deleted by the caller to avoid memory leaks.
    ///
    const KernelLookup* 
    validation_data(const std::vector<size_t>& features, 
                    const std::vector<size_t>& train, 
                    const std::vector<size_t>& validation) const;


    ///
    /// @return element at position (\a row, \a column) in the Kernel
    /// matrix
    ///
    inline double operator()(const size_t row,const size_t column) const
    { return (*kernel_)(row_index_[row],column_index_[column]); }

    ///
    /// Each column in returned MatrixLookup corresponds to the column
    /// in KernelLookup.
    ///
    /// @Note Returns a dynamically allocated MatrixLookup, which has
    /// to be deleted by the caller to avoid memory leaks.
    ///
    inline const DataLookup2D* data(void) const
    { return kernel_->data().training_data(column_index_); }


    ///
    /// Function to calculate a new Kernel element using the
    /// underlying KernelFunction. The value is calulated between @a
    /// vec and the data vector of the \f$ i \f$ th sample, in other
    /// words, the sample corresponding to the \f$ i \f$ th row or
    /// \f$ i \f$ th column. In case KernelLookup is a sub-Kernel and not
    /// symmetric, the kernel value is calculated between @a vec and
    /// the data vector corresponding to \f$ i \f$ th row.
    ///
    inline double element(const DataLookup1D& vec, const size_t i) const
    { return kernel_->element(vec, row_index_[i]); }

    ///
    /// Weighted version of element function. Using weights @a w all
    /// identical to unity results in same as using the unweighted
    /// version above. 
    ///
    inline double element(const DataLookupWeighted1D& vec, const size_t i) const
    { return kernel_->element(vec, row_index_[i]); }

    ///
    /// @todo remove
    /// @note This function will probably be removed
    ///
    /// @Note Returns a dynamically allocated KernelLookup, which has
    /// to be deleted by the caller to avoid memory leaks.
    ///
    const KernelLookup* selected(const std::vector<size_t>&) const;
    
    ///
    /// @return true if underlying Kernel is weighted
    ///
    inline bool weighted(void) const { return kernel_->weighted(); }

    ///
    /// Each column in returned MatrixLook corresponds to the column
    /// in KernelLookup. If no weights were used in construction of
    /// Kernel, each element in returned MatrixLookup is set to unity.
    ///
    /// @Note Returns a dynamically allocated MatrixLookup2D, which has
    /// to be deleted by the caller to avoid memory leaks.
    ///
    // Peter remove this
    //inline const MatrixLookup* weights(void) const
    //{ return new MatrixLookup(kernel_->weights(),column_index_, false); }


  private:
    const KernelLookup& operator=(const KernelLookup&);

    const Kernel* kernel_;
    
  }; // class KernelLookup

}} // of namespace classifier and namespace theplu

#endif