libitpp-dev 4.3.1-2 / usr / include / itpp / stat / mog_generic.h

/*!
 * \file
 * \brief Generic Mixture of Gaussians (MOG) class - header file
 * \author Conrad Sanderson
 *
 * -------------------------------------------------------------------------
 *
 * Copyright (C) 1995-2010  (see AUTHORS file for a list of contributors)
 *
 * This file is part of IT++ - a C++ library of mathematical, signal
 * processing, speech processing, and communications classes and functions.
 *
 * IT++ is free software: you can redistribute it and/or modify it under the
 * terms of the GNU General Public License as published by the Free Software
 * Foundation, either version 3 of the License, or (at your option) any
 * later version.
 *
 * IT++ is distributed in the hope that it will be useful, but WITHOUT ANY
 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
 * FOR A PARTICULAR PURPOSE.  See the GNU General Public License for more
 * details.
 *
 * You should have received a copy of the GNU General Public License along
 * with IT++.  If not, see <http://www.gnu.org/licenses/>.
 *
 * -------------------------------------------------------------------------
 */

#ifndef MOG_GENERIC_H
#define MOG_GENERIC_H

#include <itpp/base/vec.h>
#include <itpp/base/mat.h>
#include <itpp/base/array.h>
#include <itpp/itexports.h>
#include <itpp/base/base_exports.h>

namespace itpp
{

/*!
  \ingroup MOG
  \brief Generic Mixture of Gaussians (MOG) class. Used as a base for other MOG classes.
  \author Conrad Sanderson

  Used for representing a statistical distribution as a
  convex combination of multi-variate Gaussian functions.
  Also known as a Gaussian Mixture Model.
  This class handles both full and diagonal covariance matrices,
  allows loading and saving of the MOG's parameters,
  as well as calculation of likelihoods. The parameters
  are set by the user or an optimisation algorithm
  (for example, see the MOG_diag_EM class).
  For speed and space reasons, diagonal and full
  covariance matrices are stored and handled separately.
*/
class ITPP_EXPORT MOG_generic
{

public:

  /*! \brief Default constructor
      \note An empty model is created.
            The likelihood functions are not useable
            until the model's parameters are set
  */
  MOG_generic() { init(); }

  /*! \brief Construct the MOG_generic object by loading the parameters from a model file
      \param name_in The model's filename
  */
  MOG_generic(const std::string &name_in) { load(name_in); }

  /*! \brief construct a default model (all Gaussians have zero mean and unit variance for all dimensions)
      \param K_in Number of Gaussians
      \param D_in Dimensionality
      \param full_in If true, use full covariance matrices; if false, use diagonal covariance matrices. Default = false.
  */
  MOG_generic(const int &K_in, const int &D_in, bool full_in = false) { init(K_in, D_in, full_in); }

  /*! \brief Construct a model using user supplied mean vectors
      \param means_in Array of mean vectors
      \param full_in If true, use full covariance matrices; if false, use diagonal covariance matrices. Default = false.
      \note  The number of mean vectors specifies the number of Gaussians.
      The covariance matrices are set to the identity matrix.
      The weights for all Gaussians are the same, equal to 1/K, where K is the number of Gaussians
  */
  MOG_generic(Array<vec> &means_in, bool full_in = false) { init(means_in, full_in); }

  /*! \brief Construct a model using user supplied parameters (diagonal covariance version)
      \param means_in Array of mean vectors
      \param diag_covs_in Array of vectors representing diagonal covariances
      \param weights_in vector of weights
      \note  The number of mean vectors, covariance vectors and weights must be the same
  */
  MOG_generic(Array<vec> &means_in, Array<vec> &diag_covs_in, vec &weights_in) { init(means_in, diag_covs_in, weights_in); }

  /*! \brief Construct a model using user supplied parameters (full covariance version)
      \param means_in Array of mean vectors
      \param full_covs_in Array of full covariance matrices
      \param weights_in vector of weights
      \note  The number of mean vectors, covariance matrices and weights must be the same
  */
  MOG_generic(Array<vec> &means_in, Array<mat> &full_covs_in, vec &weights_in) { init(means_in, full_covs_in, weights_in); }

  //! Default destructor
  virtual ~MOG_generic() { cleanup(); }

  /*! \brief Initialise the model to be empty.
      \note The likelihood functions are not useable
            until the model's parameters are set
  */
  void init();

  /*! \brief initialise the model so that all Gaussians have zero mean and unit variance for all dimensions
      \param K_in Number of Gaussians
      \param D_in Dimensionality
      \param full_in If true, use full covariance matrices; if false, use diagonal covariance matrices. Default = false.
  */
  void init(const int &K_in, const int &D_in, bool full_in = false);

  /*! \brief Initialise the model using user supplied mean vectors
      \param means_in Array of mean vectors
      \param full_in If true, use full covariance matrices; if false, use diagonal covariance matrices. Default = false.
      \note  The number of mean vectors specifies the number of Gaussians.
      The covariance matrices are set to the identity matrix.
      The weights for all Gaussians are the same, equal to 1/K, where K is the number of Gaussians
  */
  void init(Array<vec> &means_in, bool full_in = false);

  /*! \brief Initialise the model using user supplied parameters (diagonal covariance version)
      \param means_in Array of mean vectors
      \param diag_covs_in Array of vectors representing diagonal covariances
      \param weights_in vector of weights
      \note  The number of mean vectors, covariance vectors and weights must be the same
    */
  void init(Array<vec> &means_in, Array<vec> &diag_covs_in, vec &weights_in);

  /*! \brief Initialise the model using user supplied parameters (full covariance version)
      \param means_in Array of mean vectors
      \param full_covs_in Array of covariance matrices
      \param weights_in vector of weights
      \note  The number of mean vectors, covariance matrices and weights must be the same
  */
  void init(Array<vec> &means_in, Array<mat> &full_covs_in, vec &weights_in);

  /*! \brief Release memory used by the model. The model will be empty.
      \note The likelihood functions are not useable
            until the model's parameters are re-initialised
  */
  virtual void cleanup();

  //! Returns true if the model's parameters are valid
  bool is_valid() const { return valid; }

  //! Returns true if the model has full covariance matrices
  bool is_full() const { return full; }

  //! Return the number of Gaussians
  int get_K() const { if (valid) return(K); else return(0); }

  //! Return the dimensionality
  int get_D() const { if (valid) return(D); else return(0); }

  //! Obtain a copy of the weight vector
  vec get_weights() const { vec tmp;  if (valid) { tmp = weights; } return tmp; }

  //! Obtain a copy of the array of mean vectors
  Array<vec> get_means() const { Array<vec> tmp; if (valid) { tmp = means; } return tmp; }

  //! Obtain a copy of the array of diagonal covariance vectors
  Array<vec> get_diag_covs() const { Array<vec> tmp; if (valid && !full) { tmp = diag_covs; } return tmp; }

  //! Obtain a copy of the array of full covariance matrices
  Array<mat> get_full_covs() const { Array<mat> tmp; if (valid && full) { tmp = full_covs; } return tmp; }

  /*! \brief Set the means of the model
      \note The number of means must match the number of Gaussians in the model
  */
  void set_means(Array<vec> &means_in);

  /*! \brief Set the diagonal covariance vectors of the model
      \note  The number of diagonal covariance vectors must match the number of Gaussians in the model
  */
  void set_diag_covs(Array<vec> &diag_covs_in);

  /*! \brief Set the full covariance matrices of the model
      \note  The number of covariance matrices must match the number of Gaussians in the model
  */
  void set_full_covs(Array<mat> &full_covs_in);

  /*! \brief Set the weight vector of the model
      \note  The number of elements in the weight vector must match the number of Gaussians in the model
  */
  void set_weights(vec &weights_in);

  //! Set the means in the model to be zero
  void set_means_zero();

  //! Set the diagonal covariance vectors to be unity
  void set_diag_covs_unity();

  //! Set the full covariance matrices to be unity
  void set_full_covs_unity();

  //! Set all the weights to 1/K, where K is the number of Gaussians
  void set_weights_uniform();

  /*! \brief Enable/disable internal checks for likelihood functions
      \param do_checks_in If true, checks are enabled; if false, checks are disabled
      \note Disabling checks will provide a speedup in the likelihood functions.
            Disable them only when you're happy that everything is working correctly.
  */
  void set_checks(bool do_checks_in) { do_checks = do_checks_in; }

  /*! \brief Enable/disable paranoia about numerical stability
      \param paranoid_in If true, calculate likelihoods using a safer, but slower method.
  */
  void set_paranoid(bool paranoid_in) { paranoid = paranoid_in; }

  /*! \brief Initialise the model by loading the parameters from a model file
      \param name_in The model's filename
  */
  virtual void load(const std::string &name_in);

  /*! \brief Save the model's parameters to a model file
      \param name_in The model's filename
  */
  virtual void save(const std::string &name_in) const;

  /*! \brief Mathematically join the model with a user supplied model
      \param B_in user supplied model
      \note The Arrays of mean vectors and covariance vectors/matrices from the two models
            are simply concatenated, while the weights of the resultant model are a function
            of the original weights and numbers of Gaussians from both models.
            Specifically,
            \f$ w_{new} = [ \alpha \cdot w_{A} ~~~ \beta \cdot w_{B} ]^T \f$,
            where \f$ w_{new} \f$ is the new weight vector,
            \f$ w_{A} \f$ and \f$ w_{B} \f$ are the weight vectors from model A and B,
            while \f$ \alpha = K_A / (K_A + KB_in) \f$
            and \f$ \beta = 1-\alpha \f$.
            In turn, \f$ K_A \f$ and \f$ KB_in \f$ are the numbers of Gaussians in model A and B, respectively.

            See <a href="http://dx.doi.org/10.1016/j.patcog.2005.07.001">On transforming statistical models...</a>
            for more information.
  */
  virtual void join(const MOG_generic &B_in);

  /*! \brief Convert the model to use diagonal covariances

      \note If the model is already diagonal, nothing is done.
            If the model has full covariance matrices,
            this results in irreversible information loss
            (in effect the off-diagonal covariance elements are now zero)
  */
  virtual void convert_to_diag();

  /*! \brief Convert the model to have full covariance matrices
      \note If the model has full covariance matrices, nothing is done.
            If the model has diagonal covariances, the off-diagonal
            elements in the full covariance matrices are set to zero.
  */
  virtual void convert_to_full();

  //! calculate the log likelihood of vector \c x_in using only Gaussian \c k
  virtual double log_lhood_single_gaus(const vec &x_in, const int k);

  //! calculate the log likelihood of vector \c x_in
  virtual double log_lhood(const vec &x_in);

  //! calculate the likelihood of vector \c x_in
  virtual double lhood(const vec &x_in);

  //! calculate the average log likelihood of an array of vectors \c X_in
  virtual double avg_log_lhood(const Array<vec> &X_in);

protected:

  //! indicates whether checks on input data are done
  bool do_checks;

  //! indicates whether the parameters are valid
  bool valid;

  //! indicates whether we are using full or diagonal covariance matrices
  bool full;

  //! indicates whether we are paranoid about numerical stability
  bool paranoid;

  //! number of gaussians
  int K;

  //! dimensionality
  int D;

  //! means
  Array<vec> means;

  //! diagonal covariance matrices, stored as vectors
  Array<vec> diag_covs;

  //! full covariance matrices
  Array<mat> full_covs;

  //! weights
  vec weights;

  //! Pre-calcualted std::log(std::numeric_limits<double>::max() / K), where K is the number of Gaussians
  double log_max_K;

  /*! \brief Gaussian specific pre-calcualted constants
      \note  Vector of pre-calculated \f$ -\frac{D}{2}\log(2\pi) -\frac{1}{2}\log(|\Sigma|) \f$ for each Gaussian,
             where \f$ D \f$ is the dimensionality and \f$ |\Sigma| \f$ is the determinant for the Gaussian's
             covariance matrix \f$ \Sigma \f$.
  */
  vec log_det_etc;

  //! Pre-calculated log versions of the weights
  vec log_weights;

  //! Pre-calcuated inverted version of each full covariance matrix
  Array<mat> full_covs_inv;

  //! Pre-calcuated inverted version of each diagonal covariance vector, where the covariance elements are first multiplied by two
  Array<vec> diag_covs_inv_etc;

  //! Check if vector \c x_in has the same dimensionality as the model
  bool check_size(const vec &x_in) const;

  //! Check if all vectors in Array \c X_in have the same dimensionality as the model
  bool check_size(const Array<vec> &X_in) const;

  //! Check if all vectors in Array \c X_in have the same dimensionality
  bool check_array_uniformity(const Array<vec> & A) const;

  //! ADD DOCUMENTATION HERE
  void set_means_internal(Array<vec> &means_in);
  //! ADD DOCUMENTATION HERE
  void set_diag_covs_internal(Array<vec> &diag_covs_in);
  //! ADD DOCUMENTATION HERE
  void set_full_covs_internal(Array<mat> &full_covs_in);
  //! ADD DOCUMENTATION HERE
  void set_weights_internal(vec &_weigths);

  //! ADD DOCUMENTATION HERE
  void set_means_zero_internal();
  //! ADD DOCUMENTATION HERE
  void set_diag_covs_unity_internal();
  //! ADD DOCUMENTATION HERE
  void set_full_covs_unity_internal();
  //! ADD DOCUMENTATION HERE
  void set_weights_uniform_internal();

  //! ADD DOCUMENTATION HERE
  void convert_to_diag_internal();
  //! ADD DOCUMENTATION HERE
  void convert_to_full_internal();

  //! additional processing of mean vectors, done as the last step of mean initialisation
  virtual void setup_means();

  //! additional processing of covariance vectors/matrices, done as the last step of covariance initialisation
  virtual void setup_covs();

  //! additional processing of the weight vector, done as the last step of weight initialisation
  virtual void setup_weights();

  //! additional processing of miscellaneous parameters, done as the last step of overall initialisation
  virtual void setup_misc();

  //! ADD DOCUMENTATION HERE
  virtual double log_lhood_single_gaus_internal(const vec &x_in, const int k);
  //! ADD DOCUMENTATION HERE
  virtual double log_lhood_internal(const vec &x_in);
  //! ADD DOCUMENTATION HERE
  virtual double lhood_internal(const vec &x_in);

private:
  vec tmpvecD;
  vec tmpvecK;

};

} // namespace itpp

#endif // #ifndef MOG_GENERIC_H