itpp::MOG_diag_EM_sup Class Reference

support class for MOG_diag_ML() and MOG_diag_MAP() More...

#include <mog_diag_em.h>

List of all members.

Public Member Functions

 MOG_diag_EM_sup ()
 Default constructor.
 ~MOG_diag_EM_sup ()
 Default destructor.
void ml (MOG_diag &model_in, Array< vec > &X_in, int max_iter_in=10, double var_floor_in=0.0, double weight_floor_in=0.0, bool verbose_in=false)
 ADD DOCUMENTATION HERE.
void map (MOG_diag &model_in, MOG_diag &prior_model, Array< vec > &X_in, int max_iter_in=10, double alpha_in=0.5, double var_floor_in=0.0, double weight_floor_in=0.0, bool verbose_in=false)
 ADD DOCUMENTATION HERE.
void cleanup ()
 Release memory used by the model. The model will be empty.
void load (const std::string &name_in)
 Initialise the model by loading the parameters from a model file.
void convert_to_full ()
 Do nothing. Present for compatability with the MOG_generic class.
double log_lhood_single_gaus (const double *c_x_in, const int k) const
 calculate the log likelihood of C vector c_x_in using only Gaussian k
double log_lhood_single_gaus (const vec &x_in, const int k) const
 calculate the log likelihood of IT++ vector x_in using only Gaussian k
virtual double log_lhood_single_gaus (const vec &x_in, const int k)
 calculate the log likelihood of vector x_in using only Gaussian k
double log_lhood (const double *c_x_in)
 calculate the log likelihood of C vector c_x_in
double log_lhood (const vec &x_in)
 calculate the log likelihood of IT++ vector x_in
double lhood (const double *c_x_in)
 calculate the likelihood of C vector c_x_in
double lhood (const vec &x_in)
 calculate the likelihood of IT++ vector x_in
double avg_log_lhood (const double **c_x_in, int N)
 calculate the average log likelihood of an array of C vectors ( c_x_in )
double avg_log_lhood (const Array< vec > &X_in)
 calculate the average log likelihood of an array of IT++ vectors ( X_in )
void init ()
 Initialise the model to be empty.
void init (const int &K_in, const int &D_in, bool full_in=false)
 initialise the model so that all Gaussians have zero mean and unit variance for all dimensions
void init (Array< vec > &means_in, bool full_in=false)
 Initialise the model using user supplied mean vectors.
void init (Array< vec > &means_in, Array< vec > &diag_covs_in, vec &weights_in)
 Initialise the model using user supplied parameters (diagonal covariance version).
void init (Array< vec > &means_in, Array< mat > &full_covs_in, vec &weights_in)
 Initialise the model using user supplied parameters (full covariance version).
bool is_valid () const
 Returns true if the model's parameters are valid.
bool is_full () const
 Returns true if the model has full covariance matrices.
int get_K () const
 Return the number of Gaussians.
int get_D () const
 Return the dimensionality.
vec get_weights () const
 Obtain a copy of the weight vector.
Array< vec > get_means () const
 Obtain a copy of the array of mean vectors.
Array< vec > get_diag_covs () const
 Obtain a copy of the array of diagonal covariance vectors.
Array< mat > get_full_covs () const
 Obtain a copy of the array of full covariance matrices.
void set_means (Array< vec > &means_in)
 Set the means of the model.
void set_diag_covs (Array< vec > &diag_covs_in)
 Set the diagonal covariance vectors of the model.
void set_full_covs (Array< mat > &full_covs_in)
 Set the full covariance matrices of the model.
void set_weights (vec &weights_in)
 Set the weight vector of the model.
void set_means_zero ()
 Set the means in the model to be zero.
void set_diag_covs_unity ()
 Set the diagonal covariance vectors to be unity.
void set_full_covs_unity ()
 Set the full covariance matrices to be unity.
void set_weights_uniform ()
 Set all the weights to 1/K, where K is the number of Gaussians.
void set_checks (bool do_checks_in)
 Enable/disable internal checks for likelihood functions.
void set_paranoid (bool paranoid_in)
 Enable/disable paranoia about numerical stability.
virtual void save (const std::string &name_in) const
 Save the model's parameters to a model file.
virtual void join (const MOG_generic &B_in)
 Mathematically join the model with a user supplied model.
virtual void convert_to_diag ()
 Convert the model to use diagonal covariances.

Protected Member Functions

void update_internals ()
 ADD DOCUMENTATION HERE.
void sanitise_params ()
 ADD DOCUMENTATION HERE.
double ml_update_params ()
 ADD DOCUMENTATION HERE.
void ml_iterate ()
 ADD DOCUMENTATION HERE.
void setup_means ()
 additional processing of mean vectors, done as the last step of mean initialisation
void setup_covs ()
 additional processing of covariance vectors/matrices, done as the last step of covariance initialisation
void setup_weights ()
 additional processing of the weight vector, done as the last step of weight initialisation
void setup_misc ()
 additional processing of miscellaneous parameters, done as the last step of overall initialisation
double log_lhood_single_gaus_internal (const double *c_x_in, const int k) const
 ADD DOCUMENTATION HERE.
double log_lhood_single_gaus_internal (const vec &x_in, const int k) const
 ADD DOCUMENTATION HERE.
virtual double log_lhood_single_gaus_internal (const vec &x_in, const int k)
 ADD DOCUMENTATION HERE.
double log_lhood_internal (const double *c_x_in)
 ADD DOCUMENTATION HERE.
double log_lhood_internal (const vec &x_in)
 ADD DOCUMENTATION HERE.
double lhood_internal (const double *c_x_in)
 ADD DOCUMENTATION HERE.
double lhood_internal (const vec &x_in)
 ADD DOCUMENTATION HERE.
double ** enable_c_access (Array< vec > &A_in)
 Enable C style access to an Array of vectors (vec).
int ** enable_c_access (Array< ivec > &A_in)
 Enable C style access to an Array of vectors (ivec).
double * enable_c_access (vec &v_in)
 Enable C style access to a vector (vec).
int * enable_c_access (ivec &v_in)
 Enable C style access to a vector (ivec).
double ** disable_c_access (double **A_in)
 Disable C style access to an Array of vectors (vec).
int ** disable_c_access (int **A_in)
 Disable C style access to an Array of vectors (ivec).
double * disable_c_access (double *v_in)
 Disable C style access to a vector (vec).
int * disable_c_access (int *v_in)
 Disable C style access to a vector (ivec).
void zero_all_ptrs ()
 ADD DOCUMENTATION HERE.
void free_all_ptrs ()
 ADD DOCUMENTATION HERE.
bool check_size (const vec &x_in) const
 Check if vector x_in has the same dimensionality as the model.
bool check_size (const Array< vec > &X_in) const
 Check if all vectors in Array X_in have the same dimensionality as the model.
bool check_array_uniformity (const Array< vec > &A) const
 Check if all vectors in Array X_in have the same dimensionality.
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 ()
 ADD DOCUMENTATION HERE.

Protected Attributes

bool verbose
 Whether we print the progress.
int N
 number of training vectors
int max_iter
 Maximum number of iterations.
double ** c_X
 'C' pointers to training vectors
double var_floor
 ADD DOCUMENTATION HERE.
double weight_floor
 ADD DOCUMENTATION HERE.
double ** c_means
 pointers to the mean vectors
double ** c_diag_covs
 pointers to the covariance vectors
double ** c_diag_covs_inv_etc
 pointers to the inverted covariance vectors
double * c_weights
 pointer to the weight vector
double * c_log_weights
 pointer to the log version of the weight vector
double * c_log_det_etc
 pointer to the log_det_etc vector
bool do_checks
 indicates whether checks on input data are done
bool valid
 indicates whether the parameters are valid
bool full
 indicates whether we are using full or diagonal covariance matrices
bool paranoid
 indicates whether we are paranoid about numerical stability
int K
 number of gaussians
int D
 dimensionality
Array< vec > means
 means
Array< vec > diag_covs
 diagonal covariance matrices, stored as vectors
Array< mat > full_covs
 full covariance matrices
vec weights
 weights
double log_max_K
 Pre-calcualted std::log(std::numeric_limits<double>::max() / K), where K is the number of Gaussians.
vec log_det_etc
 Gaussian specific pre-calcualted constants.
vec log_weights
 Pre-calculated log versions of the weights.
Array< mat > full_covs_inv
 Pre-calcuated inverted version of each full covariance matrix.
Array< vec > diag_covs_inv_etc
 Pre-calcuated inverted version of each diagonal covariance vector, where the covariance elements are first multiplied by two.

Detailed Description

support class for MOG_diag_ML() and MOG_diag_MAP()

Author:
Conrad Sanderson

Member Function Documentation

void itpp::MOG_diag::cleanup (  )  [inline, virtual, inherited]

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

Reimplemented from itpp::MOG_generic.

References itpp::MOG_diag::free_all_ptrs().

Referenced by ml(), itpp::MOG_diag_kmeans_sup::run(), and itpp::MOG_diag::~MOG_diag().

void itpp::MOG_generic::convert_to_diag (  )  [virtual, inherited]

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)

References itpp::MOG_generic::convert_to_diag_internal(), and itpp::MOG_generic::valid.

Referenced by itpp::MOG_diag::load(), and itpp::MOG_diag::MOG_diag().

void itpp::MOG_generic::init ( Array< vec > &  means_in,
Array< mat > &  full_covs_in,
vec &  weights_in 
) [inherited]

Initialise the model using user supplied parameters (full covariance version).

Parameters:
means_in Array of mean vectors
full_covs_in Array of covariance matrices
weights_in vector of weights
Note:
The number of mean vectors, covariance matrices and weights must be the same

References itpp::MOG_generic::check_array_uniformity(), itpp::MOG_generic::D, itpp::MOG_generic::do_checks, itpp::MOG_generic::full, it_assert, itpp::MOG_generic::K, itpp::MOG_generic::paranoid, itpp::MOG_generic::set_full_covs_internal(), itpp::MOG_generic::set_means_internal(), itpp::MOG_generic::set_weights_internal(), itpp::MOG_generic::setup_misc(), itpp::Array< T >::size(), and itpp::MOG_generic::valid.

void itpp::MOG_generic::init ( Array< vec > &  means_in,
Array< vec > &  diag_covs_in,
vec &  weights_in 
) [inherited]

Initialise the model using user supplied parameters (diagonal covariance version).

Parameters:
means_in Array of mean vectors
diag_covs_in Array of vectors representing diagonal covariances
weights_in vector of weights
Note:
The number of mean vectors, covariance vectors and weights must be the same

References itpp::MOG_generic::check_array_uniformity(), itpp::MOG_generic::D, itpp::MOG_generic::do_checks, itpp::MOG_generic::full, it_assert, itpp::MOG_generic::K, itpp::MOG_generic::paranoid, itpp::MOG_generic::set_diag_covs_internal(), itpp::MOG_generic::set_means_internal(), itpp::MOG_generic::set_weights_internal(), itpp::MOG_generic::setup_misc(), itpp::Array< T >::size(), and itpp::MOG_generic::valid.

void itpp::MOG_generic::init ( Array< vec > &  means_in,
bool  full_in = false 
) [inherited]

Initialise the model using user supplied mean vectors.

Parameters:
means_in Array of mean vectors
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

References itpp::MOG_generic::check_array_uniformity(), itpp::MOG_generic::D, itpp::MOG_generic::do_checks, itpp::MOG_generic::full, it_assert, itpp::MOG_generic::K, itpp::MOG_generic::paranoid, itpp::MOG_generic::set_diag_covs_unity_internal(), itpp::MOG_generic::set_full_covs_unity_internal(), itpp::MOG_generic::set_means(), itpp::MOG_generic::set_weights_uniform_internal(), itpp::MOG_generic::setup_misc(), itpp::Array< T >::size(), and itpp::MOG_generic::valid.

void itpp::MOG_generic::init ( const int &  K_in,
const int &  D_in,
bool  full_in = false 
) [inherited]

initialise the model so that all Gaussians have zero mean and unit variance for all dimensions

Parameters:
K_in Number of Gaussians
D_in Dimensionality
full_in If true, use full covariance matrices; if false, use diagonal covariance matrices. Default = false.

References itpp::MOG_generic::D, itpp::MOG_generic::do_checks, itpp::MOG_generic::full, it_assert, itpp::MOG_generic::K, itpp::MOG_generic::paranoid, itpp::MOG_generic::set_diag_covs_unity_internal(), itpp::MOG_generic::set_full_covs_unity_internal(), itpp::MOG_generic::set_means_zero_internal(), itpp::MOG_generic::set_weights_uniform_internal(), itpp::MOG_generic::setup_misc(), and itpp::MOG_generic::valid.

void itpp::MOG_generic::init (  )  [inherited]

Initialise the model to be empty.

Note:
The likelihood functions are not useable until the model's parameters are set

References itpp::MOG_generic::cleanup().

Referenced by itpp::MOG_generic::join(), itpp::MOG_generic::load(), ml(), itpp::MOG_diag::MOG_diag(), itpp::MOG_generic::MOG_generic(), and itpp::MOG_diag_kmeans_sup::run().

void itpp::MOG_generic::join ( const MOG_generic &  B_in  )  [virtual, inherited]

Mathematically join the model with a user supplied model.

Parameters:
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, $ w_{new} = [ \alpha \cdot w_{A} ~~~ \beta \cdot w_{B} ]^T $, where $ w_{new} $ is the new weight vector, $ w_{A} $ and $ w_{B} $ are the weight vectors from model A and B, while $ \alpha = K_A / (K_A + KB_in) $ and $ \beta = 1-\alpha $. In turn, $ K_A $ and $ KB_in $ are the numbers of Gaussians in model A and B, respectively.
See On transforming statistical models... for more information.

References itpp::MOG_generic::D, itpp::MOG_generic::diag_covs, itpp::MOG_generic::full, itpp::MOG_generic::full_covs, itpp::MOG_generic::get_D(), itpp::MOG_generic::get_diag_covs(), itpp::MOG_generic::get_full_covs(), itpp::MOG_generic::get_K(), itpp::MOG_generic::get_means(), itpp::MOG_generic::get_weights(), itpp::MOG_generic::init(), itpp::MOG_generic::is_full(), itpp::MOG_generic::is_valid(), it_assert, itpp::MOG_generic::K, itpp::MOG_generic::means, itpp::MOG_generic::valid, and itpp::MOG_generic::weights.

void itpp::MOG_diag::load ( const std::string &  name_in  )  [virtual, inherited]

Initialise the model by loading the parameters from a model file.

Parameters:
name_in The model's filename
Note:
If the model file contains a full covariance matrix model, the covariance matrices will be converted to be diagonal after loading.

Reimplemented from itpp::MOG_generic.

References itpp::MOG_generic::convert_to_diag(), and itpp::MOG_generic::full.

Referenced by itpp::MOG_diag::MOG_diag().

double itpp::MOG_diag_EM_sup::ml_update_params (  )  [protected]

ADD DOCUMENTATION HERE.

update parameters using the Maximum Likelihood version of the EM algorithm

References itpp::MOG_diag::c_diag_covs, itpp::MOG_diag::c_log_weights, itpp::MOG_diag::c_means, itpp::MOG_diag::c_weights, c_X, itpp::MOG_generic::D, itpp::exp(), itpp::MOG_generic::K, itpp::log(), itpp::log_add(), itpp::MOG_diag::log_lhood_single_gaus_internal(), itpp::MOG_generic::log_max_K, N, itpp::MOG_generic::paranoid, itpp::sum(), and itpp::trunc_exp().

Referenced by ml_iterate().

void itpp::MOG_diag_EM_sup::sanitise_params (  )  [inline, protected]

ADD DOCUMENTATION HERE.

for helping to avoid numerical instability

References itpp::MOG_diag::c_diag_covs, itpp::MOG_diag::c_weights, itpp::MOG_generic::D, itpp::MOG_generic::K, var_floor, and weight_floor.

Referenced by ml_iterate().

void itpp::MOG_generic::save ( const std::string &  name_in  )  const [virtual, inherited]

Save the model's parameters to a model file.

Parameters:
name_in The model's filename

References itpp::it_file::close(), itpp::MOG_generic::diag_covs, itpp::MOG_generic::full, itpp::MOG_generic::full_covs, itpp::MOG_generic::means, itpp::MOG_generic::valid, and itpp::MOG_generic::weights.

void itpp::MOG_generic::set_checks ( bool  do_checks_in  )  [inline, inherited]

Enable/disable internal checks for likelihood functions.

Parameters:
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.

References itpp::MOG_generic::do_checks.

void itpp::MOG_generic::set_diag_covs ( Array< vec > &  diag_covs_in  )  [inherited]

Set the diagonal covariance vectors of the model.

Note:
The number of diagonal covariance vectors must match the number of Gaussians in the model

References itpp::MOG_generic::set_diag_covs_internal(), and itpp::MOG_generic::valid.

void itpp::MOG_generic::set_full_covs ( Array< mat > &  full_covs_in  )  [inherited]

Set the full covariance matrices of the model.

Note:
The number of covariance matrices must match the number of Gaussians in the model

References itpp::MOG_generic::set_full_covs_internal(), and itpp::MOG_generic::valid.

void itpp::MOG_generic::set_means ( Array< vec > &  means_in  )  [inherited]

Set the means of the model.

Note:
The number of means must match the number of Gaussians in the model

References itpp::MOG_generic::set_means_internal(), and itpp::MOG_generic::valid.

Referenced by itpp::MOG_generic::init().

void itpp::MOG_generic::set_paranoid ( bool  paranoid_in  )  [inline, inherited]

Enable/disable paranoia about numerical stability.

Parameters:
paranoid_in If true, calculate likelihoods using a safer, but slower method.

References itpp::MOG_generic::paranoid.

void itpp::MOG_generic::set_weights ( vec &  weights_in  )  [inherited]

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

References itpp::MOG_generic::set_weights_internal(), and itpp::MOG_generic::valid.

void itpp::MOG_diag_EM_sup::update_internals (  )  [inline, protected]

ADD DOCUMENTATION HERE.

update log versions of parameters and any necessary constants

References itpp::MOG_diag::c_diag_covs, itpp::MOG_diag::c_diag_covs_inv_etc, itpp::MOG_diag::c_log_det_etc, itpp::MOG_diag::c_log_weights, itpp::MOG_diag::c_weights, itpp::MOG_generic::D, itpp::MOG_generic::K, and itpp::log().

Referenced by ml_iterate().

Member Data Documentation

vec itpp::MOG_generic::log_det_etc [protected, inherited]

Gaussian specific pre-calcualted constants.

Note:
Vector of pre-calculated $ -\frac{D}{2}\log(2\pi) -\frac{1}{2}\log(|\Sigma|) $ for each Gaussian, where $ D $ is the dimensionality and $ |\Sigma| $ is the determinant for the Gaussian's covariance matrix $ \Sigma $.

Referenced by itpp::MOG_generic::cleanup(), itpp::MOG_generic::log_lhood_single_gaus_internal(), itpp::MOG_generic::setup_covs(), and itpp::MOG_diag::setup_covs().

The documentation for this class was generated from the following files:
Generated on Tue Jun 2 10:02:19 2009 for mixpp by  doxygen 1.5.8