libeblearn
Public Member Functions | Public Attributes
ebl::classifier_meter Class Reference

#include <ebl_logger.h>

List of all members.

Public Member Functions

 classifier_meter ()
 ~classifier_meter ()
 Destructor.
void init (uint nclasses)
int correctp (ubyte co, ubyte cd)
void clear ()
void resize (intg sz)
char update (intg a, class_state *co, ubyte cd, double energy)
void update (intg age_, bool correct, double energy)
void update (intg age, uint desired, uint infered, double energy)
void test (class_state *co, ubyte cd, double energy)
double class_normalized_average_error (idx< int > &confu)
double overall_average_error (idx< int > &confu)
double class_normalized_average_success (idx< int > &confu)
int get_class_samples (idx< int > &confu, intg classid)
int get_class_errors (idx< int > &confu, intg classid)
double get_normalized_error ()
idx< int > & get_confusion ()
 Returns the confusion matrix;.
void info ()
void info_sprint ()
void info_print ()
void display (int iteration, string &dsname, std::vector< string * > *lblstr=NULL, bool ds_is_test=false)
void display_average (string &dsname, std::vector< string * > *lblstr=NULL, bool ds_is_test=false)
 Display information averaged over all iterations.
void display_positive_rates (double threshold, std::vector< string * > *lblstr=NULL)
bool save ()
bool load ()

Public Attributes

double energy
float confidence
intg size
intg age
 Number of samples seen for this iteration.
intg total_correct
intg total_error
intg total_punt
double total_energy
std::vector< uint > class_errors
std::vector< uint > class_totals
std::vector< uint > class_tpr
std::vector< uint > class_fpr
std::vector< std::string > log_fields
 Names of variables to log.
std::vector< double > log_values
 Values to log.
std::vector< double > total_values
 Totals samples used for values.

Detailed Description

Meters are classes used to measure the performance of learning machines. There are several types of meters for each specific situation. meters are generally assumed to have at least the following methods: {

update: updates the meter with the objects and values passed as argument.} { clear: resets the meter, so it can be used for a new series of measurements.} { test: simply prints performance information for the data passed as argument. This does not update any internal state.} } Methods are provided to compute and display the information measured by a meter. {

display: display performance information on the terminal} { info: returns a list of the informations printed by display} } a class that can be used to measure the performance of classifiers. This is a simple version that does not record anything but simply computes performance measures.

Constructor & Destructor Documentation

ebl::classifier_meter::classifier_meter ( )

Create a new <classifier-meter> using <comparison-function> to compare actual and desired answers. By default the <same-class?> function is used for that purpose. It takes two integer arguments, and returns 1 if they are equal, -1 if they are different, and 0 if the first argument is -1 (which means reject).

Member Function Documentation

double ebl::classifier_meter::class_normalized_average_error ( idx< int > &  confu)

Returns the average error percentage computed from the confusion matrix, i.e. taking into account the class distributions, in other words, returns the mean of the per-class mean error.

double ebl::classifier_meter::class_normalized_average_success ( idx< int > &  confu)

Returns the average success percentage computed from the confusion matrix 'confu'.

void ebl::classifier_meter::clear ( )

reset the meter. This must be called before a measurement campaign is started.

int ebl::classifier_meter::correctp ( ubyte  co,
ubyte  cd 
)

return 0 if <actual> equals -1, otherwise, return 1 if <actual> and <desired> are equal, -1 otherwise.

void ebl::classifier_meter::display ( int  iteration,
string &  dsname,
std::vector< string * > *  lblstr = NULL,
bool  ds_is_test = false 
)

Display the meter's information on the terminal. namely, the age, the number of samples (number of calls to update since the last clear), the average energy, the percentage of correctly recognize samples, the percentage of erroneously recognized samples, and the percentage of rejected samples. names of each class (lblstr) are optional.

Parameters:
ds_is_testIf true, prepend "test_" in front of varialbes.
void ebl::classifier_meter::display_positive_rates ( double  threshold,
std::vector< string * > *  lblstr = NULL 
)

display ROC points for each class. names of each class (lblstr) are optional.

int ebl::classifier_meter::get_class_errors ( idx< int > &  confu,
intg  classid 
)

Return the number of errors for class with id 'classid' in confusion matrix 'confu'.

int ebl::classifier_meter::get_class_samples ( idx< int > &  confu,
intg  classid 
)

Return the number of samples for class with id 'classid' in confusion matrix 'confu'.

double ebl::classifier_meter::get_normalized_error ( )

Returns the error rate, normalized by the number of examples of in each class.

void ebl::classifier_meter::info ( )

return a list with the age, the number of samples (number of calls to update since the last clear), the average energy, the percentage of correctly recognize samples, the percentage of erroneously recognized samples, and the percentage of rejected samples.

void ebl::classifier_meter::init ( uint  nclasses)

Initialize the meter.

Parameters:
nclassesThe number of classes.
double ebl::classifier_meter::overall_average_error ( idx< int > &  confu)

Returns the average error percentage computed regardless of the class. This measure might be bias with unbalanced datasets.

char ebl::classifier_meter::update ( intg  a,
class_state *  co,
ubyte  cd,
double  energy 
)

update the meter with results from a new sample. <age> is the number of training iterations so far, <actual> (a <class-state>) the actual output of the machine, <desired> (an idx0 of int) the desired category, and <energy> (an idx0-state) the energy.

The documentation for this class was generated from the following files: