Description of slreevallearn

Home > sltoolbox > learn > slreevallearn.m

slreevallearn

PURPOSE

SLREEVALLEARN Performs an iterative learning based on re-evaluation

SYNOPSIS

function [models, Q, info] = slreevallearn(models, Q, data, estfunctor, evalfunctor, cmpfunctor, varargin)

DESCRIPTION

SLREEVALLEARN Performs an iterative learning based on re-evaluation

 $ Syntax $
   - [models, Q] = slreevallearn(models, data, estfunctor, evalfunctor, cmpfunctor, ...)
   - [models, Q, info] = slreevallearn(models, data, estfunctor, evalfunctor, cmpfunctor, ...)

 $ Arguments $
   - models:       the models to learn
                   input one: the initial model
                   output one: the learned model
   - Q:            the evaluated quantities on the data
   - data:         the data (samples) to learn from
   - estfunctor:   the functor to estimate/update model, in the form:
                   models = f(models, data, Q, ...)
                   If the process is recorded, it is like:
                   [models, rec] = f(models, data, Q, ...)
   - evalfunctor:  the functor to evaluate/update the quantities, like:
                   Q = f(models, data, Q, ...)
   - cmpfunctor:   the functor to judge convergence, like:
                   isconverged = f(models_prev, models, Q_prev, Q, ...)
   - info:         The information of the iteration process

 $ Description $
   - [models, Q] = slreevallearn(models, data, estfunctor, evalfunctor, cmpfunctor, ...) 
     implements a skeleton of learning procedure based on re-evaluation.
     The procedure alternates the model estimation based on data and 
     corresponding evaluated quantities, and the evaluation by applying
     models to data to get quantities.
     \*
     \t   Table. The Properties of Re-evaluating Learning
     \h    name        &   description
          'iter'       & The iteration control properties. default = {}
          'isrecorded' & whether the model estimation is recorded
                         default = false;
          'verbose'    & whether to show progress, default = true
     \*
   
 $ Remarks $
   - It is implemented based on sliterproc.

   - It essentially serves as the core of many learning paradigms, 
     such as E-M method of finite mixture model learning, AdaBoost,
     K-means, robust subspace learning, etc.

 $ History $
   - Created by Dahua Lin, on Aug 31, 2006

CROSS-REFERENCE INFORMATION

This function calls:

• sliterproc SLITERPROC Runs a general iterative process
• raise_lackinput RAISE_LACKINPUT Raises an error indicating lack of input argument
• slevalfunctor SLEVALFUNCTOR Evaluates a functor
• slparseprops SLPARSEPROPS Parses input parameters
• slsharedisp SLSHAREDISP Displays message using a shared configuration
• slsharedisp_attach SLSHAREDISP_ATTACH Attachs to global display options
• slsharedisp_detach SLSHAREDISP_DETACH Detachs current function from global display

This function is called by:

• slkmeansex SLKMEANSEX Performs Generalized K-means
• slfmm SLFMM Learns a Finite Mixture Model (FMM)

SUBFUNCTIONS

• function varargout = reevallearn_iter(objects, estfunctor, evalfunctor, opts)
• function isconverged = reevallearn_cmp(objects_prev, objects, cmpfunctor)

SOURCE CODE

0001 function [models, Q, info] = slreevallearn(models, Q, data, estfunctor, evalfunctor, cmpfunctor, varargin)
0002 %SLREEVALLEARN Performs an iterative learning based on re-evaluation
0003 %
0004 % $ Syntax $
0005 %   - [models, Q] = slreevallearn(models, data, estfunctor, evalfunctor, cmpfunctor, ...)
0006 %   - [models, Q, info] = slreevallearn(models, data, estfunctor, evalfunctor, cmpfunctor, ...)
0007 %
0008 % $ Arguments $
0009 %   - models:       the models to learn
0010 %                   input one: the initial model
0011 %                   output one: the learned model
0012 %   - Q:            the evaluated quantities on the data
0013 %   - data:         the data (samples) to learn from
0014 %   - estfunctor:   the functor to estimate/update model, in the form:
0015 %                   models = f(models, data, Q, ...)
0016 %                   If the process is recorded, it is like:
0017 %                   [models, rec] = f(models, data, Q, ...)
0018 %   - evalfunctor:  the functor to evaluate/update the quantities, like:
0019 %                   Q = f(models, data, Q, ...)
0020 %   - cmpfunctor:   the functor to judge convergence, like:
0021 %                   isconverged = f(models_prev, models, Q_prev, Q, ...)
0022 %   - info:         The information of the iteration process
0023 %
0024 % $ Description $
0025 %   - [models, Q] = slreevallearn(models, data, estfunctor, evalfunctor, cmpfunctor, ...)
0026 %     implements a skeleton of learning procedure based on re-evaluation.
0027 %     The procedure alternates the model estimation based on data and
0028 %     corresponding evaluated quantities, and the evaluation by applying
0029 %     models to data to get quantities.
0030 %     \*
0031 %     \t   Table. The Properties of Re-evaluating Learning
0032 %     \h    name        &   description
0033 %          'iter'       & The iteration control properties. default = {}
0034 %          'isrecorded' & whether the model estimation is recorded
0035 %                         default = false;
0036 %          'verbose'    & whether to show progress, default = true
0037 %     \*
0038 %
0039 % $ Remarks $
0040 %   - It is implemented based on sliterproc.
0041 %
0042 %   - It essentially serves as the core of many learning paradigms,
0043 %     such as E-M method of finite mixture model learning, AdaBoost,
0044 %     K-means, robust subspace learning, etc.
0045 %
0046 % $ History $
0047 %   - Created by Dahua Lin, on Aug 31, 2006
0048 %
0049 
0050 %% parse and verify input
0051 
0052 if nargin < 6
0053     raise_lackinput('slreevallearn', 6);
0054 end
0055 
0056 opts.iter = {};
0057 opts.isrecorded = false;
0058 opts.verbose = true;
0059 opts = slparseprops(opts, varargin{:});
0060 
0061 %% Main
0062 
0063 slsharedisp_attach('slreevallearn', 'show', opts.verbose);
0064 
0065 slsharedisp('Learning by re-evaluation');
0066 
0067 objects = {models, data, Q};
0068 iterfunctor = {@reevallearn_iter, estfunctor, evalfunctor, opts};
0069 cmpfunctor = {@reevallearn_cmp, cmpfunctor};
0070 if nargout < 2
0071     objects = ...
0072         sliterproc(objects, iterfunctor, cmpfunctor, opts.isrecorded, opts.iter{:});
0073 else
0074     [objects, info] = ...
0075         sliterproc(objects, iterfunctor, cmpfunctor, opts.isrecorded, opts.iter{:});
0076 end
0077 
0078 models = objects{1};
0079 Q = objects{3};
0080 
0081 slsharedisp_detach();
0082 
0083 
0084 %% Iteration functors
0085 
0086 function varargout = reevallearn_iter(objects, estfunctor, evalfunctor, opts)
0087 % objects = {models, data, Q}
0088 
0089 [models, data, Q] = deal(objects{:});
0090 
0091 if ~opts.isrecorded
0092     models = slevalfunctor(estfunctor, models, data, Q);
0093 else
0094     [models, info] = slevalfunctor(estfunctor, models, data, Q);
0095 end
0096 Q = slevalfunctor(evalfunctor, models, data, Q);
0097 
0098 objects = {models, data, Q};
0099 
0100 if ~opts.isrecorded
0101     varargout = {objects};
0102 else
0103     varargout = {objects, info};
0104 end
0105 
0106 
0107 function isconverged = reevallearn_cmp(objects_prev, objects, cmpfunctor)
0108 
0109 models_prev = objects_prev{1};
0110 Q_prev = objects_prev{3};
0111 models = objects{1};
0112 Q = objects{3};
0113 
0114 isconverged = slevalfunctor(cmpfunctor, models_prev, models, Q_prev, Q);
0115 
0116 
0117 
0118 
0119 
0120 
0121 
0122 
0123 
0124 
0125

---

Generated on Wed 20-Sep-2006 12:43:11 by m2html © 2003