runsm.m.svn-base

Bayesian Surprise Toolkit for Matlab T. Nathan Mundhenk, Laurent Itti

%
%   References:
%
%      L. Itti, P. Baldi, A Principled Approach to Detecting Surprising 
%      Events in Video, In: Proc. IEEE Conference on Computer Vision and 
%      Pattern Recognition (CVPR), pp. 631-637, Jun 2005. 
%
%      L. Itti, P. Baldi, Bayesian Surprise Attracts Human Attention, In: 
%      Advances in Neural Information Processing Systems, Vol. 19 
%      (NIPS*2005), pp. 1-8, Cambridge, MA:MIT Press,  2006.
%
%   T. Nathan Mundhenk
%   [e-mail]
%

% //////////////////////////////////////////////////////////////////// %
% The Baysian Surprise Matlab Toolkit - Copyright (C) 2004-2007        %
% by the University of Southern California (USC) and the iLab at USC.  %
% See http://iLab.usc.edu for information about this project.          %
% //////////////////////////////////////////////////////////////////// %
% This file is part of the Baysian Surprise Matlab Toolkit             %
%                                                                      %
% The Baysian Surprise Matlab Toolkit 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 2 of the License, or (at your option) any later version.     %
%                                                                      %
% The Baysian Surprise Matlab Toolkit 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 the iBaysian Surprise Matlab Toolkit; if not, write       %
% to the Free Software Foundation, Inc., 59 Temple Place, Suite 330,   %
% Boston, MA 02111-1307 USA.                                           %
% //////////////////////////////////////////////////////////////////// %
%
% Primary maintainer for this file: T. Nathan Mundhenk <[e-mail]>
%

function smod = runsm(data,smod)

% We require that smod be reset each run. This keeps values clean
if isfield(smod,'smodisset')
    if smod.smodisset == 0
        fprintf('NOTICE: SMOD has been set once in newsm, but is no longer usable\n');
        fprintf('Most likely you have tried to call runsm in batch mode twice on\n');
        fprintf('the same smod. You must make a new smod each time you call a\n');
        fprintf('batch mode run on your data\n');
        error('SMOD is not valid');
    end
else
    error('You must create SMOD using the function newsm before you call this function');
end

% Use the asymptotic max value for beta
if strcmp(smod.options.setbetamax,'yes')
    smod.beta1    = smod.max.beta1;
    smod.beta2    = smod.max.beta2;
end

% Figure out if we are running sample by sample or in batch mode. If
% running in batch mode, we may need to transpose the vector matrix
if size(data,1) > 1
    if smod.options.debug == 2
        fprintf('Running in batch mode. Input is a column vector\n')
    end
    smod = runsmbatch(data,smod);
elseif size(data,2) > 1  
    if smod.dim == 1
        if smod.options.debug == 2
            fprintf('Running in batch mode (Univariate). Input is a row vector\n')
        end
        data = data';
        smod = runsmbatch(data,smod);
    else
        if smod.options.debug == 2
            fprintf('Running in single step mode (Multivariate). Input is a row vector\n')
        end
        smod = runsmsingle(data,smod);
    end  
else
    if smod.options.debug == 2
        fprintf('Running in single step mode (Univaraite)\n')
    end
    smod = runsmsingle(data,smod);
end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Batch Run Function 
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Call this to run surprise on a sample vector

function smod = runsmbatch(data,smod)

SX = size(data,1);
if smod.dim == 1
    SY = 1;
else
    SY = size(data,2);
end

% Create surprise values output matrix
smod.surprise = zeros(SX,SY);

% create some debug information if requested
if smod.options.debug > 0
    smod.debugdata        = struct('Description','Debug values from runsm');
    smod.debugdata.alpha1 = zeros(SX,SY);
    smod.debugdata.alpha2 = zeros(SX,SY);
    smod.debugdata.beta1  = zeros(SX,SY);
    smod.debugdata.beta2  = zeros(SX,SY);
end

% For each data item run the surprise model on it. This is the core of the
% batch mode surprise model. We compute new beta and alpha values fromt the
% sample value, then we compute the KL distance between the two Gamma PDF's
% using klgamma. We take the absolute value to support negative data
% values. However, using negative values as inputs may not in fact make
% sense. 
for n = 1:SX
    smod = newalphabeta(data(n,:),smod);
    smod.surprise(n,:) = abs(klgamma(smod.alpha1,smod.alpha2,smod.beta1,smod.beta2));
    if ~strcmp(smod.options.jointmodel,'none')
        smod.joint.surprise(n,:) = abs(klgamma(smod.joint.alpha1,smod.joint.alpha2,smod.joint.beta1,smod.joint.beta2));
    end
    
    if smod.options.debug > 0
        smod.debugdata.alpha1(n,:) = smod.alpha1;
        smod.debugdata.alpha2(n,:) = smod.alpha2;
        smod.debugdata.beta1(n,:)  = smod.beta1;
        smod.debugdata.beta2(n,:)  = smod.beta2;
        if smod.options.debug > 1
            fprintf('RUNNING INPUT %f LOOP %d ALPHA [%f,%f] BETA [%f,%f] SURPRISE VALUE %f\n',data(n,1),n,smod.alpha1,smod.alpha2,smod.beta1,smod.beta2,smod.surprise(n,1));
            fprintf('\n');    
        end
    end
        
    %smod.alpha1 = smod.alpha2; 
    %if strcmp(smod.options.setbetamax,'no')
        %smod.beta1  = smod.beta2;
        %end
    smod.epoch  = smod.epoch + 1;
    smod.smodisset = 0;
end

% Graph the surprise results if requested.
if strcmp(smod.options.graph,'surprise')
    if strcmp(smod.options.setbetamax,'yes')
        Title = 'setbetamax';
    elseif strcmp(smod.options.factordecay,'yes')
        Title = 'factordecay';
    else
        Title = 'basic';
    end
    
    res = basicsurprisegraph(smod.surprise,data,'Surprise Value','Input Value   ',Title,smod);
end
    
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Single Run Function
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Call this to run surprise on one sample at a time. 

function smod = runsmsingle(data,smod,options)
% For each data item run the surprise model on it. This is the core of the
% batch mode surprise model. We compute new beta and alpha values fromt the
% sample value, then we compute the KL distance between the two Gamma PDF's
% using klgamma. We take the absolute value to support negative data
% values. However, using negative values as inputs may not in fact make
% sense. 
smod = newalphabeta(data,smod,options);
smod.surprise = abs(klgamma(smod.alpha1,smod.alpha2,smod.beta1,smod.beta2));
if ~strcmp(smod.options.jointmodel,'none')
    smod.joint.surprise = abs(klgamma(smod.joint.alpha1,smod.joint.alpha2,smod.joint.beta1,smod.joint.beta2));
end
%smod.alpha1   = smod.alpha2; 
%if strcmp(smod.options.setbetamax,'no')
%    smod.beta1    = smod.beta2;
%end
smod.epoch    = smod.epoch + 1;