plot_chmm.m

一个隐马尔可夫模型工具箱一个隐马尔可夫模型工具箱一个隐马尔可夫模型工具箱

% -----------------------------------       
% This script was implemented to plot some of the characteristics of a CHMM.
% First, we load the hmm and the parameters of the four classes of polygons.
% 
% First, we plot the mixture of gaussians calculated for every HMM (group and class)
% and for each state.
% -------------------------------------
        load hmmpoligonos2;
       load parampoligonos vlcp;
       % Length of the window to plot the mixture of gaussians 
       longueur_fenetre=499;
       y=cell(nc,ng);
       % Variable to store the mixture of Gaussian
       mixture_of_gaussian=zeros(1,longueur_fenetre+1);
     
      for ic=1:nc
          for ig=1:ng,
              % variables for the loops to store the coefficients of the mixture of gaussian
            Np=length(agrup{ig})-1;
            Ne= length(A{ic,ig}(1,:));
            y {ic,ig}=cell(Np,1);
                for ip=1:Np;
                    y {ic,ig}{ip}=cell(Ne,1);
                    for ie=1:Ne;
                             y {ic,ig}{ip}{ie}=cell(Ngauss{ig}(ip),1);                    
                    for ngauss=1:Ngauss{ig}(ip)
                       % we extract the center of each Gaussians
                      min_mean=  min(Med{ic,ig}{ip}{ie}(:)); % we calculate the minimum of all the gaussians for this state and HMM
                      max_mean=  max(Med{ic,ig}{ip}{ie}(:)); % we calculate the maximum of all the gaussians for this state and HMM
                        % we calculate the maximum variance of the gaussians of all the gaussians for this state and HMM 
                       max_var=  max(Var{ic,ig}{ip}{ie}(:));
                       % We calculate a length to adjust the windows ro plot the mixture of gaussian
                       largeur=abs(max_var*log(0.1));
                       % The path is a variable to estimate the probabilities for "longuer_of_fenetre"+1 points for the gaussian
                       path= (max_mean-min_mean+2*largeur)/longueur_fenetre;
                      % Those points x are the same for all the gaussians from a mixture of gaussian defined by one HMM and one state 
                      x = (min_mean-largeur:path:max_mean+largeur);
%            y {ic,ig}{ip}{ie}{ngauss} is the results of the estimation of the probability for thos points
             y {ic,ig}{ip}{ie}{ngauss}= GAUSSmf(x, [Var{ic,ig}{ip}{ie}(ngauss) Med{ic,ig}{ip}{ie}(ngauss)]);
            % We store the values of the mixture_of_gaussian estimated in those x points 
            % (the results y are multiplied by the coefficient B corresponding to the coefficient of this gaussian for this mixture of gaussian) 
            mixture_of_gaussian=mixture_of_gaussian + B{ic,ig}{ip}{ie}(ngauss) * y {ic,ig}{ip}{ie}{ngauss};
                  end
                  % We plot the mixture of gaussians for each state of this HMM
                  % We make in particular a subplot to show the 12 mixture of gaussians (one for each state) for this HMM and this parameter
                  h3=figure(3);
                  subplot(3,4,ie)
                  plot(mixture_of_gaussian) ; 
                  mixture_of_gaussian=zeros(1,longueur_fenetre+1);
                  title([' cl',num2str(ic),' gr',num2str(ig),' pr',num2str(ip),' state ',num2str(ie)]);
            end
           pause;
           % we could save the figure obtained
          % saveas(h3,'fig3.jpg','jpg');
                
         end 
     end
 end
 
 for ir=1:nr
        for ic=1:nc
            for ig=1:ng,
               
                         
                % we evaluate the probability for the 4 HMM
                for ihmm=1:nc;
                    O=vlcp{ic,ig}{ir};
                   salida(ihmm,1)=probsecc(A{ihmm,ig},B{ihmm,ig},Med{ihmm,ig},Var{ihmm,ig},Pi{ihmm,ig},O,agrup{ig});
                    % we calculate the alfa matrices tanks to the forward backward process
                    [alfa,beta,c]=alfabetac(A{ihmm,ig},B{ihmm,ig},Med{ihmm,ig},Var{ihmm,ig},Pi{ihmm,ig},O,agrup{ig});
                    T=size(O);
                    % We calculate the most probable state sequence thanks to the viterbi algorithm
                    qP=viterbic(A{ihmm,ig},B{ihmm,ig},Med{ihmm,ig},Var{ihmm,ig},Pi{ihmm,ig},O,agrup{ig});
                    h4=figure(4)
                    subplot(211)
                    plot(qP);
                    title([' class real: ',num2str(ic),' HMM model of the class: ',num2str(ihmm),' group and repetition: ', num2str(ig),' ',num2str(ir),])
                    subplot(212)
                    imagesc(alfa);
                    title(['graph 1: most probable state sequence with viterbi, graph 2: Alpha']);
                    pause

%        saveas(h4,'fig4.jpg','jpg');
                end
            end;  
        end
    end
    
end