dhmm_def.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

% ------------------------------------------------------------	
% 
% 
%                  %%%%%%%%%%%%
%                  % DHMM_DEF % 
%                  %%%%%%%%%%%%
% 
% 
% 
% This function is described only as example to set up the HMM
% 
% This function defines the parameters of the discrete HMM. We can split the parameters in 4 groups:
% 
% Entry:	fhmm is the HMM磗 file name.
% 
% 	vDB defines the database:
% vDB=[' nc ng agrup Np'];
% where nc is the number of class, ng is the number of group,
% agrup defines how to gather the parameters together
% NP is a vector with the number of each group of parameter.
% We have the following relations:
% [nc,ng]=size(vl);
% agrup{ig}=[1 .... size(vl{1,ig}{1},2)+1];
% Np(ig)=length(agrup{ig})-1;
% 
% 	vHMM defines the HMM :
% 	vHMM=[' BAKIS salto maxiter umbral maxitermi TOPN Ne Ns A B Pi'];
% if Bakis = 1 implements a Bakis HMM (also called left-right), else an ergodic HMM is defined 
% Salto is the maximum path authorized in the Bakis HMM from a state to another.
% Maxiter is the maximum iteration number in the Bakis HMM (condition to stop the algorithm)
% Umbral is the threshold condition to stop the HMM (the error is calculated with the maximum likelihood criterion) 
% TOPN number of labels to take into account for the multi-labelling during the training phase
% Ne states number 
% Ns symbols number for  every state
% A, B and Pi are the matrices of the HMM (it will be calculated thanks to the HMM) but must de defined here.
% Salhmm is a matrix used to store the probabilities values
% We have the following relations:
% TOPN=cell(ng,1);
% 
% 	 vQ is the matrix parameter for the library :
% vVQ=[' LBG dpztoLBG maxiterVQ umbralVQ men biblio'];
% where LBG is the choice of the algorithm to compile the library
% LBG = 1 the algorithm is LBG, in the other case we choose kMedias
% dptzoLBG percentage maximum for the distance of the code vector in the algorithm LBG (see gen_bib) 
% maxiterVQ is the maximum number of iteration to compute the library (condition to stop the algorithm)
% umbralVQ is the threshold condition to stop the library computation (condition to stop the algorithm)
% men=1 or 0 if men =1 the library generates message else no.
% biblio is a matrix parameter for the library computation (we define it here but calculate in the gen_bib function)
% 
% 	 vTEST is the matrix parameter for the test :
% vTEST=[' TOPNtest salhmm'];
% where TOPNtest is the number of labels to take into account for the multi-labeling during the test phase
% Salhmm is a matrix used to store the probabilities values
% We have the following relations:
% TOPNtest=cell(ng,1);
% salhmm=cell(nc,ng);
% ---------------------------------------------------------
function dhmm_def(fhmm)

% Name of the set-up file
%fhmm='hmm';


% vDB defines the database:
vDB=[' nc ng agrup Np'];
% where nc is the number of class, ng is the number of group,
% agrup defines how to gather the parameters together
% NP is a vector with the number of each group of parameter.
% with [nc,ng]=size(vl);
nc=4; % we have 4 classes of polygons
ng=2;  % two groups                                          

% Definition of the gathering
% with agrup{ig}=[1 .... size(vl{1,ig}{1},2)+1];
agrup=cell(ng,1);
Np=zeros(ng,1);
for ig=1:ng
   agrup{ig}=[1 2 3];% 1 2 3
end
% agrup{1}=[1 2 3 4 ];% if we put only one group, we can gather the parameters by [1 2 3 4] for instance
% agrup{2}=[1 2 3 4];% 1 2
for ig=1:ng
  Np(ig)=length(agrup{ig})-1;

end

% VARIABLeS OF THE HMM
vHMM=[' BAKIS salto maxiter umbral maxitermi TOPN Ne Ns A B Pi'];
% if Bakis = 1 implements a Bakis HMM (also called left-right), else an ergodic HMM is defined 
% Salto is the maximum path authorized in the Bakis HMM from a state to another.
% Maxiter is the maximum iteration number in the Bakis HMM (condition to stop the algorithm)
% Umbral is the threshold condition to stop the HMM (the error is calculated with the maximum likelihood criterion) 
% TOPN number of labels to take into account for the multi-labelling during the training phase
% Ne states number 
% Ns symbols number for  every state
% A, B and Pi are the matrices of the HMM (it will be calculated thanks to the HMM) but must de defined here.
% Salhmm is a matrix used to store the probabilities values
% We have the following relations:
% TOPN=cell(ng,1);
TOPN=cell(ng,1);
for ig=1:ng   
   TOPN{ig}=1.*ones(Np(ig),1); %2.*ones(Np(ig),1) if we want to put two labels
end


% Number of sates.
Ne=10.*ones(nc,ng); % 10 states are been chosen, if we want to change it: Ne=20.*ones(nc,ng)
% Number of parameters per state.
Ns=cell(ng,1);
for ig=1:ng   
   Ns{ig}=32.*ones(Np(ig),1); %32 symbols are chosen
end

BAKIS=1; % if we wnat to choose an ergodic HMM Bakis=0

salto=1; % if we want to increase the path alloweb between two states salto=2 or salto=3

maxiter=30; % maximum iteration of the baum welch is here 30 we can change it maxiter =40 for example

umbral=0.005; % to change the threshold umbral=0.001 for instance

maxitermi=10; % the iteration number to find the initial model, we can chenge it by for instance maxitermi=20;
% Matrices of the HMM
A=cell(nc,ng);
B=cell(nc,ng);
Pi=cell(nc,ng);
%for ic=1:nc,
%   for ig=1:ng
%      A{ic,ig}=zeros(Ne(ic,ig),Ne(ic,ig));
%      B{ic,ig}=cell(Np(ig));
%      for ip=1:Np(ig)
%         B{ic,ig}{ip}=zeros(Ne(ic,ig),Ns{ig}(ip));
%      end
%      P{ic,ig}=zeros(Ne(ic,ig),1);
%   end
%end


% vQ is the matrix parameter for the library :
vVQ=[' LBG dpztoLBG maxiterVQ umbralVQ men biblio'];

% where LBG is the choice of the algorithm to compile the library
% LBG = 1 the algorithm is LBG, in the other case we choose kMedias
% dptzoLBG percentage maximum for the distance of the code vector in the algorithm LBG (see gen_bib) 
% maxiterVQ is the maximum number of iteration to compute the library (condition to stop the algorithm)
% umbralVQ is the threshold condition to stop the library computation (condition to stop the algorithm)
% men=1 or 0 if men =1 the library generates message else no.
% biblio is a matrix parameter for the library computation (we define it here but calculate in the gen_bib function)

LBG=1; % Put LBG=0 to use only the k-mean algorithm.

dpztoLBG=0.1; % This can be changed by dpztoLBG=value

maxiterVQ=40; %maxiterVQ=value to change the maximum of iteration of the VQ

umbralVQ=0.001; % Put umbralVQ=value to change the threshold

men=1; % Put men =0 if we want to avoid messages
% we definate the varaiables for the Libray
biblio=cell(ng,1);
%for ig=1:ng   
%   biblio{ig}=cell(Np(ig),1);
%   for ip=1:Np(ig)
%      dinic=agrup{ig}(ip);dfinal=agrup{ig}(ip+1)-1;
%      dim=dfinal-dinic+1;
%      biblio{ig}{ip}=zeros(Ns{ig}(ip),dim);
%   end
%end

% VARIABLES FOR THE TEST
vTEST=[' TOPNtest salhmm'];
% vTEST is the matrix parameter for the test :
% where TOPNtest is the number of labels to take into account for the multi-labeling during the test phase
% Salhmm is a matrix used to store the probabilities values
% We have the following relations:
% TOPNtest=cell(ng,1);
% salhmm=cell(nc,ng);

TOPNtest=cell(ng,1);
for ig=1:ng   
   TOPNtest{ig}=1.*ones(Np(ig),1); % To take into account two labels: TOPNtest{ig}=2.*ones(Np(ig),1)
end
% Salhmm is a matrix used to store the probabilities values for the output per test/model
salhmm=cell(nc,ng);

% we save the file and the parameters of the HMM
guardar=['save ',fhmm,vDB,vHMM,vVQ,vTEST,' vDB vHMM vVQ vTEST guardar'];
eval([guardar]);