annotate toolboxes/FullBNT-1.0.7/bnt/examples/static/Zoubin/mfa.m @ 0:e9a9cd732c1e tip

first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
rev   line source
wolffd@0 1 % function [Lh,Ph,Mu,Pi,LL]=mfa(X,M,K,cyc,tol);
wolffd@0 2 %
wolffd@0 3 % Maximum Likelihood Mixture of Factor Analysis using EM
wolffd@0 4 %
wolffd@0 5 % X - data matrix
wolffd@0 6 % M - number of mixtures (default 1)
wolffd@0 7 % K - number of factors in each mixture (default 2)
wolffd@0 8 % cyc - maximum number of cycles of EM (default 100)
wolffd@0 9 % tol - termination tolerance (prop change in likelihood) (default 0.0001)
wolffd@0 10 %
wolffd@0 11 % Lh - factor loadings
wolffd@0 12 % Ph - diagonal uniquenesses matrix
wolffd@0 13 % Mu - mean vectors
wolffd@0 14 % Pi - priors
wolffd@0 15 % LL - log likelihood curve
wolffd@0 16 %
wolffd@0 17 % Iterates until a proportional change < tol in the log likelihood
wolffd@0 18 % or cyc steps of EM
wolffd@0 19
wolffd@0 20 function [Lh, Ph, Mu, Pi, LL] = mfa(X,M,K,cyc,tol)
wolffd@0 21
wolffd@0 22 if nargin<5 tol=0.0001; end;
wolffd@0 23 if nargin<4 cyc=100; end;
wolffd@0 24 if nargin<3 K=2; end;
wolffd@0 25 if nargin<2 M=1; end;
wolffd@0 26
wolffd@0 27 N=length(X(:,1));
wolffd@0 28 D=length(X(1,:));
wolffd@0 29 tiny=exp(-700);
wolffd@0 30
wolffd@0 31 %rand('state',0);
wolffd@0 32
wolffd@0 33 fprintf('\n');
wolffd@0 34
wolffd@0 35 if (M==1)
wolffd@0 36 [Lh,Ph,LL]=ffa(X,K,cyc,tol);
wolffd@0 37 Mu=mean(X);
wolffd@0 38 Pi=1;
wolffd@0 39 else
wolffd@0 40 if N==1
wolffd@0 41 mX = X;
wolffd@0 42 else
wolffd@0 43 mX=mean(X);
wolffd@0 44 end
wolffd@0 45 cX=cov(X);
wolffd@0 46 scale=det(cX)^(1/D);
wolffd@0 47 randn('state',0);
wolffd@0 48 Lh=randn(D*M,K)*sqrt(scale/K);
wolffd@0 49 Ph=diag(cX)+tiny;
wolffd@0 50 Pi=ones(M,1)/M;
wolffd@0 51 %randn('state',0);
wolffd@0 52 Mu=randn(M,D)*sqrtm(cX)+ones(M,1)*mX;
wolffd@0 53 oldMu=Mu;
wolffd@0 54 I=eye(K);
wolffd@0 55
wolffd@0 56 lik=0;
wolffd@0 57 LL=[];
wolffd@0 58
wolffd@0 59 H=zeros(N,M); % E(w|x)
wolffd@0 60 EZ=zeros(N*M,K);
wolffd@0 61 EZZ=zeros(K*M,K);
wolffd@0 62 XX=zeros(D*M,D);
wolffd@0 63 s=zeros(M,1);
wolffd@0 64 const=(2*pi)^(-D/2);
wolffd@0 65 %%%%%%%%%%%%%%%%%%%%
wolffd@0 66 for i=1:cyc;
wolffd@0 67
wolffd@0 68 %%%% E Step %%%%
wolffd@0 69
wolffd@0 70 Phi=1./Ph;
wolffd@0 71 Phid=diag(Phi);
wolffd@0 72 for k=1:M
wolffd@0 73 Lht=Lh((k-1)*D+1:k*D,:);
wolffd@0 74 LP=Phid*Lht;
wolffd@0 75 MM=Phid-LP*inv(I+Lht'*LP)*LP';
wolffd@0 76 dM=sqrt(det(MM));
wolffd@0 77 Xk=(X-ones(N,1)*Mu(k,:));
wolffd@0 78 XM=Xk*MM;
wolffd@0 79 H(:,k)=const*Pi(k)*dM*exp(-0.5*rsum(XM.*Xk));
wolffd@0 80 EZ((k-1)*N+1:k*N,:)=XM*Lht;
wolffd@0 81 end;
wolffd@0 82
wolffd@0 83 Hsum=rsum(H);
wolffd@0 84 oldlik=lik;
wolffd@0 85 lik=sum(log(Hsum+(Hsum==0)*exp(-744)));
wolffd@0 86
wolffd@0 87 Hzero=(Hsum==0); Nz=sum(Hzero);
wolffd@0 88 H(Hzero,:)=tiny*ones(Nz,M)/M;
wolffd@0 89 Hsum(Hzero)=tiny*ones(Nz,1);
wolffd@0 90
wolffd@0 91 H=rdiv(H,Hsum);
wolffd@0 92 s=csum(H);
wolffd@0 93 s=s+(s==0)*tiny;
wolffd@0 94 s2=sum(s)+tiny;
wolffd@0 95
wolffd@0 96 for k=1:M
wolffd@0 97 kD=(k-1)*D+1:k*D;
wolffd@0 98 Lht=Lh(kD,:);
wolffd@0 99 LP=Phid*Lht;
wolffd@0 100 MM=Phid-LP*inv(I+Lht'*LP)*LP';
wolffd@0 101 Xk=(X-ones(N,1)*Mu(k,:));
wolffd@0 102 XX(kD,:)=rprod(Xk,H(:,k))'*Xk/s(k);
wolffd@0 103 beta=Lht'*MM;
wolffd@0 104 EZZ((k-1)*K+1:k*K,:)=I-beta*Lht +beta*XX(kD,:)*beta';
wolffd@0 105 end;
wolffd@0 106
wolffd@0 107 %%%% log likelihood %%%%
wolffd@0 108
wolffd@0 109 LL=[LL lik];
wolffd@0 110 fprintf('cycle %g \tlog likelihood %g ',i,lik);
wolffd@0 111
wolffd@0 112 if (i<=2)
wolffd@0 113 likbase=lik;
wolffd@0 114 elseif (lik<oldlik)
wolffd@0 115 fprintf(' violation');
wolffd@0 116 elseif ((lik-likbase)<(1 + tol)*(oldlik-likbase)||~isfinite(lik))
wolffd@0 117 break;
wolffd@0 118 end;
wolffd@0 119
wolffd@0 120 fprintf('\n');
wolffd@0 121
wolffd@0 122 %%%% M Step %%%%
wolffd@0 123
wolffd@0 124 % means and covariance structure
wolffd@0 125
wolffd@0 126 Ph=zeros(D,1);
wolffd@0 127 for k=1:M
wolffd@0 128 kD=(k-1)*D+1:k*D;
wolffd@0 129 kK=(k-1)*K+1:k*K;
wolffd@0 130 kN=(k-1)*N+1:k*N;
wolffd@0 131
wolffd@0 132 T0=rprod(X,H(:,k));
wolffd@0 133 T1=T0'*[EZ(kN,:) ones(N,1)];
wolffd@0 134 XH=EZ(kN,:)'*H(:,k);
wolffd@0 135 T2=inv([s(k)*EZZ(kK,:) XH; XH' s(k)]);
wolffd@0 136 T3=T1*T2;
wolffd@0 137 Lh(kD,:)=T3(:,1:K);
wolffd@0 138 Mu(k,:)=T3(:,K+1)';
wolffd@0 139 T4=diag(T0'*X-T3*T1')/s2;
wolffd@0 140 Ph=Ph+T4.*(T4>0);
wolffd@0 141 end;
wolffd@0 142
wolffd@0 143 Phmin=exp(-700);
wolffd@0 144 Ph=Ph.*(Ph>Phmin)+(Ph<=Phmin)*Phmin; % to avoid zero variances
wolffd@0 145
wolffd@0 146 % priors
wolffd@0 147 Pi=s'/s2;
wolffd@0 148
wolffd@0 149 end;
wolffd@0 150 fprintf('\n');
wolffd@0 151 end;
wolffd@0 152
wolffd@0 153