Daniel@0: % Factor analysis
Daniel@0: % Z -> X,  Z in R^k, X in R^D, k << D (high dimensional observations explained by small source)
Daniel@0: % Z ~ N(0,I),   X|Z ~ N(L z, Psi), where Psi is diagonal.
Daniel@0: %
Daniel@0: % Mixtures of FA
Daniel@0: % Now X|Z,W=i ~ N(mu(i) + L(i) Z, Psi(i))
Daniel@0: %
Daniel@0: % We compare to Zoubin Ghahramani's code.
Daniel@0: 
Daniel@0: randn('state', 0);
Daniel@0: max_iter = 3;
Daniel@0: M = 2;
Daniel@0: k = 3;
Daniel@0: D = 5;
Daniel@0: 
Daniel@0: n = 5;
Daniel@0: X1 = randn(n, D);
Daniel@0: X2 = randn(n, D) + 2; % move the mean to (2,2,2...)
Daniel@0: X = [X1; X2];
Daniel@0: N = size(X, 1);
Daniel@0: 
Daniel@0: % initialise as in mfa
Daniel@0: tiny=exp(-700);
Daniel@0: mX = mean(X);
Daniel@0: cX=cov(X);
Daniel@0: scale=det(cX)^(1/D);
Daniel@0: randn('state',0); % must reset seed here so initial params are identical to mfa
Daniel@0: L0=randn(D*M,k)*sqrt(scale/k);
Daniel@0: W0 = permute(reshape(L0, [D M k]), [1 3 2]); % use D,K,M 
Daniel@0: Psi0=diag(cX)+tiny;
Daniel@0: Pi0=ones(M,1)/M;
Daniel@0: Mu0=randn(M,D)*sqrtm(cX)+ones(M,1)*mX;
Daniel@0: 
Daniel@0: [Lh1, Ph1, Mu1, Pi1, LL1] = mfa(X,M,k,max_iter);
Daniel@0: Lh1 = permute(reshape(Lh1, [D M k]), [1 3 2]); % use D,K,M 
Daniel@0: 
Daniel@0: 
Daniel@0: ns = [M k D];
Daniel@0: dag = zeros(3);
Daniel@0: dag(1,3) = 1;
Daniel@0: dag(2,3) = 1;
Daniel@0: dnodes = 1;
Daniel@0: onodes = 3;
Daniel@0: 
Daniel@0: bnet = mk_bnet(dag, ns, 'discrete', dnodes, 'observed', onodes);
Daniel@0: bnet.CPD{1} = tabular_CPD(bnet, 1, Pi0);
Daniel@0: 
Daniel@0: %bnet.CPD{2} = gaussian_CPD(bnet, 2, zeros(k, 1), eye(k), [], 'diag', 'untied', 'clamp_mean',  'clamp_cov');
Daniel@0: 
Daniel@0: bnet.CPD{2} = gaussian_CPD(bnet, 2, 'mean', zeros(k, 1), 'cov', eye(k), 'cov_type', 'diag', ...
Daniel@0: 			   'cov_prior_weight', 0, 'clamp_mean', 1, 'clamp_cov', 1);
Daniel@0: 
Daniel@0: %bnet.CPD{3} = gaussian_CPD(bnet, 3, Mu0', repmat(diag(Psi0), [1 1 M]), W0, 'diag', 'tied');
Daniel@0: 
Daniel@0: bnet.CPD{3} = gaussian_CPD(bnet, 3, 'mean', Mu0', 'cov', repmat(diag(Psi0), [1 1 M]), ...
Daniel@0: 			   'weights', W0, 'cov_type', 'diag', 'cov_prior_weight', 0, 'tied_cov', 1);
Daniel@0: 
Daniel@0: engine = jtree_inf_engine(bnet);
Daniel@0: evidence = cell(3, N);
Daniel@0: evidence(3,:) = num2cell(X', 1);
Daniel@0: 
Daniel@0: [bnet2, LL2, engine2] = learn_params_em(engine, evidence, max_iter);
Daniel@0: 
Daniel@0: s = struct(bnet2.CPD{1});
Daniel@0: Pi2 = s.CPT(:);
Daniel@0: s = struct(bnet2.CPD{3});
Daniel@0: Mu2 = s.mean;
Daniel@0: W2 = s.weights;
Daniel@0: Sigma2 = s.cov;
Daniel@0: 
Daniel@0: 
Daniel@0: % Compare to Zoubin's code
Daniel@0: assert(approxeq(LL1,LL2));
Daniel@0: for i=1:M
Daniel@0:   assert(approxeq(W2(:,:,i), Lh1(:,:,i)));
Daniel@0:   assert(approxeq(Sigma2(:,:,i), diag(Ph1)));
Daniel@0:   assert(approxeq(Mu2(:,i), Mu1(i,:)));
Daniel@0:   assert(approxeq(Pi2(:), Pi1(:)));
Daniel@0: end
Daniel@0: