Mercurial > hg > camir-aes2014
diff toolboxes/FullBNT-1.0.7/KPMstats/mixgauss_Mstep.m @ 0:e9a9cd732c1e tip
first hg version after svn
author | wolffd |
---|---|
date | Tue, 10 Feb 2015 15:05:51 +0000 |
parents | |
children |
line wrap: on
line diff
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/toolboxes/FullBNT-1.0.7/KPMstats/mixgauss_Mstep.m Tue Feb 10 15:05:51 2015 +0000 @@ -0,0 +1,106 @@ +function [mu, Sigma] = mixgauss_Mstep(w, Y, YY, YTY, varargin) +% MSTEP_COND_GAUSS Compute MLEs for mixture of Gaussians given expected sufficient statistics +% function [mu, Sigma] = Mstep_cond_gauss(w, Y, YY, YTY, varargin) +% +% We assume P(Y|Q=i) = N(Y; mu_i, Sigma_i) +% and w(i,t) = p(Q(t)=i|y(t)) = posterior responsibility +% See www.ai.mit.edu/~murphyk/Papers/learncg.pdf. +% +% INPUTS: +% w(i) = sum_t w(i,t) = responsibilities for each mixture component +% If there is only one mixture component (i.e., Q does not exist), +% then w(i) = N = nsamples, and +% all references to i can be replaced by 1. +% YY(:,:,i) = sum_t w(i,t) y(:,t) y(:,t)' = weighted outer product +% Y(:,i) = sum_t w(i,t) y(:,t) = weighted observations +% YTY(i) = sum_t w(i,t) y(:,t)' y(:,t) = weighted inner product +% You only need to pass in YTY if Sigma is to be estimated as spherical. +% +% Optional parameters may be passed as 'param_name', param_value pairs. +% Parameter names are shown below; default values in [] - if none, argument is mandatory. +% +% 'cov_type' - 'full', 'diag' or 'spherical' ['full'] +% 'tied_cov' - 1 (Sigma) or 0 (Sigma_i) [0] +% 'clamped_cov' - pass in clamped value, or [] if unclamped [ [] ] +% 'clamped_mean' - pass in clamped value, or [] if unclamped [ [] ] +% 'cov_prior' - Lambda_i, added to YY(:,:,i) [0.01*eye(d,d,Q)] +% +% If covariance is tied, Sigma has size d*d. +% But diagonal and spherical covariances are represented in full size. + +[cov_type, tied_cov, clamped_cov, clamped_mean, cov_prior, other] = ... + process_options(varargin,... + 'cov_type', 'full', 'tied_cov', 0, 'clamped_cov', [], 'clamped_mean', [], ... + 'cov_prior', []); + +[Ysz Q] = size(Y); +N = sum(w); +if isempty(cov_prior) + %cov_prior = zeros(Ysz, Ysz, Q); + %for q=1:Q + % cov_prior(:,:,q) = 0.01*cov(Y(:,q)'); + %end + cov_prior = repmat(0.01*eye(Ysz,Ysz), [1 1 Q]); +end +%YY = reshape(YY, [Ysz Ysz Q]) + cov_prior; % regularize the scatter matrix +YY = reshape(YY, [Ysz Ysz Q]); + +% Set any zero weights to one before dividing +% This is valid because w(i)=0 => Y(:,i)=0, etc +w = w + (w==0); + +if ~isempty(clamped_mean) + mu = clamped_mean; +else + % eqn 6 + %mu = Y ./ repmat(w(:)', [Ysz 1]);% Y may have a funny size + mu = zeros(Ysz, Q); + for i=1:Q + mu(:,i) = Y(:,i) / w(i); + end +end + +if ~isempty(clamped_cov) + Sigma = clamped_cov; + return; +end + +if ~tied_cov + Sigma = zeros(Ysz,Ysz,Q); + for i=1:Q + if cov_type(1) == 's' + % eqn 17 + s2 = (1/Ysz)*( (YTY(i)/w(i)) - mu(:,i)'*mu(:,i) ); + Sigma(:,:,i) = s2 * eye(Ysz); + else + % eqn 12 + SS = YY(:,:,i)/w(i) - mu(:,i)*mu(:,i)'; + if cov_type(1)=='d' + SS = diag(diag(SS)); + end + Sigma(:,:,i) = SS; + end + end +else % tied cov + if cov_type(1) == 's' + % eqn 19 + s2 = (1/(N*Ysz))*(sum(YTY,2) + sum(diag(mu'*mu) .* w)); + Sigma = s2*eye(Ysz); + else + SS = zeros(Ysz, Ysz); + % eqn 15 + for i=1:Q % probably could vectorize this... + SS = SS + YY(:,:,i)/N - mu(:,i)*mu(:,i)'; + end + if cov_type(1) == 'd' + Sigma = diag(diag(SS)); + else + Sigma = SS; + end + end +end + +if tied_cov + Sigma = repmat(Sigma, [1 1 Q]); +end +Sigma = Sigma + cov_prior;