refined doc for ch11. demo needs check

chapter11/Gauss.m

@@ -1,3 +1,4 @@
+% Class for Gaussian distribution used by Dirichlet process
 classdef Gauss
      properties
          n_

chapter11/GaussWishart.m

@@ -1,3 +1,5 @@
+% Class for Gaussian-Wishart distribution used by Dirichlet process
+
 classdef GaussWishart
         properties
          kappa_
@@ -18,23 +20,23 @@
          function obj = clone(obj)
          end
 
-%          function obj = addData(obj, X)
-%              kappa0 = obj.kappa_;
-%              m0 = obj.m_;
-%              nu0 = obj.nu_;
-%              U0 = obj.U_;
-%              
-%              n = size(X,2);
-%              kappa = kappa0+n;
-%              m = (kappa0*m0+sum(X,2))/kappa;
-%              nu = nu0+n;
-%              U = chol(U0'*U0+X*X');
-% 
-%              obj.kappa_ = kappa;
-%              obj.m_ = m;
-%              obj.nu_ = nu;
-%              obj.U_ = U;
-%          end
+         function obj = addData(obj, X)
+             kappa0 = obj.kappa_;
+             m0 = obj.m_;
+             nu0 = obj.nu_;
+             U0 = obj.U_;
+
+             n = size(X,2);
+             kappa = kappa0+n;
+             m = (kappa0*m0+sum(X,2))/kappa;
+             nu = nu0+n;
+             U = chol(U0'*U0+X*X');
+
+             obj.kappa_ = kappa;
+             obj.m_ = m;
+             obj.nu_ = nu;
+             obj.U_ = U;
+         end
 
          function obj = addSample(obj, x)
              kappa = obj.kappa_;

chapter11/demo.m

@@ -1,5 +1,17 @@
 
-%% demo for sequential update Gaussian 
+% demos for ch11
+%% DPGM
+close all; clear;
+d = 2;
+k = 3;
+n = 1000;
+[X,label] = mixGaussRnd(d,k,n);
+plotClass(X,label);
+
+[y,model] = mixGaussGb(X);
+figure
+plotClass(X,y);
+%% Sequential update for Gaussian 
 % close all; clear;
 % d = 2;
 % n = 100;
@@ -11,10 +23,10 @@
 % Sigma = Xo*Xo'/n;
 % p1 = logGauss(x,mu,Sigma);
 % 
-% gauss = Gaussian(X(:,3:end)).addSample(X(:,1)).addSample(X(:,2)).addSample(X(:,3)).delSample(X(:,3));
+% gauss = Gauss(X(:,3:end)).addSample(X(:,1)).addSample(X(:,2)).addSample(X(:,3)).delSample(X(:,3));
 % p2 = gauss.logPdf(x);
-% abs(p1-p2)
-%% Gaussian Wishart
+% maxdiff(p1,p2)
+% %% Sequential update for Gaussian-Wishart
 % close all; clear;
 % d = 2;
 % n = 100;
@@ -33,7 +45,6 @@
 % Xo = bsxfun(@minus,X,m);
 % X0 = m0-m;
 % S = S0+Xo*Xo'+kappa0*(X0*X0');
-% % S = S0+X*X'+kappa0*m0*m0'-kappa*m*m';
 % 
 % v = (nu-d+1);
 % r = (1+1/kappa)/v;
@@ -48,26 +59,5 @@
 %     gw = gw.addSample(X(:,i));
 % end
 % p2 = gw.logPredPdf(x);
-% abs(p1-p2)
-%% Demo for DPGM
-% close all; clear;
-% d = 2;
-% k = 3;
-% n = 1000;
-% [X,label] = mixGaussRnd(d,k,n);
-% plotClass(X,label);
+% maxdiff(p1,p2)
 % 
-% [y,model] = mixGaussGb(X);
-% figure
-% plotClass(X,y);
-%% Demo for DPGM
-close all; clear;
-d = 2;
-k = 3;
-n = 200;
-[X,label] = mixGaussRnd(d,k,n);
-plotClass(X,label);
-
-[y,theta,w,llh] = mixGaussGb(X);
-figure
-plotClass(X,y);

chapter11/dirichletRnd.m

@@ -1,7 +1,10 @@
 function x = dirichletRnd(a, m)
-% Sampling from a Dirichlet distribution.
+% Generate samples from a Dirichlet distribution.
+% Input:
 %   a: k dimensional vector
 %   m: k dimensional mean vector
+% Outpet:
+%   x: generated sample x~Dir(a,m)
 % Written by Mo Chen ([email protected]).
 if nargin == 2
     a = a*m;

chapter11/discreteRnd.m

@@ -1,5 +1,10 @@
 function x = discreteRnd(p, n)
-% Sampling from a discrete distribution (multinomial).
+% Generate samples from a discrete distribution (multinomial).
+% Input:
+%   p: k dimensional probability vector
+%   n: number of samples
+% Ouput:
+%   x: k x n generated samples x~Mul(p)
 % Written by Mo Chen ([email protected]).
 if nargin == 1
     n = 1;

chapter11/gaussRnd.m

@@ -1,5 +1,11 @@
-function x = gaussRnd(mu,Sigma,n)
-% Sampling from a Gaussian distribution.
+function x = gaussRnd(mu, Sigma, n)
+% Generate samples from a Gaussian distribution.
+% Input:
+%   mu: d x 1 mean vector
+%   Sigma: d x d covariance matrix
+%   n: number of samples
+% Outpet:
+%   x: d x n generated sample x~Gauss(mu,Sigma)
 % Written by Mo Chen ([email protected]).
 if nargin == 2
     n = 1;

chapter11/mixDpGb.m

@@ -1,6 +1,16 @@
 function [label, Theta, w, llh] = mixDpGb(X, alpha, theta)
-% Collapsed Gibbs sampling for Dirichlet process (infinite) mixture model (a.k.a.
-% DPGM). Any component model can be used, such as Gaussian
+% Collapsed Gibbs sampling for Dirichlet process (infinite) mixture model. 
+% Any component model can be used, such as Gaussian.
+% Input: 
+%   X: d x n data matrix
+%   alpha: parameter for Dirichlet process prior
+%   theta: class object for prior of component distribution (such as Gauss)
+% Output:
+%   label: 1 x n cluster label
+%   Theta: 1 x k structure of trained components
+%   w: 1 x k component weight vector
+%   llh: loglikelihood
+% Written by Mo Chen ([email protected]).
 n = size(X,2);
 [label,Theta,w] = mixDpGbOl(X,alpha,theta);
 nk = n*w;

chapter11/mixDpGbOl.m

@@ -1,6 +1,15 @@
 function [label, Theta, w, llh] = mixDpGbOl(X, alpha, theta)
-% Online collapsed Gibbs sampling for Dirichlet process (infinite) mixture model (a.k.a.
-% DPGM). Any component model can be used, such as Gaussian
+% Online collapsed Gibbs sampling for Dirichlet process (infinite) mixture model. 
+% Input: 
+%   X: d x n data matrix
+%   alpha: parameter for Dirichlet process prior
+%   theta: class object for prior of component distribution (such as Gauss)
+% Output:
+%   label: 1 x n cluster label
+%   Theta: 1 x k structure of trained components
+%   w: 1 x k component weight vector
+%   llh: loglikelihood
+% Written by Mo Chen ([email protected]).
 n = size(X,2);
 Theta = {};
 nk = [];

chapter11/mixGaussGb.m

@@ -1,6 +1,15 @@
 function [label, Theta, w, llh] = mixGaussGb(X, opt)
-% Collapsed Gibbs sampling for Dirichlet process (infinite) Gaussian mixture model (a.k.a.
-% DPGM). 
+% Collapsed Gibbs sampling for Dirichlet process (infinite) Gaussian mixture model (a.k.a. DPGM). 
+% This is a wrapper function which calls underlying Dirichlet process mixture model.
+% Input: 
+%   X: d x n data matrix
+%   opt(optional): prior parameters
+% Output:
+%   label: 1 x n cluster label
+%   Theta: 1 x k structure of trained Gaussian components
+%   w: 1 x k component weight vector
+%   llh: loglikelihood
+% Written by Mo Chen ([email protected]).
 [d,n] = size(X);
 mu = mean(X,2);
 Xo = bsxfun(@minus,X,mu);

chapter12/demo.m

@@ -35,6 +35,6 @@
 [model, energy] = ppcaVb(X);
 plot(energy);
 % 
-%% factor analysis
+%% Factor analysis
 [W, mu, psi, llh] = fa(X, m);
 plot(llh);