git-svn-id: https://pmtk3.googlecode.com/svn/trunk@2791 b6abd7f4-f95b…

…-11de-aa3c-59de0406b4f5

config-local.txt

@@ -1,17 +1,4 @@
-% This file stores hard coded paths and configuration variables
-% Use getConfigValue(varname) to access a value, e.g. 
-% getConfigValue('PMTKsupportLink')
-%
-% If a config-local.txt file is found, its values will override these.
-%
-% PMTKmetaDirs             meta, docs, tmp
-% PMTKred                  #990000
-% PMTKauthors              Kevin Murphy, Matt Dunham
-% PMTKcodeDirs             toolbox, demos, localUtil, matlabTools
-% PMTKgvizPath             C:\Program Files\Graphviz2.26\bin
-% PMTKlocalWikiPath        C:\path\to\pmtk3Wiki
 % PMTKlocalDataPath        /Users/kpmurphy/GoogleCode/pmtkdata
 % PMTKlocalSupportPath    /Users/kpmurphy/GoogleCode/pmtksupport
-% PMTKpmlBookSource        /Users/kpmurphy/Dropbox/MLbook/Text
-% PMTKpmlFigures         /Users/kpmurphy/Dropbox/MLbook/Figures
-% PMTKlightSpeedDir        lightspeed2.3
+% PMTKgvizPath             C:\Program Files\Graphviz2.26\bin
+

config.txt

@@ -4,10 +4,11 @@
 % If a config-local.txt file is found, its values will override these.
 %
 %
-% Used by initPmtk3.m (UPDATE THESE DIRECTORIES TO MATCH YOUR INSTALLATION)
+% Used by initPmtk3.m 
 % PMTKgvizPath             C:\Program Files\Graphviz2.26\bin  
 % PMTKlocalDataPath        path/to/data
 % PMTKlocalSupportPath     path/to/support
+% PMTKlightSpeedDir        lightspeed2.3
 %
 % Used by generateAuthorReport ,  pmtkTagReport, etc
 % PMTKred                  #990000
@@ -16,7 +17,8 @@
 % PMTKmetaDirs             meta, docs, tmp
 %
 % Used by pmlFigureCodeReport
-% PMTKpmlBookSource        C:\path\to\PML\Text
+% PMTKpmlBookSource        /Users/kpmurphy/Dropbox/MLbook/Text
+% PMTKpmlFigures         /Users/kpmurphy/Dropbox/MLbook/Figures
 %
 % used by publishDemos.m, generateSynopses.m
 % PMTKlocalWikiPath        C:\path\to\pmtk3Wiki

demos/binaryFaDemoNewsgroups.m

@@ -1,6 +1,8 @@
 % Demo of factor analysis applied to binary newsgroups bag of words
 % We compute 2d embedding
 
+%PMTKreallySlow
+
 % This file is from pmtk3.googlecode.com
 
 requireStatsToolbox; % cmdscale
@@ -45,12 +47,13 @@
   m = m + 1;
   methods(m).modelname = 'JJ';
   methods(m).fitFn = @(data) binaryFAfit(data, K, 'maxIter', 6, ...
-    'verbose', truesize, 'computeLoglik', false);
+    'verbose', true, 'computeLoglik', false);
   methods(m).infFn = @(model, labels) binaryFAinferLatent(model, labels);
   methods(m).nlatent = K;
 end
 
-Ks = [];
+
+%Ks = [];
 for kk=1:numel(Ks)
   K = Ks(kk);
   m = m + 1;
@@ -61,6 +64,7 @@
   methods(m).nlatent = K;
 end
 
+
 Nmethods = numel(methods);
 for m=1:Nmethods
   fitFn = methods(m).fitFn;
@@ -87,7 +91,7 @@
   end
   ndx = 1:1:D;
   for d=ndx(:)'
-    text(mdsCoords(d,1), mdsCoords(d,2), wordlist{d}, 'fontsize', 10);
+    %text(mdsCoords(d,1), mdsCoords(d,2), wordlist{d}, 'fontsize', 10);
   end
   nlatent = methods(m).nlatent;
   title(sprintf('L=%d, N=%d, method = %s', nlatent, N, methodname))

demos/catFAdemo.m

@@ -0,0 +1,117 @@
+function catFAdemo()
+% Factor analysis with categorical and continuous data
+% We reproduce the demo from
+% http://www.cs.ubc.ca/~emtiyaz/software/mixedDataFA.html
+% This just check the code runs, it is not intrinsically interesting.
+
+clear all
+setSeed(16)
+% generate data - data is stored along columns not rows!
+[trainData,testData,simParams] = makeSimDataMixedDataFA(100);
+nClass = simParams.nClass;
+% introduce missing variables in train data
+missProb = 0.1;
+trainData.continuousTruth = trainData.continuous;
+trainData.discreteTruth = trainData.discrete;
+[D,N] = size(trainData.continuous);
+miss = rand(D,N)<missProb;
+trainData.continuous(miss) = NaN;
+[D,N] = size(trainData.discrete);
+miss = rand(D,N)<missProb;
+trainData.discrete(miss) = NaN;
+
+
+Dz = 2;
+[model, loglikTrace] = catFAfit(trainData.discrete',  trainData.continuous', Dz);
+
+[mu, Sigma, loglik] = catFAinferLatent(model, testData.discrete', testData.continuous')
+
+[predD, predC] = catFApredictMissing(model,testData.discrete', testData.continuous')
+
+
+end
+
+function [trainData, testData, params] = makeSimDataMixedDataFA(N)
+% [TRAINDATA, TESTDATA, PARAMS] = makeSimDataMixedDataFA(N) makes simulated data
+% for mixedDataFA with N data points
+%
+% Written by Emtiyaz, CS, UBC
+% Modified on June 09, 2010
+
+  missProb = 0.3;
+  Dz = 2;
+  mean_ = [5 -5]';% -5 -5]';
+  covMat = eye(Dz);
+  z = repmat(mean_,1,N) + chol(covMat)*randn(Dz,N);
+  Dc = 5;
+  nClass = 2*ones(10,1);%[3 2 4];
+  noiseCovMat = 0.01*eye(Dc);%abs(diag(randn(Dc,1)));
+  Bc = rand(Dc,Dz);
+
+  % generate data
+  yc = Bc*z + chol(noiseCovMat)*randn(Dc,N);
+  %yc = yc - repmat(mean(yc,2), 1, N);
+
+  Bm = [];
+  for c = 1:length(nClass)
+    Bmc = rand(nClass(c)-1,Dz);
+    p = [exp(Bmc*z); ones(1,N)];
+    pMult = p./repmat(sum(p,1),nClass(c),1);
+    yd(c,:) = sum(repmat([1:nClass(c)],N,1).*mnrnd(1,pMult'),2);
+    Bm = [Bm; Bmc];
+  end
+
+  % model parameters structure
+  params.nClass = nClass;
+  params.mean = mean_;
+  params.covMat = covMat;
+  params.noiseCovMat = noiseCovMat;
+  params.betaMult = Bm;
+  params.betaCont = Bc;
+
+  % split test and train data
+  ratio = .7;
+  [trainData, testData] = splitData(yc,yd,ratio);
+
+%{
+  % introduce missing variables in test data 
+  testData.continuousTruth = testData.continuous;
+  testData.discreteTruth = testData.discrete;
+  [D,N] = size(testData.continuous);
+  miss = rand(D,N)<missProb;
+  testData.continuous(miss) = NaN;
+  [D,N] = size(testData.discrete);
+  miss = rand(D,N)<missProb;
+  testData.discrete(miss) = NaN;
+  %}
+
+
+end
+
+function [trainData, testData, idx] = splitData(yc, yd, ratio) 
+% splits data into training and testing set
+% yc is the continuous data, yd is discrete data
+% ratio is the split ratio
+
+  [Dc,Nc] = size(yc);
+  [Dd,Nd] = size(yd);
+  N = max(Nc,Nd);
+  nTrain = ceil(ratio*N);
+  idx = randperm(N);
+  if Dc>0
+    testData.continuous = yc(:,idx(nTrain+1:end));
+    trainData.continuous = yc(:,idx(1:nTrain));
+  else
+    testData.continuous = [];
+    trainData.continuous = [];
+  end
+  if Dd>0
+    testData.discrete = yd(:,idx(nTrain+1:end));
+    trainData.discrete = yd(:,idx(1:nTrain));
+  else
+    testData.discrete = [];
+    trainData.discrete = [];
+  end
+
+end
+

demos/classificationShootout.m

@@ -22,6 +22,10 @@
 nDataSets = numel(dataSets);
 
 methods = {'SVM', 'RVM', 'SMLR', 'RMLR'};
+if ~svmInstalled
+  methods = {'SVM', 'RVM', 'SMLR', 'RMLR'};
+end
+
 nMethods = numel(methods);
 results = cell(nDataSets, nMethods);
 for i=1:nDataSets

demos/classificationShootoutCvLambdaOnly.m

@@ -23,6 +23,9 @@
 nDataSets = numel(dataSets);
 
 methods = {'SVM', 'RVM', 'SMLR', 'RMLR'};
+if ~svmInstalled
+  methods = {'SVM', 'RVM', 'SMLR', 'RMLR'};
+end
 nMethods = numel(methods);
 results = cell(nDataSets, nMethods);
 for i=1:nDataSets

demos/demoMinfuncHighdim.m

@@ -9,23 +9,22 @@ function demoMinfuncHighdim()
 %methods = {'sd', 'cg',  'bb'};
 methods = {'sd', 'cg',  'bb', 'newton', 'newtoncg', 'bfgs', 'lbfgs'};
 %methods = {'sd', 'cg',  'bb', 'newton', 'newtoncg', 'bfgs', 'lbfgs', 'newton0', 'newton0lbfgs'};
-[styles] =  plotColors();
+[styles, colors, symbols, plotstyles]  = plotColors();
 
-for i=1:length(methods)
- names{i} = methods{i}(1:min(5,length(methods{i})))
-end
 names = {'sd', 'cg',  'bb', 'n', 'ncg', 'bfgs', 'lbfgs'};
 
 
 seed = 1; setSeed(seed);
 dims = [10 100 500];
+dims = [10 100];
+clear str
 for j=1:length(dims)
   d = dims(j);
   x0 = randn(d,1);
   figure;
   for i=1:length(methods)
     hold on
-    [fx(i,j), fcount(i,j),tim(i,j)] = helper(x0, methods{i}, styles{i});
+    [fx(i,j), fcount(i,j),tim(i,j)] = helper(x0, methods{i}, plotstyles{i});
 
     str{i} = sprintf('%s (f* = %5.2f, %d feval, %3.2f sec)', ...
       names{i}, fx(i,j), fcount(i,j), tim(i,j));
@@ -55,6 +54,7 @@ function demoMinfuncHighdim()
   bar(tim(:,j))
   set(gca,'xticklabel',names)
   title(sprintf('time d=%d', d))
+  drawnow
 end
 printPmtkFigure minfuncRosenBar
 
@@ -75,7 +75,7 @@ function demoMinfuncHighdim()
 options.display = 'none';
 options.maxFunEvals = 500;
 options.tolFun = 1e-2;
-options.outputFn = @optimstore;
+options.outputFcn = @optimstore;
 options.Method = method;
 options.HessianModify = 1;
 if strcmp(method, 'newtoncg')

demos/demoRosenConstrained.m

@@ -9,7 +9,7 @@ function demoRosenConstrained()
 % minimize 2d rosenbrock st x1^2 + x^2 <= 1
 % Example from p1-8 of Mathworks Optimization Toolbox manual
 
-requiresOptimToolbox
+requireOptimToolbox
 
 xstart = [-1 2];
  % Hessian is ignored by quasi-Newton so we use interior point

demos/demoRosenHighDim.m

@@ -3,14 +3,14 @@
 %%
 %PMTKinteractive
 %PMTKneedsOptimToolbox
+
+
+% This file is from pmtk3.googlecode.com
+
 % "A note on the extended rosenbrock function" Evol. Comp. 2006
 % claims that for d=4 to 30 dims there are 2 local minima, at [1,1,...1] and
 % and near [-1,1,...,1].
 % Let us verify this for d=4 and d=5
-
-% This file is from pmtk3.googlecode.com
-
-
 %xstart = [-0.77565923 0.61309337 0.38206285 0.14597202]';
 xstart = [-0.96205109 0.93573953 0.88071386 0.77787813 0.60509438]';
 [f g H] = rosenbrock(xstart);
@@ -22,21 +22,25 @@
 % So the claim seems dubious...
 %
 %%
+
 requireOptimToolbox
 x = rand(10,1);
 [f g H] = rosenbrock(x);
 figure;spy(H)
 title(sprintf('sparsity pattern of Hessian for extended Rosenbrock'))
-
+  printPmtkFigure rosen10dSpy
+
+
 % Now compare speed of using Hessian or approximating it
 
-d = 200;
+d = 20; % 200;
 seed = 0;
 setSeed(seed);
 xstart = 2*rand(d,1)-1;
-opts = optimset('display', 'off', 'DerivativeCheck', 'on');
+opts = optimset('display', 'off', 'DerivativeCheck', 'off');
 [f g H] = rosenbrock(xstart);
 
+clear options
 options{1} = optimset(opts, 'GradObj', 'on', 'Hessian', 'on'); % analtyic Hessian
 options{2} = optimset(opts, 'GradObj', 'on', 'Hessian', []); % dense numerical Hessian
 options{3} = optimset(opts, 'GradObj', 'on', 'HessPattern', H); % sparse numerical Hessian
@@ -52,5 +56,3 @@
 
 final
 t
-
-printPmtkFigure rosen10dSpy

demos/discrimAnalysisHeightWeightDemo.m

@@ -4,6 +4,7 @@
 
 % This file is from pmtk3.googlecode.com
 
+clear all
 rawdata = loadData('heightWeight');
 data.Y = rawdata(:,1); % 1=male, 2=female
 data.X = [rawdata(:,2) rawdata(:,3)]; % height, weight

demos/gpcDemo2d.m

@@ -13,6 +13,7 @@
 x = [x1 x2]';
 y = [repmat(-1,1,n1) repmat(1,1,n2)]';
 [t1 t2] = meshgrid(-4:0.1:4,-4:0.1:4);
+t = [t1(:) t2(:)]; % test
 
 % training
 loghyper = [0; 0]; % initial guess

demos/hclustYeastDemo.m

@@ -1,12 +1,11 @@
 %% Hierarchical Clustering Demo
 %PMTKneedsStatsToobox cluster, pdist, linkage
-%PMTKneedsBioToolbox clustergram
 %%
 
 % This file is from pmtk3.googlecode.com
 
 requireStatsToolbox
-requireBioinfoToolbox
+%requireBioinfoToolbox
 loadData('yeastData310') % 'X', 'genes', 'times');
 
 corrDist = pdist(X, 'corr');
@@ -22,11 +21,12 @@
 printPmtkFigure('clusterYeastHier16')
 
 
+if bioinfoToolboxInstalled
 figure(5);clf
 clustergram(X(:,2:end),'RowLabels',genes, 'ColumnLabels',times(2:end))
 title('hierarchical clustering')
 printPmtkFigure('clusterYeastRowPerm')
-
+end
 
 figure(6); clf
 dendrogram(linkage(corrDist, 'average'));

demos/hmmNbestDemo.m

@@ -3,6 +3,8 @@
 % Nilsson and Goldberger, IJCAI 2001
 % We do the example in sec 3.2
 
+%PMTKinprogress
+
 initDist = [0.6; 0.4];
 transmat = [0.6 0.4; 0.2 0.8];
 obsmat = [0.9 0.1; 0.3 0.7];