mainfile_labelme.m

function [] = mainfile_labelme (MAXCOUNT, MaxFun, MAXESTEPITER, MAXMSTEPITER, filename, p1, p2, p3, k2, option, troption, svmoptionval, svmcval, minvtopic, pathname, otherindex)


%% 1,2,3,4,5,6,7 debugging done

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% file for simulating synthetic data and checking the performance of the model on the same
% call variational_EM() directly with appropriate arguments for any other dataset
%%%%%%%%%%%
% Input:
% MAXCOUNT: maximum number of EM iteration
% MaxFun: maximum number of function evaluations for fmincon
% MAXESTEPITER: maximum number of E-step iteration
% MAXMSTEPITER: maximum number of M-step iteration
% filename: name of the file containing conference names
% p1: % of training data from source classes
% p2: % of training data from target class
% option: 1,2,3,4,5,6,7,8 -- use 4 for DSLDA; 1 for unsupervised LDA, 2 for labeled LDA, 3 for Med-LDA, 
% 5 for DSLDA-NSLT-Ray, 6 for DSLDA-NSLT-Ayan, 7 for DSLDA-OSST, 8-NPDSLDA.
% troption: 1 for test on training data, 0 for test on test data
% svmoptionval: should be 4 for multi-class svm
% svmcval: value of the margin error penalizer term c in svm
% pathname: path to the saved files of the form "<conference name>_train.mat" and "<conference name>_test.mat"
% otherindex: any string that distinguishes the current run from others
%%%%%%%%%%%
% output: saved in savefilename -- look for savefilename towards end of the code
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% example input:
%% mainfile_labelme (5, 2, 4, 4, 'LabelMeCategoryNames.txt', 0.3, 0.3, 120, 4, 0, 4, 10, 4, '/lusr/u/ayan/MLDisk/labelme_files/', '3rdtry');
%% mainfile_labelme (6, 3, 4, 4, 'LabelMeCategoryNames.txt', 0.4, 0.4, 100, 4, 0, 4, 10, 3, '/lusr/u/ayan/MLDisk/labelme_files/', '3rdtry');
%% mainfile_labelme (7, 5, 5, 5, 'LabelMeCategoryNames.txt', 0.5, 0.5, 0.3, 100, 4, 0, 4, 1, 4, '/lusr/u/ayan/MLDisk/labelme_files/', '3rdtry');
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

format long;
warning('off');
clc;

if(isdeployed)
    MAXCOUNT = str2num(MAXCOUNT);
    MaxFun = str2num(MaxFun);
    MAXESTEPITER = str2num(MAXESTEPITER);
    MAXMSTEPITER = str2num(MAXMSTEPITER);
    p1 = str2num(p1);
    p2 = str2num(p2);
    k2 = str2num(k2);
    svmoptionval = str2num(svmoptionval);
    svmcval = str2num(svmcval);
    option   = str2num(option);
    troption = str2num(troption);
    minvtopic = str2num(minvtopic);
end

if(~isdeployed)
    %% these C++ files should be compiled first
    mex sum_phi.cpp;
    mex update_beta_cpp.cpp;
    mex update_phi_cpp.cpp;
    addpath(genpath('/lusr/u/ayan/Documents/DSLDA_SDM/medlda/'));
    addpath(genpath('/lusr/u/ayan/Documents/DSLDA_SDM/liblinear-v0.1/'));
    addpath(genpath('/lusr/u/ayan/Documents/DSLDA_SDM/forconfdata/'));
    addpath(genpath('/lusr/u/ayan/Documents/DSLDA_SDM/siftpp/'));
end

p    = 0.9; %(should be > 0) multiplier for Dirichlet Distribution; change this to vary initialization

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% data is a structure containing the following fields:
% annotations: N*k1 binary matrix with 1 indicating presence of an annotation in a document, N indicates number of documents and k1 indicates maximum number of visible topics/annotations
% windex: an N*1 cell array with w{n} containing the indicies of the words in the vocabulary appearing in nth document.
% wcount: an N*1 cell array with w{n} containing the counts of the words in the vocabulary appearing in nth document.
% classlabels: N*1 matrix with each element indicating the class label of corresponding document
% Y: number of classes (at least one document from each of the classes should be present in the training data)
% k1: number of visible topics/annotations
% k2: number of latent topics -- note that K = k1+K2
% nwordspdoc: N*1 matrix with each element indicating the number of words in corresponding document
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% initlaize performance structures to null

svmperf = [];
trperf = []; 
testperf = []; 
trperfmedLDA = []; 
testperfmedLDA = [];
trperfDSLDANSLT1 = [];
testperfDSLDANSLT1 = [];
trperfDSLDANSLT2 = [];
testperfDSLDANSLT2 = [];
trperfDSLDAOSST = [];
testperfDSLDAOSST = [];
trperfmedLDAova = [];
testperfmedLDAova = [];

%% get the data in proper format
[trdata, testdata, selindex, classnames] = get_LabelMedata(pathname, p1, p2, p3, filename, k2, minvtopic);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

disp('*************** SVM starts ****************');
%% run SVM using the liblinear package
[svmperf] = getLiblinearperf(pathname, selindex, classnames, svmoptionval, svmcval);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

disp('*************** medLDA-OVA starts ****************');
%% run medLDA in one-vs-all setting (with the same training and test data)

%if(troption==1)
% medLDAovaprobassign = zeros(length(trdata.classlabels),trdata.Y);
%else
% medLDAovaprobassign = zeros(length(testdata.classlabels),testdata.Y);
%end

%for i=1:trdata.Y
%   i
%   [trdataova, testdataova] = get_medLDAovadata(trdata, testdata, i); 
%   [trmodelmedLDAova{i}, trperfmedLDAova{i}] = variational_EM(trdataova, MAXCOUNT, MAXESTEPITER, MAXMSTEPITER, MaxFun, p, [], 3, 1, []);
%   if(troption==1)  %% test on training data
%       testdataova = trdataova;
%   end
%   [testmodelmedLDAova{i}, testperfmedLDAova{i}, medLDAovaprobassign(:,i)] = InferenceOnTest(trmodelmedLDAova{i}, testdataova, MAXESTEPITER, 3, p);
%end
%[~, tempacc] = max(medLDAovaprobassign, [], 2);

%if(troption==1)
% testperfmedLDAova{1}.multiclassaccuracy = mean(tempacc==testdata.classlabels);
%else
% testperfmedLDAova{1}.multiclassaccuracy = mean(tempacc==testdata.classlabels);
%end

%testperfmedLDAova{1}.multiclassaccuracy 
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
disp('*************** DSLDA starts ****************');
%% run DSLDA (with the same training and test data)
[trmodel, trperf] = variational_EM(trdata, MAXCOUNT, MAXESTEPITER, MAXMSTEPITER, MaxFun, p, [], 4, 1, []);
if(troption==1)  %% test on training data
    testdata = trdata;
end
[testmodel, testperf] = InferenceOnTest(trmodel, testdata, MAXESTEPITER, 4, p);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


disp('*************** DSLDA-OSST starts ****************');
%%run labeled LDA with class labels (with the same training and test data)
[trmodelDSLDAOSST, trperfDSLDAOSST] = variational_EM (trdata, MAXCOUNT, MAXESTEPITER, MAXMSTEPITER, MaxFun, p, [], 7, 1, []);
if(troption==1)  %% test on training data
    testdata = trdata;
end
[testmodelDSLDAOSST, testperfDSLDAOSST] = InferenceOnTest (trmodelDSLDAOSST, testdata, MAXESTEPITER, 7, p);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


% % disp('*************** DSLDA-NSLT-Ayan starts ****************');
% % %% run labeled LDA with class labels (with the same training and test data)
% % [trmodelDSLDANSLT2, trperfDSLDANSLT2] = variational_EM(trdata, MAXCOUNT, MAXESTEPITER, MAXMSTEPITER, MaxFun, p, [], 6, 1, []);
% % if(troption==1)  %% test on training data
% %     testdata = trdata;
% % end
% % [testmodelDSLDANSLT2, testperfDSLDANSLT2] = InferenceOnTest(trmodelDSLDANSLT2, testdata, MAXESTEPITER, 6, p);
% % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%


disp('*************** DSLDA-NSLT-Ray starts ****************');
%% run labeled LDA with class labels (with the same training and test data)
[trmodelDSLDANSLT1, trperfDSLDANSLT1] = variational_EM(trdata, MAXCOUNT, MAXESTEPITER, MAXMSTEPITER, MaxFun, p, [], 5, 1, []);
if(troption==1)  %% test on training data
    testdata = trdata;
end
[testmodelDSLDANSLT1, testperfDSLDANSLT1] = InferenceOnTest(trmodelDSLDANSLT1, testdata, MAXESTEPITER, 5, p);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

disp('*************** medLDA-MTL starts ****************');
%% run medLDA with all the classes together (with the same training and test data)
[trmodelmedLDA, trperfmedLDA] = variational_EM(trdata, MAXCOUNT, MAXESTEPITER, MAXMSTEPITER, MaxFun, p, [], 3, 1, []);
if(troption==1)  %% test on training data
     testdata = trdata;
end
[testmodelmedLDA, testperfmedLDA] = InferenceOnTest(trmodelmedLDA, testdata, MAXESTEPITER, 3, p);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%% modify this pathname and filename if you are running on/off condor

dirdate = date;
dirdate = regexprep(dirdate, '-', '_');
dirname = ['/lusr/u/ayan/MLDisk/LabelMeData/LabelMeResults_' dirdate];
if(exist(dirname, 'dir')==0)
    system(['mkdir ' dirname]);
end
savepath = [dirname '/'];
savefilename = [savepath classnames{end} '_' num2str(minvtopic) '_' num2str(p1) '_' num2str(p2) '_' num2str(k2) '_' num2str(option) '_' num2str(troption) '_' num2str(svmoptionval)];
savefilename = [savefilename '_' num2str(svmcval) '_' otherindex '.mat'];
save(savefilename, 'svmperf', 'trperf', 'testperf', 'trperfmedLDA', 'testperfmedLDA', 'trperfDSLDANSLT1', 'testperfDSLDANSLT1', 'trperfDSLDANSLT2', 'testperfDSLDANSLT2', 'trperfmedLDAova', 'testperfmedLDAova', 'trperfDSLDAOSST', 'testperfDSLDAOSST');

%% accuracies are determined by F-scores..
%% for per-class accuracies, you have the confusion matrices
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

disp('svm medLDA MTL, DSLDAOSST, DSLDANSLT1, DSLDA');
svmacc = sum(diag(svmperf.confmat_test))/sum(sum(svmperf.confmat_test));

svmacc
testperfmedLDA.multiclassacc
testperfDSLDAOSST.multiclassacc
testperfDSLDANSLT1.multiclassacc
testperf.multiclassacc
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

disp('********** experiments done ********');

end