Merge branch 'master' of github.com:uShadow/tabletop-private

matlab/hyperparam.m

@@ -1,14 +1,18 @@
 function hyperParam = hyperparam(paramFromData, varargin)
 
 % Default values.
-hyperParam.nS = 37;
+hyperParam.nS = 11;
 hyperParam.nhandFet = 3;
 hyperParam.cov_type = 'diag';
+hyperParam.resetS = false;
+hyperParam.inferMethod = 'viterbi';
 
 for i = 1 : 2 : length(varargin)
   switch varargin{i}
     case 'nS'
       hyperParam.nS = varargin{i + 1};
+    case 'resetS'
+      hyperParam.resetS = varargin{i + 1};
     otherwise
       error(['Unrecognized parameter: ' varargin{i}]);
   end
@@ -24,13 +28,15 @@
 for i = 1 : nmodel
   param = paramFromData;
   param.learner = 'ahmm';
+  param.inferMethod = hyperParam.inferMethod;
   param.preprocess = {@denoise @eigenhand @standardizefeature};
   param.nS = hyperParam.nS(i);
   param.nhandFet = hyperParam.nhandFet;
   param.startHandFetNDX = hyperParam.startHandFetNDX;
   param.dir = hyperParam.dir;
   param.maxIter = hyperParam.maxIter;
   param.cov_type = hyperParam.cov_type;
+  param.resetS = hyperParam.resetS;
   hyperParam.model{i} = param;
 end
 end

matlab/jobparam.m

@@ -1,6 +1,6 @@
 function jobParam = jobparam
 jobParam.verbose = true;
-jobParam.destroy = true;
+jobParam.destroy = false;
 jobParam.type = 'local';
 jobParam.path = [genpath('c:/users/yingyin/workspace/tabletop/matlab-lib/lib') ';' genpath(pwd)];
 end

matlab/meet/evaluate/errorperframe.m

@@ -0,0 +1,3 @@
+function error = errorperframe(Ytrue, Ystar)
+error = sum(1 - (cell2mat(Ytrue) == cell2mat(Ystar)), 2);
+end

matlab/meet/evaluate/evalclassification.m

@@ -1,30 +1,30 @@
-function stat = evalclassification(Y, R, loss)
+function stat = evalclassification(Y, R, evalfun)
 % Args:
-% - loss: loss function.
+% - evalfun: evaluation function.
 if iscell(R)
   nfold = size(R, 2);
   stat = cell(1, nfold);
   for i = 1 : nfold
     split1 = R{i}.split;
     Ytrue.train = Y(split1{1});
     Ytrue.validate = Y(split1{2});
-    stat{i} = evalOneFold(Ytrue, R{i}.prediction, loss);
+    stat{i} = evalOneFold(Ytrue, R{i}.prediction, evalfun);
   end
 else
-  stat = evalOneFold(Y, R, loss);
+  stat = evalOneFold(Y, R, evalfun);
 end
 end
 
-function stat = evalOneFold(Y, R, loss)
+function stat = evalOneFold(Y, R, evalfun)
 stat = containers.Map();
-stat('train') = evaluate(Y.train, R.train, loss);
-stat('validate') = evaluate(Y.validate, R.validate, loss);
+stat('train') = evaluate(Y.train, R.train, evalfun);
+stat('validate') = evaluate(Y.validate, R.validate, evalfun);
 if isfield(Y, 'test')
-  stat('test') = evaluate(Y.test, R.test, loss);
+  stat('test') = evaluate(Y.test, R.test, evalfun);
 end
 end
 
-function r = evaluate(Ytrue, Ystar, loss)
+function r = evaluate(Ytrue, Ystar, evalfun)
 %
 % Args:
 % - Ytrue: cell array of sequences.
@@ -33,12 +33,8 @@
 r = containers.Map();
 error = zeros(nlabel, 1);
 for i = 1 : length(Ytrue)
-  for j = 1 : size(Ytrue{i}, 2)
-    total = total + 1;
-    for k = 1 : nlabel
-      error(k) = error(k) + loss(Ytrue{i}{k, j}, Ystar{i}{k, j});
-    end
-  end
+  error = error + evalfun(Ytrue{i}, Ystar{i});
+  total = total + size(Ytrue{i}, 2);
 end
 r('error') = error / total;
 end

matlab/meet/evaluate/evalseqclass.m

@@ -0,0 +1,9 @@
+function stat = evalseqclass(Ytrue, Ystar)
+% EVALSEQCLASS evaluates the classification as a sequence not by frame.
+% Args:
+% - Ytrue: a vector.
+% - Ystar: a vector.
+trueLabel = unique(Ytrue);
+actualLabel = unqiue(Ystar);
+stat = edit_distance_levenshtein(trueLabel, actualLabel);
+end

matlab/meet/learn/ahmm/createahmm.m

@@ -40,6 +40,9 @@
 inter(G1, G1) = 1; % Node G1 in slice t - 1 connects to node G1 in slice t.
 inter(F1, G1) = 1;
 inter(S1, S1) = 1;
+if param.resetS
+  inter(F1, S1) = 1;
+end
 
 nodeSize = double([param.nG param.nS param.nF param.nX]);
 dnodes = [G1 S1 F1];
@@ -78,5 +81,11 @@
 % Tablular CPD are stored as multidimentional arrays where the dimensions
 % are arranged in the same order as the nodes. Nodes in the 2nd slice is 
 % is after the ndoes in the 1st slice.
-ahmm.CPD{S2} = tabular_CPD(ahmm, S2, param.Stransprob);
+if param.resetS
+  ahmm.CPD{S2} = hhmmQ_CPD(ahmm, S2, 'Fself', F1, 'Fbelow', [], ...
+                            'Qps', G2, 'startprob', param.Sstartprob, ...
+                            'transprob', param.Stransprob);
+else
+  ahmm.CPD{S2} = tabular_CPD(ahmm, S2, param.Stransprob);
+end
 end

matlab/meet/learn/ahmm/inferenceahmm.m

@@ -1,16 +1,21 @@
-function R = inferenceahmm(ahmm, data, predictNode)
+function R = inferenceahmm(ahmm, data, predictNode, method)
 
 engine = smoother_engine(jtree_2TBN_inf_engine(ahmm));
 
 nseq = length(data);
 R = cell(1, nseq);
 for i = 1 : nseq
   evidence = data{i};
-  engine = enter_evidence(engine, evidence);
-  R{i} = mapest(engine, predictNode, length(evidence));
-
-  % Find the most probable explanation (Viterbi).
-  %mpe = find_mpe(engine, evidence);
-  %R{i} = mpe(predictNode, :);
+  switch method
+    case 'fixed-interval-smoothing'
+      engine = enter_evidence(engine, evidence);
+      R{i} = mapest(engine, predictNode, length(evidence));
+    case 'viterbi'
+      % Find the most probable explanation (Viterbi).
+      mpe = find_mpe(engine, evidence);
+      R{i} = mpe(predictNode, :);
+    otherwise
+      error(['Inference method not implemented: ' method]);
+  end
 end
 end

matlab/meet/learn/ahmm/initahmmparam.m

@@ -4,6 +4,7 @@
 % - param: struct with the following fields
 %   -- fold: the current fold.
 
+param.nF = 2;
 param.nX = param.nconFet + param.nhandFet;
 assert(all(size(mean) == [param.nX param.nS]));
 
@@ -30,7 +31,7 @@
 param.Gtransprob(3 : param.nG, 2 : param.nG) = ...
     (1 - delta) / (param.nG - 1);
 
-% Parameters related to S.
+% Parameters related to S. Uniform initialization.
 param.Sstartprob = ones(param.nG, param.nS) / param.nS;
 param.Stransprob = ones(param.nS, param.nG, param.nS) / param.nS;
 param.Stermprob = ones(param.nG, param.nS, param.nF) / param.nF;

matlab/meet/learn/ahmm/learnahmm.m

@@ -7,8 +7,8 @@
 % case{l}{i, t} is the value of node i in slice t in sequel l, or [] if
 % unobserved.
 nfeature = param.nconFet + param.nhandFet;
-filename = sprintf('%sfeature-%d-%d-mean-%d.csv', param.dir, nfeature, ...
-                   param.fold, param.nS); 
+filename = sprintf('%s%s-feature-%d-%d-mean-%d.csv', param.dir, ...
+                   param.userId, nfeature, param.fold, param.nS);
 logdebug('learnahmm', 'read file', filename);
 imported = importdata(filename, ',', 1);
 mean = imported.data;
@@ -24,10 +24,12 @@
 checkahmm(finalAhmm);
 
 trainData = createInputData(Y.train, X.train, param);
-R.train = inferenceahmm(finalAhmm, trainData, predictNode);
+R.train = inferenceahmm(finalAhmm, trainData, predictNode, ...
+                        param.inferMethod);
 
 validateData = createInputData(Y.validate, X.validate, param);
-R.validate = inferenceahmm(finalAhmm, validateData, predictNode);
+R.validate = inferenceahmm(finalAhmm, validateData, predictNode, ...
+                           param.inferMethod);
 end
 
 function data = createInputData(Y, X, param)

matlab/meet/runexperiment.m

@@ -9,6 +9,8 @@
 %   - R      : struct, result
 
   R.split = split;
+  R.param = param;
+  param.userId = data.userId;
 
   % Step 1: Prepare data
   if exist('data','var')
@@ -47,7 +49,7 @@
     case 'svr'
       R = learn_svr(Y, X, param);
     case 'ahmm'
-      [R.prediction R.learnedModel] = learnahmm(Y, X, param);
+      [R.prediction, R.learnedModel] = learnahmm(Y, X, param);
     otherwise
         error('%s Not implemented yet', param.model);
   end
@@ -66,11 +68,10 @@
     case 'svr'
       R.stat = eval_svr(Y, R.prediction, param);
     case 'ahmm'
-      R.stat = evalclassification(Y, R.prediction, @zerooneloss);
+      R.stat = evalclassification(Y, R.prediction, @errorperframe);
     otherwise
-      error('%s Not implemented yet', param.model);
+      error('%s Not implemented yet', param.learner);
   end
-
 end
 
 

matlab/meet/runexperimentparallel.m

@@ -1,4 +1,4 @@
-function R = runexperimentparallel(data, split, modelParam, jobParam)
+function R = runexperimentparallel(data, modelParam, jobParam)
 %
 % Args
 % - modelParam: a cell array of model parameter.
@@ -13,7 +13,7 @@
             distributed = false;
     end
 
-
+    split = data.split;
     R = cell(numel(modelParam),size(split, 2));    
     nargout = 1; % to be returned from each task 
 
@@ -59,17 +59,14 @@
         if verbose, fprintf('Collect results'); tid=tic(); end    
         rows = cellfun(@(x) getfield(x,'row'), job.JobData.job_log);
         cols = cellfun(@(x) getfield(x,'col'), job.JobData.job_log);
-        for r=unique(rows)
-            for c=unique(cols)
-                tasks = find(rows==r & cols==c);
-                R{r,c} = cell(1,numel(tasks));
-                for t=1:numel(tasks)
-                    if verbose, fprintf('.'); end 
-                    if ~isempty(job.Tasks(tasks(t)).OutputArguments)
-                        R{r,c}{t} = job.Tasks(tasks(t)).OutputArguments{1};
-                    end
-                end
+        for r = unique(rows)
+          for c = unique(cols)
+            task = rows == r & cols == c;
+            if verbose, fprintf('.'); end 
+            if ~isempty(job.Tasks(task).OutputArguments)
+              R{r,c} = job.Tasks(task).OutputArguments{1};
             end
+          end
         end
         if verbose, t=toc(tid); fprintf('done (%.2f secs)\n', t); end
 

matlab/prepdata.m

@@ -1,7 +1,12 @@
-function [data segFrameId dataParam] = prepdata(dirname, suffix)
+function [data, segFrameId, dataParam] = prepdata(dirname, suffix, userId)
 % PREPTRAININGDATA prepares the training data into right structure for 
 % preprocesssing.
 %
+% Args:
+% - suffix: suffix of the feature file, e.g. '-100'.
+% - userID: an optional string or cell array of userIDs indicating the user 
+%           data to be imported.
+%
 % Return
 % - data: a struct. data.Y is the label and data.X is the feature.
 %         Each field is a cell array.
@@ -12,9 +17,9 @@
     name = file.name;
     indices = strfind(name, '.');
     last_index = indices(1);
-    basename = name(1 : last_index - 1);
+    basename = name(1 : last_index - 1); % without extension.
     ext = name(last_index + 1 : end);
-    if strcmp(ext, 'glab.csv')
+    if strcmp(ext, 'glab.csv') && isuserfile(basename, userId)
       labelFile = [dirname name];
       label = importdata(labelFile, ',', 1);
       logdebug('prepdata', 'label file', labelFile);
@@ -26,28 +31,31 @@
       % Init parameters.
       dataParam.nconFet = header{3};
       dataParam.nG = length(unique(label.data(:, 2)));
-      dataParam.nF = length(unique(label.data(:, 3)));
       imageWidth = header{5};
       dataParam.imageSize = imageWidth * imageWidth; 
       dataParam.dir = dirname;
-
-      [data segFrameId] = combinelabelfeature(label.data, feature.data, ...
-                                              dataParam);
+      [data, segFrameId] = combinelabelfeature(label.data, feature.data, ...
+                                               dataParam);
+      data.userId = userId;
+      data.split = getnfoldindex(numel(data.X), 0, 10);
     end
   end
 end
 end
 
-function [data segFrameId] = combinelabelfeature(label, feature, param)
+function b = isuserfile(basename, userID)
+  b = isempty(userID) || any(strncmp(basename, userID, length(basename)));
+end
+
+function [data, segFrameId] = combinelabelfeature(label, feature, param)
 labelFrameId = label(:, 1);
 featureFrameId = feature(:, 1);
 [frameids, labelNDX, featureNDX] = intersect(labelFrameId, featureFrameId);
 logdebug('preptrainingdata', 'number of frames', length(frameids));
-[segment segFrameId] = createsegment(frameids); % cell array
+[segment, segFrameId] = createsegment(frameids); % cell array
 label = label(labelNDX, 2 : end); % Removes frame ID.
 feature = feature(featureNDX, 2 : end); % Each row is a feature vector.
 
-assert(checklabel(label), 'Label is not valid.');
 assert(size(label, 1) == size(feature, 1));
 
 data.Y = cell(1, length(segment));
@@ -65,7 +73,7 @@
 end
 end
 
-function [seg segFrameId] = createsegment(frameid)
+function [seg, segFrameId] = createsegment(frameid)
 % segments = segment(frameids) finds the segments in a vector of frame 
 % IDs. A segment is a sequence of continuous frame IDs. 
 % 

matlab/savefeature.m

@@ -14,7 +14,8 @@
   standardized = standardizefeature(result, 0, 'retMat', true);
   nfeature = size(standardized, 1);
   assert(nfeature == hyperParam.nhandFet + hyperParam.startHandFetNDX - 1);
-  filename = sprintf('%sfeature-%d-%d.csv', hyperParam.dir, nfeature, i);
+  filename = sprintf('%s%s-feature-%d-%d.csv', hyperParam.dir, ...
+                     hyperParam.userId, nfeature, i);
   csvwrite(filename, standardized');
 end
 end