Merge branch 'develop'

evaluate.m

@@ -56,16 +56,17 @@
     end
 
     % Embed the test attributes representation (attRepreTe_emb)
-    mat = emb.rndmat(1:emb.M,:);
-    tmp = mat*data.attReprTe;
+    matx = emb.rndmatx(1:emb.M,:);
+    tmp = matx*data.attReprTe;
     attReprTe_emb = 1/sqrt(emb.M) * [ cos(tmp); sin(tmp)];
     attReprTe_emb = bsxfun(@minus, attReprTe_emb, emb.matts);
     attReprTe_emb = emb.Wx(:,1:emb.K)' * attReprTe_emb;
     attReprTe_emb = (bsxfun(@rdivide, attReprTe_emb, sqrt(sum(attReprTe_emb.*attReprTe_emb))));
 
     % Embed the dictionary
     phocs = single(lexicon.phocs);
-    tmp = mat*phocs;
+    maty = emb.rndmaty(1:emb.M,:);
+    tmp = maty*phocs;
     phocs_cca = 1/sqrt(emb.M) * [ cos(tmp); sin(tmp)];
     phocs_cca=bsxfun(@minus, phocs_cca, emb.mphocs);
     phocs_cca = emb.Wy(:,1:emb.K)' * phocs_cca;

prepare_opts.m

@@ -8,7 +8,7 @@
 rng('default');
 
 % Select the dataset
-opts.dataset = 'ICDAR11';
+opts.dataset = 'SVT';
 
 opts.path_datasets = 'datasets';
 opts.pathDataset = sprintf('%s/%s/',opts.path_datasets,opts.dataset);
@@ -54,7 +54,7 @@
 % Options embedding
 opts.RemoveStopWords = 0;
 opts.TestFV = 0;
-opts.TestDirect = 1;
+opts.TestDirect = 0;
 
 opts.TestPlatts = 0;
 opts.Platts.verbose = 1;

util/evaluateKCCA.m

@@ -6,11 +6,12 @@
 disp('***************  KCSR   **************');
 disp('**************************************');
 
-mat = embedding.rndmat(1:embedding.M,:);
+matx = embedding.rndmatx(1:embedding.M,:);
+maty = embedding.rndmaty(1:embedding.M,:);
 
-tmp = mat*DATA.attReprTe;
+tmp = matx*DATA.attReprTe;
 attReprTe_emb = 1/sqrt(embedding.M) * [ cos(tmp); sin(tmp)];
-tmp = mat*DATA.phocsTe;
+tmp = maty*DATA.phocsTe;
 phocsTe_emb = 1/sqrt(embedding.M) * [ cos(tmp); sin(tmp)];
 
 % Mean center

util/learnKCCA.m

@@ -18,26 +18,30 @@
 bestG = opts.KCCA.G(end);
 bestM = opts.KCCA.M(end);
 
-D = size(DATA.attReprTr,1);
+Dx = size(DATA.attReprTr,1);
+Dy = size(DATA.phocsTr,1);
 for G=opts.KCCA.G
     % Create random projections matrix
     RandStream.setGlobalStream(RandStream('mt19937ar','seed',0));
-    rndmat = normrnd(0,1/G, 2500,D);
+    rndmatx = normrnd(0,1/G, 2500,Dx);
+    rndmaty = normrnd(0,1/G, 2500,Dy);
+
     for M=opts.KCCA.M
         % Cut slice
-        mat = rndmat(1:M,:);
+        matx = rndmatx(1:M,:);
+        maty = rndmaty(1:M,:);
 
         % Project attributes and phocs using the explicit exponential
         % embedding. Project train, val, and test.
         % train
-        tmp = mat*DATA.attReprTr;
+        tmp = matx*DATA.attReprTr;
         attReprTr_emb = 1/sqrt(M) * [ cos(tmp); sin(tmp)];
-        tmp = mat*DATA.phocsTr;
+        tmp = maty*DATA.phocsTr;
         phocsTr_emb = 1/sqrt(M) * [ cos(tmp); sin(tmp)];
         % val
-        tmp = mat*DATA.attReprVa;
+        tmp = matx*DATA.attReprVa;
         attReprVal_emb = 1/sqrt(M) * [ cos(tmp); sin(tmp)];
-        tmp = mat*DATA.phocsVa;
+        tmp = maty*DATA.phocsVa;
         phocsVal_emb = 1/sqrt(M) * [ cos(tmp); sin(tmp)];
 
         % Mean center
@@ -106,15 +110,17 @@
 
 % Get random matrix
 RandStream.setGlobalStream(RandStream('mt19937ar','seed',0));
-rndmat = normrnd(0,1/bestG, 2500,D);
-mat = rndmat(1:bestM,:);
+rndmatx = normrnd(0,1/bestG, 2500,Dx);
+matx = rndmatx(1:bestM,:);
+rndmaty = normrnd(0,1/bestG, 2500,Dy);
+maty = rndmaty(1:bestM,:);
 % Embed train (full) and test
 % Project attributes and phocs using the explicit exponential
 % embedding. Project train, val, and test.
 % train
-tmp = mat*DATA.attReprTrFull;
+tmp = matx*DATA.attReprTrFull;
 attReprTrFull_emb = 1/sqrt(bestM) * [ cos(tmp); sin(tmp)];
-tmp = mat*DATA.phocsTrFull;
+tmp = maty*DATA.phocsTrFull;
 phocsTrFull_emb = 1/sqrt(bestM) * [ cos(tmp); sin(tmp)];
 
 % Mean center
@@ -133,7 +139,8 @@
 embedding.K = bestK;
 embedding.M = bestM;
 embedding.reg = bestReg;
-embedding.rndmat = rndmat;
+embedding.rndmatx = rndmatx;
+embedding.rndmaty = rndmaty;
 embedding.matts = ma;
 embedding.mphocs = mh;