WARNING: this change list changed GauPost to GaussPost, and also made it pdf-id indexed instead of transition-id indexed. Details: 1) Modified those posterior related programs that use TransitionIdToPdf to get a pdf-id, and later on only use the pdf-id. We merge the posteriors that corresponds to the same pdf-id to avoid redundant computation. 2) Modified phone lattice determinization, added a wrapper for the lattice type determinization to reduce redundant code in the decoding binaries.

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@3588 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

src/bin/acc-lda.cc

@@ -1,6 +1,7 @@
 // bin/acc-lda.cc
 
 // Copyright 2009-2011  Microsoft Corporation, Go-Vivace Inc.
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -95,14 +96,15 @@ int main(int argc, char *argv[]) {
         continue;
       }
 
+      Posterior pdf_post;
+      ConvertPosteriorToPdfs(trans_model, post, &pdf_post);
       for (int32 i = 0; i < feats.NumRows(); i++) {
         SubVector<BaseFloat> feat(feats, i);
-        for (size_t j = 0; j < post[i].size(); j++) {
+        for (size_t j = 0; j < pdf_post[i].size(); j++) {
-          int32 tid = post[i][j].first;
+          int32 pdf_id = pdf_post[i][j].first;
-          BaseFloat weight = RandPrune(post[i][j].second, rand_prune);
+          BaseFloat weight = RandPrune(pdf_post[i][j].second, rand_prune);
           if (weight != 0.0) {
-            int32 pdf = trans_model.TransitionIdToPdf(tid);
+            lda.Accumulate(feat, pdf_id, weight);
-            lda.Accumulate(feat, pdf, weight);
           }
         }
       }

src/decoder/lattice-faster-decoder.cc

@@ -790,20 +790,9 @@ void DecodeUtteranceLatticeFasterClass::operator () () {
     KALDI_ERR << "Unexpected problem getting lattice for utterance " << utt_;
   fst::Connect(lat_);
   if (determinize_) {
-    Invert(lat_);
-    if (!TopSort(lat_)) {
-      // Cannot topologically sort the lattice -- determinization will fail.
-      KALDI_WARN << "Topological sorting of state-level lattice failed "
-                 << "(probably your lexicon has empty words or your LM has "
-                 << "epsilon cycles).";
-      delete lat_;    // Delete it here.
-      success_ = false;
-      return;
-    }
-    fst::ILabelCompare<LatticeArc> ilabel_comp;
-    ArcSort(lat_, ilabel_comp);
     clat_ = new CompactLattice;
-    if (!DeterminizeLatticePhonePruned(*trans_model_,
+    if (!DeterminizeLatticePhonePrunedWrapper(
+            *trans_model_,
             lat_,
             decoder_->GetOptions().lattice_beam,
             clat_,
@@ -811,7 +800,6 @@ void DecodeUtteranceLatticeFasterClass::operator () () {
       KALDI_WARN << "Determinization finished earlier than the beam for "
                  << "utterance " << utt_;
     delete lat_;
-    fst::Connect(clat_);
     // We'll write the lattice without acoustic scaling.
     if (acoustic_scale_ != 0.0)
       fst::ScaleLattice(fst::AcousticLatticeScale(1.0 / acoustic_scale_), clat_);
@@ -971,25 +959,15 @@ bool DecodeUtteranceLatticeFaster(
     KALDI_ERR << "Unexpected problem getting lattice for utterance " << utt;
   fst::Connect(&lat);
   if (determinize) {
-    Invert(&lat);
-    if (!TopSort(&lat)) {
-      // Cannot topologically sort the lattice -- determinization will fail.
-      KALDI_WARN << "Topological sorting of state-level lattice failed "
-                 << "(probably your lexicon has empty words or your LM has "
-                 << "epsilon cycles).";
-      return false;
-    }
-    fst::ILabelCompare<LatticeArc> ilabel_comp;
-    ArcSort(&lat, ilabel_comp);
     CompactLattice clat;
-    if (!DeterminizeLatticePhonePruned(trans_model,
+    if (!DeterminizeLatticePhonePrunedWrapper(
+            trans_model,
             &lat,
             decoder.GetOptions().lattice_beam,
             &clat,
             decoder.GetOptions().det_opts))
       KALDI_WARN << "Determinization finished earlier than the beam for "
                  << "utterance " << utt;
-    fst::Connect(&clat);
     // We'll write the lattice without acoustic scaling.
     if (acoustic_scale != 0.0)
       fst::ScaleLattice(fst::AcousticLatticeScale(1.0 / acoustic_scale), &clat);

src/decoder/lattice-simple-decoder.cc

@@ -656,25 +656,15 @@ bool DecodeUtteranceLatticeSimple(
     KALDI_ERR << "Unexpected problem getting lattice for utterance " << utt;
   fst::Connect(&lat);
   if (determinize) {
-    Invert(&lat);
-    if (!TopSort(&lat)) {
-      // Cannot topologically sort the lattice -- determinization will fail.
-      KALDI_WARN << "Topological sorting of state-level lattice failed "
-                 << "(probably your lexicon has empty words or your LM has "
-                 << "epsilon cycles).";
-      return false;
-    }
-    fst::ILabelCompare<LatticeArc> ilabel_comp;
-    ArcSort(&lat, ilabel_comp);
     CompactLattice clat;
-    if (!DeterminizeLatticePhonePruned(trans_model,
+    if (!DeterminizeLatticePhonePrunedWrapper(
+            trans_model,
             &lat,
             decoder.GetOptions().lattice_beam,
             &clat,
             decoder.GetOptions().det_opts))
       KALDI_WARN << "Determinization finished earlier than the beam for "
                  << "utterance " << utt;
-    fst::Connect(&clat);
     // We'll write the lattice without acoustic scaling.
     if (acoustic_scale != 0.0)
       fst::ScaleLattice(fst::AcousticLatticeScale(1.0 / acoustic_scale), &clat);

src/decoder/lattice-tracking-decoder.cc

@@ -938,25 +938,15 @@ bool DecodeUtteranceLatticeTracking(
     KALDI_ERR << "Unexpected problem getting lattice for utterance " << utt;
   fst::Connect(&lat);
   if (determinize) {
-    Invert(&lat);
-    if (!TopSort(&lat)) {
-      // Cannot topologically sort the lattice -- determinization will fail.
-      KALDI_WARN << "Topological sorting of state-level lattice failed "
-                 << "(probably your lexicon has empty words or your LM has "
-                 << "epsilon cycles).";
-      return false;
-    }
-    fst::ILabelCompare<LatticeArc> ilabel_comp;
-    ArcSort(&lat, ilabel_comp);
     CompactLattice clat;
-    if (!DeterminizeLatticePhonePruned(trans_model,
+    if (!DeterminizeLatticePhonePrunedWrapper(
+            trans_model,
             &lat,
             decoder.GetOptions().lattice_beam,
             &clat,
             decoder.GetOptions().det_opts))
       KALDI_WARN << "Determinization finished earlier than the beam for "
                  << "utterance " << utt;
-    fst::Connect(&clat);
     // We'll write the lattice without acoustic scaling.
     if (acoustic_scale != 0.0)
       fst::ScaleLattice(fst::AcousticLatticeScale(1.0 / acoustic_scale), &clat);

src/gmmbin/gmm-acc-hlda.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-acc-hlda.cc
 
 // Copyright 2009-2011  Microsoft Corporation
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -106,11 +107,12 @@ int main(int argc, char *argv[]) {
         num_done++;
         BaseFloat tot_like_this_file = 0.0, tot_weight = 0.0;
 
+        Posterior pdf_posterior;
+        ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
         for (size_t i = 0; i < posterior.size(); i++) {
-          for (size_t j = 0; j < posterior[i].size(); j++) {
+          for (size_t j = 0; j < pdf_posterior[i].size(); j++) {
-            int32 tid = posterior[i][j].first,  // transition identifier.
+            int32 pdf_id = pdf_posterior[i][j].first;
-                pdf_id = trans_model.TransitionIdToPdf(tid);
+            BaseFloat weight = pdf_posterior[i][j].second;
-            BaseFloat weight = posterior[i][j].second;
 
             Vector<BaseFloat> posteriors;
             const DiagGmm &gmm = am_gmm.GetPdf(pdf_id);

src/gmmbin/gmm-acc-mllt.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-acc-mllt.cc
 
 // Copyright 2009-2011  Microsoft Corporation
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -90,11 +91,12 @@ int main(int argc, char *argv[]) {
         num_done++;
         BaseFloat tot_like_this_file = 0.0, tot_weight = 0.0;
 
+        Posterior pdf_posterior;
+        ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
         for (size_t i = 0; i < posterior.size(); i++) {
-          for (size_t j = 0; j < posterior[i].size(); j++) {
+          for (size_t j = 0; j < pdf_posterior[i].size(); j++) {
-            int32 tid = posterior[i][j].first,  // transition identifier.
+            int32 pdf_id = pdf_posterior[i][j].first;
-                pdf_id = trans_model.TransitionIdToPdf(tid);
+            BaseFloat weight = pdf_posterior[i][j].second;
-            BaseFloat weight = posterior[i][j].second;
 
             tot_like_this_file += mllt_accs.AccumulateFromGmm(am_gmm.GetPdf(pdf_id),
                                                               mat.Row(i),

src/gmmbin/gmm-acc-stats-twofeats.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-acc-stats-twofeats.cc
 
 // Copyright 2009-2011  Microsoft Corporation
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -111,12 +112,13 @@ int main(int argc, char *argv[]) {
         BaseFloat tot_like_this_file = 0.0,
             tot_weight_this_file = 0.0;
 
+        Posterior pdf_posterior;
+        ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
         for (size_t i = 0; i < posterior.size(); i++) {
-          for (size_t j = 0; j < posterior[i].size(); j++) {
+          // Accumulates for GMM.
-            int32 tid = posterior[i][j].first,  // transition identifier.
+          for (size_t j = 0; j <pdf_posterior[i].size(); j++) {
-                pdf_id = trans_model.TransitionIdToPdf(tid);
+            int32 pdf_id = pdf_posterior[i][j].first;
-            BaseFloat weight = posterior[i][j].second;
+            BaseFloat weight = pdf_posterior[i][j].second;
-            trans_model.Accumulate(weight, tid, &transition_accs);
             tot_like_this_file += weight *
                 gmm_accs.AccumulateForGmmTwofeats(am_gmm,
                                                   mat1.Row(i),
@@ -125,6 +127,13 @@ int main(int argc, char *argv[]) {
                                                   weight);
             tot_weight_this_file += weight;
           }
+
+          // Accumulates for transitions.
+          for (size_t j = 0; j < posterior[i].size(); j++) {
+            int32 tid = posterior[i][j].first;
+            BaseFloat weight = posterior[i][j].second;
+            trans_model.Accumulate(weight, tid, &transition_accs);
+          }
         }
         KALDI_LOG << "Average like for this file is "
                   << (tot_like_this_file/tot_weight_this_file) << " over "

src/gmmbin/gmm-acc-stats.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-acc-stats.cc
 
 // Copyright 2009-2012  Microsoft Corporation  Johns Hopkins University (Author: Daniel Povey)
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -98,16 +99,24 @@ int main(int argc, char *argv[]) {
         num_done++;
         BaseFloat tot_like_this_file = 0.0, tot_weight = 0.0;
 
+        Posterior pdf_posterior;
+        ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
         for (size_t i = 0; i < posterior.size(); i++) {
-          for (size_t j = 0; j < posterior[i].size(); j++) {
+          // Accumulates for GMM.
-            int32 tid = posterior[i][j].first,  // transition identifier.
+          for (size_t j = 0; j < pdf_posterior[i].size(); j++) {
-                pdf_id = trans_model.TransitionIdToPdf(tid);
+            int32 pdf_id = pdf_posterior[i][j].first;
-            BaseFloat weight = posterior[i][j].second;
+            BaseFloat weight = pdf_posterior[i][j].second;
-            trans_model.Accumulate(weight, tid, &transition_accs);
             tot_like_this_file += gmm_accs.AccumulateForGmm(am_gmm, mat.Row(i), pdf_id, weight)
                 * weight;
             tot_weight += weight;
           }
+
+          // Accumulates for transitions.
+          for (size_t j = 0; j < posterior[i].size(); j++) {
+            int32 tid = posterior[i][j].first;
+            BaseFloat weight = posterior[i][j].second;
+            trans_model.Accumulate(weight, tid, &transition_accs);
+          }
         }
         if (num_done % 50 == 0) {
           KALDI_LOG << "Processed " << num_done << " utterances; for utterance "

src/gmmbin/gmm-adapt-map.cc

@@ -2,6 +2,7 @@
 
 // Copyright 2012  Cisco Systems (author: Neha Agrawal)
 //                 Johns Hopkins University (author: Daniel Povey)
+//           2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -119,11 +120,12 @@ int main(int argc, char *argv[]) {
           }
 
           BaseFloat file_like = 0.0, file_t = 0.0;
+          Posterior pdf_posterior;
+          ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
           for ( size_t i = 0; i < posterior.size(); i++ ) {
-            for ( size_t j = 0; j < posterior[i].size(); j++ ) {
+            for ( size_t j = 0; j < pdf_posterior[i].size(); j++ ) {
-              int32 pdf_id = 
+              int32 pdf_id = pdf_posterior[i][j].first; 
-                  trans_model.TransitionIdToPdf(posterior[i][j].first); 
+              BaseFloat weight = pdf_posterior[i][j].second; 
-              BaseFloat weight = posterior[i][j].second; 
               file_like += map_accs.AccumulateForGmm(copy_am_gmm,
                                                      feats.Row(i),
                                                      pdf_id, weight);
@@ -183,11 +185,12 @@ int main(int argc, char *argv[]) {
         }
         num_done++;
         BaseFloat file_like = 0.0, file_t = 0.0;
+        Posterior pdf_posterior;
+        ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
         for ( size_t i = 0; i < posterior.size(); i++ ) {
-          for ( size_t j = 0; j < posterior[i].size(); j++ ) {
+          for ( size_t j = 0; j < pdf_posterior[i].size(); j++ ) {
-            int32 pdf_id = 
+            int32 pdf_id = pdf_posterior[i][j].first;
-                trans_model.TransitionIdToPdf(posterior[i][j].first);
+            BaseFloat prob = pdf_posterior[i][j].second;
-            BaseFloat prob = posterior[i][j].second;
             file_like += map_accs.AccumulateForGmm(copy_am_gmm,feats.Row(i),
                                                    pdf_id, prob);
             file_t += prob;

src/gmmbin/gmm-basis-fmllr-accs-gpost.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-basis-fmllr-accs-gpost.cc
 
 // Copyright 2012  Carnegie Mellon University (author: Yajie Miao)
+//           2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -32,13 +33,13 @@ using std::vector;
 
 namespace kaldi {
 void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
-                            const GauPost &gpost,
+                            const GaussPost &gpost,
                             const TransitionModel &trans_model,
                             const AmDiagGmm &am_gmm,
                             FmllrDiagGmmAccs *spk_stats) {
   for (size_t i = 0; i < gpost.size(); i++) {
     for (size_t j = 0; j < gpost[i].size(); j++) {
-      int32 pdf_id = trans_model.TransitionIdToPdf(gpost[i][j].first);
+      int32 pdf_id = gpost[i][j].first;
       const Vector<BaseFloat> & posterior(gpost[i][j].second);
       spk_stats->AccumulateFromPosteriors(am_gmm.GetPdf(pdf_id),
                                           feats.Row(i), posterior);
@@ -89,7 +90,7 @@ int main(int argc, char *argv[]) {
       am_gmm.Read(ki.Stream(), binary);
     }
 
-    RandomAccessGauPostReader gpost_reader(gpost_rspecifier);
+    RandomAccessGaussPostReader gpost_reader(gpost_rspecifier);
     BasisFmllrAccus basis_accs(am_gmm.Dim());
 
     int32 num_done = 0, num_no_post = 0, num_other_error = 0;
@@ -115,9 +116,9 @@ int main(int argc, char *argv[]) {
             continue;
           }
           const Matrix<BaseFloat> &feats = feature_reader.Value(utt);
-          const GauPost &gpost = gpost_reader.Value(utt);
+          const GaussPost &gpost = gpost_reader.Value(utt);
           if (static_cast<int32>(gpost.size()) != feats.NumRows()) {
-            KALDI_WARN << "GauPost has wrong size " << (gpost.size())
+            KALDI_WARN << "GaussPost has wrong size " << (gpost.size())
                        << " vs. " << (feats.NumRows());
             num_other_error++;
             continue;
@@ -143,10 +144,10 @@ int main(int argc, char *argv[]) {
           continue;
         }
         const Matrix<BaseFloat> &feats = feature_reader.Value();
-        const GauPost &gpost = gpost_reader.Value(utt);
+        const GaussPost &gpost = gpost_reader.Value(utt);
 
         if (static_cast<int32>(gpost.size()) != feats.NumRows()) {
-          KALDI_WARN << "GauPost has wrong size " << (gpost.size())
+          KALDI_WARN << "GaussPost has wrong size " << (gpost.size())
                      << " vs. " << (feats.NumRows());
           num_other_error++;
           continue;

src/gmmbin/gmm-basis-fmllr-accs.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-basis-fmllr-accs.cc
 
 // Copyright 2012  Carnegie Mellon University (author: Yajie Miao)
+//           2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -36,12 +37,14 @@ void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
                             const TransitionModel &trans_model,
                             const AmDiagGmm &am_gmm,
                             FmllrDiagGmmAccs *spk_stats) {
+  Posterior pdf_post;
+  ConvertPosteriorToPdfs(trans_model, post, &pdf_post);
   for (size_t i = 0; i < post.size(); i++) {
-    for (size_t j = 0; j < post[i].size(); j++) {
+    for (size_t j = 0; j < pdf_post[i].size(); j++) {
-      int32 pdf_id = trans_model.TransitionIdToPdf(post[i][j].first);
+      int32 pdf_id = pdf_post[i][j].first;
       spk_stats->AccumulateForGmm(am_gmm.GetPdf(pdf_id),
                                   feats.Row(i),
-                                  post[i][j].second);
+                                  pdf_post[i][j].second);
     }
   }
 }

src/gmmbin/gmm-est-basis-fmllr-gpost.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-est-basis-fmllr-gpost.cc
 
 // Copyright 2012  Carnegie Mellon University (author: Yajie Miao)
+//           2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -32,13 +33,13 @@ using std::vector;
 
 namespace kaldi {
 void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
-                            const GauPost &gpost,
+                            const GaussPost &gpost,
                             const TransitionModel &trans_model,
                             const AmDiagGmm &am_gmm,
                             FmllrDiagGmmAccs *spk_stats) {
   for (size_t i = 0; i < gpost.size(); i++) {
     for (size_t j = 0; j < gpost[i].size(); j++) {
-      int32 pdf_id = trans_model.TransitionIdToPdf(gpost[i][j].first);
+      int32 pdf_id = gpost[i][j].first;
       const Vector<BaseFloat> & posterior(gpost[i][j].second);
       spk_stats->AccumulateFromPosteriors(am_gmm.GetPdf(pdf_id),
                                           feats.Row(i), posterior);
@@ -102,7 +103,7 @@ int main(int argc, char *argv[]) {
       basis_est.ReadBasis(ki.Stream(), binary, false);
     }
 
-    RandomAccessGauPostReader gpost_reader(gpost_rspecifier);
+    RandomAccessGaussPostReader gpost_reader(gpost_rspecifier);
 
     double tot_impr = 0.0, tot_t = 0.0;
 
@@ -134,9 +135,9 @@ int main(int argc, char *argv[]) {
             continue;
           }
           const Matrix<BaseFloat> &feats = feature_reader.Value(utt);
-          const GauPost &gpost = gpost_reader.Value(utt);
+          const GaussPost &gpost = gpost_reader.Value(utt);
           if (static_cast<int32>(gpost.size()) != feats.NumRows()) {
-        	KALDI_WARN << "GauPost has wrong size " << (gpost.size())
+        	KALDI_WARN << "GaussPost has wrong size " << (gpost.size())
         	           << " vs. " << (feats.NumRows());
             num_other_error++;
             continue;
@@ -178,10 +179,10 @@ int main(int argc, char *argv[]) {
           continue;
         }
         const Matrix<BaseFloat> &feats = feature_reader.Value();
-        const GauPost &gpost = gpost_reader.Value(utt);
+        const GaussPost &gpost = gpost_reader.Value(utt);
 
         if (static_cast<int32>(gpost.size()) != feats.NumRows()) {
-          KALDI_WARN << "GauPost has wrong size " << (gpost.size())
+          KALDI_WARN << "GaussPost has wrong size " << (gpost.size())
                      << " vs. " << (feats.NumRows());
           num_other_error++;
           continue;

src/gmmbin/gmm-est-basis-fmllr.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-est-basis-fmllr.cc
 
 // Copyright 2012  Carnegie Mellon University (author: Yajie Miao)
+//           2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -36,12 +37,14 @@ void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
                             const TransitionModel &trans_model,
                             const AmDiagGmm &am_gmm,
                             FmllrDiagGmmAccs *spk_stats) {
+  Posterior pdf_post;
+  ConvertPosteriorToPdfs(trans_model, post, &pdf_post);
   for (size_t i = 0; i < post.size(); i++) {
-    for (size_t j = 0; j < post[i].size(); j++) {
+    for (size_t j = 0; j < pdf_post[i].size(); j++) {
-      int32 pdf_id = trans_model.TransitionIdToPdf(post[i][j].first);
+      int32 pdf_id = pdf_post[i][j].first;
       spk_stats->AccumulateForGmm(am_gmm.GetPdf(pdf_id),
                                   feats.Row(i),
-                                  post[i][j].second);
+                                  pdf_post[i][j].second);
     }
   }
 }

src/gmmbin/gmm-est-fmllr-gpost.cc

@@ -2,6 +2,7 @@
 
 // Copyright 2009-2011  Microsoft Corporation; Saarland University
 //                2013  Johns Hopkins University (author: Daniel Povey)
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -32,13 +33,13 @@ using std::vector;
 
 namespace kaldi {
 void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
-                            const GauPost &gpost,
+                            const GaussPost &gpost,
                             const TransitionModel &trans_model,
                             const AmDiagGmm &am_gmm,
                             FmllrDiagGmmAccs *spk_stats) {
   for (size_t i = 0; i < gpost.size(); i++) {
     for (size_t j = 0; j < gpost[i].size(); j++) {
-      int32 pdf_id = trans_model.TransitionIdToPdf(gpost[i][j].first);
+      int32 pdf_id = gpost[i][j].first;
       const Vector<BaseFloat> & posterior(gpost[i][j].second);
       spk_stats->AccumulateFromPosteriors(am_gmm.GetPdf(pdf_id),
                                           feats.Row(i), posterior);
@@ -89,7 +90,7 @@ int main(int argc, char *argv[]) {
       am_gmm.Read(ki.Stream(), binary);
     }
 
-    RandomAccessGauPostReader gpost_reader(gpost_rspecifier);
+    RandomAccessGaussPostReader gpost_reader(gpost_rspecifier);
 
     double tot_impr = 0.0, tot_t = 0.0;
 
@@ -117,9 +118,9 @@ int main(int argc, char *argv[]) {
             continue;
           }
           const Matrix<BaseFloat> &feats = feature_reader.Value(utt);
-          const GauPost &gpost = gpost_reader.Value(utt);
+          const GaussPost &gpost = gpost_reader.Value(utt);
           if (static_cast<int32>(gpost.size()) != feats.NumRows()) {
-            KALDI_WARN << "GauPost vector has wrong size " << (gpost.size())
+            KALDI_WARN << "GaussPost vector has wrong size " << (gpost.size())
                        << " vs. " << (feats.NumRows());
             num_other_error++;
             continue;
@@ -153,10 +154,10 @@ int main(int argc, char *argv[]) {
           continue;
         }
         const Matrix<BaseFloat> &feats = feature_reader.Value();
-        const GauPost &gpost = gpost_reader.Value(utt);
+        const GaussPost &gpost = gpost_reader.Value(utt);
 
         if (static_cast<int32>(gpost.size()) != feats.NumRows()) {
-          KALDI_WARN << "GauPost has wrong size " << (gpost.size())
+          KALDI_WARN << "GaussPost has wrong size " << (gpost.size())
               << " vs. " << (feats.NumRows());
           num_other_error++;
           continue;

src/gmmbin/gmm-est-fmllr-raw-gpost.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-est-fmllr-raw-gpost.cc
 
 // Copyright 2013  Johns Hopkins University (author: Daniel Povey)
+//           2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -29,12 +30,12 @@ namespace kaldi {
 
 void AccStatsForUtterance(const TransitionModel &trans_model,
                           const AmDiagGmm &am_gmm,
-                          const GauPost &gpost,
+                          const GaussPost &gpost,
                           const Matrix<BaseFloat> &feats,
                           FmllrRawAccs *accs) {
   for (size_t t = 0; t < gpost.size(); t++) {
     for (size_t i = 0; i < gpost[t].size(); i++) {
-      int32 pdf = trans_model.TransitionIdToPdf(gpost[t][i].first);
+      int32 pdf = gpost[t][i].first;
       const Vector<BaseFloat> &posterior(gpost[t][i].second);      
       accs->AccumulateFromPosteriors(am_gmm.GetPdf(pdf),
                                      feats.Row(t), posterior);
@@ -94,7 +95,7 @@ int main(int argc, char *argv[]) {
     Matrix<BaseFloat> full_lda_mat;
     ReadKaldiObject(full_lda_mat_rxfilename, &full_lda_mat);
     
-    RandomAccessGauPostReader gpost_reader(gpost_rspecifier);
+    RandomAccessGaussPostReader gpost_reader(gpost_rspecifier);
     BaseFloatMatrixWriter transform_writer(transform_wspecifier);
     
     double tot_auxf_impr = 0.0, tot_count = 0.0;
@@ -121,7 +122,7 @@ int main(int argc, char *argv[]) {
             continue;
           }
           const Matrix<BaseFloat> &feats = feature_reader.Value(utt);
-          const GauPost &gpost = gpost_reader.Value(utt);
+          const GaussPost &gpost = gpost_reader.Value(utt);
           if (static_cast<int32>(gpost.size()) != feats.NumRows()) {
             KALDI_WARN << "Size mismatch between gposteriors " << gpost.size()
                        << " and features " << feats.NumRows();
@@ -155,7 +156,7 @@ int main(int argc, char *argv[]) {
           continue;
         }
         const Matrix<BaseFloat> &feats = feature_reader.Value();
-        const GauPost &gpost = gpost_reader.Value(utt);
+        const GaussPost &gpost = gpost_reader.Value(utt);
 
         if (static_cast<int32>(gpost.size()) != feats.NumRows()) {
           KALDI_WARN << "Size mismatch between posteriors " << gpost.size()

src/gmmbin/gmm-est-fmllr-raw.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-est-fmllr-raw.cc
 
 // Copyright 2013  Johns Hopkins University (author: Daniel Povey)
+//           2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -32,13 +33,14 @@ void AccStatsForUtterance(const TransitionModel &trans_model,
                           const Posterior &post,
                           const Matrix<BaseFloat> &feats,
                           FmllrRawAccs *accs) {
+  Posterior pdf_post;
+  ConvertPosteriorToPdfs(trans_model, post, &pdf_post);
   for (size_t t = 0; t < post.size(); t++) {
-    for (size_t i = 0; i < post[t].size(); i++) {
+    for (size_t i = 0; i < pdf_post[t].size(); i++) {
-      int32 pdf = trans_model.TransitionIdToPdf(post[t][i].first);
+      int32 pdf = pdf_post[t][i].first;
-      BaseFloat weight = post[t][i].second;
+      BaseFloat weight = pdf_post[t][i].second;
       accs->AccumulateForGmm(am_gmm.GetPdf(pdf),
-                                  feats.Row(t),
+                             feats.Row(t), weight);
-                                  weight);
     }
   }
 }

src/gmmbin/gmm-est-fmllr.cc

@@ -2,6 +2,7 @@
 
 // Copyright 2009-2011  Microsoft Corporation;  Saarland University
 //                2013  Johns Hopkins University (author: Daniel Povey)
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -36,12 +37,14 @@ void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
                             const TransitionModel &trans_model,
                             const AmDiagGmm &am_gmm,
                             FmllrDiagGmmAccs *spk_stats) {
+  Posterior pdf_post;
+  ConvertPosteriorToPdfs(trans_model, post, &pdf_post);
   for (size_t i = 0; i < post.size(); i++) {
-    for (size_t j = 0; j < post[i].size(); j++) {
+    for (size_t j = 0; j < pdf_post[i].size(); j++) {
-      int32 pdf_id = trans_model.TransitionIdToPdf(post[i][j].first);
+      int32 pdf_id = pdf_post[i][j].first;
       spk_stats->AccumulateForGmm(am_gmm.GetPdf(pdf_id),
                                   feats.Row(i),
-                                  post[i][j].second);
+                                  pdf_post[i][j].second);
     }
   }
 }

src/gmmbin/gmm-est-regtree-fmllr.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-est-regtree-fmllr.cc
 
 // Copyright 2009-2011  Saarland University;  Microsoft Corporation
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -117,10 +118,12 @@ int main(int argc, char *argv[]) {
           }
 
           BaseFloat file_like = 0.0, file_t = 0.0;
+          Posterior pdf_posterior;
+          ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
           for (size_t i = 0; i < posterior.size(); i++) {
-            for (size_t j = 0; j < posterior[i].size(); j++) {
+            for (size_t j = 0; j < pdf_posterior[i].size(); j++) {
-              int32 pdf_id = trans_model.TransitionIdToPdf(posterior[i][j].first);
+              int32 pdf_id = pdf_posterior[i][j].first;
-              BaseFloat prob = posterior[i][j].second;
+              BaseFloat prob = pdf_posterior[i][j].second;
               file_like += fmllr_accs.AccumulateForGmm(regtree, am_gmm,
                                                        feats.Row(i), pdf_id,
                                                        prob);
@@ -168,10 +171,12 @@ int main(int argc, char *argv[]) {
         num_done++;
         BaseFloat file_like = 0.0, file_t = 0.0;
         fmllr_accs.SetZero();
+        Posterior pdf_posterior;
+        ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
         for (size_t i = 0; i < posterior.size(); i++) {
-          for (size_t j = 0; j < posterior[i].size(); j++) {
+          for (size_t j = 0; j < pdf_posterior[i].size(); j++) {
-            int32 pdf_id = trans_model.TransitionIdToPdf(posterior[i][j].first);
+            int32 pdf_id = pdf_posterior[i][j].first;
-            BaseFloat prob = posterior[i][j].second;
+            BaseFloat prob = pdf_posterior[i][j].second;
             file_like += fmllr_accs.AccumulateForGmm(regtree, am_gmm,
                                                      feats.Row(i), pdf_id,
                                                      prob);

src/gmmbin/gmm-est-regtree-mllr.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-est-regtree-mllr.cc
 
 // Copyright 2009-2011  Saarland University;  Microsoft Corporation
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -117,10 +118,12 @@ int main(int argc, char *argv[]) {
           }
 
           BaseFloat file_like = 0.0, file_t = 0.0;
+          Posterior pdf_posterior;
+          ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
           for (size_t i = 0; i < posterior.size(); i++) {
-            for (size_t j = 0; j < posterior[i].size(); j++) {
+            for (size_t j = 0; j < pdf_posterior[i].size(); j++) {
-              int32 pdf_id = trans_model.TransitionIdToPdf(posterior[i][j].first);
+              int32 pdf_id = pdf_posterior[i][j].first;
-              BaseFloat prob = posterior[i][j].second;
+              BaseFloat prob = pdf_posterior[i][j].second;
               file_like += mllr_accs.AccumulateForGmm(regtree, am_gmm,
                                                       feats.Row(i), pdf_id,
                                                       prob);
@@ -168,10 +171,12 @@ int main(int argc, char *argv[]) {
         num_done++;
         BaseFloat file_like = 0.0, file_t = 0.0;
         mllr_accs.SetZero();
+        Posterior pdf_posterior;
+        ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
         for (size_t i = 0; i < posterior.size(); i++) {
-          for (size_t j = 0; j < posterior[i].size(); j++) {
+          for (size_t j = 0; j < pdf_posterior[i].size(); j++) {
-            int32 pdf_id = trans_model.TransitionIdToPdf(posterior[i][j].first);
+            int32 pdf_id = pdf_posterior[i][j].first;
-            BaseFloat prob = posterior[i][j].second;
+            BaseFloat prob = pdf_posterior[i][j].second;
             file_like += mllr_accs.AccumulateForGmm(regtree, am_gmm,
                                                     feats.Row(i), pdf_id,
                                                     prob);

src/gmmbin/gmm-latgen-biglm-faster.cc

@@ -100,25 +100,15 @@ bool DecodeUtterance(LatticeBiglmFasterDecoder &decoder, // not const but is rea
     KALDI_ERR << "Unexpected problem getting lattice for utterance " << utt;
   fst::Connect(&lat);
   if (determinize) {
-    Invert(&lat);
-    if (!TopSort(&lat)) {
-      // Cannot topologically sort the lattice -- determinization will fail.
-      KALDI_WARN << "Topological sorting of state-level lattice failed "
-                 << "(probably your lexicon has empty words or your LM has "
-                 << "epsilon cycles).";
-      return false;
-    }
-    fst::ILabelCompare<LatticeArc> ilabel_comp;
-    ArcSort(&lat, ilabel_comp);
     CompactLattice clat;
-    if (!DeterminizeLatticePhonePruned(trans_model,
+    if (!DeterminizeLatticePhonePrunedWrapper(
+            trans_model,
             &lat,
             decoder.GetOptions().lattice_beam,
             &clat,
             decoder.GetOptions().det_opts))
       KALDI_WARN << "Determinization finished earlier than the beam for "
                  << "utterance " << utt;
-    fst::Connect(&clat);
     // We'll write the lattice without acoustic scaling.
     if (acoustic_scale != 0.0)
       fst::ScaleLattice(fst::AcousticLatticeScale(1.0 / acoustic_scale), &clat);

src/gmmbin/gmm-post-to-gpost.cc

@@ -1,6 +1,7 @@
 // gmmbin/gmm-post-to-gpost.cc
 
 // Copyright 2009-2011  Microsoft Corporation
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -69,7 +70,7 @@ int main(int argc, char *argv[]) {
     SequentialBaseFloatMatrixReader feature_reader(feature_rspecifier);
     RandomAccessPosteriorReader posteriors_reader(posteriors_rspecifier);
 
-    GauPostWriter gpost_writer(gpost_wspecifier);
+    GaussPostWriter gpost_writer(gpost_wspecifier);
 
     int32 num_done = 0, num_no_posterior = 0, num_other_error = 0;
     for (; !feature_reader.Done(); feature_reader.Next()) {
@@ -79,7 +80,7 @@ int main(int argc, char *argv[]) {
       } else {
         const Matrix<BaseFloat> &mat = feature_reader.Value();
         const Posterior &posterior = posteriors_reader.Value(key);
-        GauPost gpost(posterior.size());
+        GaussPost gpost(posterior.size());
 
         if (posterior.size() != mat.NumRows()) {
           KALDI_WARN << "Posterior vector has wrong size "<< (posterior.size()) << " vs. "<< (mat.NumRows());
@@ -90,12 +91,13 @@ int main(int argc, char *argv[]) {
         num_done++;
         BaseFloat tot_like_this_file = 0.0, tot_weight = 0.0;
 
+        Posterior pdf_posterior;
+        ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
         for (size_t i = 0; i < posterior.size(); i++) {
-          gpost[i].reserve(posterior[i].size());
+          gpost[i].reserve(pdf_posterior[i].size());
-          for (size_t j = 0; j < posterior[i].size(); j++) {
+          for (size_t j = 0; j < pdf_posterior[i].size(); j++) {
-            int32 tid = posterior[i][j].first,  // transition identifier.
+            int32 pdf_id = pdf_posterior[i][j].first;
-                pdf_id = trans_model.TransitionIdToPdf(tid);
+            BaseFloat weight = pdf_posterior[i][j].second;
-            BaseFloat weight = posterior[i][j].second;
             const DiagGmm &gmm = am_gmm.GetPdf(pdf_id);
             Vector<BaseFloat> this_post_vec;           
             BaseFloat like =
@@ -106,7 +108,7 @@ int main(int argc, char *argv[]) {
                 this_post_vec(k) = RandPrune(this_post_vec(k),
                                              rand_prune);
             if (!this_post_vec.IsZero())
-              gpost[i].push_back(std::make_pair(tid, this_post_vec));
+              gpost[i].push_back(std::make_pair(pdf_id, this_post_vec));
             tot_like_this_file += like * weight;
             tot_weight += weight;
           }

src/hmm/posterior.cc

@@ -2,6 +2,7 @@
 
 // Copyright 2009-2011  Microsoft Corporation
 //                2013  Johns Hopkins University (author: Daniel Povey)
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -136,13 +137,13 @@ bool PosteriorHolder::Read(std::istream &is) {
 }
 
 // static
-bool GauPostHolder::Write(std::ostream &os, bool binary, const T &t) {
+bool GaussPostHolder::Write(std::ostream &os, bool binary, const T &t) {
   InitKaldiOutputStream(os, binary);  // Puts binary header if binary mode.
   try {
     // We don't bother making this a one-line format.
     int32 sz = t.size();
     WriteBasicType(os, binary, sz);
-    for (GauPost::const_iterator iter = t.begin(); iter != t.end(); ++iter) {
+    for (GaussPost::const_iterator iter = t.begin(); iter != t.end(); ++iter) {
       int32 sz2 = iter->size();
       WriteBasicType(os, binary, sz2);
       for (std::vector<std::pair<int32, Vector<BaseFloat> > >::const_iterator iter2=iter->begin();
@@ -161,7 +162,7 @@ bool GauPostHolder::Write(std::ostream &os, bool binary, const T &t) {
   }
 }
 
-bool GauPostHolder::Read(std::istream &is) {
+bool GaussPostHolder::Read(std::istream &is) {
   t_.clear();
 
   bool is_binary;
@@ -175,7 +176,7 @@ bool GauPostHolder::Read(std::istream &is) {
     if (sz < 0)
       KALDI_ERR << "Reading posteriors: got negative size\n";
     t_.resize(sz);
-    for (GauPost::iterator iter = t_.begin(); iter != t_.end(); ++iter) {
+    for (GaussPost::iterator iter = t_.begin(); iter != t_.end(); ++iter) {
       int32 sz2;
       ReadBasicType(is, is_binary, &sz2);
       if (sz2 < 0)
@@ -270,6 +271,27 @@ void AlignmentToPosterior(const std::vector<int32> &ali,
   }
 }
 
+struct ComparePosteriorByPdfs {
+  const TransitionModel *tmodel_;
+  ComparePosteriorByPdfs(const TransitionModel &tmodel): tmodel_(&tmodel) {}
+  bool operator() (const std::pair<int32, BaseFloat> &a,
+                   const std::pair<int32, BaseFloat> &b) {
+    if (tmodel_->TransitionIdToPdf(a.first)
+        < tmodel_->TransitionIdToPdf(b.first))
+      return true;
+    else
+      return false;
+  }
+};
+
+void SortPosteriorByPdfs(const TransitionModel &tmodel,
+                         Posterior *post) {
+  ComparePosteriorByPdfs compare(tmodel);
+  for (size_t i = 0; i < post->size(); i++) {
+    sort((*post)[i].begin(), (*post)[i].end(), compare);
+  }
+}
+
 void ConvertPosteriorToPdfs(const TransitionModel &tmodel,
                             const Posterior &post_in,
                             Posterior *post_out) {

src/hmm/posterior.h

@@ -2,6 +2,7 @@
 
 // Copyright 2009-2011     Microsoft Corporation
 //           2013-2014     Johns Hopkins University (author: Daniel Povey)
+//                2014     Guoguo Chen
 
 
 // See ../../COPYING for clarification regarding multiple authors
@@ -40,10 +41,14 @@ namespace kaldi {
 /// is a probability (typically between zero and one).
 typedef std::vector<std::vector<std::pair<int32, BaseFloat> > > Posterior;
 
-/// GauPost is a typedef for storing Gaussian-level posteriors for an utterance.
+/// GaussPost is a typedef for storing Gaussian-level posteriors for an utterance.
 /// the "int32" is a transition-id, and the Vector<BaseFloat> is a vector of
 /// Gaussian posteriors.
-typedef std::vector<std::vector<std::pair<int32, Vector<BaseFloat> > > > GauPost;
+/// WARNING: We changed "int32" from transition-id to pdf-id, and the change is
+/// applied for all programs using GaussPost. This is for efficiency purpose. We
+/// also changed the name slightly from GauPost to GaussPost to reduce the
+/// chance that the change will go un-noticed in downstream code.
+typedef std::vector<std::vector<std::pair<int32, Vector<BaseFloat> > > > GaussPost;
 
 
 // PosteriorHolder is a holder for Posterior, which is
@@ -75,19 +80,19 @@ class PosteriorHolder {
 };
 
 
-// GauPostHolder is a holder for GauPost, which is
+// GaussPostHolder is a holder for GaussPost, which is
 // std::vector<std::vector<std::pair<int32, Vector<BaseFloat> > > >
 // This is used for storing posteriors of transition id's for an
 // utterance.
-class GauPostHolder {
+class GaussPostHolder {
  public:
-  typedef GauPost T;
+  typedef GaussPost T;
 
-  GauPostHolder() { }
+  GaussPostHolder() { }
 
   static bool Write(std::ostream &os, bool binary, const T &t);  
 
-  void Clear() {  GauPost tmp;  std::swap(tmp, t_); }
+  void Clear() {  GaussPost tmp;  std::swap(tmp, t_); }
 
   // Reads into the holder.
   bool Read(std::istream &is);
@@ -99,7 +104,7 @@ class GauPostHolder {
   const T &Value() const { return t_; }
   
  private:
-  KALDI_DISALLOW_COPY_AND_ASSIGN(GauPostHolder);
+  KALDI_DISALLOW_COPY_AND_ASSIGN(GaussPostHolder);
   T t_;
 };
 
@@ -112,10 +117,10 @@ typedef SequentialTableReader<PosteriorHolder> SequentialPosteriorReader;
 typedef RandomAccessTableReader<PosteriorHolder> RandomAccessPosteriorReader;
 
 
-// typedef std::vector<std::vector<std::pair<int32, Vector<BaseFloat> > > > GauPost;
+// typedef std::vector<std::vector<std::pair<int32, Vector<BaseFloat> > > > GaussPost;
-typedef TableWriter<GauPostHolder> GauPostWriter;
+typedef TableWriter<GaussPostHolder> GaussPostWriter;
-typedef SequentialTableReader<GauPostHolder> SequentialGauPostReader;
+typedef SequentialTableReader<GaussPostHolder> SequentialGaussPostReader;
-typedef RandomAccessTableReader<GauPostHolder> RandomAccessGauPostReader;
+typedef RandomAccessTableReader<GaussPostHolder> RandomAccessGaussPostReader;
 
 
 /// Scales the BaseFloat (weight) element in the posterior entries.
@@ -146,6 +151,11 @@ int32 MergePosteriors(const Posterior &post1,
 void AlignmentToPosterior(const std::vector<int32> &ali,
                           Posterior *post);
 
+/// Sorts posterior entries so that transition-ids with same pdf-id are next to
+/// each other.
+void SortPosteriorByPdfs(const TransitionModel &tmodel,
+                         Posterior *post);
+
 /// Converts a posterior over transition-ids to be a posterior
 /// over pdf-ids.
 void ConvertPosteriorToPdfs(const TransitionModel &tmodel,

src/lat/determinize-lattice-pruned.cc

@@ -1464,6 +1464,30 @@ bool DeterminizeLatticePhonePruned(
                                        beam, ofst, opts);
 }
 
+bool DeterminizeLatticePhonePrunedWrapper(
+    const kaldi::TransitionModel &trans_model,
+    MutableFst<kaldi::LatticeArc> *ifst,
+    double beam,
+    MutableFst<kaldi::CompactLatticeArc> *ofst,
+    DeterminizeLatticePhonePrunedOptions opts) {
+  bool ans = true;
+  Invert(ifst);
+  if (ifst->Properties(fst::kTopSorted, true) == 0) {
+    if (!TopSort(ifst)) {
+      // Cannot topologically sort the lattice -- determinization will fail.
+      KALDI_ERR << "Topological sorting of state-level lattice failed (probably"
+                << " your lexicon has empty words or your LM has epsilon cycles"
+                << ").";
+    }
+  }
+  ILabelCompare<kaldi::LatticeArc> ilabel_comp;
+  ArcSort(ifst, ilabel_comp);
+  ans = DeterminizeLatticePhonePruned<kaldi::LatticeWeight, kaldi::int32>(
+      trans_model, ifst, beam, ofst, opts);
+  Connect(ofst);
+  return ans;
+}
+
 // Instantiate the templates for the types we might need.
 // Note: there are actually four templates, each of which
 // we instantiate for a single type.

src/lat/determinize-lattice-pruned.h

@@ -30,6 +30,7 @@
 #include "fstext/lattice-weight.h"
 #include "hmm/transition-model.h"
 #include "itf/options-itf.h"
+#include "lat/kaldi-lattice.h"
 
 namespace fst {
 
@@ -256,6 +257,19 @@ bool DeterminizeLatticePhonePruned(
     DeterminizeLatticePhonePrunedOptions opts
       = DeterminizeLatticePhonePrunedOptions());
 
+/** This function is a wrapper of DeterminizeLatticePhonePruned() that works for
+    Lattice type FSTs. Unlike other determinization routines, the function
+    requires "ifst" to have transition-id's on the input side and words on the
+    output side.
+*/
+bool DeterminizeLatticePhonePrunedWrapper(
+    const kaldi::TransitionModel &trans_model,
+    MutableFst<kaldi::LatticeArc> *ifst,
+    double prune,
+    MutableFst<kaldi::CompactLatticeArc> *ofst,
+    DeterminizeLatticePhonePrunedOptions opts
+      = DeterminizeLatticePhonePrunedOptions());
+
 /// @} end "addtogroup fst_extensions"
 
 } // end namespace fst

src/latbin/lattice-determinize-phone-pruned-parallel.cc

@@ -40,34 +40,17 @@ class DeterminizeLatticeTask {
       bool minimize,
       Lattice *lat,
       CompactLatticeWriter *clat_writer,
-      int32 *num_warn,
+      int32 *num_warn):
-      int32 *num_fail):
       trans_model_(&trans_model), opts_(opts), key_(key),
       acoustic_scale_(acoustic_scale), beam_(beam), minimize_(minimize),
-      lat_(lat), clat_writer_(clat_writer), num_warn_(num_warn),
+      lat_(lat), clat_writer_(clat_writer), num_warn_(num_warn) { }
-      num_fail_(num_fail), skip_writting_(false) { }
 
   void operator () () {
-    // Put word labels on the input side.
-    Invert(lat_);
-
     // We apply the acoustic scale before determinization and will undo it
     // afterward, since it can affect the result.
     fst::ScaleLattice(fst::AcousticLatticeScale(acoustic_scale_), lat_);
 
-    if (!TopSort(lat_)) {
+    if (!DeterminizeLatticePhonePrunedWrapper(
-      KALDI_WARN << "Could not topologically sort lattice: this probably means "
-          "it has bad properties e.g. epsilon cycles. Your LM or lexicon might "
-          "be broken, e.g. LM with epsilon cycles or lexicon with empty words.";
-      (*num_fail_)++;
-      skip_writting_ = true;
-
-      delete lat_;
-      lat_ = NULL;
-    } else {
-      fst::ArcSort(lat_, fst::ILabelCompare<LatticeArc>());
-
-      if (!DeterminizeLatticePhonePruned(
             *trans_model_, lat_, beam_, &det_clat_, opts_)) {
       KALDI_WARN << "For key " << key_ << ", determinization did not succeed"
           "(partial output will be pruned tighter than the specified beam.)";
@@ -86,15 +69,12 @@ class DeterminizeLatticeTask {
     fst::ScaleLattice(fst::AcousticLatticeScale(1.0/acoustic_scale_),
                       &det_clat_);
   }
-  }
 
   ~DeterminizeLatticeTask() {
-    if (!skip_writting_) {
     KALDI_VLOG(2) << "Wrote lattice with " << det_clat_.NumStates()
                   << " for key " << key_;
     clat_writer_->Write(key_, det_clat_);
   }
-  }
  private:
   const TransitionModel *trans_model_;
   const fst::DeterminizeLatticePhonePrunedOptions &opts_;
@@ -109,8 +89,6 @@ class DeterminizeLatticeTask {
   CompactLattice det_clat_;
   CompactLatticeWriter *clat_writer_;
   int32 *num_warn_;
-  int32 *num_fail_;
-  bool skip_writting_;
 
 };
 
@@ -175,7 +153,7 @@ int main(int argc, char *argv[]) {
 
     TaskSequencer<DeterminizeLatticeTask> sequencer(sequencer_opts);
 
-    int32 n_done = 0, n_warn = 0, n_fail = 0;
+    int32 n_done = 0, n_warn = 0;
 
     if (acoustic_scale == 0.0)
       KALDI_ERR << "Do not use a zero acoustic scale (cannot be inverted)";
@@ -190,7 +168,7 @@ int main(int argc, char *argv[]) {
 
       DeterminizeLatticeTask *task = new DeterminizeLatticeTask(
           trans_model, determinize_opts, key, acoustic_scale, beam, minimize,
-          lat, &compact_lat_writer, &n_warn, &n_fail);
+          lat, &compact_lat_writer, &n_warn);
       sequencer.Run(task);
 
       n_done++;
@@ -198,7 +176,7 @@ int main(int argc, char *argv[]) {
     sequencer.Wait();
     KALDI_LOG << "Done " << n_done << " lattices, determinization finished "
               << "earlier than specified by the beam on " << n_warn << " of "
-              << "these, failed for " << n_fail;
+              << "these.";
     return (n_done != 0 ? 0 : 1);
   } catch(const std::exception &e) {
     std::cerr << e.what();

src/latbin/lattice-determinize-phone-pruned.cc

@@ -77,7 +77,7 @@ int main(int argc, char *argv[]) {
     // Writes as compact lattice.
     CompactLatticeWriter compact_lat_writer(lats_wspecifier); 
 
-    int32 n_done = 0, n_warn = 0, n_fail = 0;
+    int32 n_done = 0, n_warn = 0;
 
     if (acoustic_scale == 0.0)
       KALDI_ERR << "Do not use a zero acoustic scale (cannot be inverted)";
@@ -89,23 +89,11 @@ int main(int argc, char *argv[]) {
 
       KALDI_VLOG(2) << "Processing lattice " << key;
 
-      Invert(&lat); // so word labels are on the input side.
-
       fst::ScaleLattice(fst::AcousticLatticeScale(acoustic_scale), &lat);
 
-      if (!TopSort(&lat)) {
-        KALDI_WARN << "Could not topologically sort lattice: this probably "
-            "means it has bad properties e.g. epsilon cycles.  Your LM or "
-            "lexicon might be broken, e.g. LM with epsilon cycles or lexicon "
-            "with empty words.";
-        n_fail++;
-        continue;
-      }
-      fst::ArcSort(&lat, fst::ILabelCompare<LatticeArc>());
-
       CompactLattice det_clat;
-      if (!DeterminizeLatticePhonePruned(
+      if (!DeterminizeLatticePhonePrunedWrapper(
-              trans_model, lat, beam, &det_clat, opts)) {
+              trans_model, &lat, beam, &det_clat, opts)) {
         KALDI_WARN << "For key " << key << ", determinization did not succeed"
             "(partial output will be pruned tighter than the specified beam.)";
         n_warn++;
@@ -124,7 +112,7 @@ int main(int argc, char *argv[]) {
 
     KALDI_LOG << "Done " << n_done << " lattices, determinization finished "
               << "earlier than specified by the beam on " << n_warn << " of "
-              << "these, failed for " << n_fail;
+              << "these.";
     return (n_done != 0 ? 0 : 1);
   } catch(const std::exception &e) {
     std::cerr << e.what();

src/sgmm2bin/sgmm2-acc-stats.cc

@@ -2,6 +2,7 @@
 
 // Copyright 2009-2012   Saarland University (Author:  Arnab Ghoshal),
 //                       Johns Hopkins University (Author:  Daniel Povey)
+//                2014   Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -161,20 +162,27 @@ int main(int argc, char *argv[]) {
       
         BaseFloat tot_like_this_file = 0.0, tot_weight = 0.0;
 
+        Posterior pdf_posterior;
+        ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
         for (size_t i = 0; i < posterior.size(); i++) {
           am_sgmm.ComputePerFrameVars(features.Row(i), gselect[i], spk_vars,
                                       &per_frame_vars);
-
+          // Accumulates for SGMM.
-          for (size_t j = 0; j < posterior[i].size(); j++) {
+          for (size_t j = 0; j < pdf_posterior[i].size(); j++) {
-            int32 tid = posterior[i][j].first,  // transition identifier.
+            int32 pdf_id = pdf_posterior[i][j].first;
-                pdf_id = trans_model.TransitionIdToPdf(tid);
+            BaseFloat weight = pdf_posterior[i][j].second;
-            BaseFloat weight = posterior[i][j].second;
-            trans_model.Accumulate(weight, tid, &transition_accs);
             tot_like_this_file += sgmm_accs.Accumulate(am_sgmm, per_frame_vars,
                                                        pdf_id, weight, &spk_vars)
                 * weight;
             tot_weight += weight;
           }
+
+          // Accumulates for transitions.
+          for (size_t j = 0; j < posterior[i].size(); j++) {
+            int32 tid = posterior[i][j].first;
+            BaseFloat weight = posterior[i][j].second;
+            trans_model.Accumulate(weight, tid, &transition_accs);
+          }
         }
         
         KALDI_VLOG(2) << "Average like for this file is "

src/sgmm2bin/sgmm2-est-fmllr.cc

@@ -1,6 +1,7 @@
 // sgmm2bin/sgmm2-est-fmllr.cc
 
 // Copyright 2009-2012  Saarland University   Microsoft Corporation  Johns Hopkins University (Author: Daniel Povey)
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -42,6 +43,8 @@ void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
                             FmllrSgmm2Accs *spk_stats) {
   kaldi::Sgmm2PerFrameDerivedVars per_frame_vars;
 
+  Posterior pdf_post;
+  ConvertPosteriorToPdfs(trans_model, post, &pdf_post);
   for (size_t t = 0; t < post.size(); t++) {
     // per-frame vars only used for computing posteriors... use the
     // transformed feats for this, if available.
@@ -49,12 +52,12 @@ void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
                                 *spk_vars, &per_frame_vars);
     
 
-    for (size_t j = 0; j < post[t].size(); j++) {
+    for (size_t j = 0; j < pdf_post[t].size(); j++) {
-      int32 pdf_id = trans_model.TransitionIdToPdf(post[t][j].first);
+      int32 pdf_id = pdf_post[t][j].first;
       Matrix<BaseFloat> posteriors;
       am_sgmm.ComponentPosteriors(per_frame_vars, pdf_id,
                                   spk_vars, &posteriors);
-      posteriors.Scale(post[t][j].second);
+      posteriors.Scale(pdf_post[t][j].second);
       spk_stats->AccumulateFromPosteriors(am_sgmm, *spk_vars, feats.Row(t),
                                           gselect[t], posteriors, pdf_id);
     }

src/sgmm2bin/sgmm2-est-spkvecs.cc

@@ -2,6 +2,7 @@
 
 // Copyright 2009-2012  Saarland University  Microsoft Corporation
 //                      Johns Hopkins University (Author: Daniel Povey)
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -42,14 +43,16 @@ void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
   kaldi::Sgmm2PerFrameDerivedVars per_frame_vars;
 
   KALDI_ASSERT(gselect.size() == feats.NumRows());
+  Posterior pdf_post;
+  ConvertPosteriorToPdfs(trans_model, post, &pdf_post);
   for (size_t i = 0; i < post.size(); i++) {
     am_sgmm.ComputePerFrameVars(feats.Row(i), gselect[i],
                                 *spk_vars, &per_frame_vars);
     
-    for (size_t j = 0; j < post[i].size(); j++) {
+    for (size_t j = 0; j < pdf_post[i].size(); j++) {
-      int32 pdf_id = trans_model.TransitionIdToPdf(post[i][j].first);
+      int32 pdf_id = pdf_post[i][j].first;
       spk_stats->Accumulate(am_sgmm, per_frame_vars, pdf_id,
-                            post[i][j].second, spk_vars);
+                            pdf_post[i][j].second, spk_vars);
     }
   }
 }

src/sgmm2bin/sgmm2-post-to-gpost.cc

@@ -2,6 +2,7 @@
 
 // Copyright 2009-2012   Saarland University  Microsoft Corporation
 //                       Johns Hopkins University (Author: Daniel Povey)
+//                2014   Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -94,7 +95,7 @@ int main(int argc, char *argv[]) {
         num_err++;
         continue;
       }
-      const Posterior &posterior = posteriors_reader.Value(utt);
+      Posterior posterior = posteriors_reader.Value(utt);
 
       if (!gselect_reader.HasKey(utt) ||
           gselect_reader.Value(utt).size() != mat.NumRows()) {
@@ -123,6 +124,10 @@ int main(int argc, char *argv[]) {
 
       Sgmm2GauPost gpost(posterior.size());  // posterior.size() == T.
 
+      SortPosteriorByPdfs(trans_model, &posterior);
+      int32 prev_pdf_id = -1;
+      BaseFloat prev_like = 0;
+      Matrix<BaseFloat> prev_posterior;
       for (size_t i = 0; i < posterior.size(); i++) {
         am_sgmm.ComputePerFrameVars(mat.Row(i), gselect[i],
                                     spk_vars, &per_frame_vars);
@@ -131,17 +136,24 @@ int main(int argc, char *argv[]) {
         gpost[i].tids.resize(posterior[i].size());
         gpost[i].posteriors.resize(posterior[i].size());
 
+        prev_pdf_id = -1;       // Only cache for the same frame.
         for (size_t j = 0; j < posterior[i].size(); j++) {
           int32 tid = posterior[i][j].first,  // transition identifier.
               pdf_id = trans_model.TransitionIdToPdf(tid);
           BaseFloat weight = posterior[i][j].second;
           gpost[i].tids[j] = tid;
 
-          tot_like_this_file +=
+          if (pdf_id != prev_pdf_id) {
-              am_sgmm.ComponentPosteriors(per_frame_vars, pdf_id,
+            // First time see this pdf-id for this frame, update the cached
+            // variables.
+            prev_pdf_id = pdf_id;
+            prev_like = am_sgmm.ComponentPosteriors(per_frame_vars, pdf_id,
                                                     &spk_vars,
-                                          &(gpost[i].posteriors[j]))
+                                                    &prev_posterior);
-              * weight;
+          }
+
+          gpost[i].posteriors[j] = prev_posterior;
+          tot_like_this_file += prev_like * weight;
           tot_weight += weight;
           gpost[i].posteriors[j].Scale(weight);
         }

src/sgmmbin/sgmm-acc-stats.cc

@@ -1,6 +1,7 @@
 // sgmmbin/sgmm-acc-stats.cc
 
 // Copyright 2009-2011   Saarland University (Author:  Arnab Ghoshal),
+//                2014   Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -138,6 +139,8 @@ int main(int argc, char *argv[]) {
         num_done++;
         BaseFloat tot_like_this_file = 0.0, tot_weight = 0.0;
 
+        Posterior pdf_posterior;
+        ConvertPosteriorToPdfs(trans_model, posterior, &pdf_posterior);
         for (size_t i = 0; i < posterior.size(); i++) {
           if (posterior[i].empty())
             continue;
@@ -147,18 +150,25 @@ int main(int argc, char *argv[]) {
           am_sgmm.ComputePerFrameVars(mat.Row(i), this_gselect, spk_vars, 0.0,
                                       &per_frame_vars);
 
-          for (size_t j = 0; j < posterior[i].size(); j++) {
+          // Accumulates for SGMM.
-            int32 tid = posterior[i][j].first,  // transition identifier.
+          for (size_t j = 0; j < pdf_posterior[i].size(); j++) {
-                pdf_id = trans_model.TransitionIdToPdf(tid);
+            int32 pdf_id = pdf_posterior[i][j].first;
-            BaseFloat weight = posterior[i][j].second;
+            BaseFloat weight = pdf_posterior[i][j].second;
-            if (acc_flags & kaldi::kSgmmTransitions)
-              trans_model.Accumulate(weight, tid, &transition_accs);
             tot_like_this_file += sgmm_accs.Accumulate(am_sgmm, per_frame_vars,
                                                        spk_vars.v_s, pdf_id,
                                                        weight, acc_flags)
                                                        * weight;
             tot_weight += weight;
           }
+
+          // Accumulates for transitions.
+          for (size_t j = 0; j < posterior[i].size(); j++) {
+            if (acc_flags & kaldi::kSgmmTransitions) {
+              int32 tid = posterior[i][j].first;
+              BaseFloat weight = posterior[i][j].second;
+              trans_model.Accumulate(weight, tid, &transition_accs);
+            }
+          }
         }
 
         sgmm_accs.CommitStatsForSpk(am_sgmm, spk_vars.v_s);  // no harm doing it per utterance.

src/sgmmbin/sgmm-est-fmllr.cc

@@ -1,6 +1,7 @@
 // sgmmbin/sgmm-est-fmllr.cc
 
 // Copyright 2009-2012  Saarland University   Microsoft Corporation  Johns Hopkins University (Author: Daniel Povey)
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -43,6 +44,8 @@ void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
                             FmllrSgmmAccs *spk_stats) {
   kaldi::SgmmPerFrameDerivedVars per_frame_vars;
 
+  Posterior pdf_post;
+  ConvertPosteriorToPdfs(trans_model, post, &pdf_post);
   for (size_t t = 0; t < post.size(); t++) {
     std::vector<int32> this_gselect;
     if (!gselect.empty()) {
@@ -57,12 +60,12 @@ void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
                                 0.0 /*fMLLR logdet*/, &per_frame_vars);
 
 
-    for (size_t j = 0; j < post[t].size(); j++) {
+    for (size_t j = 0; j < pdf_post[t].size(); j++) {
-      int32 pdf_id = trans_model.TransitionIdToPdf(post[t][j].first);
+      int32 pdf_id = pdf_post[t][j].first;
       Matrix<BaseFloat> posteriors;
       am_sgmm.ComponentPosteriors(per_frame_vars, pdf_id,
                                   &posteriors);
-      posteriors.Scale(post[t][j].second);
+      posteriors.Scale(pdf_post[t][j].second);
       spk_stats->AccumulateFromPosteriors(am_sgmm, spk_vars, feats.Row(t),
                                           this_gselect,
                                           posteriors, pdf_id);

src/sgmmbin/sgmm-est-spkvecs.cc

@@ -1,6 +1,7 @@
 // sgmmbin/sgmm-est-spkvecs.cc
 
 // Copyright 2009-2012  Saarland University  Microsoft Corporation  Johns Hopkins University (Author: Daniel Povey)
+//                2014  Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -41,6 +42,8 @@ void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
                             MleSgmmSpeakerAccs *spk_stats) {
   kaldi::SgmmPerFrameDerivedVars per_frame_vars;
 
+  Posterior pdf_post;
+  ConvertPosteriorToPdfs(trans_model, post, &pdf_post);
   for (size_t i = 0; i < post.size(); i++) {
     std::vector<int32> this_gselect;
     if (!gselect.empty())
@@ -49,9 +52,9 @@ void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
       am_sgmm.GaussianSelection(gselect_opts, feats.Row(i), &this_gselect);
     am_sgmm.ComputePerFrameVars(feats.Row(i), this_gselect, spk_vars, 0.0, &per_frame_vars);
 
-    for (size_t j = 0; j < post[i].size(); j++) {
+    for (size_t j = 0; j < pdf_post[i].size(); j++) {
-      int32 pdf_id = trans_model.TransitionIdToPdf(post[i][j].first);
+      int32 pdf_id = pdf_post[i][j].first;
-      spk_stats->Accumulate(am_sgmm, per_frame_vars, pdf_id, post[i][j].second);
+      spk_stats->Accumulate(am_sgmm, per_frame_vars, pdf_id, pdf_post[i][j].second);
     }
   }
 }

src/sgmmbin/sgmm-post-to-gpost.cc

@@ -1,6 +1,7 @@
 // sgmmbin/sgmm-post-to-gpost.cc
 
 // Copyright 2009-2012   Saarland University  Microsoft Corporation  Johns Hopkins University (Author: Daniel Povey)
+//                2014   Guoguo Chen
 
 // See ../../COPYING for clarification regarding multiple authors
 //
@@ -87,7 +88,7 @@ int main(int argc, char *argv[]) {
         num_no_posterior++;
       } else {
         const Matrix<BaseFloat> &mat = feature_reader.Value();
-        const Posterior &posterior = posteriors_reader.Value(utt);
+        Posterior posterior = posteriors_reader.Value(utt);
 
         bool have_gselect  = !gselect_rspecifier.empty()
             && gselect_reader.HasKey(utt)
@@ -123,6 +124,10 @@ int main(int argc, char *argv[]) {
 
         SgmmGauPost gpost(posterior.size());  // posterior.size() == T.
 
+        SortPosteriorByPdfs(trans_model, &posterior);
+        int32 prev_pdf_id = -1;
+        BaseFloat prev_like = 0;
+        Matrix<BaseFloat> prev_posterior;
         for (size_t i = 0; i < posterior.size(); i++) {
 
           std::vector<int32> this_gselect;
@@ -134,16 +139,23 @@ int main(int argc, char *argv[]) {
           gpost[i].tids.resize(posterior[i].size());
           gpost[i].posteriors.resize(posterior[i].size());
 
+          prev_pdf_id = -1;       // Only cache for the same frame.
           for (size_t j = 0; j < posterior[i].size(); j++) {
             int32 tid = posterior[i][j].first,  // transition identifier.
                 pdf_id = trans_model.TransitionIdToPdf(tid);
             BaseFloat weight = posterior[i][j].second;
             gpost[i].tids[j] = tid;
 
-            tot_like_this_file +=
+            if (pdf_id != prev_pdf_id) {
-                am_sgmm.ComponentPosteriors(per_frame_vars, pdf_id,
+              // First time see this pdf-id for this frame, update the cached
-                                            &(gpost[i].posteriors[j]))
+              // variables.
-                * weight;
+              prev_pdf_id = pdf_id;
+              prev_like = am_sgmm.ComponentPosteriors(per_frame_vars, pdf_id,
+                                                      &prev_posterior);
+            }
+
+            gpost[i].posteriors[j] = prev_posterior;
+            tot_like_this_file += prev_like * weight;
             tot_weight += weight;
             gpost[i].posteriors[j].Scale(weight);
           }