// gmmbin/gmm-est-fmllr.cc
  
  // Copyright 2009-2011  Microsoft Corporation;  Saarland University
  //                2013  Johns Hopkins University (author: Daniel Povey)
  //                2014  Guoguo Chen
  
  // See ../../COPYING for clarification regarding multiple authors
  //
  // Licensed under the Apache License, Version 2.0 (the "License");
  // you may not use this file except in compliance with the License.
  // You may obtain a copy of the License at
  //
  //  http://www.apache.org/licenses/LICENSE-2.0
  //
  // THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
  // KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
  // WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
  // MERCHANTABLITY OR NON-INFRINGEMENT.
  // See the Apache 2 License for the specific language governing permissions and
  // limitations under the License.
  
  #include <string>
  using std::string;
  #include <vector>
  using std::vector;
  
  #include "base/kaldi-common.h"
  #include "util/common-utils.h"
  #include "gmm/am-diag-gmm.h"
  #include "hmm/transition-model.h"
  #include "transform/fmllr-diag-gmm.h"
  #include "hmm/posterior.h"
  
  namespace kaldi {
  void AccumulateForUtterance(const Matrix<BaseFloat> &feats,
                              const Posterior &post,
                              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 < pdf_post[i].size(); j++) {
        int32 pdf_id = pdf_post[i][j].first;
        spk_stats->AccumulateForGmm(am_gmm.GetPdf(pdf_id),
                                    feats.Row(i),
                                    pdf_post[i][j].second);
      }
    }
  }
  
  
  }
  
  int main(int argc, char *argv[]) {
    try {
      typedef kaldi::int32 int32;
      using namespace kaldi;
      const char *usage =
          "Estimate global fMLLR transforms, either per utterance or for the supplied
  "
          "set of speakers (spk2utt option).  Reads posteriors (on transition-ids).  Writes
  "
          "to a table of matrices.
  "
          "Usage: gmm-est-fmllr [options] <model-in> "
          "<feature-rspecifier> <post-rspecifier> <transform-wspecifier>
  ";
  
      ParseOptions po(usage);
      FmllrOptions fmllr_opts;
      string spk2utt_rspecifier;
      po.Register("spk2utt", &spk2utt_rspecifier, "rspecifier for speaker to "
                  "utterance-list map");
      fmllr_opts.Register(&po);
  
      po.Read(argc, argv);
  
      if (po.NumArgs() != 4) {
        po.PrintUsage();
        exit(1);
      }
  
      string
          model_rxfilename = po.GetArg(1),
          feature_rspecifier = po.GetArg(2),
          post_rspecifier = po.GetArg(3),
          trans_wspecifier = po.GetArg(4);
  
      TransitionModel trans_model;
      AmDiagGmm am_gmm;
      {
        bool binary;
        Input ki(model_rxfilename, &binary);
        trans_model.Read(ki.Stream(), binary);
        am_gmm.Read(ki.Stream(), binary);
      }
  
      RandomAccessPosteriorReader post_reader(post_rspecifier);
  
      double tot_impr = 0.0, tot_t = 0.0;
  
      BaseFloatMatrixWriter transform_writer(trans_wspecifier);
  
      int32 num_done = 0, num_no_post = 0, num_other_error = 0;
      if (spk2utt_rspecifier != "") {  // per-speaker adaptation
        SequentialTokenVectorReader spk2utt_reader(spk2utt_rspecifier);
        RandomAccessBaseFloatMatrixReader feature_reader(feature_rspecifier);
  
        for (; !spk2utt_reader.Done(); spk2utt_reader.Next()) {
          FmllrDiagGmmAccs spk_stats(am_gmm.Dim(), fmllr_opts);
          string spk = spk2utt_reader.Key();
          const vector<string> &uttlist = spk2utt_reader.Value();
          for (size_t i = 0; i < uttlist.size(); i++) {
            std::string utt = uttlist[i];
            if (!feature_reader.HasKey(utt)) {
              KALDI_WARN << "Did not find features for utterance " << utt;
              num_other_error++;
              continue;
            }
            if (!post_reader.HasKey(utt)) {
              KALDI_WARN << "Did not find posteriors for utterance " << utt;
              num_no_post++;
              continue;
            }
            const Matrix<BaseFloat> &feats = feature_reader.Value(utt);
            const Posterior &post = post_reader.Value(utt);
            if (static_cast<int32>(post.size()) != feats.NumRows()) {
              KALDI_WARN << "Posterior vector has wrong size " << (post.size())
                         << " vs. " << (feats.NumRows());
              num_other_error++;
              continue;
            }
  
            AccumulateForUtterance(feats, post, trans_model, am_gmm, &spk_stats);
  
            num_done++;
          }  // end looping over all utterances of the current speaker
  
          BaseFloat impr, spk_tot_t;
          {  // Compute the transform and write it out.
            Matrix<BaseFloat> transform(am_gmm.Dim(), am_gmm.Dim()+1);
            transform.SetUnit();
            spk_stats.Update(fmllr_opts, &transform, &impr, &spk_tot_t);
            transform_writer.Write(spk, transform);
          }
          KALDI_LOG << "For speaker " << spk << ", auxf-impr from fMLLR is "
                    << (impr/spk_tot_t) << ", over " << spk_tot_t << " frames.";
          tot_impr += impr;
          tot_t += spk_tot_t;
        }  // end looping over speakers
      } else {  // per-utterance adaptation
        SequentialBaseFloatMatrixReader feature_reader(feature_rspecifier);
        for (; !feature_reader.Done(); feature_reader.Next()) {
          string utt = feature_reader.Key();
          if (!post_reader.HasKey(utt)) {
            KALDI_WARN << "Did not find posts for utterance "
                       << utt;
            num_no_post++;
            continue;
          }
          const Matrix<BaseFloat> &feats = feature_reader.Value();
          const Posterior &post = post_reader.Value(utt);
  
          if (static_cast<int32>(post.size()) != feats.NumRows()) {
            KALDI_WARN << "Posterior has wrong size " << (post.size())
                << " vs. " << (feats.NumRows());
            num_other_error++;
            continue;
          }
          num_done++;
  
          FmllrDiagGmmAccs spk_stats(am_gmm.Dim(), fmllr_opts);
  
          AccumulateForUtterance(feats, post, trans_model, am_gmm,
                                 &spk_stats);
  
          BaseFloat impr, utt_tot_t;
          {  // Compute the transform and write it out.
            Matrix<BaseFloat> transform(am_gmm.Dim(), am_gmm.Dim()+1);
            transform.SetUnit();
            spk_stats.Update(fmllr_opts, &transform, &impr, &utt_tot_t);
            transform_writer.Write(utt, transform);
          }
          KALDI_LOG << "For utterance " << utt << ", auxf-impr from fMLLR is "
                    << (impr/utt_tot_t) << ", over " << utt_tot_t << " frames.";
          tot_impr += impr;
          tot_t += utt_tot_t;
        }
      }
  
      KALDI_LOG << "Done " << num_done << " files, " << num_no_post
                << " with no posts, " << num_other_error << " with other errors.";
      KALDI_LOG << "Overall fMLLR auxf impr per frame is "
                << (tot_impr / tot_t) << " over " << tot_t << " frames.";
      return (num_done != 0 ? 0 : 1);
    } catch(const std::exception &e) {
      std::cerr << e.what();
      return -1;
    }
  }