nnet-latgen-faster-parallel.cc File Reference

#include "base/kaldi-common.h"
#include "util/common-utils.h"
#include "tree/context-dep.h"
#include "hmm/transition-model.h"
#include "fstext/kaldi-fst-io.h"
#include "decoder/decoder-wrappers.h"
#include "nnet2/decodable-am-nnet.h"
#include "base/timer.h"
#include "util/kaldi-thread.h"

Include dependency graph for nnet-latgen-faster-parallel.cc:

Go to the source code of this file.

Functions
int	main (int argc, char *argv[])

Function Documentation

main()

int main	(	int	argc,
		char *	argv[]
	)

Definition at line 33 of file nnet-latgen-faster-parallel.cc.

References kaldi::ClassifyRspecifier(), LatticeFasterDecoderConfig::determinize_lattice, SequentialTableReader< Holder >::Done(), Timer::Elapsed(), ParseOptions::GetArg(), ParseOptions::GetOptArg(), RandomAccessTableReader< Holder >::HasKey(), KALDI_ERR, KALDI_LOG, KALDI_WARN, SequentialTableReader< Holder >::Key(), kaldi::kNoRspecifier, SequentialTableReader< Holder >::Next(), TaskSequencerConfig::num_threads, ParseOptions::NumArgs(), MatrixBase< Real >::NumRows(), TableWriter< Holder >::Open(), ParseOptions::PrintUsage(), AmNnet::Read(), ParseOptions::Read(), TransitionModel::Read(), fst::ReadFstKaldiGeneric(), LatticeFasterDecoderConfig::Register(), ParseOptions::Register(), TaskSequencerConfig::Register(), Timer::Reset(), TaskSequencer< C >::Run(), Input::Stream(), RandomAccessTableReader< Holder >::Value(), SequentialTableReader< Holder >::Value(), and TaskSequencer< C >::Wait().

                                 {
  try {
    using namespace kaldi;
    using namespace kaldi::nnet2;
    typedef kaldi::int32 int32;
    using fst::SymbolTable;
    using fst::Fst;
    using fst::StdArc;

    const char *usage =
        "Generate lattices using neural net model.\n"
        "Usage: nnet-latgen-faster-parallel [options] <nnet-in> <fst-in|fsts-rspecifier> <features-rspecifier>"
        " <lattice-wspecifier> [ <words-wspecifier> [<alignments-wspecifier>] ]\n";
    ParseOptions po(usage);
    Timer timer;
    bool allow_partial = false;
    BaseFloat acoustic_scale = 0.1;
    LatticeFasterDecoderConfig config;
    TaskSequencerConfig sequencer_config; // has --num-threads option

    std::string word_syms_filename;
    sequencer_config.Register(&po);
    config.Register(&po);
    po.Register("acoustic-scale", &acoustic_scale, "Scaling factor for acoustic likelihoods");
    po.Register("word-symbol-table", &word_syms_filename, "Symbol table for words [for debug output]");
    po.Register("allow-partial", &allow_partial, "If true, produce output even if end state was not reached.");

    po.Read(argc, argv);

    if (po.NumArgs() < 4 || po.NumArgs() > 6) {
      po.PrintUsage();
      exit(1);
    }

    std::string model_in_filename = po.GetArg(1),
        fst_in_str = po.GetArg(2),
        feature_rspecifier = po.GetArg(3),
        lattice_wspecifier = po.GetArg(4),
        words_wspecifier = po.GetOptArg(5),
        alignment_wspecifier = po.GetOptArg(6);

    TransitionModel trans_model;
    AmNnet am_nnet;
    {
      bool binary;
      Input ki(model_in_filename, &binary);
      trans_model.Read(ki.Stream(), binary);
      am_nnet.Read(ki.Stream(), binary);
    }

    bool determinize = config.determinize_lattice;
    CompactLatticeWriter compact_lattice_writer;
    LatticeWriter lattice_writer;
    if (! (determinize ? compact_lattice_writer.Open(lattice_wspecifier)
           : lattice_writer.Open(lattice_wspecifier)))
      KALDI_ERR << "Could not open table for writing lattices: "
                 << lattice_wspecifier;

    TaskSequencer<DecodeUtteranceLatticeFasterClass> sequencer(sequencer_config);

    Int32VectorWriter words_writer(words_wspecifier);

    Int32VectorWriter alignment_writer(alignment_wspecifier);

    fst::SymbolTable *word_syms = NULL;
    if (word_syms_filename != "")
      if (!(word_syms = fst::SymbolTable::ReadText(word_syms_filename)))
        KALDI_ERR << "Could not read symbol table from file "
                   << word_syms_filename;

    // We support reading in a vector to describe each speaker, if the neural
    // net requires this (i.e. it was trained with this).

    double tot_like = 0.0;
    kaldi::int64 frame_count = 0;
    int num_done = 0, num_err = 0;
    Fst<StdArc> *decode_fst = NULL;
    if (ClassifyRspecifier(fst_in_str, NULL, NULL) == kNoRspecifier) {
      SequentialBaseFloatMatrixReader feature_reader(feature_rspecifier);

      decode_fst = fst::ReadFstKaldiGeneric(fst_in_str);
      timer.Reset();

      {

        for (; !feature_reader.Done(); feature_reader.Next()) {
          std::string utt = feature_reader.Key();
          const Matrix<BaseFloat> &features (feature_reader.Value());
          if (features.NumRows() == 0) {
            KALDI_WARN << "Zero-length utterance: " << utt;
            num_err++;
            continue;
          }
          bool pad_input = true;
          DecodableAmNnetParallel *nnet_decodable = new DecodableAmNnetParallel(
              trans_model, am_nnet,
              new CuMatrix<BaseFloat>(features),
              pad_input, acoustic_scale);

          LatticeFasterDecoder *decoder = new LatticeFasterDecoder(*decode_fst,
                                                                   config);

          DecodeUtteranceLatticeFasterClass *task =
              new DecodeUtteranceLatticeFasterClass(
                  decoder, nnet_decodable, // takes ownership of these two.
                  trans_model, word_syms, utt, acoustic_scale, determinize,
                  allow_partial, &alignment_writer, &words_writer,
                  &compact_lattice_writer, &lattice_writer,
                  &tot_like, &frame_count, &num_done, &num_err, NULL);

          sequencer.Run(task); // takes ownership of "task",
                               // and will delete it when done.
        }
      }
    } else { // We have different FSTs for different utterances.
      SequentialTableReader<fst::VectorFstHolder> fst_reader(fst_in_str);
      RandomAccessBaseFloatMatrixReader feature_reader(feature_rspecifier);
      for (; !fst_reader.Done(); fst_reader.Next()) {
        std::string utt = fst_reader.Key();
        if (!feature_reader.HasKey(utt)) {
          KALDI_WARN << "Not decoding utterance " << utt
                     << " because no features available.";
          num_err++;
          continue;
        }
        const Matrix<BaseFloat> &features = feature_reader.Value(utt);
        if (features.NumRows() == 0) {
          KALDI_WARN << "Zero-length utterance: " << utt;
          num_err++;
          continue;
        }

        // This constructor of LatticeFasterDecoder takes ownership of the FST.
        LatticeFasterDecoder *decoder =
            new LatticeFasterDecoder(config, fst_reader.Value().Copy());

        bool pad_input = true;
        DecodableAmNnetParallel *nnet_decodable = new DecodableAmNnetParallel(
            trans_model, am_nnet,
            new CuMatrix<BaseFloat>(features),
            pad_input, acoustic_scale);

        DecodeUtteranceLatticeFasterClass *task =
            new DecodeUtteranceLatticeFasterClass(
                decoder, nnet_decodable, // takes ownership of these two.
                trans_model, word_syms, utt, acoustic_scale, determinize,
                allow_partial, &alignment_writer, &words_writer,
                &compact_lattice_writer, &lattice_writer,
                &tot_like, &frame_count, &num_done, &num_err, NULL);

        sequencer.Run(task); // takes ownership of "task",
                             // and will delete it when done.
      }
    }
    sequencer.Wait(); // Waits for all tasks to be done.
    delete decode_fst;

    double elapsed = timer.Elapsed();
    KALDI_LOG << "Time taken "<< elapsed
              << "s: real-time factor per thread assuming 100 frames/sec is "
              << (sequencer_config.num_threads * elapsed * 100.0 / frame_count);
    KALDI_LOG << "Done " << num_done << " utterances, failed for "
              << num_err;
    KALDI_LOG << "Overall log-likelihood per frame is "
              << (tot_like / frame_count) << " over " << frame_count
              << " frames.";

    delete word_syms;
    if (num_done != 0) return 0;
    else return 1;
  } catch(const std::exception &e) {
    std::cerr << e.what();
    return -1;
  }
}