lattice-determinize-non-compact.cc File Reference

#include "base/kaldi-common.h"
#include "util/common-utils.h"
#include "util/stl-utils.h"
#include "fstext/fstext-lib.h"
#include "lat/kaldi-lattice.h"
#include "lat/lattice-functions.h"
#include "lat/push-lattice.h"
#include "lat/minimize-lattice.h"

Include dependency graph for lattice-determinize-non-compact.cc:

Go to the source code of this file.

Namespaces
	kaldi
	This code computes Goodness of Pronunciation (GOP) and extracts phone-level pronunciation feature for mispronunciations detection tasks, the reference:

Typedefs
typedef Lattice::StateId	StateId
typedef Lattice::Arc	Arc

Functions
bool	DeterminizeLatticeWrapper (const Lattice &lat, const std::string &key, bool prune, BaseFloat beam, BaseFloat beam_ratio, int32 max_mem, int32 max_loop, BaseFloat delta, int32 num_loops, CompactLattice *clat)
int	main (int argc, char *argv[])

Function Documentation

main()

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

Definition at line 95 of file lattice-determinize-non-compact.cc.

References fst::AcousticLatticeScale(), kaldi::CompactLatticeDepth(), kaldi::ComputeAcousticScoresMap(), fst::ConvertLattice(), kaldi::DeterminizeLatticeWrapper(), SequentialTableReader< Holder >::Done(), SequentialTableReader< Holder >::FreeCurrent(), ParseOptions::GetArg(), KALDI_ERR, KALDI_LOG, SequentialTableReader< Holder >::Key(), fst::MinimizeCompactLattice(), SequentialTableReader< Holder >::Next(), ParseOptions::NumArgs(), ParseOptions::PrintUsage(), fst::PushCompactLatticeStrings(), fst::PushCompactLatticeWeights(), ParseOptions::Read(), ParseOptions::Register(), kaldi::ReplaceAcousticScoresFromMap(), fst::ScaleLattice(), kaldi::TopSortCompactLatticeIfNeeded(), SequentialTableReader< Holder >::Value(), and TableWriter< Holder >::Write().

                                 {
  try {
    using namespace kaldi;
    typedef kaldi::int32 int32;
    typedef kaldi::int64 int64;
    using fst::SymbolTable;
    using fst::VectorFst;
    using fst::StdArc;

    const char *usage =
        "lattice-determinize lattices (and apply a pruning beam)\n"
        " (see http://kaldi-asr.org/doc/lattices.html for more explanation)\n"
        "This version of the program retains the original "
        "acoustic scores of arcs in the determinized lattice and writes it "
        "as a normal (non-compact) lattice. \n"
        " note: this program is tyically only useful if you generated state-level\n"
        " lattices, e.g. called gmm-latgen-simple with --determinize=false\n"
        "\n"
        "Usage: lattice-determinize-non-compact [options] lattice-rspecifier lattice-wspecifier\n"
        " e.g.: lattice-determinize-non-compact --acoustic-scale=0.1 --beam=15.0 ark:1.lats ark:det.lats\n";

    ParseOptions po(usage);
    BaseFloat acoustic_scale = 1.0;
    BaseFloat beam = 10.0;
    BaseFloat beam_ratio = 0.9;
    int32 num_loops = 20;
    int32 max_mem = 50000000; // 50 MB
    int32 max_loop = 500000;
    BaseFloat delta = fst::kDelta;
    bool prune = false;
    bool minimize = false;

    po.Register("acoustic-scale", &acoustic_scale,
                "Scaling factor for acoustic likelihoods");
    po.Register("beam", &beam,
                "Pruning beam [applied after acoustic scaling]-- also used "
                "to handle determinization failures, set --prune=false to "
                "disable routine pruning");
    po.Register("delta", &delta, "Tolerance used in determinization");
    po.Register("prune", &prune, "If true, prune determinized lattices "
                "with the --beam option.");
    po.Register("max-mem", &max_mem, "Maximum approximate memory usage in "
                "determinization (real usage might be many times this)");
    po.Register("max-loop", &max_loop, "Option to detect a certain "
                "type of failure in lattice determinization (not critical)");
    po.Register("beam-ratio", &beam_ratio, "Ratio by which to "
                "decrease beam if we reach the max-arcs.");
    po.Register("num-loops", &num_loops, "Number of times to "
                "decrease beam by beam-ratio if determinization fails.");
    po.Register("minimize", &minimize,
                "If true, push and minimize after determinization");

    po.Read(argc, argv);

    if (po.NumArgs() != 2) {
      po.PrintUsage();
      exit(1);
    }

    std::string lats_rspecifier = po.GetArg(1),
        lats_wspecifier = po.GetArg(2);

    // Read as regular lattice-- this is the form we need it in for efficient
    // pruning.
    SequentialLatticeReader lattice_reader(lats_rspecifier);

    // Write as regular lattice.
    LatticeWriter lattice_writer(lats_wspecifier);

    int32 n_done = 0, n_error = 0;

    // depth stats (for diagnostics).
    double sum_depth_in = 0.0,
          sum_depth_out = 0.0, sum_t = 0.0;

    if (acoustic_scale == 0.0)
      KALDI_ERR << "Do not use a zero acoustic scale (cannot be inverted)";
    LatticeWeight beam_weight(beam, static_cast<BaseFloat>(0.0));

    for (; !lattice_reader.Done(); lattice_reader.Next()) {
      std::string key = lattice_reader.Key();
      Lattice lat = lattice_reader.Value();

      lattice_reader.FreeCurrent();

      fst::TopSort(&lat);

      fst::ScaleLattice(fst::AcousticLatticeScale(acoustic_scale), &lat);


      // Compute a map from each (t, tid) to (sum_of_acoustic_scores, count)
      unordered_map<std::pair<int32,int32>, std::pair<BaseFloat, int32>,
                                          PairHasher<int32> > acoustic_scores;
      ComputeAcousticScoresMap(lat, &acoustic_scores);

      Invert(&lat); // make it so word labels are on the input.

      CompactLattice clat;
      if (DeterminizeLatticeWrapper(lat, key, prune,
                                    beam, beam_ratio, max_mem, max_loop,
                                    delta, num_loops, &clat)) {
        if (minimize) {
          PushCompactLatticeStrings(&clat);
          PushCompactLatticeWeights(&clat);
          MinimizeCompactLattice(&clat);
        }

        int32 t;
        TopSortCompactLatticeIfNeeded(&clat);
        double depth = CompactLatticeDepth(clat, &t);
        sum_depth_in += lat.NumStates();
        sum_depth_out += depth * t;
        sum_t += t;

        Lattice out_lat;
        fst::ConvertLattice(clat, &out_lat);
        fst::TopSort(&out_lat);

        // Replace each arc (t, tid) with the averaged acoustic score from
        // the computed map
        ReplaceAcousticScoresFromMap(acoustic_scores, &out_lat);

        fst::ScaleLattice(fst::AcousticLatticeScale(1.0/acoustic_scale),
                          &out_lat);
        lattice_writer.Write(key, out_lat);
        n_done++;
      } else {
        n_error++; // will have already printed warning.
      }
    }

    if (sum_t != 0.0) {
      KALDI_LOG << "Average input-lattice depth (measured at at state level) is "
                << (sum_depth_in / sum_t) << ", output depth is "
                << (sum_depth_out / sum_t) << ", over " << sum_t <<  "frames "
                << " (average num-frames = " << (sum_t / n_done) << ").";
    }
    KALDI_LOG << "Done " << n_done << " lattices, errors on " << n_error;
    return (n_done != 0 ? 0 : 1);
  } catch(const std::exception &e) {
    std::cerr << e.what();
    return -1;
  }
}