DiscriminativeExampleSplitter Class Reference

For each frame, judge: More...

Collaboration diagram for DiscriminativeExampleSplitter:

Classes
struct	FrameInfo

Public Member Functions
	DiscriminativeExampleSplitter (const SplitDiscriminativeExampleConfig &config, const TransitionModel &tmodel, const DiscriminativeNnetExample &eg, std::vector< DiscriminativeNnetExample > *egs_out)
void	Excise (SplitExampleStats *stats)
void	Split (SplitExampleStats *stats)

Private Types
typedef LatticeArc	Arc
typedef Arc::StateId	StateId
typedef Arc::Label	Label

Private Member Functions
void	PrepareLattice (bool first_time)
void	CollapseTransitionIds ()
bool	ComputeFrameInfo ()
void	OutputOneSplit (int32 seg_begin, int32 seg_end)
void	DoSplit (SplitExampleStats *stats)
void	DoExcise (SplitExampleStats *stats)
int32	NumFrames () const
int32	RightContext ()
void	CreateOutputLattice (int32 seg_begin, int32 seg_end, CompactLattice *clat_out)
StateId	GetOutputStateId (StateId s, unordered_map< StateId, StateId > *state_map, Lattice *lat_out)

Static Private Member Functions
static void	RemoveAllOutputSymbols (Lattice *lat)

Private Attributes
const SplitDiscriminativeExampleConfig &	config_
const TransitionModel &	tmodel_
const DiscriminativeNnetExample &	eg_
std::vector< DiscriminativeNnetExample > *	egs_out_
Lattice	lat_
std::vector< FrameInfo >	frame_info_
std::vector< int32 >	state_times_

Detailed Description

For each frame, judge:

• does it produce a nonzero derivative? [this differs MMI vs MPE]
• can it be split here [or what is the penalty for splitting here.]
  • depends whether lattice has just one path at that point.

Time taken to process segment of a certain length: [must be sub-linear.] [use quadratic function that's max at specified segment length and zero at zero.]

No penalty for processing frames we don't need to process (already implicit in segment-processing time above.)

Penalty for splitting where we should not split. [Make it propto log(#paths).]

Definition at line 100 of file nnet-example-functions.cc.

Member Typedef Documentation

Arc

typedef LatticeArc Arc

private

Definition at line 134 of file nnet-example-functions.cc.

Label

typedef Arc::Label Label

private

Definition at line 136 of file nnet-example-functions.cc.

StateId

typedef Arc::StateId StateId

private

Definition at line 135 of file nnet-example-functions.cc.

Constructor & Destructor Documentation

DiscriminativeExampleSplitter()

DiscriminativeExampleSplitter ( const SplitDiscriminativeExampleConfig &  config, const TransitionModel &  tmodel, const DiscriminativeNnetExample &  eg, std::vector< DiscriminativeNnetExample > *  egs_out  )

inline

Definition at line 102 of file nnet-example-functions.cc.

                                                    :
      config_(config), tmodel_(tmodel), eg_(eg), egs_out_(egs_out) { }

Member Function Documentation

CollapseTransitionIds()

void CollapseTransitionIds ( )

private

Definition at line 236 of file nnet-example-functions.cc.

References count, KALDI_ASSERT, DiscriminativeExampleSplitter::lat_, kaldi::LatticeStateTimes(), DiscriminativeExampleSplitter::tmodel_, and TransitionModel::TransitionIdToPdf().

Referenced by DiscriminativeExampleSplitter::PrepareLattice().

                                                          {
  std::vector<int32> times;
  TopSort(&lat_); // Topologically sort the lattice (required by
                  // LatticeStateTimes)
  int32 num_frames = LatticeStateTimes(lat_, &times);  
  StateId num_states = lat_.NumStates();

  std::vector<std::map<int32, int32> > pdf_to_tid(num_frames);
  for (StateId s = 0; s < num_states; s++) {
    int32 t = times[s];
    for (fst::MutableArcIterator<Lattice> aiter(&lat_, s);
         !aiter.Done(); aiter.Next()) {
      KALDI_ASSERT(t >= 0 && t < num_frames);
      Arc arc = aiter.Value();
      KALDI_ASSERT(arc.ilabel != 0 && arc.ilabel == arc.olabel);
      int32 pdf = tmodel_.TransitionIdToPdf(arc.ilabel);
      if (pdf_to_tid[t].count(pdf) != 0) {
        arc.ilabel = arc.olabel = pdf_to_tid[t][pdf];
        aiter.SetValue(arc);
      } else {
        pdf_to_tid[t][pdf] = arc.ilabel;
      }
    }
  }    
}

ComputeFrameInfo()

bool ComputeFrameInfo ( )

private

Definition at line 299 of file nnet-example-functions.cc.

References DiscriminativeExampleSplitter::FrameInfo::can_excise_frame, DiscriminativeExampleSplitter::config_, SplitDiscriminativeExampleConfig::criterion, DiscriminativeExampleSplitter::FrameInfo::den_pdf_count, DiscriminativeExampleSplitter::FrameInfo::den_state_count, SplitDiscriminativeExampleConfig::drop_frames, DiscriminativeExampleSplitter::eg_, DiscriminativeExampleSplitter::FrameInfo::end_state, DiscriminativeExampleSplitter::frame_info_, rnnlm::i, KALDI_ASSERT, DiscriminativeExampleSplitter::lat_, kaldi::LatticeStateTimes(), DiscriminativeExampleSplitter::FrameInfo::multiple_transition_ids, DiscriminativeExampleSplitter::FrameInfo::nonzero_derivative, DiscriminativeNnetExample::num_ali, DiscriminativeExampleSplitter::FrameInfo::num_den_overlap, DiscriminativeExampleSplitter::NumFrames(), DiscriminativeExampleSplitter::FrameInfo::start_state, DiscriminativeExampleSplitter::state_times_, DiscriminativeExampleSplitter::tmodel_, and TransitionModel::TransitionIdToPdf().

Referenced by DiscriminativeExampleSplitter::Excise(), and DiscriminativeExampleSplitter::Split().

                                                     {
  
  int32 num_frames = NumFrames();

  frame_info_.clear();
  frame_info_.resize(num_frames + 1);
  
  LatticeStateTimes(lat_, &state_times_);

  std::vector<std::set<int32> > pdfs_per_frame(num_frames),
      tids_per_frame(num_frames);
  
  int32 num_states = lat_.NumStates();
  
  for (int32 state = 0; state < num_states; state++) {
    int32 t = state_times_[state];
    KALDI_ASSERT(t >= 0 && t <= num_frames);
    frame_info_[t].den_state_count++;
    for (fst::ArcIterator<Lattice> aiter(lat_, state); !aiter.Done();
         aiter.Next()) {
      const LatticeArc &arc = aiter.Value();
      KALDI_ASSERT(arc.ilabel != 0 && arc.ilabel == arc.olabel); 
      int32 transition_id = arc.ilabel,
          pdf_id = tmodel_.TransitionIdToPdf(transition_id);
      tids_per_frame[t].insert(transition_id);
      pdfs_per_frame[t].insert(pdf_id);
    }
    if (t < num_frames)
      frame_info_[t+1].start_state = std::min(state,
                                              frame_info_[t+1].start_state);
    frame_info_[t].end_state = std::max(state,
                                        frame_info_[t].end_state);
  }

  for (int32 i = 1; i <= NumFrames(); i++)
    frame_info_[i].end_state = std::max(frame_info_[i-1].end_state,
                                        frame_info_[i].end_state);
  for (int32 i = NumFrames() - 1; i >= 0; i--)
    frame_info_[i].start_state = std::min(frame_info_[i+1].start_state,
                                          frame_info_[i].start_state);
  
  for (int32 t = 0; t < num_frames; t++) {
    FrameInfo &frame_info = frame_info_[t];
    int32 transition_id = eg_.num_ali[t],
        pdf_id = tmodel_.TransitionIdToPdf(transition_id);
    frame_info.num_den_overlap = (pdfs_per_frame[t].count(pdf_id) != 0);
    frame_info.multiple_transition_ids = (tids_per_frame[t].size() > 1);
    KALDI_ASSERT(!pdfs_per_frame[t].empty());
    frame_info.den_pdf_count = pdfs_per_frame[t].size();

    if (config_.criterion == "mpfe" || config_.criterion == "smbr") {
      frame_info.nonzero_derivative = (frame_info.den_pdf_count > 1);
    } else {
      KALDI_ASSERT(config_.criterion == "mmi");
      if (config_.drop_frames) {
        // With frame dropping, we'll get nonzero derivative only
        // if num and den overlap, *and* den has >1 active pdf.
        frame_info.nonzero_derivative = frame_info.num_den_overlap  &&
            frame_info.den_state_count > 1;
      } else {
        // Without frame dropping, we'll get nonzero derivative if num and den
        // do not overlap , or den has >1 active pdf.
        frame_info.nonzero_derivative = !frame_info.num_den_overlap ||
            frame_info.den_state_count > 1;
      }
    }
    // If a frame is part of a segment, but it's not going to contribute
    // to the derivative and the den lattice has only one pdf active
    // at that time, then this frame can be excised from the lattice
    // because it will not affect the posteriors around it.
    if (config_.criterion == "mpfe") {
      frame_info.can_excise_frame =
          !frame_info.nonzero_derivative && \
          !frame_info.multiple_transition_ids;
      // in the mpfe case, if there are multiple transition-ids on a
      // frame there may be multiple phones on a frame, which could
      // contribute to the objective function even if they share pdf-ids.
      // (this was an issue that came up during testing).
    } else {      
      frame_info.can_excise_frame =
          !frame_info.nonzero_derivative && frame_info.den_pdf_count == 1;
    }
  }
  return true;
}

CreateOutputLattice()

void CreateOutputLattice ( int32  seg_begin, int32  seg_end, CompactLattice *  clat_out  )

private

Definition at line 644 of file nnet-example-functions.cc.

References fst::ConvertLattice(), DiscriminativeExampleSplitter::frame_info_, DiscriminativeExampleSplitter::GetOutputStateId(), KALDI_ASSERT, DiscriminativeExampleSplitter::lat_, DiscriminativeExampleSplitter::NumFrames(), LatticeWeightTpl< BaseFloat >::One(), DiscriminativeExampleSplitter::RemoveAllOutputSymbols(), DiscriminativeExampleSplitter::FrameInfo::start_state, and DiscriminativeExampleSplitter::state_times_.

Referenced by DiscriminativeExampleSplitter::OutputOneSplit(), and DiscriminativeExampleSplitter::RightContext().

                              {
  Lattice lat_out;

  // Below, state_map will map from states in the original lattice
  // lat_ to ones in the new lattice lat_out.
  unordered_map<StateId, StateId> state_map;

  // The range of the loop over s could be made over the
  // entire lattice, but we limit it for efficiency.
  
  for (StateId s = frame_info_[seg_begin].start_state;
       s <= frame_info_[seg_end].end_state; s++) {
    int32 t = state_times_[s];

    if (t < seg_begin || t > seg_end) // state out of range.
      continue;

    int32 this_state = GetOutputStateId(s, &state_map, &lat_out);

    if (t == seg_begin) // note: we only split on frames with just one
      lat_out.SetStart(this_state); // state, so we reach this only once.
    
    if (t == seg_end) { // Make it final and don't process its arcs out.
      if (seg_end == NumFrames()) {
        lat_out.SetFinal(this_state, lat_.Final(s));
      } else {
        lat_out.SetFinal(this_state, LatticeWeight::One());
      }
      continue; // don't process arcs out of this state.
    }
    
    for (fst::ArcIterator<Lattice> aiter(lat_, s); !aiter.Done(); aiter.Next()) {
      const Arc &arc = aiter.Value();
      StateId next_state = GetOutputStateId(arc.nextstate,
                                            &state_map, &lat_out);
      KALDI_ASSERT(arc.ilabel != 0 && arc.ilabel == arc.olabel); // We expect no epsilons.
      lat_out.AddArc(this_state, Arc(arc.ilabel, arc.olabel, arc.weight,
                                      next_state));
    }
  }
  Connect(&lat_out); // this is not really necessary, it's only to make sure
                     // the assert below fails when it should. TODO: remove it.
  KALDI_ASSERT(lat_out.NumStates() > 0);
  RemoveAllOutputSymbols(&lat_out);
  ConvertLattice(lat_out, clat_out);
}

DoExcise()

void DoExcise ( SplitExampleStats *  stats )

private

Definition at line 398 of file nnet-example-functions.cc.

References DiscriminativeNnetExample::Check(), fst::ConvertLattice(), VectorBase< Real >::CopyFromVec(), DiscriminativeNnetExample::den_lat, DiscriminativeExampleSplitter::eg_, DiscriminativeExampleSplitter::egs_out_, DiscriminativeExampleSplitter::frame_info_, rnnlm::i, DiscriminativeNnetExample::input_frames, KALDI_ASSERT, KALDI_WARN, DiscriminativeExampleSplitter::lat_, DiscriminativeNnetExample::left_context, SplitExampleStats::longest_segment_after_excise, DiscriminativeNnetExample::num_ali, SplitExampleStats::num_frames_kept_after_excise, MatrixBase< Real >::NumCols(), DiscriminativeExampleSplitter::NumFrames(), DiscriminativeExampleSplitter::RemoveAllOutputSymbols(), Matrix< Real >::Resize(), DiscriminativeExampleSplitter::RightContext(), DiscriminativeNnetExample::spk_info, DiscriminativeExampleSplitter::state_times_, and DiscriminativeNnetExample::weight.

Referenced by DiscriminativeExampleSplitter::Excise().

                                                                     {
  int32 left_context = eg_.left_context,
      right_context = RightContext(),
      num_frames = NumFrames();
  // Compute, for each frame, whether we can excise it.
  // 
  std::vector<bool> can_excise(num_frames, false);
  
  bool need_some_frame = false;
  for (int32 t = 0; t < num_frames; t++) {
    can_excise[t] = frame_info_[t].can_excise_frame;
    if (!can_excise[t])
      need_some_frame = true;
  }
  if (!need_some_frame) { // We don't need any frame within this file, so simply
                          // delete the segment.
    KALDI_WARN << "Example completely removed when excising."; // unexpected,
    // as the segment should have been deleted when splitting.
    egs_out_->clear();
    return;
  }
  egs_out_->resize(1);
  DiscriminativeNnetExample &eg_out = (*egs_out_)[0];

  // start_t and end_t will be the central part of the segment, excluding any
  // frames at the edges that we can excise.
  int32 start_t, end_t;
  for (start_t = 0; can_excise[start_t]; start_t++);
  for (end_t = num_frames; can_excise[end_t-1]; end_t--);

  // for frames from start_t to end_t-1, do not excise them if
  // they are within the context-window of a frame that we need to keep.
  // Note: we do t2 = t - right_context to t + left_context, because we're
  // concerned whether frame t2 has frame t in its window... it might
  // seem a bit backwards.
  std::vector<bool> will_excise(can_excise);
  for (int32 t = start_t; t < end_t; t++) {
    for (int32 t2 = t - right_context; t2 <= t + left_context; t2++)
      if (t2 >= start_t && t2 < end_t && !can_excise[t2])
        will_excise[t] = false; // can't excise this frame, it's needed for
                                // context.
  }

  // Remove all un-needed frames from the lattice by replacing the
  // symbols with epsilon and then removing the epsilons.
  // Note, this operation is destructive (it changes lat_).
  int32 num_states = lat_.NumStates();
  for (int32 state = 0; state < num_states; state++) {
    int32 t = state_times_[state];
    for (::fst::MutableArcIterator<Lattice> aiter(&lat_, state); !aiter.Done();
         aiter.Next()) {
      Arc arc = aiter.Value();
      if (will_excise[t]) {
        arc.ilabel = arc.olabel = 0;
        aiter.SetValue(arc);
      }
    }
  }
  RmEpsilon(&lat_);
  RemoveAllOutputSymbols(&lat_);
  ConvertLattice(lat_, &eg_out.den_lat);

  eg_out.num_ali.clear();
  int32 num_frames_kept = 0;
  for (int32 t = 0; t < num_frames; t++) {
    if (!will_excise[t]) {
      eg_out.num_ali.push_back(eg_.num_ali[t]);
      num_frames_kept++;
    }
  }

  stats->num_frames_kept_after_excise += num_frames_kept;
  stats->longest_segment_after_excise = std::max(stats->longest_segment_after_excise,
                                                 num_frames_kept);
  
  int32 num_frames_kept_plus = num_frames_kept + left_context + right_context;
  eg_out.input_frames.Resize(num_frames_kept_plus,
                             eg_.input_frames.NumCols());

  // the left-context of the output will be shifted to the right by
  // start_t.
  for (int32 i = 0; i < left_context; i++) {
    SubVector<BaseFloat> dst(eg_out.input_frames, i);
    SubVector<BaseFloat> src(eg_.input_frames, start_t + i);
    dst.CopyFromVec(src);
  }
  // the right-context will also be shifted, we take the frames
  // to the right of end_t.
  for (int32 i = 0; i < right_context; i++) {
    SubVector<BaseFloat> dst(eg_out.input_frames,
                             num_frames_kept + left_context + i);
    SubVector<BaseFloat> src(eg_.input_frames,
                             end_t + left_context + i);
    dst.CopyFromVec(src);
  }
  // now copy the central frames (those that were not excised).
  int32 dst_t = 0;
  for (int32 t = start_t; t < end_t; t++) {
    if (!will_excise[t]) {
      SubVector<BaseFloat> dst(eg_out.input_frames,
                               left_context + dst_t);
      SubVector<BaseFloat> src(eg_.input_frames,
                               left_context + t);
      dst.CopyFromVec(src);
      dst_t++;
    }
  }
  KALDI_ASSERT(dst_t == num_frames_kept);


  eg_out.weight = eg_.weight;
  eg_out.left_context = eg_.left_context;
  eg_out.spk_info = eg_.spk_info;

  eg_out.Check();
}

DoSplit()

void DoSplit ( SplitExampleStats *  stats )

private

Definition at line 516 of file nnet-example-functions.cc.

References DiscriminativeExampleSplitter::FrameInfo::den_state_count, DiscriminativeExampleSplitter::egs_out_, DiscriminativeExampleSplitter::frame_info_, SplitExampleStats::longest_lattice, SplitExampleStats::longest_segment_after_split, DiscriminativeExampleSplitter::FrameInfo::nonzero_derivative, SplitExampleStats::num_frames_kept_after_split, SplitExampleStats::num_frames_must_keep, SplitExampleStats::num_frames_orig, SplitExampleStats::num_kept_segments, SplitExampleStats::num_lattices, SplitExampleStats::num_segments, DiscriminativeExampleSplitter::NumFrames(), and DiscriminativeExampleSplitter::OutputOneSplit().

Referenced by DiscriminativeExampleSplitter::Split().

                                                                    {
  std::vector<int32> split_points;
  int32 num_frames = NumFrames();
  {
    // Make the "split points" 0 and num_frames, and
    // any frame that has just one state on it and the previous
    // frame had >1 state.  This gives us one split for each
    // "pinch point" in the lattice.  Later we may move each split
    // to a more optimal location.
    split_points.push_back(0);
    for (int32 t = 1; t < num_frames; t++) {
      if (frame_info_[t].den_state_count == 1 &&
          frame_info_[t-1].den_state_count > 1)
        split_points.push_back(t);
    }
    split_points.push_back(num_frames);
  }

  std::vector<bool> is_kept(split_points.size() - 1);
  { // A "split" is a pair of successive split points.  Work out for each split
    // whether we must keep it (we must if it contains at least one frame for
    // which "nonzero_derivative" == true.)
    for (size_t s = 0; s < is_kept.size(); s++) {
      int32 start = split_points[s], end = split_points[s+1];
      bool keep_this_split = false;
      for (int32 t = start; t < end; t++)
        if (frame_info_[t].nonzero_derivative)
          keep_this_split = true;
      is_kept[s] = keep_this_split;
    }
  }

  egs_out_->clear();
  egs_out_->reserve(is_kept.size());

  stats->num_lattices++;
  stats->longest_lattice = std::max(stats->longest_lattice, num_frames);
  stats->num_segments += is_kept.size();
  stats->num_frames_orig += num_frames;
  for (int32 t = 0; t < num_frames; t++)
    if (frame_info_[t].nonzero_derivative)
      stats->num_frames_must_keep++;
  
  for (size_t s = 0; s < is_kept.size(); s++) {
    if (is_kept[s]) {
      stats->num_kept_segments++;
      OutputOneSplit(split_points[s], split_points[s+1]);
      int32 segment_len = split_points[s+1] - split_points[s];
      stats->num_frames_kept_after_split += segment_len;
      stats->longest_segment_after_split =
          std::max(stats->longest_segment_after_split, segment_len);
    }
  }
}

Excise()

void Excise ( SplitExampleStats *  stats )

inline

Definition at line 109 of file nnet-example-functions.cc.

References DiscriminativeNnetExample::Check(), DiscriminativeExampleSplitter::ComputeFrameInfo(), DiscriminativeExampleSplitter::config_, DiscriminativeExampleSplitter::DoExcise(), DiscriminativeExampleSplitter::eg_, DiscriminativeExampleSplitter::egs_out_, SplitDiscriminativeExampleConfig::excise, and DiscriminativeExampleSplitter::PrepareLattice().

Referenced by kaldi::nnet2::ExciseDiscriminativeExample().

                                        {
    eg_.Check();
    PrepareLattice(false);
    ComputeFrameInfo();
    if (!config_.excise) {
      egs_out_->resize(1);
      (*egs_out_)[0] = eg_;
    } else {
      DoExcise(stats);
    }
  }

GetOutputStateId()

DiscriminativeExampleSplitter::StateId GetOutputStateId ( StateId  s, unordered_map< StateId, StateId > *  state_map, Lattice *  lat_out  )

private

Definition at line 635 of file nnet-example-functions.cc.

Referenced by DiscriminativeExampleSplitter::CreateOutputLattice(), and DiscriminativeExampleSplitter::RightContext().

                                                                             {
  if (state_map->count(s) == 0) {
    return ((*state_map)[s] = lat_out->AddState());
  } else {
    return (*state_map)[s];
  }
}

NumFrames()

int32 NumFrames ( ) const

inlineprivate

Definition at line 159 of file nnet-example-functions.cc.

References DiscriminativeExampleSplitter::eg_, and DiscriminativeNnetExample::num_ali.

Referenced by DiscriminativeExampleSplitter::ComputeFrameInfo(), DiscriminativeExampleSplitter::CreateOutputLattice(), DiscriminativeExampleSplitter::DoExcise(), DiscriminativeExampleSplitter::DoSplit(), DiscriminativeExampleSplitter::OutputOneSplit(), and DiscriminativeExampleSplitter::RightContext().

{ return static_cast<int32>(eg_.num_ali.size()); }

OutputOneSplit()

void OutputOneSplit ( int32  seg_begin, int32  seg_end  )

private

Definition at line 596 of file nnet-example-functions.cc.

References DiscriminativeNnetExample::Check(), DiscriminativeExampleSplitter::CreateOutputLattice(), DiscriminativeNnetExample::den_lat, DiscriminativeExampleSplitter::eg_, DiscriminativeExampleSplitter::egs_out_, DiscriminativeNnetExample::input_frames, KALDI_ASSERT, DiscriminativeNnetExample::left_context, DiscriminativeNnetExample::num_ali, MatrixBase< Real >::NumCols(), DiscriminativeExampleSplitter::NumFrames(), MatrixBase< Real >::Range(), DiscriminativeExampleSplitter::RightContext(), DiscriminativeNnetExample::spk_info, and DiscriminativeNnetExample::weight.

Referenced by DiscriminativeExampleSplitter::DoSplit().

                                                                  {
  KALDI_ASSERT(seg_begin >= 0 && seg_end > seg_begin && seg_end <= NumFrames());
  egs_out_->resize(egs_out_->size() + 1);
  int32 left_context = eg_.left_context, right_context = RightContext(),
      tot_context = left_context + right_context;
  DiscriminativeNnetExample &eg_out = egs_out_->back();
  eg_out.weight = eg_.weight;

  eg_out.num_ali.insert(eg_out.num_ali.end(),
                        eg_.num_ali.begin() + seg_begin,
                        eg_.num_ali.begin() + seg_end);

  CreateOutputLattice(seg_begin, seg_end, &(eg_out.den_lat));
  
  eg_out.input_frames = eg_.input_frames.Range(seg_begin, seg_end - seg_begin +
                                               tot_context,
                                               0, eg_.input_frames.NumCols());

  eg_out.left_context = eg_.left_context;

  eg_out.spk_info = eg_.spk_info;
  
  eg_out.Check();  
}

PrepareLattice()

void PrepareLattice ( bool  first_time )

private

Definition at line 263 of file nnet-example-functions.cc.

References SplitDiscriminativeExampleConfig::collapse_transition_ids, DiscriminativeExampleSplitter::CollapseTransitionIds(), DiscriminativeExampleSplitter::config_, fst::ConvertLattice(), SplitDiscriminativeExampleConfig::criterion, DiscriminativeNnetExample::den_lat, SplitDiscriminativeExampleConfig::determinize, DiscriminativeExampleSplitter::eg_, DiscriminativeExampleSplitter::lat_, and SplitDiscriminativeExampleConfig::minimize.

Referenced by DiscriminativeExampleSplitter::Excise(), and DiscriminativeExampleSplitter::Split().

                                                                  {
  ::fst::ConvertLattice(eg_.den_lat, &lat_);

  Project(&lat_, fst::PROJECT_INPUT); // Get rid of the word labels and put the
                                      // transition-ids on both sides.
  
  RmEpsilon(&lat_); // Remove epsilons.. this simplifies
                    // certain things.

  if (first_time) {
    if (config_.collapse_transition_ids && config_.criterion != "mpfe")
      CollapseTransitionIds();
  
    if (config_.determinize) {
      if (!config_.minimize) {
        Lattice det_lat;
        Determinize(lat_, &det_lat);
        lat_ = det_lat;
      } else {
        Lattice tmp_lat;
        Reverse(lat_, &tmp_lat);
        Determinize(tmp_lat, &lat_);
        Reverse(lat_, &tmp_lat);
        Determinize(tmp_lat, &lat_);
        RmEpsilon(&lat_);
        // Previously we determinized, then did
        // Minimize(&lat_);
        // but this was too slow.
      }
    }
  }
  TopSort(&lat_); // Topologically sort the lattice.
}

RemoveAllOutputSymbols()

void RemoveAllOutputSymbols ( Lattice *  lat )

staticprivate

Definition at line 623 of file nnet-example-functions.cc.

Referenced by DiscriminativeExampleSplitter::CreateOutputLattice(), and DiscriminativeExampleSplitter::DoExcise().

                                                                       {
  for (StateId s = 0; s < lat->NumStates(); s++) {
    for (::fst::MutableArcIterator<Lattice> aiter(lat, s); !aiter.Done();
         aiter.Next()) {
      Arc arc = aiter.Value();
      arc.olabel = 0;
      aiter.SetValue(arc);
    }
  }  
}

RightContext()

int32 RightContext ( )

inlineprivate

Definition at line 161 of file nnet-example-functions.cc.

References DiscriminativeExampleSplitter::CreateOutputLattice(), DiscriminativeExampleSplitter::eg_, DiscriminativeExampleSplitter::GetOutputStateId(), DiscriminativeNnetExample::input_frames, DiscriminativeNnetExample::left_context, DiscriminativeExampleSplitter::NumFrames(), and MatrixBase< Real >::NumRows().

Referenced by DiscriminativeExampleSplitter::DoExcise(), and DiscriminativeExampleSplitter::OutputOneSplit().

{ return eg_.input_frames.NumRows() - NumFrames() - eg_.left_context; }

Split()

void Split ( SplitExampleStats *  stats )

inline

Definition at line 121 of file nnet-example-functions.cc.

References DiscriminativeNnetExample::Check(), DiscriminativeExampleSplitter::ComputeFrameInfo(), DiscriminativeExampleSplitter::config_, DiscriminativeExampleSplitter::DoSplit(), DiscriminativeExampleSplitter::eg_, DiscriminativeExampleSplitter::egs_out_, DiscriminativeExampleSplitter::PrepareLattice(), and SplitDiscriminativeExampleConfig::split.

Referenced by kaldi::nnet2::SplitDiscriminativeExample().

                                       {
    if (!config_.split) {
      egs_out_->resize(1);
      (*egs_out_)[0] = eg_;
    } else {
      eg_.Check();    
      PrepareLattice(true);
      ComputeFrameInfo();
      DoSplit(stats);
    }
  }

Member Data Documentation

config_

const SplitDiscriminativeExampleConfig& config_

private

Definition at line 213 of file nnet-example-functions.cc.

Referenced by DiscriminativeExampleSplitter::ComputeFrameInfo(), DiscriminativeExampleSplitter::Excise(), DiscriminativeExampleSplitter::PrepareLattice(), and DiscriminativeExampleSplitter::Split().

eg_

const DiscriminativeNnetExample& eg_

private

Definition at line 215 of file nnet-example-functions.cc.

Referenced by DiscriminativeExampleSplitter::ComputeFrameInfo(), DiscriminativeExampleSplitter::DoExcise(), DiscriminativeExampleSplitter::Excise(), DiscriminativeExampleSplitter::NumFrames(), DiscriminativeExampleSplitter::OutputOneSplit(), DiscriminativeExampleSplitter::PrepareLattice(), DiscriminativeExampleSplitter::RightContext(), and DiscriminativeExampleSplitter::Split().

egs_out_

std::vector<DiscriminativeNnetExample>* egs_out_

private

Definition at line 216 of file nnet-example-functions.cc.

Referenced by DiscriminativeExampleSplitter::DoExcise(), DiscriminativeExampleSplitter::DoSplit(), DiscriminativeExampleSplitter::Excise(), DiscriminativeExampleSplitter::OutputOneSplit(), and DiscriminativeExampleSplitter::Split().

frame_info_

std::vector<FrameInfo> frame_info_

private

Definition at line 223 of file nnet-example-functions.cc.

Referenced by DiscriminativeExampleSplitter::ComputeFrameInfo(), DiscriminativeExampleSplitter::CreateOutputLattice(), DiscriminativeExampleSplitter::DoExcise(), and DiscriminativeExampleSplitter::DoSplit().

lat_

Lattice lat_

private

Definition at line 218 of file nnet-example-functions.cc.

Referenced by DiscriminativeExampleSplitter::CollapseTransitionIds(), DiscriminativeExampleSplitter::ComputeFrameInfo(), DiscriminativeExampleSplitter::CreateOutputLattice(), DiscriminativeExampleSplitter::DoExcise(), and DiscriminativeExampleSplitter::PrepareLattice().

state_times_

std::vector<int32> state_times_

private

Definition at line 226 of file nnet-example-functions.cc.

Referenced by DiscriminativeExampleSplitter::ComputeFrameInfo(), DiscriminativeExampleSplitter::CreateOutputLattice(), and DiscriminativeExampleSplitter::DoExcise().

tmodel_

const TransitionModel& tmodel_

private

Definition at line 214 of file nnet-example-functions.cc.

Referenced by DiscriminativeExampleSplitter::CollapseTransitionIds(), and DiscriminativeExampleSplitter::ComputeFrameInfo().

---

The documentation for this class was generated from the following file:

• nnet2/nnet-example-functions.cc