This struct is used to store the information we need for discriminative training (MMI or MPE). More...

#include <nnet-example.h>

Collaboration diagram for DiscriminativeNnetExample:

[legend]

Public Member Functions
void	Check () const

void	Write (std::ostream &os, bool binary) const

void	Read (std::istream &is, bool binary)

Public Attributes
BaseFloat	weight
	The weight we assign to this example; this will typically be one, but we include it for the sake of generality. More...

std::vector< int32 >	num_ali
	The numerator alignment. More...

CompactLattice	den_lat
	The denominator lattice. More...

Matrix< BaseFloat >	input_frames
	The input data– typically with a number of frames [NumRows()] larger than labels.size(), because it includes features to the left and right as needed for the temporal context of the network. More...

int32	left_context
	The number of frames of left context in the features (we can work out the #frames of right context from input_frames.NumRows(), num_ali.size(), and this). More...

Vector< BaseFloat >	spk_info
	spk_info contains any component of the features that varies slowly or not at all with time (and hence, we would lose little by averaging it over time and storing the average). More...

Detailed Description

This struct is used to store the information we need for discriminative training (MMI or MPE).

Each example corresponds to one chunk of a file (for better randomization and to prevent instability, we may split files in the middle). The example contains the numerator alignment, the denominator lattice, and the input features (extended at the edges according to the left-context and right-context the network needs). It may also contain a speaker-vector (note: this is not part of any standard recipe right now but is included in case it's useful in the future).

Definition at line 136 of file nnet-example.h.

Member Function Documentation

◆ Check()

void Check ( ) const

Definition at line 295 of file nnet-example.cc.

References kaldi::CompactLatticeStateTimes(), NnetExample::input_frames, KALDI_ASSERT, NnetExample::left_context, and CompressedMatrix::NumRows().

Referenced by DiscriminativeExampleSplitter::DoExcise(), DiscriminativeExampleSplitter::Excise(), kaldi::nnet2::LatticeToDiscriminativeExample(), DiscriminativeExampleSplitter::OutputOneSplit(), and DiscriminativeExampleSplitter::Split().

                                             {
   KALDI_ASSERT(weight > 0.0);
   KALDI_ASSERT(!num_ali.empty());
   int32 num_frames = static_cast<int32>(num_ali.size());
 
 
   std::vector<int32> times;
   int32 num_frames_den = CompactLatticeStateTimes(den_lat, &times);
   KALDI_ASSERT(num_frames == num_frames_den);
   KALDI_ASSERT(input_frames.NumRows() >= left_context + num_frames);
 }

◆ Read()

void Read	(	std::istream &	is,
		bool	binary
	)

Definition at line 269 of file nnet-example.cc.

References kaldi::ExpectToken(), NnetExample::input_frames, KALDI_ERR, NnetExample::left_context, CompressedMatrix::Read(), kaldi::ReadBasicType(), kaldi::ReadCompactLattice(), kaldi::ReadIntegerVector(), and NnetExample::spk_info.

                                                           {
   // Note: weight, num_ali, den_lat, input_frames, left_context and spk_info are
   // members.  This is a struct.
   ExpectToken(is, binary, "<DiscriminativeNnetExample>");
   ExpectToken(is, binary, "<Weight>");
   ReadBasicType(is, binary, &weight);
   ExpectToken(is, binary, "<NumAli>");
   ReadIntegerVector(is, binary, &num_ali);
   CompactLattice *den_lat_tmp = NULL;
   if (!ReadCompactLattice(is, binary, &den_lat_tmp) || den_lat_tmp == NULL) {
     // We can't return error status from this function so we
     // throw an exception. 
     KALDI_ERR << "Error reading CompactLattice from stream";
   }
   den_lat = *den_lat_tmp;
   delete den_lat_tmp;
   ExpectToken(is, binary, "<InputFrames>");
   input_frames.Read(is, binary);
   ExpectToken(is, binary, "<LeftContext>");
   ReadBasicType(is, binary, &left_context);
   ExpectToken(is, binary, "<SpkInfo>");
   spk_info.Read(is, binary);
   ExpectToken(is, binary, "</DiscriminativeNnetExample>");
 }

◆ Write()

void Write	(	std::ostream &	os,
		bool	binary
	)		const

Definition at line 242 of file nnet-example.cc.

References NnetExample::input_frames, KALDI_ERR, NnetExample::left_context, NnetExample::spk_info, CompressedMatrix::Write(), kaldi::WriteBasicType(), kaldi::WriteCompactLattice(), kaldi::WriteIntegerVector(), and kaldi::WriteToken().

                                                                  {
   // Note: weight, num_ali, den_lat, input_frames, left_context and spk_info are
   // members.  This is a struct.
   WriteToken(os, binary, "<DiscriminativeNnetExample>");
   WriteToken(os, binary, "<Weight>");
   WriteBasicType(os, binary, weight);
   WriteToken(os, binary, "<NumAli>");
   WriteIntegerVector(os, binary, num_ali);
   if (!WriteCompactLattice(os, binary, den_lat)) {
     // We can't return error status from this function so we
     // throw an exception. 
     KALDI_ERR << "Error writing CompactLattice to stream";
   }
   WriteToken(os, binary, "<InputFrames>");
   {
     CompressedMatrix cm(input_frames); // Note: this can be read as a regular
                                        // matrix.
     cm.Write(os, binary);
   }
   WriteToken(os, binary, "<LeftContext>");
   WriteBasicType(os, binary, left_context);
   WriteToken(os, binary, "<SpkInfo>");
   spk_info.Write(os, binary);
   WriteToken(os, binary, "</DiscriminativeNnetExample>");
 }

Member Data Documentation

◆ den_lat

CompactLattice den_lat

The denominator lattice.

Note: any acoustic likelihoods in the denominator lattice will be recomputed at the time we train.

Definition at line 148 of file nnet-example.h.

Referenced by kaldi::nnet2::AppendDiscriminativeExamples(), DiscriminativeExampleSplitter::DoExcise(), kaldi::nnet2::ExampleToPdfPost(), NnetDiscriminativeUpdater::LatticeComputations(), kaldi::nnet2::LatticeToDiscriminativeExample(), main(), DiscriminativeExampleSplitter::OutputOneSplit(), and DiscriminativeExampleSplitter::PrepareLattice().

◆ input_frames

Matrix<BaseFloat> input_frames

The input data– typically with a number of frames [NumRows()] larger than labels.size(), because it includes features to the left and right as needed for the temporal context of the network.

(see also the left_context variable). Caution: when we write this to disk, we do so as CompressedMatrix. Because we do various manipulations on these things in memory, such as splitting, we don't want it to be a CompressedMatrix in memory as this would be wasteful in time and also would lead to further loss of accuracy.

Definition at line 159 of file nnet-example.h.

Referenced by kaldi::nnet2::AppendDiscriminativeExamples(), kaldi::nnet2::AverageConstPart(), DiscriminativeExampleSplitter::DoExcise(), NnetDiscriminativeUpdater::GetInputFeatures(), kaldi::nnet2::LatticeToDiscriminativeExample(), main(), NnetDiscriminativeUpdater::NnetDiscriminativeUpdater(), DiscriminativeExampleSplitter::OutputOneSplit(), DiscriminativeExampleSplitter::RightContext(), and kaldi::nnet2::UpdateHash().

◆ left_context

int32 left_context

The number of frames of left context in the features (we can work out the #frames of right context from input_frames.NumRows(), num_ali.size(), and this).

Definition at line 164 of file nnet-example.h.

Referenced by kaldi::nnet2::AppendDiscriminativeExamples(), DiscriminativeExampleSplitter::DoExcise(), NnetDiscriminativeUpdater::GetInputFeatures(), kaldi::nnet2::LatticeToDiscriminativeExample(), DiscriminativeExampleSplitter::OutputOneSplit(), DiscriminativeExampleSplitter::RightContext(), and kaldi::nnet2::UpdateHash().

◆ num_ali

std::vector<int32> num_ali

The numerator alignment.

Definition at line 143 of file nnet-example.h.

Referenced by kaldi::nnet2::AppendDiscriminativeExamples(), DiscriminativeExampleSplitter::ComputeFrameInfo(), DiscriminativeExampleSplitter::DoExcise(), kaldi::nnet2::ExampleToPdfPost(), NnetDiscriminativeUpdater::GetDiscriminativePosteriors(), NnetDiscriminativeUpdater::GetInputFeatures(), NnetDiscriminativeUpdater::LatticeComputations(), kaldi::nnet2::LatticeToDiscriminativeExample(), DiscriminativeExampleSplitter::NumFrames(), DiscriminativeExampleSplitter::OutputOneSplit(), and kaldi::nnet2::UpdateHash().

◆ spk_info

Vector<BaseFloat> spk_info

spk_info contains any component of the features that varies slowly or not at all with time (and hence, we would lose little by averaging it over time and storing the average).

We'll append this to each of the input features, if used.

Definition at line 171 of file nnet-example.h.

Referenced by kaldi::nnet2::AppendDiscriminativeExamples(), kaldi::nnet2::AverageConstPart(), DiscriminativeExampleSplitter::DoExcise(), DiscriminativeExampleSplitter::OutputOneSplit(), and NnetDiscriminativeUpdater::Propagate().