kaldi::discriminative Namespace Reference

Classes
class	DiscriminativeComputation
struct	DiscriminativeObjectiveInfo
struct	DiscriminativeOptions
struct	DiscriminativeSupervision
class	DiscriminativeSupervisionSplitter
struct	SplitDiscriminativeSupervisionOptions

Functions
void	MergeSupervision (const std::vector< const DiscriminativeSupervision * > &input, DiscriminativeSupervision *output_supervision)
	This function appends a list of supervision objects to create what will usually be a single such object, but if the weights and num-frames are not all the same it will only append Supervision objects where successive ones have the same weight and num-frames, and if 'compactify' is true. More...
void	ComputeDiscriminativeObjfAndDeriv (const DiscriminativeOptions &opts, const TransitionModel &tmodel, const CuVectorBase< BaseFloat > &log_priors, const DiscriminativeSupervision &supervision, const CuMatrixBase< BaseFloat > &nnet_output, DiscriminativeObjectiveInfo *stats, CuMatrixBase< BaseFloat > *nnet_output_deriv, CuMatrixBase< BaseFloat > *xent_output_deriv)
	This function does forward-backward on the numerator and denominator lattices and computes derivates wrt to the output for the specified objective function. More...

Function Documentation

ComputeDiscriminativeObjfAndDeriv()

void ComputeDiscriminativeObjfAndDeriv	(	const DiscriminativeOptions &	opts,
		const TransitionModel &	tmodel,
		const CuVectorBase< BaseFloat > &	log_priors,
		const DiscriminativeSupervision &	supervision,
		const CuMatrixBase< BaseFloat > &	nnet_output,
		DiscriminativeObjectiveInfo *	stats,
		CuMatrixBase< BaseFloat > *	nnet_output_deriv,
		CuMatrixBase< BaseFloat > *	xent_output_deriv
	)

This function does forward-backward on the numerator and denominator lattices and computes derivates wrt to the output for the specified objective function.

Parameters

[in]	opts	Struct containing options
[in]	tmodel	Transition model
[in]	log_priors	Vector of log-priors for pdfs
[in]	supervision	The supervision object, containing the numerator and denominator paths. The denominator is always a lattice. The numerator is an alignment.
[in]	nnet_output	The output of the neural net; dimension must equal ((supervision.num_sequences * supervision.frames_per_sequence) by tmodel.NumPdfs()).
[out]	stats	Statistics accumulated during training such as the objective function and the total weight.
[out]	xent_output_deriv	If non-NULL, then the xent objective derivative (which equals a posterior from the numerator forward-backward, scaled by the supervision weight) is written to here. This will be used in the cross-entropy regularization code.

Definition at line 546 of file discriminative-training.cc.

References DiscriminativeComputation::Compute().

Referenced by DiscriminativeObjectiveInfo::AccumulateOutput(), NnetDiscriminativeComputeObjf::ProcessOutputs(), and NnetDiscriminativeTrainer::ProcessOutputs().

                                                                                   {
  DiscriminativeComputation computation(opts, tmodel, log_priors, supervision,
                                        nnet_output, stats,
                                        nnet_output_deriv, xent_output_deriv);
  computation.Compute();
}

MergeSupervision()

void MergeSupervision	(	const std::vector< const DiscriminativeSupervision * > &	input,
		DiscriminativeSupervision *	output_supervision
	)

This function appends a list of supervision objects to create what will usually be a single such object, but if the weights and num-frames are not all the same it will only append Supervision objects where successive ones have the same weight and num-frames, and if 'compactify' is true.

The normal use-case for this is when you are combining neural-net examples for training; appending them like this helps to simplify the training process.

Definition at line 402 of file discriminative-supervision.cc.

References DiscriminativeSupervision::Check(), DiscriminativeSupervision::den_lat, DiscriminativeSupervision::frames_per_sequence, rnnlm::i, KALDI_ASSERT, KALDI_ERR, DiscriminativeSupervision::num_ali, DiscriminativeSupervision::num_sequences, and DiscriminativeSupervision::weight.

Referenced by kaldi::nnet3::MergeSupervision().

                                                   {
  KALDI_ASSERT(!input.empty());
  int32 num_inputs = input.size();
  if (num_inputs == 1) {
    *output_supervision = *(input[0]);
    return;
  }
  *output_supervision = *(input[num_inputs-1]);
  for (int32 i = num_inputs - 2; i >= 0; i--) {
    const DiscriminativeSupervision &src = *(input[i]);
    KALDI_ASSERT(src.num_sequences == 1);
    if (output_supervision->weight == src.weight &&
        output_supervision->frames_per_sequence ==
        src.frames_per_sequence) {
      // Combine with current output
      // append src.den_lat to output_supervision->den_lat.
      fst::Concat(src.den_lat, &output_supervision->den_lat);

      output_supervision->num_ali.insert(
          output_supervision->num_ali.begin(),
          src.num_ali.begin(), src.num_ali.end());

      output_supervision->num_sequences++;
    } else {
      KALDI_ERR << "Mismatch weight or frames_per_sequence  between inputs";
    }
  }
  DiscriminativeSupervision &out_sup = *output_supervision;
  fst::TopSort(&(out_sup.den_lat));
  out_sup.Check();
}