Stats for fMLLR subspace estimation. More...

#include <basis-fmllr-diag-gmm.h>

Collaboration diagram for BasisFmllrAccus:

Public Member Functions
	BasisFmllrAccus ()

	BasisFmllrAccus (int32 dim)

void	ResizeAccus (int32 dim)

void	Write (std::ostream &out_stream, bool binary) const
	Routines for reading and writing stats. More...

void	Read (std::istream &in_stream, bool binary, bool add=false)

void	AccuGradientScatter (const AffineXformStats &spk_stats)
	Accumulate gradient scatter for one (training) speaker. More...

Public Attributes
SpMatrix< BaseFloat >	grad_scatter_
	Gradient scatter. Dim is [(D+1)D] [(D+1)D]. More...

int32	dim_
	Feature dimension. More...

double	beta_
	Occupancy count. More...

Detailed Description

Stats for fMLLR subspace estimation.

This class is only to estimate the "basis", which is done in training time. Class BasisFmllrEstimate contains the functions that are used in test time. (see the function BasisFmllrCoefficients()).

Definition at line 73 of file basis-fmllr-diag-gmm.h.

Constructor & Destructor Documentation

◆ BasisFmllrAccus() [1/2]

BasisFmllrAccus ( )

inline

Definition at line 76 of file basis-fmllr-diag-gmm.h.

76 { }

◆ BasisFmllrAccus() [2/2]

BasisFmllrAccus ( int32 dim )

inlineexplicit

Definition at line 77 of file basis-fmllr-diag-gmm.h.

                                       {
       dim_ = dim;
       beta_ = 0;
       ResizeAccus(dim);
   }

Member Function Documentation

◆ AccuGradientScatter()

void AccuGradientScatter ( const AffineXformStats & spk_stats )

Accumulate gradient scatter for one (training) speaker.

To finish the process, we need to traverse the whole training set. Parallelization works if the speaker list is splitted, and stats are summed up by setting add=true in BasisFmllrEstimate:: ReadBasis. See section 5.2 of the paper.

Definition at line 91 of file basis-fmllr-diag-gmm.cc.

References MatrixBase< Real >::AddMat(), AffineXformStats::beta_, BasisFmllrAccus::beta_, VectorBase< Real >::CopyRowsFromMat(), rnnlm::d, BasisFmllrAccus::dim_, AffineXformStats::G_, BasisFmllrAccus::grad_scatter_, AffineXformStats::K_, MatrixBase< Real >::Row(), MatrixBase< Real >::Scale(), and MatrixBase< Real >::SetUnit().

                                                          {
 
   // Gradient of auxf w.r.t. xform_spk
   // Eq. (33)
   Matrix<double> grad_mat(dim_, dim_ + 1);
   grad_mat.SetUnit();
   grad_mat.Scale(spk_stats.beta_);
   grad_mat.AddMat(1.0, spk_stats.K_);
   for (int d = 0; d < dim_; ++d) {
       Matrix<double> G_d_mat(spk_stats.G_[d]);
        grad_mat.Row(d).AddVec(-1.0, G_d_mat.Row(d));
   }
   // Row stack of gradient matrix
   Vector<BaseFloat> grad_vec((dim_+1) * dim_);
   grad_vec.CopyRowsFromMat(grad_mat);
   // The amount of data beta_ is likely to be ZERO, especially
   // when silence-weight is set to be 0 and we are using the
   // per-utt mode.
   if (spk_stats.beta_ > 0) {
     beta_ += spk_stats.beta_;
     grad_scatter_.AddVec2(BaseFloat(1.0 / spk_stats.beta_), grad_vec);
   }
 }

◆ Read()

void Read	(	std::istream &	in_stream,
		bool	binary,
		bool	add = `false`
	)

Definition at line 66 of file basis-fmllr-diag-gmm.cc.

References BasisFmllrAccus::beta_, kaldi::ExpectToken(), BasisFmllrAccus::grad_scatter_, and kaldi::ReadBasicType().

                                      {
   ExpectToken(is, binary, "<BASISFMLLRACCUS>");
   ExpectToken(is, binary, "<BETA>");
   double tmp_beta = 0;
   ReadBasicType(is, binary, &tmp_beta);
   if (add) {
     beta_ += tmp_beta;
   } else {
     beta_ = tmp_beta;
   }
   ExpectToken(is, binary, "<GRADSCATTER>");
   grad_scatter_.Read(is, binary, add);
   ExpectToken(is, binary, "</BASISFMLLRACCUS>");
 }

◆ ResizeAccus()

void ResizeAccus ( int32 dim )

Definition at line 82 of file basis-fmllr-diag-gmm.cc.

References BasisFmllrAccus::grad_scatter_, KALDI_ERR, and kaldi::kSetZero.

                                            {
   if (dim <= 0) {
     KALDI_ERR << "Invalid feature dimension " << dim; // dim=0 is not allowed
   } else {
     // 'kSetZero' may not be necessary, but makes computation safe
     grad_scatter_.Resize((dim + 1) * dim, kSetZero);
   }
 }

◆ Write()

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

Routines for reading and writing stats.

Definition at line 53 of file basis-fmllr-diag-gmm.cc.

References BasisFmllrAccus::beta_, BasisFmllrAccus::grad_scatter_, kaldi::WriteBasicType(), and kaldi::WriteToken().

                                                              {
 
   WriteToken(os, binary, "<BASISFMLLRACCUS>");
   WriteToken(os, binary, "<BETA>");
   WriteBasicType(os, binary, beta_);
   if (!binary) os << '\n';
   if (grad_scatter_.NumCols() != 0) {
     WriteToken(os, binary, "<GRADSCATTER>");
     grad_scatter_.Write(os, binary);
   }
   WriteToken(os, binary, "</BASISFMLLRACCUS>");
 }