LinearVtln Class Reference

#include <lvtln.h>

Collaboration diagram for LinearVtln:

Public Member Functions
	LinearVtln ()
	LinearVtln (int32 dim, int32 num_classes, int32 default_class)
void	SetTransform (int32 i, const MatrixBase< BaseFloat > &transform)
void	SetWarp (int32 i, BaseFloat warp)
BaseFloat	GetWarp (int32 i) const
void	GetTransform (int32 i, MatrixBase< BaseFloat > *transform) const
void	ComputeTransform (const FmllrDiagGmmAccs &accs, std::string norm_type, BaseFloat logdet_scale, MatrixBase< BaseFloat > *Ws, int32 *class_idx, BaseFloat *logdet_out, BaseFloat *objf_impr=NULL, BaseFloat *count=NULL)
	Compute the transform for the speaker. More...
void	Read (std::istream &is, bool binary)
void	Write (std::ostream &os, bool binary) const
int32	Dim () const
int32	NumClasses () const
void	GetOffset (const FmllrDiagGmmAccs &speaker_stats, int32 class_idx, VectorBase< BaseFloat > *offset) const

Protected Attributes
int32	default_class_
std::vector< Matrix< BaseFloat > >	A_
std::vector< BaseFloat >	logdets_
std::vector< BaseFloat >	warps_

Friends
class	LinearVtlnStats

Detailed Description

Definition at line 40 of file lvtln.h.

Constructor & Destructor Documentation

LinearVtln() [1/2]

LinearVtln ( )

inline

Definition at line 42 of file lvtln.h.

References LinearVtln::ComputeTransform(), count, LinearVtln::GetTransform(), LinearVtln::GetWarp(), rnnlm::i, LinearVtln::Read(), LinearVtln::SetTransform(), LinearVtln::SetWarp(), and LinearVtln::Write().

{ } // This initializer will probably be used prior to calling

LinearVtln() [2/2]

LinearVtln	(	int32	dim,
		int32	num_classes,
		int32	default_class
	)

Definition at line 29 of file lvtln.cc.

References LinearVtln::A_, LinearVtln::default_class_, rnnlm::i, KALDI_ASSERT, LinearVtln::logdets_, and LinearVtln::warps_.

                                                                        {
  default_class_ = default_class;
  KALDI_ASSERT(default_class >= 0 && default_class < num_classes);
  A_.resize(num_classes);
  for (int32 i = 0; i < num_classes; i++) {
    A_[i].Resize(dim, dim);
    A_[i].SetUnit();
  }
  logdets_.clear();
  logdets_.resize(num_classes, 0.0);
  warps_.clear();
  warps_.resize(num_classes, 1.0);
} // namespace kaldi

Member Function Documentation

ComputeTransform()

void ComputeTransform	(	const FmllrDiagGmmAccs &	accs,
		std::string	norm_type,
		BaseFloat	logdet_scale,
		MatrixBase< BaseFloat > *	Ws,
		int32 *	class_idx,
		BaseFloat *	logdet_out,
		BaseFloat *	objf_impr = NULL,
		BaseFloat *	count = NULL
	)

Compute the transform for the speaker.

Definition at line 97 of file lvtln.cc.

References LinearVtln::A_, kaldi::ApplyFeatureTransformToStats(), AffineXformStats::beta_, kaldi::ComposeTransforms(), kaldi::ComputeFmllrMatrixDiagGmm(), MatrixBase< Real >::CopyFromMat(), LinearVtln::default_class_, LinearVtln::Dim(), kaldi::FmllrAuxFuncDiagGmm(), rnnlm::i, KALDI_ASSERT, KALDI_ERR, KALDI_WARN, LinearVtln::logdets_, LinearVtln::NumClasses(), MatrixBase< Real >::NumCols(), MatrixBase< Real >::NumRows(), MatrixBase< Real >::Range(), and MatrixBase< Real >::SetUnit().

Referenced by LinearVtln::LinearVtln(), and main().

                                                    {
  int32 dim = Dim();
  KALDI_ASSERT(dim != 0);
  if (norm_type != "none"  && norm_type != "offset" && norm_type != "diag")
    KALDI_ERR << "LinearVtln::ComputeTransform, norm_type should be "
        "one of \"none\", \"offset\" or \"diag\"";
  
  if (accs.beta_ == 0.0) {
    KALDI_WARN << "no stats, returning default transform";
    int32 dim = Dim();
    if (Ws) {
      KALDI_ASSERT(Ws->NumRows() == dim && Ws->NumCols() == dim+1);
      Ws->Range(0, dim, 0, dim).CopyFromMat(A_[default_class_]);
      Ws->Range(0, dim, dim, 1).SetZero();  // Set last column to zero.
    }
    if (class_idx) *class_idx = default_class_;
    if (logdet_out) *logdet_out = logdets_[default_class_];
    if (objf_impr) *objf_impr = 0;
    if (count) *count = 0;
    return;
  }
  
  Matrix<BaseFloat> best_transform(dim, dim+1);
  best_transform.SetUnit();
  BaseFloat old_objf = FmllrAuxFuncDiagGmm(best_transform, accs),
      best_objf = -1.0e+100;
  int32 best_class = -1;

  for (int32 i = 0; i < NumClasses(); i++) {
    FmllrDiagGmmAccs accs_tmp(accs);
    ApplyFeatureTransformToStats(A_[i], &accs_tmp);
    // "old_trans" just needed by next function as "initial" transform.
    Matrix<BaseFloat> old_trans(dim, dim+1); old_trans.SetUnit();
    Matrix<BaseFloat> trans(dim, dim+1);
    ComputeFmllrMatrixDiagGmm(old_trans, accs_tmp, norm_type,
                              100,  // num iters.. don't care since norm_type != "full"
                              &trans);
    Matrix<BaseFloat> product(dim, dim+1);
    // product = trans * A_[i] (modulo messing about with offsets)
    ComposeTransforms(trans, A_[i], false, &product);

    BaseFloat objf = FmllrAuxFuncDiagGmm(product, accs);

    if (logdet_scale != 1.0)
      objf += accs.beta_ * (logdet_scale - 1.0) * logdets_[i];
    
    if (objf > best_objf) {
      best_objf = objf;
      best_class = i;
      best_transform.CopyFromMat(product);
    }
  }
  KALDI_ASSERT(best_class != -1);
  if (Ws) Ws->CopyFromMat(best_transform);
  if (class_idx) *class_idx = best_class;
  if (logdet_out) *logdet_out = logdets_[best_class];
  if (objf_impr) *objf_impr = best_objf - old_objf;
  if (count) *count = accs.beta_;
}

Dim()

int32 Dim ( ) const

inline

Definition at line 77 of file lvtln.h.

References LinearVtln::A_, and KALDI_ASSERT.

Referenced by LinearVtln::ComputeTransform(), LinearVtln::GetTransform(), main(), and LinearVtln::SetTransform().

{ KALDI_ASSERT(!A_.empty()); return A_[0].NumRows(); }

GetOffset()

void GetOffset	(	const FmllrDiagGmmAccs &	speaker_stats,
		int32	class_idx,
		VectorBase< BaseFloat > *	offset
	)		const

Referenced by LinearVtln::NumClasses().

GetTransform()

void GetTransform	(	int32	i,
		MatrixBase< BaseFloat > *	transform
	)		const

Definition at line 185 of file lvtln.cc.

References LinearVtln::A_, MatrixBase< Real >::CopyFromMat(), LinearVtln::Dim(), KALDI_ASSERT, LinearVtln::NumClasses(), MatrixBase< Real >::NumCols(), and MatrixBase< Real >::NumRows().

Referenced by LinearVtln::LinearVtln().

                                                                             {
  KALDI_ASSERT(i >= 0 && i < NumClasses());
  KALDI_ASSERT(transform->NumRows() == transform->NumCols() &&
               static_cast<int32>(transform->NumRows()) == Dim());
  transform->CopyFromMat(A_[i]);
}

GetWarp()

BaseFloat GetWarp

(

int32

i

)

const

Definition at line 180 of file lvtln.cc.

References rnnlm::i, KALDI_ASSERT, LinearVtln::NumClasses(), and LinearVtln::warps_.

Referenced by LinearVtln::LinearVtln(), and main().

                                           {
  KALDI_ASSERT(i >= 0 && i < NumClasses());
  return warps_[i];
}

NumClasses()

int32 NumClasses ( ) const

inline

Definition at line 78 of file lvtln.h.

References LinearVtln::A_, and LinearVtln::GetOffset().

Referenced by LinearVtln::ComputeTransform(), LinearVtln::GetTransform(), LinearVtln::GetWarp(), main(), LinearVtln::SetTransform(), and LinearVtln::SetWarp().

{ return A_.size(); }

Read()

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

Definition at line 45 of file lvtln.cc.

References LinearVtln::A_, LinearVtln::default_class_, kaldi::ExpectToken(), rnnlm::i, KALDI_ASSERT, LinearVtln::logdets_, kaldi::ReadBasicType(), kaldi::ReadToken(), and LinearVtln::warps_.

Referenced by LinearVtln::LinearVtln().

                                                 {
  int32 sz;
  ExpectToken(is, binary, "<LinearVtln>");
  ReadBasicType(is, binary, &sz);
  A_.resize(sz);
  logdets_.resize(sz);
  warps_.resize(sz);
  for (int32 i = 0; i < sz; i++) {
    ExpectToken(is, binary, "<A>");
    A_[i].Read(is, binary);
    ExpectToken(is, binary, "<logdet>");
    ReadBasicType(is, binary, &(logdets_[i]));
    ExpectToken(is, binary, "<warp>");
    ReadBasicType(is, binary, &(warps_[i]));
  }
  std::string token;
  ReadToken(is, binary, &token);
  if (token == "</LinearVtln>") {
    // the older code had a bug in that it wasn't writing or reading
    // default_class_.  The following guess at its value is likely to be
    // correct.
    default_class_ = (sz + 1) / 2;
  } else {
    KALDI_ASSERT(token == "<DefaultClass>");
    ReadBasicType(is, binary, &default_class_);
    ExpectToken(is, binary, "</LinearVtln>");
  }
}

SetTransform()

void SetTransform	(	int32	i,
		const MatrixBase< BaseFloat > &	transform
	)

Definition at line 166 of file lvtln.cc.

References LinearVtln::A_, LinearVtln::Dim(), rnnlm::i, KALDI_ASSERT, LinearVtln::logdets_, LinearVtln::NumClasses(), MatrixBase< Real >::NumCols(), and MatrixBase< Real >::NumRows().

Referenced by LinearVtln::LinearVtln(), and main().

                                                                             {
  KALDI_ASSERT(i >= 0 && i < NumClasses());
  KALDI_ASSERT(transform.NumRows() == transform.NumCols() &&
               static_cast<int32>(transform.NumRows()) == Dim());
  A_[i].CopyFromMat(transform);
  logdets_[i] = A_[i].LogDet();
}

SetWarp()

void SetWarp	(	int32	i,
		BaseFloat	warp
	)

Definition at line 174 of file lvtln.cc.

References rnnlm::i, KALDI_ASSERT, LinearVtln::NumClasses(), and LinearVtln::warps_.

Referenced by LinearVtln::LinearVtln(), and main().

                                                {
  KALDI_ASSERT(i >= 0 && i < NumClasses());
  KALDI_ASSERT(warps_.size() == static_cast<size_t>(NumClasses()));
  warps_[i] = warp;
}

Write()

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

Definition at line 74 of file lvtln.cc.

References LinearVtln::A_, LinearVtln::default_class_, rnnlm::i, KALDI_ASSERT, LinearVtln::logdets_, LinearVtln::warps_, kaldi::WriteBasicType(), and kaldi::WriteToken().

Referenced by LinearVtln::LinearVtln(), and main().

                                                        {
  WriteToken(os, binary, "<LinearVtln>");
  if(!binary) os << "\n";
  int32 sz = A_.size();
  KALDI_ASSERT(static_cast<size_t>(sz) == logdets_.size());
  KALDI_ASSERT(static_cast<size_t>(sz) == warps_.size());
  WriteBasicType(os, binary, sz);
  for (int32 i = 0; i < sz; i++) {
    WriteToken(os, binary, "<A>");
    A_[i].Write(os, binary);
    WriteToken(os, binary, "<logdet>");
    WriteBasicType(os, binary, logdets_[i]);
    WriteToken(os, binary, "<warp>");
    WriteBasicType(os, binary, warps_[i]);
    if(!binary) os << "\n";
  }
  WriteToken(os, binary, "<DefaultClass>");
  WriteBasicType(os, binary, default_class_);
  WriteToken(os, binary, "</LinearVtln>");
}

Friends And Related Function Documentation

LinearVtlnStats

friend class LinearVtlnStats

friend

Definition at line 85 of file lvtln.h.

Member Data Documentation

A_

std::vector<Matrix<BaseFloat> > A_

protected

Definition at line 88 of file lvtln.h.

Referenced by LinearVtln::ComputeTransform(), LinearVtln::Dim(), LinearVtln::GetTransform(), LinearVtln::LinearVtln(), LinearVtln::NumClasses(), LinearVtln::Read(), LinearVtln::SetTransform(), and LinearVtln::Write().

default_class_

int32 default_class_

protected

Definition at line 87 of file lvtln.h.

Referenced by LinearVtln::ComputeTransform(), LinearVtln::LinearVtln(), LinearVtln::Read(), and LinearVtln::Write().

logdets_

std::vector<BaseFloat> logdets_

protected

Definition at line 89 of file lvtln.h.

Referenced by LinearVtln::ComputeTransform(), LinearVtln::LinearVtln(), LinearVtln::Read(), LinearVtln::SetTransform(), and LinearVtln::Write().

warps_

std::vector<BaseFloat> warps_

protected

Definition at line 90 of file lvtln.h.

Referenced by LinearVtln::GetWarp(), LinearVtln::LinearVtln(), LinearVtln::Read(), LinearVtln::SetWarp(), and LinearVtln::Write().

---

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

• transform/lvtln.h
• transform/lvtln.cc