GaussClusterable Class Reference

GaussClusterable wraps Gaussian statistics in a form accessible to generic clustering algorithms. More...

#include <clusterable-classes.h>

Inheritance diagram for GaussClusterable:

Collaboration diagram for GaussClusterable:

Public Member Functions
	GaussClusterable ()
	GaussClusterable (int32 dim, BaseFloat var_floor)
	GaussClusterable (const Vector< BaseFloat > &x_stats, const Vector< BaseFloat > &x2_stats, BaseFloat var_floor, BaseFloat count)
virtual std::string	Type () const
	Return a string that describes the inherited type. More...
void	AddStats (const VectorBase< BaseFloat > &vec, BaseFloat weight=1.0)
virtual BaseFloat	Objf () const
	Return the objective function associated with the stats [assuming ML estimation]. More...
virtual void	SetZero ()
	Set stats to empty. More...
virtual void	Add (const Clusterable &other_in)
	Add other stats. More...
virtual void	Sub (const Clusterable &other_in)
	Subtract other stats. More...
virtual BaseFloat	Normalizer () const
	Return the normalizer (typically, count) associated with the stats. More...
virtual Clusterable *	Copy () const
	Return a copy of this object. More...
virtual void	Scale (BaseFloat f)
	Scale the stats by a positive number f [not mandatory to supply this]. More...
virtual void	Write (std::ostream &os, bool binary) const
	Write data to stream. More...
virtual Clusterable *	ReadNew (std::istream &is, bool binary) const
	Read data from a stream and return the corresponding object (const function; it's a class member because we need access to the vtable so generic code can read derived types). More...
virtual	~GaussClusterable ()
BaseFloat	count () const
SubVector< double >	x_stats () const
SubVector< double >	x2_stats () const
Public Member Functions inherited from Clusterable
virtual	~Clusterable ()
virtual BaseFloat	ObjfPlus (const Clusterable &other) const
	Return the objective function of the combined object this + other. More...
virtual BaseFloat	ObjfMinus (const Clusterable &other) const
	Return the objective function of the subtracted object this - other. More...
virtual BaseFloat	Distance (const Clusterable &other) const
	Return the objective function decrease from merging the two clusters, negated to be a positive number (or zero). More...

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

Private Attributes
double	count_
Matrix< double >	stats_
double	var_floor_

Detailed Description

GaussClusterable wraps Gaussian statistics in a form accessible to generic clustering algorithms.

Definition at line 65 of file clusterable-classes.h.

Constructor & Destructor Documentation

GaussClusterable ( )

inline

Definition at line 67 of file clusterable-classes.h.

Referenced by GaussClusterable::Copy(), and GaussClusterable::ReadNew().

: count_(0.0), var_floor_(0.0) {}

GaussClusterable ( int32  dim, BaseFloat  var_floor  )

inline

Definition at line 68 of file clusterable-classes.h.

                                                  :
      count_(0.0), stats_(2, dim), var_floor_(var_floor) {}

GaussClusterable ( const Vector< BaseFloat > &  x_stats, const Vector< BaseFloat > &  x2_stats, BaseFloat  var_floor, BaseFloat  count  )

inline

Definition at line 107 of file clusterable-classes.h.

References MatrixBase< Real >::Row(), and GaussClusterable::stats_.

                                                                               :
    count_(count), stats_(2, x_stats.Dim()), var_floor_(var_floor) {
  stats_.Row(0).CopyFromVec(x_stats);
  stats_.Row(1).CopyFromVec(x2_stats);
}

virtual ~GaussClusterable ( )

inlinevirtual

Definition at line 86 of file clusterable-classes.h.

{}

Member Function Documentation

void Add ( const Clusterable &  other )

virtual

Add other stats.

Implements Clusterable.

Definition at line 144 of file clusterable-classes.cc.

References MatrixBase< Real >::AddMat(), GaussClusterable::count_, KALDI_ASSERT, GaussClusterable::stats_, and Clusterable::Type().

                                                      {
  KALDI_ASSERT(other_in.Type() == "gauss");
  const GaussClusterable *other =
      static_cast<const GaussClusterable*>(&other_in);
  count_ += other->count_;
  stats_.AddMat(1.0, other->stats_);
}

void AddStats	(	const VectorBase< BaseFloat > &	vec,
		BaseFloat	weight = 1.0
	)

Definition at line 137 of file clusterable-classes.cc.

References GaussClusterable::count_, MatrixBase< Real >::Row(), and GaussClusterable::stats_.

Referenced by kaldi::GenRandStats().

                                                  {
  count_ += weight;
  stats_.Row(0).AddVec(weight, vec);
  stats_.Row(1).AddVec2(weight, vec);
}

Clusterable * Copy ( ) const

virtual

Return a copy of this object.

Implements Clusterable.

Definition at line 160 of file clusterable-classes.cc.

References GaussClusterable::GaussClusterable(), KALDI_ASSERT, MatrixBase< Real >::NumCols(), MatrixBase< Real >::NumRows(), GaussClusterable::stats_, and GaussClusterable::var_floor_.

                                          {
  KALDI_ASSERT(stats_.NumRows() == 2);
  GaussClusterable *ans = new GaussClusterable(stats_.NumCols(), var_floor_);
  ans->Add(*this);
  return ans;
}

BaseFloat count ( ) const

inline

Definition at line 88 of file clusterable-classes.h.

References GaussClusterable::count_.

Referenced by kaldi::ClusterGaussiansToUbm(), DiagGmm::DiagGmm(), kaldi::InitAmGmm(), and DiagGmm::MergeKmeans().

{ return count_; }

virtual BaseFloat Normalizer ( ) const

inlinevirtual

Return the normalizer (typically, count) associated with the stats.

Implements Clusterable.

Definition at line 81 of file clusterable-classes.h.

References GaussClusterable::count_.

{ return count_; }

BaseFloat Objf ( ) const

virtual

Return the objective function associated with the stats [assuming ML estimation].

Implements Clusterable.

Definition at line 193 of file clusterable-classes.cc.

References GaussClusterable::count_, rnnlm::d, KALDI_ISNAN, KALDI_WARN, M_LOG_2PI, MatrixBase< Real >::NumCols(), GaussClusterable::stats_, VectorBase< Real >::SumLog(), and GaussClusterable::var_floor_.

                                       {
  if (count_ <= 0.0) {
    if (count_ < -0.1) {
      KALDI_WARN << "GaussClusterable::Objf(), count is negative " << count_;
    }
    return 0.0;
  } else {
    size_t dim = stats_.NumCols();
    Vector<double> vars(dim);
    double objf_per_frame = 0.0;
    for (size_t d = 0; d < dim; d++) {
      double mean(stats_(0, d) / count_), var = stats_(1, d) / count_ - mean
          * mean, floored_var = std::max(var, var_floor_);
      vars(d) = floored_var;
      objf_per_frame += -0.5 * var / floored_var;
    }
    objf_per_frame += -0.5 * (vars.SumLog() + M_LOG_2PI * dim);
    if (KALDI_ISNAN(objf_per_frame)) {
      KALDI_WARN << "GaussClusterable::Objf(), objf is NaN";
      return 0.0;
    }
    // KALDI_VLOG(2) << "count = " << count_ << ", objf_per_frame = "<< objf_per_frame
    //   << ", returning " << (objf_per_frame*count_) << ", floor = " << var_floor_;
    return objf_per_frame * count_;
  }
}

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

private

Definition at line 186 of file clusterable-classes.cc.

References GaussClusterable::count_, kaldi::ExpectToken(), Matrix< Real >::Read(), kaldi::ReadBasicType(), GaussClusterable::stats_, and GaussClusterable::var_floor_.

Referenced by GaussClusterable::ReadNew().

                                                       {
  ExpectToken(is, binary, "GCL");  // magic string.
  ReadBasicType(is, binary, &count_);
  ReadBasicType(is, binary, &var_floor_);
  stats_.Read(is, binary);
}

Clusterable * ReadNew ( std::istream &  os, bool  binary  ) const

virtual

Read data from a stream and return the corresponding object (const function; it's a class member because we need access to the vtable so generic code can read derived types).

Implements Clusterable.

Definition at line 180 of file clusterable-classes.cc.

References GaussClusterable::GaussClusterable(), and GaussClusterable::Read().

                                                                        {
  GaussClusterable *gc = new GaussClusterable();
  gc->Read(is, binary);
  return gc;
}

void Scale ( BaseFloat  f )

virtual

Scale the stats by a positive number f [not mandatory to supply this].

Reimplemented from Clusterable.

Definition at line 167 of file clusterable-classes.cc.

References GaussClusterable::count_, KALDI_ASSERT, MatrixBase< Real >::Scale(), and GaussClusterable::stats_.

                                        {
  KALDI_ASSERT(f >= 0.0);
  count_ *= f;
  stats_.Scale(f);
}

void SetZero ( )

inlinevirtual

Set stats to empty.

Implements Clusterable.

Definition at line 102 of file clusterable-classes.h.

References GaussClusterable::count_, MatrixBase< Real >::SetZero(), and GaussClusterable::stats_.

                                      {
  count_ = 0;
  stats_.SetZero();
}

void Sub ( const Clusterable &  other )

virtual

Subtract other stats.

Implements Clusterable.

Definition at line 152 of file clusterable-classes.cc.

References MatrixBase< Real >::AddMat(), GaussClusterable::count_, KALDI_ASSERT, GaussClusterable::stats_, and Clusterable::Type().

                                                      {
  KALDI_ASSERT(other_in.Type() == "gauss");
  const GaussClusterable *other =
      static_cast<const GaussClusterable*>(&other_in);
  count_ -= other->count_;
  stats_.AddMat(-1.0, other->stats_);
}

virtual std::string Type ( ) const

inlinevirtual

Return a string that describes the inherited type.

Implements Clusterable.

Definition at line 75 of file clusterable-classes.h.

{  return "gauss"; }

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

virtual

Write data to stream.

Implements Clusterable.

Definition at line 173 of file clusterable-classes.cc.

References GaussClusterable::count_, GaussClusterable::stats_, GaussClusterable::var_floor_, MatrixBase< Real >::Write(), kaldi::WriteBasicType(), and kaldi::WriteToken().

                                                              {
  WriteToken(os, binary, "GCL");  // magic string.
  WriteBasicType(os, binary, count_);
  WriteBasicType(os, binary, var_floor_);
  stats_.Write(os, binary);
}

SubVector<double> x2_stats ( ) const

inline

Definition at line 91 of file clusterable-classes.h.

References MatrixBase< Real >::Row(), and GaussClusterable::stats_.

Referenced by kaldi::ClusterGaussiansToUbm(), DiagGmm::DiagGmm(), and DiagGmm::MergeKmeans().

{ return stats_.Row(1); }

SubVector<double> x_stats ( ) const

inline

Definition at line 90 of file clusterable-classes.h.

References MatrixBase< Real >::Row(), and GaussClusterable::stats_.

Referenced by kaldi::ClusterGaussiansToUbm(), DiagGmm::DiagGmm(), and DiagGmm::MergeKmeans().

{ return stats_.Row(0); }

Member Data Documentation

double count_

private

Definition at line 93 of file clusterable-classes.h.

Referenced by GaussClusterable::Add(), GaussClusterable::AddStats(), GaussClusterable::count(), GaussClusterable::Normalizer(), GaussClusterable::Objf(), GaussClusterable::Read(), GaussClusterable::Scale(), GaussClusterable::SetZero(), GaussClusterable::Sub(), and GaussClusterable::Write().

Matrix<double> stats_

private

Definition at line 94 of file clusterable-classes.h.

Referenced by GaussClusterable::Add(), GaussClusterable::AddStats(), GaussClusterable::Copy(), GaussClusterable::GaussClusterable(), GaussClusterable::Objf(), GaussClusterable::Read(), GaussClusterable::Scale(), GaussClusterable::SetZero(), GaussClusterable::Sub(), GaussClusterable::Write(), GaussClusterable::x2_stats(), and GaussClusterable::x_stats().

double var_floor_

private

Definition at line 95 of file clusterable-classes.h.

Referenced by GaussClusterable::Copy(), GaussClusterable::Objf(), GaussClusterable::Read(), and GaussClusterable::Write().

---

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

• tree/clusterable-classes.h
• tree/clusterable-classes.cc