Classes | Public Types | Public Member Functions | Static Public Member Functions | Static Public Attributes | Protected Member Functions | Protected Attributes | Static Private Member Functions | List of all members
Component Class Referenceabstract

Abstract class, building block of the network. More...

#include <nnet-component.h>

Inheritance diagram for Component:
Inheritance graph
[legend]
Collaboration diagram for Component:
Collaboration graph
[legend]

Classes

struct  key_value
 A pair of type and marker,. More...
 

Public Types

enum  ComponentType {
  kUnknown = 0x0, kUpdatableComponent = 0x0100, kAffineTransform, kLinearTransform,
  kConvolutionalComponent, kLstmProjected, kBlstmProjected, kRecurrentComponent,
  kActivationFunction = 0x0200, kSoftmax, kHiddenSoftmax, kBlockSoftmax,
  kSigmoid, kTanh, kParametricRelu, kDropout,
  kLengthNormComponent, kTranform = 0x0400, kRbm, kSplice,
  kCopy, kTranspose, kBlockLinearity, kAddShift,
  kRescale, kKlHmm = 0x0800, kSentenceAveragingComponent, kSimpleSentenceAveragingComponent,
  kAveragePoolingComponent, kMaxPoolingComponent, kFramePoolingComponent, kParallelComponent,
  kMultiBasisComponent
}
 Component type identification mechanism,. More...
 

Public Member Functions

 Component (int32 input_dim, int32 output_dim)
 Generic interface of a component,. More...
 
virtual ~Component ()
 
virtual Component * Copy () const =0
 Copy component (deep copy),. More...
 
virtual ComponentType GetType () const =0
 Get Type Identification of the component,. More...
 
virtual bool IsUpdatable () const
 Check if componeny has 'Updatable' interface (trainable components),. More...
 
virtual bool IsMultistream () const
 Check if component has 'Recurrent' interface (trainable and recurrent),. More...
 
int32 InputDim () const
 Get the dimension of the input,. More...
 
int32 OutputDim () const
 Get the dimension of the output,. More...
 
void Propagate (const CuMatrixBase< BaseFloat > &in, CuMatrix< BaseFloat > *out)
 Perform forward-pass propagation 'in' -> 'out',. More...
 
void Backpropagate (const CuMatrixBase< BaseFloat > &in, const CuMatrixBase< BaseFloat > &out, const CuMatrixBase< BaseFloat > &out_diff, CuMatrix< BaseFloat > *in_diff)
 Perform backward-pass propagation 'out_diff' -> 'in_diff'. More...
 
void Write (std::ostream &os, bool binary) const
 Write the component to a stream,. More...
 
virtual std::string Info () const
 Print some additional info (after <ComponentName> and the dims),. More...
 
virtual std::string InfoGradient () const
 Print some additional info about gradient (after <...> and dims),. More...
 

Static Public Member Functions

static const char * TypeToMarker (ComponentType t)
 Converts component type to marker,. More...
 
static ComponentType MarkerToType (const std::string &s)
 Converts marker to component type (case insensitive),. More...
 
static Component * Init (const std::string &conf_line)
 Initialize component from a line in config file,. More...
 
static Component * Read (std::istream &is, bool binary)
 Read the component from a stream (static method),. More...
 

Static Public Attributes

static const struct key_value kMarkerMap []
 The table with pairs of Component types and markers (defined in nnet-component.cc),. More...
 

Protected Member Functions

virtual void PropagateFnc (const CuMatrixBase< BaseFloat > &in, CuMatrixBase< BaseFloat > *out)=0
 Abstract interface for propagation/backpropagation. More...
 
virtual void BackpropagateFnc (const CuMatrixBase< BaseFloat > &in, const CuMatrixBase< BaseFloat > &out, const CuMatrixBase< BaseFloat > &out_diff, CuMatrixBase< BaseFloat > *in_diff)=0
 Backward pass transformation (to be implemented by descending class...) More...
 
virtual void InitData (std::istream &is)
 Virtual interface for initialization and I/O,. More...
 
virtual void ReadData (std::istream &is, bool binary)
 Reads the component content. More...
 
virtual void WriteData (std::ostream &os, bool binary) const
 Writes the component content. More...
 

Protected Attributes

int32 input_dim_
 Data members,. More...
 
int32 output_dim_
 Dimension of the output of the Component,. More...
 

Static Private Member Functions

static Component * NewComponentOfType (ComponentType t, int32 input_dim, int32 output_dim)
 Private members (descending classes cannot call this),. More...
 

Detailed Description

Abstract class, building block of the network.

It is able to propagate (PropagateFnc: compute the output based on its input) and backpropagate (BackpropagateFnc: i.e. transform loss derivative w.r.t. output to derivative w.r.t. the input) the formulas are implemented in descendant classes (AffineTransform,Sigmoid,Softmax,...).

Definition at line 51 of file nnet-component.cc.

Member Enumeration Documentation

◆ ComponentType

enum ComponentType

Component type identification mechanism,.

Types of Components,

Enumerator
kUnknown 
kUpdatableComponent 
kAffineTransform 
kLinearTransform 
kConvolutionalComponent 
kLstmProjected 
kBlstmProjected 
kRecurrentComponent 
kActivationFunction 
kSoftmax 
kHiddenSoftmax 
kBlockSoftmax 
kSigmoid 
kTanh 
kParametricRelu 
kDropout 
kLengthNormComponent 
kTranform 
kRbm 
kSplice 
kCopy 
kTranspose 
kBlockLinearity 
kAddShift 
kRescale 
kKlHmm 
kSentenceAveragingComponent 
kSimpleSentenceAveragingComponent 
kAveragePoolingComponent 
kMaxPoolingComponent 
kFramePoolingComponent 
kParallelComponent 
kMultiBasisComponent 

Definition at line 47 of file nnet-component.h.

47  {
48  kUnknown = 0x0,
49 
50  kUpdatableComponent = 0x0100,
51  kAffineTransform,
52  kLinearTransform,
53  kConvolutionalComponent,
54  kLstmProjected,
55  kBlstmProjected,
56  kRecurrentComponent,
57 
58  kActivationFunction = 0x0200,
59  kSoftmax,
60  kHiddenSoftmax,
61  kBlockSoftmax,
62  kSigmoid,
63  kTanh,
64  kParametricRelu,
65  kDropout,
66  kLengthNormComponent,
67 
68  kTranform = 0x0400,
69  kRbm,
70  kSplice,
71  kCopy,
72  kTranspose,
73  kBlockLinearity,
74  kAddShift,
75  kRescale,
76 
77  kKlHmm = 0x0800,
78  kSentenceAveragingComponent, /* deprecated */
79  kSimpleSentenceAveragingComponent,
80  kAveragePoolingComponent,
81  kMaxPoolingComponent,
82  kFramePoolingComponent,
83  kParallelComponent,
84  kMultiBasisComponent
85  } ComponentType;
kaldi::nnet1::Component::ComponentType
ComponentType
Component type identification mechanism,.
Definition: nnet-component.h:47

Constructor & Destructor Documentation

◆ Component()

Component ( int32  input_dim,
int32  output_dim 
)
inline

Generic interface of a component,.

Definition at line 105 of file nnet-component.h.

105  :
106  input_dim_(input_dim),
107  output_dim_(output_dim)
108  { }
kaldi::nnet1::Component::input_dim_
int32 input_dim_
Data members,.
Definition: nnet-component.h:190
kaldi::nnet1::Component::output_dim_
int32 output_dim_
Dimension of the output of the Component,.
Definition: nnet-component.h:191

◆ ~Component()

virtual ~Component ( )
inlinevirtual

Definition at line 110 of file nnet-component.h.

References Component::Copy(), and Component::GetType().

111  { }

Member Function Documentation

◆ Backpropagate()

void Backpropagate ( const CuMatrixBase< BaseFloat > &  in,
const CuMatrixBase< BaseFloat > &  out,
const CuMatrixBase< BaseFloat > &  out_diff,
CuMatrix< BaseFloat > *  in_diff 
)
inline

Perform backward-pass propagation 'out_diff' -> 'in_diff'.

Note: 'in' and 'out' will be used only sometimes...

Definition at line 325 of file nnet-component.h.

References Component::BackpropagateFnc(), Component::InputDim(), KALDI_ASSERT, KALDI_ERR, kaldi::kSetZero, CuMatrixBase< Real >::NumCols(), CuMatrixBase< Real >::NumRows(), Component::OutputDim(), and CuMatrix< Real >::Resize().

Referenced by Component::OutputDim(), kaldi::nnet1::UnitTestConvolutionalComponent3x3(), kaldi::nnet1::UnitTestConvolutionalComponentUnity(), kaldi::nnet1::UnitTestDropoutComponent(), kaldi::nnet1::UnitTestMaxPoolingComponent(), and kaldi::nnet1::UnitTestSimpleSentenceAveragingComponent().

328  {
329  // Check the dims,
330  if (OutputDim() != out_diff.NumCols()) {
331  KALDI_ERR << "Non-matching dims! Component output dim " << OutputDim()
332  << ", the dim of output derivatives " << out_diff.NumCols();
333  }
334 
335  int32 num_frames = out_diff.NumRows();
336  KALDI_ASSERT(num_frames == in.NumRows());
337  KALDI_ASSERT(num_frames == out.NumRows());
338 
339  KALDI_ASSERT(InputDim() == in.NumCols());
340  KALDI_ASSERT(OutputDim() == out.NumCols());
341 
342  // Allocate target buffer,
343  KALDI_ASSERT(in_diff != NULL);
344  in_diff->Resize(num_frames, InputDim(), kSetZero); // reset,
345 
346  // Call the 'virtual' backprop function,
347  BackpropagateFnc(in, out, out_diff, in_diff);
348 }
kaldi::int32
kaldi::int32 int32
Definition: online-tcp-source.cc:27
kaldi::nnet1::Component::InputDim
int32 InputDim() const
Get the dimension of the input,.
Definition: nnet-component.h:130
KALDI_ERR
#define KALDI_ERR
Definition: kaldi-error.h:147
KALDI_ASSERT
#define KALDI_ASSERT(cond)
Definition: kaldi-error.h:185
kaldi::nnet1::Component::OutputDim
int32 OutputDim() const
Get the dimension of the output,.
Definition: nnet-component.h:135
kaldi::nnet1::Component::BackpropagateFnc
virtual void BackpropagateFnc(const CuMatrixBase< BaseFloat > &in, const CuMatrixBase< BaseFloat > &out, const CuMatrixBase< BaseFloat > &out_diff, CuMatrixBase< BaseFloat > *in_diff)=0
Backward pass transformation (to be implemented by descending class...)

◆ BackpropagateFnc()

virtual void BackpropagateFnc ( const CuMatrixBase< BaseFloat > &  in,
const CuMatrixBase< BaseFloat > &  out,
const CuMatrixBase< BaseFloat > &  out_diff,
CuMatrixBase< BaseFloat > *  in_diff 
)
protectedpure virtual

Backward pass transformation (to be implemented by descending class...)

Implemented in BlstmProjected, Rescale, LstmProjected, AddShift, ConvolutionalComponent, Dropout, MultiBasisComponent, LengthNormComponent, SentenceAveragingComponent, ParallelComponent, Tanh, RecurrentComponent, FramePoolingComponent, Sigmoid, CopyComponent, BlockSoftmax, AffineTransform, LinearTransform, ParametricRelu, SimpleSentenceAveragingComponent, AveragePoolingComponent, MaxPoolingComponent, Splice, KlHmm, HiddenSoftmax, RbmBase, and Softmax.

Referenced by Component::Backpropagate(), and Component::InfoGradient().

◆ Copy()

virtual Component* Copy ( ) const
pure virtual

Copy component (deep copy),.

Implemented in Rescale, AddShift, Dropout, LengthNormComponent, Tanh, Sigmoid, SentenceAveragingComponent, CopyComponent, BlockSoftmax, Rbm, ConvolutionalComponent, HiddenSoftmax, BlstmProjected, LstmProjected, FramePoolingComponent, SimpleSentenceAveragingComponent, AveragePoolingComponent, MaxPoolingComponent, RecurrentComponent, Splice, ParametricRelu, KlHmm, Softmax, MultiBasisComponent, ParallelComponent, AffineTransform, and LinearTransform.

Referenced by Nnet::AppendComponent(), Nnet::Nnet(), Nnet::operator=(), Nnet::ReplaceComponent(), and Component::~Component().

◆ GetType()

virtual ComponentType GetType ( ) const
pure virtual

Get Type Identification of the component,.

Implemented in Rescale, AddShift, Dropout, LengthNormComponent, Tanh, Sigmoid, SentenceAveragingComponent, CopyComponent, BlockSoftmax, Rbm, ConvolutionalComponent, HiddenSoftmax, BlstmProjected, LstmProjected, SimpleSentenceAveragingComponent, FramePoolingComponent, AveragePoolingComponent, MaxPoolingComponent, RecurrentComponent, Splice, ParametricRelu, KlHmm, Softmax, MultiBasisComponent, ParallelComponent, AffineTransform, and LinearTransform.

Referenced by kaldi::ConvertComponent(), main(), Component::Propagate(), Nnet::SetDropoutRate(), Component::Write(), and Component::~Component().

◆ Info()

virtual std::string Info ( ) const
inlinevirtual

Print some additional info (after <ComponentName> and the dims),.

Reimplemented in Rescale, BlstmProjected, AddShift, Dropout, LstmProjected, ConvolutionalComponent, MultiBasisComponent, SentenceAveragingComponent, CopyComponent, BlockSoftmax, ParallelComponent, FramePoolingComponent, RecurrentComponent, AffineTransform, LinearTransform, ParametricRelu, SimpleSentenceAveragingComponent, and Splice.

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

159 { return ""; }

◆ InfoGradient()

virtual std::string InfoGradient ( ) const
inlinevirtual

Print some additional info about gradient (after <...> and dims),.

Reimplemented in Rescale, BlstmProjected, AddShift, LstmProjected, MultiBasisComponent, ConvolutionalComponent, SentenceAveragingComponent, ParallelComponent, FramePoolingComponent, RecurrentComponent, AffineTransform, LinearTransform, ParametricRelu, and SimpleSentenceAveragingComponent.

Definition at line 162 of file nnet-component.h.

References Component::BackpropagateFnc(), and Component::PropagateFnc().

162 { return ""; }

◆ Init()

Component * Init ( const std::string &  conf_line)
static

Initialize component from a line in config file,.

Definition at line 203 of file nnet-component.cc.

References kaldi::ExpectToken(), Component::InitData(), Component::MarkerToType(), Component::NewComponentOfType(), kaldi::ReadBasicType(), and kaldi::ReadToken().

Referenced by Nnet::Init(), Component::OutputDim(), and kaldi::nnet1::UnitTestMaxPoolingComponent().

203  {
204  std::istringstream is(conf_line);
205  std::string component_type_string;
206  int32 input_dim, output_dim;
207 
208  // initialize component w/o internal data
209  ReadToken(is, false, &component_type_string);
210  ComponentType component_type = MarkerToType(component_type_string);
211  ExpectToken(is, false, "<InputDim>");
212  ReadBasicType(is, false, &input_dim);
213  ExpectToken(is, false, "<OutputDim>");
214  ReadBasicType(is, false, &output_dim);
215  Component *ans = NewComponentOfType(component_type, input_dim, output_dim);
216 
217  // initialize internal data with the remaining part of config line
218  ans->InitData(is);
219 
220  return ans;
221 }
kaldi::ReadBasicType
void ReadBasicType(std::istream &is, bool binary, T *t)
ReadBasicType is the name of the read function for bool, integer types, and floating-point types...
Definition: io-funcs-inl.h:55
kaldi::nnet1::Component::Component
Component(int32 input_dim, int32 output_dim)
Generic interface of a component,.
Definition: nnet-component.h:105
kaldi::int32
kaldi::int32 int32
Definition: online-tcp-source.cc:27
kaldi::ReadToken
void ReadToken(std::istream &is, bool binary, std::string *str)
ReadToken gets the next token and puts it in str (exception on failure).
Definition: io-funcs.cc:154
kaldi::nnet1::Component::ComponentType
ComponentType
Component type identification mechanism,.
Definition: nnet-component.h:47
kaldi::ExpectToken
void ExpectToken(std::istream &is, bool binary, const char *token)
ExpectToken tries to read in the given token, and throws an exception on failure. ...
Definition: io-funcs.cc:191
kaldi::nnet1::Component::MarkerToType
static ComponentType MarkerToType(const std::string &s)
Converts marker to component type (case insensitive),.
Definition: nnet-component.cc:94
kaldi::nnet1::Component::NewComponentOfType
static Component * NewComponentOfType(ComponentType t, int32 input_dim, int32 output_dim)
Private members (descending classes cannot call this),.
Definition: nnet-component.cc:110

◆ InitData()

virtual void InitData ( std::istream &  is)
inlineprotectedvirtual

Virtual interface for initialization and I/O,.

Initialize internal data of a component

Reimplemented in Rescale, AddShift, Dropout, UpdatableComponent, SentenceAveragingComponent, CopyComponent, BlockSoftmax, Rbm, ConvolutionalComponent, BlstmProjected, LstmProjected, SimpleSentenceAveragingComponent, FramePoolingComponent, AveragePoolingComponent, MaxPoolingComponent, RecurrentComponent, Splice, ParametricRelu, ParallelComponent, MultiBasisComponent, AffineTransform, and LinearTransform.

Definition at line 180 of file nnet-component.h.

Referenced by Component::Init(), and UpdatableComponent::SetBiasLearnRateCoef().

180 { }

◆ InputDim()

int32 InputDim ( ) const
inline

Get the dimension of the input,.

Definition at line 130 of file nnet-component.h.

References Component::input_dim_.

Referenced by Component::Backpropagate(), kaldi::ConvertSigmoidComponent(), kaldi::ConvertSoftmaxComponent(), kaldi::ConvertSpliceComponent(), LinearTransform::InitData(), AffineTransform::InitData(), MultiBasisComponent::InitData(), ParallelComponent::InitData(), Splice::InitData(), FramePoolingComponent::InitData(), Rbm::InitData(), CopyComponent::InitData(), SentenceAveragingComponent::InitData(), AddShift::InitData(), Rescale::InitData(), SimpleSentenceAveragingComponent::PropagateFnc(), Nnet::Read(), Splice::ReadData(), ParallelComponent::ReadData(), FramePoolingComponent::ReadData(), MultiBasisComponent::ReadData(), SentenceAveragingComponent::ReadData(), AddShift::Update(), Rescale::Update(), Component::Write(), and Rbm::WriteAsNnet().

130  {
131  return input_dim_;
132  }
kaldi::nnet1::Component::input_dim_
int32 input_dim_
Data members,.
Definition: nnet-component.h:190

◆ IsMultistream()

virtual bool IsMultistream ( ) const
inlinevirtual

Check if component has 'Recurrent' interface (trainable and recurrent),.

Reimplemented in MultistreamComponent.

Definition at line 125 of file nnet-component.h.

Referenced by Nnet::ResetStreams(), and Nnet::SetSeqLengths().

125  {
126  return false;
127  }

◆ IsUpdatable()

virtual bool IsUpdatable ( ) const
inlinevirtual

Check if componeny has 'Updatable' interface (trainable components),.

Reimplemented in UpdatableComponent.

Definition at line 120 of file nnet-component.h.

Referenced by main(), and Nnet::SetTrainOptions().

120  {
121  return false;
122  }

◆ MarkerToType()

Component::ComponentType MarkerToType ( const std::string &  s)
static

Converts marker to component type (case insensitive),.

Definition at line 94 of file nnet-component.cc.

References rnnlm::i, KALDI_ERR, Component::kMarkerMap, Component::kUnknown, and Component::key_value::value.

Referenced by Component::Init(), and Component::Read().

94  {
95  std::string s_lowercase(s);
96  std::transform(s.begin(), s.end(), s_lowercase.begin(), ::tolower); // lc
97  int32 N = sizeof(kMarkerMap) / sizeof(kMarkerMap[0]);
98  for (int i = 0; i < N; i++) {
99  std::string m(kMarkerMap[i].value);
100  std::string m_lowercase(m);
101  std::transform(m.begin(), m.end(), m_lowercase.begin(), ::tolower);
102  if (s_lowercase == m_lowercase) return kMarkerMap[i].key;
103  }
104  KALDI_ERR << "Unknown 'Component' marker : '" << s << "'\n"
105  << "(isn't the model 'too old' or incompatible?)";
106  return kUnknown;
107 }
kaldi::int32
kaldi::int32 int32
Definition: online-tcp-source.cc:27
kaldi::nnet1::Component::kMarkerMap
static const struct key_value kMarkerMap[]
The table with pairs of Component types and markers (defined in nnet-component.cc),.
Definition: nnet-component.h:95
KALDI_ERR
#define KALDI_ERR
Definition: kaldi-error.h:147

◆ NewComponentOfType()

Component * NewComponentOfType ( ComponentType  t,
int32  input_dim,
int32  output_dim 
)
staticprivate

Private members (descending classes cannot call this),.

Create a new intance of component,

Definition at line 110 of file nnet-component.cc.

References Component::kAddShift, Component::kAffineTransform, KALDI_ERR, Component::kAveragePoolingComponent, Component::kBlockSoftmax, Component::kBlstmProjected, Component::kConvolutionalComponent, Component::kCopy, Component::kDropout, Component::kFramePoolingComponent, Component::kHiddenSoftmax, Component::kKlHmm, Component::kLengthNormComponent, Component::kLinearTransform, Component::kLstmProjected, Component::kMaxPoolingComponent, Component::kMultiBasisComponent, Component::kParallelComponent, Component::kParametricRelu, Component::kRbm, Component::kRecurrentComponent, Component::kRescale, Component::kSentenceAveragingComponent, Component::kSigmoid, Component::kSimpleSentenceAveragingComponent, Component::kSoftmax, Component::kSplice, Component::kTanh, Component::kUnknown, kaldi::cu::Splice(), and Component::TypeToMarker().

Referenced by Component::Init(), and Component::Read().

111  {
112  Component *ans = NULL;
113  switch (comp_type) {
114  case Component::kAffineTransform :
115  ans = new AffineTransform(input_dim, output_dim);
116  break;
117  case Component::kLinearTransform :
118  ans = new LinearTransform(input_dim, output_dim);
119  break;
120  case Component::kConvolutionalComponent :
121  ans = new ConvolutionalComponent(input_dim, output_dim);
122  break;
123  case Component::kLstmProjected :
124  ans = new LstmProjected(input_dim, output_dim);
125  break;
126  case Component::kBlstmProjected :
127  ans = new BlstmProjected(input_dim, output_dim);
128  break;
129  case Component::kRecurrentComponent :
130  ans = new RecurrentComponent(input_dim, output_dim);
131  break;
132  case Component::kSoftmax :
133  ans = new Softmax(input_dim, output_dim);
134  break;
135  case Component::kHiddenSoftmax :
136  ans = new HiddenSoftmax(input_dim, output_dim);
137  break;
138  case Component::kBlockSoftmax :
139  ans = new BlockSoftmax(input_dim, output_dim);
140  break;
141  case Component::kSigmoid :
142  ans = new Sigmoid(input_dim, output_dim);
143  break;
144  case Component::kTanh :
145  ans = new Tanh(input_dim, output_dim);
146  break;
147  case Component::kParametricRelu :
148  ans = new ParametricRelu(input_dim, output_dim);
149  break;
150  case Component::kDropout :
151  ans = new Dropout(input_dim, output_dim);
152  break;
153  case Component::kLengthNormComponent :
154  ans = new LengthNormComponent(input_dim, output_dim);
155  break;
156  case Component::kRbm :
157  ans = new Rbm(input_dim, output_dim);
158  break;
159  case Component::kSplice :
160  ans = new Splice(input_dim, output_dim);
161  break;
162  case Component::kCopy :
163  ans = new CopyComponent(input_dim, output_dim);
164  break;
165  case Component::kAddShift :
166  ans = new AddShift(input_dim, output_dim);
167  break;
168  case Component::kRescale :
169  ans = new Rescale(input_dim, output_dim);
170  break;
171  case Component::kKlHmm :
172  ans = new KlHmm(input_dim, output_dim);
173  break;
174  case Component::kSentenceAveragingComponent :
175  ans = new SentenceAveragingComponent(input_dim, output_dim);
176  break;
177  case Component::kSimpleSentenceAveragingComponent :
178  ans = new SimpleSentenceAveragingComponent(input_dim, output_dim);
179  break;
180  case Component::kAveragePoolingComponent :
181  ans = new AveragePoolingComponent(input_dim, output_dim);
182  break;
183  case Component::kMaxPoolingComponent :
184  ans = new MaxPoolingComponent(input_dim, output_dim);
185  break;
186  case Component::kFramePoolingComponent :
187  ans = new FramePoolingComponent(input_dim, output_dim);
188  break;
189  case Component::kParallelComponent :
190  ans = new ParallelComponent(input_dim, output_dim);
191  break;
192  case Component::kMultiBasisComponent :
193  ans = new MultiBasisComponent(input_dim, output_dim);
194  break;
195  case Component::kUnknown :
196  default :
197  KALDI_ERR << "Missing type: " << TypeToMarker(comp_type);
198  }
199  return ans;
200 }
kaldi::nnet1::Component::Component
Component(int32 input_dim, int32 output_dim)
Generic interface of a component,.
Definition: nnet-component.h:105
kaldi::nnet1::Component::TypeToMarker
static const char * TypeToMarker(ComponentType t)
Converts component type to marker,.
Definition: nnet-component.cc:84
KALDI_ERR
#define KALDI_ERR
Definition: kaldi-error.h:147
kaldi::cu::Splice
void Splice(const CuMatrixBase< Real > &src, const CuArray< int32 > &frame_offsets, CuMatrixBase< Real > *tgt)
Splice concatenates frames of src as specified in frame_offsets into tgt.
Definition: cu-math.cc:132

◆ OutputDim()

int32 OutputDim ( ) const
inline

Get the dimension of the output,.

Definition at line 135 of file nnet-component.h.

References Component::Backpropagate(), Component::Init(), Component::output_dim_, Component::Propagate(), Component::Read(), and Component::Write().

Referenced by Component::Backpropagate(), SimpleSentenceAveragingComponent::BackpropagateFnc(), MultiBasisComponent::BackpropagateFnc(), AffineTransform::InitData(), LinearTransform::InitData(), MultiBasisComponent::InitData(), ParallelComponent::InitData(), Splice::InitData(), Rbm::InitData(), BlockSoftmax::InitData(), CopyComponent::InitData(), SentenceAveragingComponent::InitData(), Splice::ReadData(), ParallelComponent::ReadData(), MultiBasisComponent::ReadData(), BlockSoftmax::ReadData(), CopyComponent::ReadData(), SentenceAveragingComponent::ReadData(), AffineTransform::Update(), RecurrentComponent::Update(), Component::Write(), and Rbm::WriteAsNnet().

135  {
136  return output_dim_;
137  }
kaldi::nnet1::Component::output_dim_
int32 output_dim_
Dimension of the output of the Component,.
Definition: nnet-component.h:191

◆ Propagate()

void Propagate ( const CuMatrixBase< BaseFloat > &  in,
CuMatrix< BaseFloat > *  out 
)
inline

Perform forward-pass propagation 'in' -> 'out',.

Definition at line 311 of file nnet-component.h.

References Component::GetType(), Component::input_dim_, KALDI_ERR, kaldi::kSetZero, CuMatrixBase< Real >::NumCols(), CuMatrixBase< Real >::NumRows(), Component::output_dim_, Component::PropagateFnc(), CuMatrix< Real >::Resize(), and Component::TypeToMarker().

Referenced by Component::OutputDim(), kaldi::nnet1::UnitTestConvolutionalComponent3x3(), kaldi::nnet1::UnitTestConvolutionalComponentUnity(), kaldi::nnet1::UnitTestDropoutComponent(), kaldi::nnet1::UnitTestLengthNorm(), kaldi::nnet1::UnitTestMaxPoolingComponent(), and kaldi::nnet1::UnitTestSimpleSentenceAveragingComponent().

312  {
313  // Check the dims
314  if (input_dim_ != in.NumCols()) {
315  KALDI_ERR << "Non-matching dims on the input of " << TypeToMarker(GetType())
316  << " component. The input-dim is " << input_dim_
317  << ", the data had " << in.NumCols() << " dims.";
318  }
319  // Allocate target buffer
320  out->Resize(in.NumRows(), output_dim_, kSetZero); // reset
321  // Call the propagation implementation of the component
322  PropagateFnc(in, out);
323 }
kaldi::nnet1::Component::input_dim_
int32 input_dim_
Data members,.
Definition: nnet-component.h:190
kaldi::nnet1::Component::TypeToMarker
static const char * TypeToMarker(ComponentType t)
Converts component type to marker,.
Definition: nnet-component.cc:84
KALDI_ERR
#define KALDI_ERR
Definition: kaldi-error.h:147
kaldi::nnet1::Component::output_dim_
int32 output_dim_
Dimension of the output of the Component,.
Definition: nnet-component.h:191
kaldi::nnet1::Component::PropagateFnc
virtual void PropagateFnc(const CuMatrixBase< BaseFloat > &in, CuMatrixBase< BaseFloat > *out)=0
Abstract interface for propagation/backpropagation.
kaldi::nnet1::Component::GetType
virtual ComponentType GetType() const =0
Get Type Identification of the component,.

◆ PropagateFnc()

virtual void PropagateFnc ( const CuMatrixBase< BaseFloat > &  in,
CuMatrixBase< BaseFloat > *  out 
)
protectedpure virtual

Abstract interface for propagation/backpropagation.

Forward pass transformation (to be implemented by descending class...)

Implemented in BlstmProjected, Rescale, AddShift, LstmProjected, Dropout, MultiBasisComponent, ConvolutionalComponent, LengthNormComponent, Tanh, SentenceAveragingComponent, ParallelComponent, Rbm, Sigmoid, FramePoolingComponent, CopyComponent, RecurrentComponent, AffineTransform, BlockSoftmax, LinearTransform, ParametricRelu, SimpleSentenceAveragingComponent, Splice, AveragePoolingComponent, MaxPoolingComponent, HiddenSoftmax, KlHmm, and Softmax.

Referenced by Component::InfoGradient(), and Component::Propagate().

◆ Read()

Component * Read ( std::istream &  is,
bool  binary 
)
static

Read the component from a stream (static method),.

Definition at line 224 of file nnet-component.cc.

References kaldi::ExpectToken(), Component::MarkerToType(), Component::NewComponentOfType(), kaldi::Peek(), kaldi::PeekToken(), kaldi::ReadBasicType(), Component::ReadData(), and kaldi::ReadToken().

Referenced by Component::OutputDim(), Nnet::Read(), and kaldi::nnet1::ReadComponentFromString().

224  {
225  int32 dim_out, dim_in;
226  std::string token;
227 
228  int first_char = Peek(is, binary);
229  if (first_char == EOF) return NULL;
230 
231  ReadToken(is, binary, &token);
232  // Skip the optional initial token,
233  if (token == "<Nnet>") {
234  ReadToken(is, binary, &token);
235  }
236  // Network ends after terminal token appears,
237  if (token == "</Nnet>") {
238  return NULL;
239  }
240 
241  // Read the dims,
242  ReadBasicType(is, binary, &dim_out);
243  ReadBasicType(is, binary, &dim_in);
244 
245  // Create the component,
246  Component *ans = NewComponentOfType(MarkerToType(token), dim_in, dim_out);
247 
248  // Read the content,
249  ans->ReadData(is, binary);
250 
251  // 'Eat' the component separtor (can be already consumed by 'ReadData(.)'),
252  if ('<' == Peek(is, binary) && '!' == PeekToken(is, binary)) {
253  ExpectToken(is, binary, "<!EndOfComponent>");
254  }
255 
256  return ans;
257 }
kaldi::ReadBasicType
void ReadBasicType(std::istream &is, bool binary, T *t)
ReadBasicType is the name of the read function for bool, integer types, and floating-point types...
Definition: io-funcs-inl.h:55
kaldi::nnet1::Component::Component
Component(int32 input_dim, int32 output_dim)
Generic interface of a component,.
Definition: nnet-component.h:105
kaldi::int32
kaldi::int32 int32
Definition: online-tcp-source.cc:27
kaldi::ReadToken
void ReadToken(std::istream &is, bool binary, std::string *str)
ReadToken gets the next token and puts it in str (exception on failure).
Definition: io-funcs.cc:154
kaldi::Peek
int Peek(std::istream &is, bool binary)
Peek consumes whitespace (if binary == false) and then returns the peek() value of the stream...
Definition: io-funcs.cc:145
kaldi::ExpectToken
void ExpectToken(std::istream &is, bool binary, const char *token)
ExpectToken tries to read in the given token, and throws an exception on failure. ...
Definition: io-funcs.cc:191
kaldi::nnet1::Component::MarkerToType
static ComponentType MarkerToType(const std::string &s)
Converts marker to component type (case insensitive),.
Definition: nnet-component.cc:94
kaldi::PeekToken
int PeekToken(std::istream &is, bool binary)
PeekToken will return the first character of the next token, or -1 if end of file.
Definition: io-funcs.cc:170
kaldi::nnet1::Component::NewComponentOfType
static Component * NewComponentOfType(ComponentType t, int32 input_dim, int32 output_dim)
Private members (descending classes cannot call this),.
Definition: nnet-component.cc:110

◆ ReadData()

virtual void ReadData ( std::istream &  is,
bool  binary 
)
inlineprotectedvirtual

Reads the component content.

Reimplemented in Rescale, AddShift, Dropout, SentenceAveragingComponent, CopyComponent, Rbm, BlockSoftmax, MultiBasisComponent, BlstmProjected, ConvolutionalComponent, FramePoolingComponent, KlHmm, ParallelComponent, LstmProjected, LinearTransform, Splice, RecurrentComponent, AffineTransform, SimpleSentenceAveragingComponent, ParametricRelu, AveragePoolingComponent, and MaxPoolingComponent.

Definition at line 183 of file nnet-component.h.

Referenced by Component::Read().

183 { }

◆ TypeToMarker()

const char * TypeToMarker ( ComponentType  t)
static

Converts component type to marker,.

Definition at line 84 of file nnet-component.cc.

References rnnlm::i, KALDI_ERR, Component::key_value::key, and Component::kMarkerMap.

Referenced by CopyComponent::BackpropagateFnc(), Nnet::Check(), kaldi::ConvertComponent(), Nnet::Info(), Nnet::InfoBackPropagate(), Nnet::InfoGradient(), Nnet::InfoPropagate(), main(), Component::NewComponentOfType(), Component::Propagate(), Component::Write(), and Rbm::WriteAsNnet().

84  {
85  // Retuns the 1st '<string>' corresponding to the type in 'kMarkerMap',
86  int32 N = sizeof(kMarkerMap) / sizeof(kMarkerMap[0]);
87  for (int i = 0; i < N; i++) {
88  if (kMarkerMap[i].key == t) return kMarkerMap[i].value;
89  }
90  KALDI_ERR << "Unknown type : " << t;
91  return NULL;
92 }
kaldi::int32
kaldi::int32 int32
Definition: online-tcp-source.cc:27
kaldi::nnet1::Component::kMarkerMap
static const struct key_value kMarkerMap[]
The table with pairs of Component types and markers (defined in nnet-component.cc),.
Definition: nnet-component.h:95
KALDI_ERR
#define KALDI_ERR
Definition: kaldi-error.h:147

◆ Write()

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

Write the component to a stream,.

Definition at line 260 of file nnet-component.cc.

References Component::GetType(), Component::InputDim(), Component::OutputDim(), Component::TypeToMarker(), kaldi::WriteBasicType(), Component::WriteData(), and kaldi::WriteToken().

Referenced by Component::OutputDim().

260  {
261  WriteToken(os, binary, Component::TypeToMarker(GetType()));
262  WriteBasicType(os, binary, OutputDim());
263  WriteBasicType(os, binary, InputDim());
264  if (!binary) os << "\n";
265  this->WriteData(os, binary);
266  WriteToken(os, binary, "<!EndOfComponent>"); // Write component separator.
267  if (!binary) os << "\n";
268 }
kaldi::nnet1::Component::TypeToMarker
static const char * TypeToMarker(ComponentType t)
Converts component type to marker,.
Definition: nnet-component.cc:84
kaldi::nnet1::Component::InputDim
int32 InputDim() const
Get the dimension of the input,.
Definition: nnet-component.h:130
kaldi::nnet1::Component::WriteData
virtual void WriteData(std::ostream &os, bool binary) const
Writes the component content.
Definition: nnet-component.h:186
kaldi::WriteToken
void WriteToken(std::ostream &os, bool binary, const char *token)
The WriteToken functions are for writing nonempty sequences of non-space characters.
Definition: io-funcs.cc:134
kaldi::nnet1::Component::GetType
virtual ComponentType GetType() const =0
Get Type Identification of the component,.
kaldi::WriteBasicType
void WriteBasicType(std::ostream &os, bool binary, T t)
WriteBasicType is the name of the write function for bool, integer types, and floating-point types...
Definition: io-funcs-inl.h:34
kaldi::nnet1::Component::OutputDim
int32 OutputDim() const
Get the dimension of the output,.
Definition: nnet-component.h:135

◆ WriteData()

virtual void WriteData ( std::ostream &  os,
bool  binary 
) const
inlineprotectedvirtual

Writes the component content.

Reimplemented in Rescale, AddShift, Dropout, SentenceAveragingComponent, Rbm, CopyComponent, BlstmProjected, ConvolutionalComponent, MultiBasisComponent, BlockSoftmax, FramePoolingComponent, LstmProjected, ParallelComponent, KlHmm, LinearTransform, RecurrentComponent, AffineTransform, SimpleSentenceAveragingComponent, ParametricRelu, Splice, AveragePoolingComponent, and MaxPoolingComponent.

Definition at line 186 of file nnet-component.h.

Referenced by Component::Write().

186 { }

Member Data Documentation

◆ input_dim_

int32 input_dim_
protected

Data members,.

Dimension of the input of the Component,

Definition at line 190 of file nnet-component.h.

Referenced by MaxPoolingComponent::BackpropagateFnc(), AveragePoolingComponent::BackpropagateFnc(), LstmProjected::BackpropagateFnc(), BlstmProjected::BackpropagateFnc(), RecurrentComponent::InitData(), LstmProjected::InitData(), BlstmProjected::InitData(), ConvolutionalComponent::InitData(), Component::InputDim(), Component::Propagate(), AveragePoolingComponent::PropagateFnc(), MaxPoolingComponent::PropagateFnc(), ConvolutionalComponent::PropagateFnc(), Rbm::RbmUpdate(), AveragePoolingComponent::ReadData(), MaxPoolingComponent::ReadData(), AffineTransform::ReadData(), LinearTransform::ReadData(), KlHmm::ReadData(), FramePoolingComponent::ReadData(), ConvolutionalComponent::ReadData(), Rbm::ReadData(), Rbm::Reconstruct(), ConvolutionalComponent::ReverseIndexes(), and KlHmm::SetStats().

◆ kMarkerMap

const struct key_value kMarkerMap[]
static

The table with pairs of Component types and markers (defined in nnet-component.cc),.

Definition at line 95 of file nnet-component.h.

Referenced by Component::MarkerToType(), and Component::TypeToMarker().

◆ output_dim_

int32 output_dim_
protected

Dimension of the output of the Component,.

Definition at line 191 of file nnet-component.h.

Referenced by RecurrentComponent::InitData(), ConvolutionalComponent::InitData(), Component::OutputDim(), Component::Propagate(), Rbm::RbmUpdate(), MaxPoolingComponent::ReadData(), AveragePoolingComponent::ReadData(), ParametricRelu::ReadData(), AffineTransform::ReadData(), LinearTransform::ReadData(), KlHmm::ReadData(), FramePoolingComponent::ReadData(), ConvolutionalComponent::ReadData(), Rbm::ReadData(), Rbm::Reconstruct(), and KlHmm::SetStats().

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