DiscriminativeOptions Struct Reference

#include <discriminative-training.h>

Collaboration diagram for DiscriminativeOptions:

Public Member Functions

 DiscriminativeOptions ()
void Register (OptionsItf *opts)

Public Attributes

std::string criterion
BaseFloat acoustic_scale
bool drop_frames
bool one_silence_class
BaseFloat boost
std::string silence_phones_str
BaseFloat xent_regularize
BaseFloat l2_regularize
bool accumulate_gradients
bool accumulate_output
int32 num_pdfs

Detailed Description

Definition at line 48 of file discriminative-training.h.

Constructor & Destructor Documentation

◆ DiscriminativeOptions()

Definition at line 83 of file discriminative-training.h.

Member Function Documentation

◆ Register()

void Register ( OptionsItf opts)

Definition at line 94 of file discriminative-training.h.

References OptionsItf::Register().

Referenced by main(), and NnetDiscriminativeOptions::Register().

94  {
95  opts->Register("criterion", &criterion, "Criterion, 'mmi'|'mpfe'|'smbr', "
96  "determines the objective function to use. Should match "
97  "option used when we created the examples.");
98  opts->Register("acoustic-scale", &acoustic_scale, "Weighting factor to "
99  "apply to acoustic likelihoods.");
100  opts->Register("drop-frames", &drop_frames, "For MMI, if true we drop frames "
101  "with no overlap of num and den pdf-ids");
102  opts->Register("boost", &boost, "Boosting factor for boosted MMI (e.g. 0.1)");
103  opts->Register("one-silence-class", &one_silence_class, "If true, newer "
104  "behavior which will tend to reduce insertions "
105  "when using MPFE or SMBR objective");
106  opts->Register("silence-phones", &silence_phones_str,
107  "For MPFE or SMBR objectives, colon-separated list of "
108  "integer ids of silence phones, e.g. 1:2:3");
109  opts->Register("l2-regularize", &l2_regularize, "l2 regularization "
110  "constant for 'chain' output "
111  "of the neural net.");
112  opts->Register("xent-regularize", &xent_regularize, "Cross-entropy "
113  "regularization constant for sequence training. If "
114  "nonzero, the network is expected to have an output "
115  "named 'output-xent', which should have a softmax as "
116  "its final nonlinearity.");
117  opts->Register("accumulate-gradients", &accumulate_gradients,
118  "Accumulate gradients wrt nnet output "
119  "for debugging discriminative training");
120  opts->Register("accumulate-output", &accumulate_output,
121  "Accumulate nnet output "
122  "for debugging discriminative training");
123  opts->Register("num-pdfs", &num_pdfs,
124  "Number of pdfs; "
125  "applicable when accumulate-output or accumulate-gradients "
126  "is true for discriminative training");
127  }

Member Data Documentation

◆ accumulate_gradients

bool accumulate_gradients

Definition at line 74 of file discriminative-training.h.

Referenced by DiscriminativeObjectiveInfo::Configure().

◆ accumulate_output

bool accumulate_output

Definition at line 77 of file discriminative-training.h.

Referenced by DiscriminativeObjectiveInfo::Configure().

◆ acoustic_scale

BaseFloat acoustic_scale

◆ boost

BaseFloat boost

Definition at line 56 of file discriminative-training.h.

Referenced by DiscriminativeComputation::Compute().

◆ criterion

◆ drop_frames

bool drop_frames

◆ l2_regularize

BaseFloat l2_regularize

Definition at line 69 of file discriminative-training.h.

Referenced by DiscriminativeComputation::Compute().

◆ num_pdfs

◆ one_silence_class

bool one_silence_class

◆ silence_phones_str

std::string silence_phones_str

◆ xent_regularize

The documentation for this struct was generated from the following file: