NnetCombineFastConfig Struct Reference

Configuration class that controls neural net combination, where we combine a number of neural nets, trying to find for each layer the optimal weighted combination of the different neural-net parameters. More...

#include <combine-nnet-fast.h>

Collaboration diagram for NnetCombineFastConfig:

Public Member Functions

 NnetCombineFastConfig ()
 
void Register (OptionsItf *opts)
 

Public Attributes

int32 initial_model
 
int32 num_lbfgs_iters
 
int32 num_threads
 
BaseFloat initial_impr
 
BaseFloat fisher_floor
 
BaseFloat alpha
 
int32 fisher_minibatch_size
 
int32 minibatch_size
 
int32 max_lbfgs_dim
 
BaseFloat regularizer
 

Detailed Description

Configuration class that controls neural net combination, where we combine a number of neural nets, trying to find for each layer the optimal weighted combination of the different neural-net parameters.

Definition at line 50 of file combine-nnet-fast.h.

Constructor & Destructor Documentation

◆ NnetCombineFastConfig()

Member Function Documentation

◆ Register()

void Register ( OptionsItf opts)
inline

Definition at line 73 of file combine-nnet-fast.h.

References kaldi::nnet2::CombineNnetsFast(), and OptionsItf::Register().

Referenced by main().

73  {
74  opts->Register("initial-model", &initial_model, "Specifies where to start the "
75  "optimization from. If 0 ... #models-1, then specifies the model; "
76  "if >= #models, then the average of all inputs; if <0, chosen "
77  "automatically from the previous options.");
78  opts->Register("num-lbfgs-iters", &num_lbfgs_iters, "Maximum number of function "
79  "evaluations for L-BFGS to use when optimizing combination weights");
80  opts->Register("initial-impr", &initial_impr, "Amount of objective-function change "
81  "We aim for on the first iteration.");
82  opts->Register("num-threads", &num_threads, "Number of threads to use in "
83  "multi-core computation");
84  opts->Register("fisher-floor", &fisher_floor,
85  "Floor for diagonal of Fisher matrix (used in preconditioning)");
86  opts->Register("alpha", &alpha, "Value we use in smoothing the Fisher matrix "
87  "with its diagonal, in preconditioning the update.");
88  opts->Register("fisher-minibatch-size", &fisher_minibatch_size, "Size of minibatch "
89  "used in computation of Fisher matrix (smaller -> better "
90  "preconditioning");
91  opts->Register("minibatch-size", &minibatch_size, "Minibatch size used in computing "
92  "gradients (only affects speed)");
93  opts->Register("max-lbfgs-dim", &max_lbfgs_dim, "Maximum dimension to use in "
94  "L-BFGS (will not get higher than this even if the dimension "
95  "of the space gets higher.)");
96  opts->Register("regularizer", &regularizer, "Add to the objective "
97  "function (which is average log-like per frame), -0.5 * "
98  "regularizer * square of parameters.");
99  }

Member Data Documentation

◆ alpha

BaseFloat alpha

Definition at line 59 of file combine-nnet-fast.h.

◆ fisher_floor

BaseFloat fisher_floor

Definition at line 56 of file combine-nnet-fast.h.

◆ fisher_minibatch_size

int32 fisher_minibatch_size

Definition at line 60 of file combine-nnet-fast.h.

◆ initial_impr

BaseFloat initial_impr

Definition at line 55 of file combine-nnet-fast.h.

◆ initial_model

int32 initial_model

Definition at line 51 of file combine-nnet-fast.h.

◆ max_lbfgs_dim

int32 max_lbfgs_dim

Definition at line 65 of file combine-nnet-fast.h.

◆ minibatch_size

int32 minibatch_size

Definition at line 63 of file combine-nnet-fast.h.

◆ num_lbfgs_iters

int32 num_lbfgs_iters

Definition at line 53 of file combine-nnet-fast.h.

◆ num_threads

int32 num_threads

Definition at line 54 of file combine-nnet-fast.h.

Referenced by main().

◆ regularizer

BaseFloat regularizer

Definition at line 66 of file combine-nnet-fast.h.


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