#include <nnet-discriminative-training.h>

Collaboration diagram for DiscriminativeObjectiveFunctionInfo:

[legend]

Public Member Functions
	DiscriminativeObjectiveFunctionInfo ()

void	UpdateStats (const std::string &output_name, const std::string &criterion, int32 minibatches_per_phase, int32 minibatch_counter, discriminative::DiscriminativeObjectiveInfo stats)

void	PrintStatsForThisPhase (const std::string &output_name, const std::string &criterion, int32 minibatches_per_phase) const

bool	PrintTotalStats (const std::string &output_name, const std::string &criterion) const

Public Attributes
int32	current_phase

discriminative::DiscriminativeObjectiveInfo	stats

discriminative::DiscriminativeObjectiveInfo	stats_this_phase

Detailed Description

Definition at line 55 of file nnet-discriminative-training.h.

Constructor & Destructor Documentation

◆ DiscriminativeObjectiveFunctionInfo()

DiscriminativeObjectiveFunctionInfo ( )

inline

Definition at line 61 of file nnet-discriminative-training.h.

61 :

62 current_phase(0) { }

kaldi::nnet3::DiscriminativeObjectiveFunctionInfo::current_phase

int32 current_phase

Definition: nnet-discriminative-training.h:56

Member Function Documentation

◆ PrintStatsForThisPhase()

void PrintStatsForThisPhase	(	const std::string &	output_name,
		const std::string &	criterion,
		int32	minibatches_per_phase
	)		const

Definition at line 225 of file nnet-discriminative-training.cc.

References KALDI_LOG.

                                        {
   int32 start_minibatch = current_phase * minibatches_per_phase,
       end_minibatch = start_minibatch + minibatches_per_phase - 1;
 
   BaseFloat objf = (stats_this_phase.TotalObjf(criterion) / stats_this_phase.tot_t_weighted);
   KALDI_LOG << "Average objective function for '" << output_name
             << "' for minibatches " << start_minibatch
             << '-' << end_minibatch << " is " << objf
             << " over " << stats_this_phase.tot_t_weighted << " frames.";
 }

◆ PrintTotalStats()

bool PrintTotalStats	(	const std::string &	output_name,
		const std::string &	criterion
	)		const

Definition at line 239 of file nnet-discriminative-training.cc.

References KALDI_LOG.

Referenced by NnetDiscriminativeTrainer::PrintTotalStats().

                                                   {
   BaseFloat objf = stats.TotalObjf(criterion) /stats.tot_t_weighted;
 
   double avg_gradients = (stats.tot_num_count + stats.tot_den_count) /
                          stats.tot_t_weighted;
   KALDI_LOG << "Average num+den count of stats is " << avg_gradients
               << " per frame, over "
               << stats.tot_t_weighted << " frames.";
   if (stats.tot_l2_term != 0.0) {
     KALDI_LOG << "Average l2 norm of output per frame is "
               << (stats.tot_l2_term / stats.tot_t_weighted) << " over "
               << stats.tot_t_weighted << " frames.";
   }
 
 
   KALDI_LOG << "Overall average objective function for '" << name << "' is "
             << objf << " over " << stats.tot_t_weighted << " frames.";
   KALDI_LOG << "[this line is to be parsed by a script:] "
             << criterion << "-per-frame="
             << objf;
   return (stats.tot_t_weighted != 0.0);
 }

◆ UpdateStats()

void UpdateStats	(	const std::string &	output_name,
		const std::string &	criterion,
		int32	minibatches_per_phase,
		int32	minibatch_counter,
		discriminative::DiscriminativeObjectiveInfo	stats
	)

Definition at line 208 of file nnet-discriminative-training.cc.

References KALDI_ASSERT.

                                                                     {
   int32 phase = minibatch_counter / minibatches_per_phase;
   if (phase != current_phase) {
     KALDI_ASSERT(phase == current_phase + 1); // or doesn't really make sense.
     PrintStatsForThisPhase(output_name, criterion, minibatches_per_phase);
     current_phase = phase;
     stats_this_phase.Reset();
   }
   stats_this_phase.Add(this_minibatch_stats);
   stats.Add(this_minibatch_stats);
 }