SmartEngine::LossTrainingMethodInfo Struct Reference

#include <LossTrainer.h>

Public Attributes
LossTrainingMethod	method = LossTrainingMethod::Stochastic
	The method to use while training More...
const int64 *	sessionIndices = nullptr
	Session indices divide the graph input into logical chunks. These are row indicies in the input buffer. More...
int	sessionCount = 0
	The session index array count More...
int	batchSize = 32
	The number of rows to train at a time. More...
int	sequenceLength = 1
	How many sequences should be trained every step. Only applies to stepped trainable layers (such as LSTM). Effectively, this determines how long of a memory the layer (LSTM) has. More...

Member Data Documentation

batchSize

int SmartEngine::LossTrainingMethodInfo::batchSize = 32

The number of rows to train at a time.

method

LossTrainingMethod SmartEngine::LossTrainingMethodInfo::method = LossTrainingMethod::Stochastic

The method to use while training

sequenceLength

int SmartEngine::LossTrainingMethodInfo::sequenceLength = 1

How many sequences should be trained every step. Only applies to stepped trainable layers (such as LSTM). Effectively, this determines how long of a memory the layer (LSTM) has.

sessionCount

int SmartEngine::LossTrainingMethodInfo::sessionCount = 0

The session index array count

sessionIndices

const int64* SmartEngine::LossTrainingMethodInfo::sessionIndices = nullptr

Session indices divide the graph input into logical chunks. These are row indicies in the input buffer.

When training an LSTM, sequence runs won't go past session lines. When training stochastically, one row from each session will be batched together.