Package org.deeplearning4j.nn.conf.memory

Class LayerMemoryReport.Builder

java.lang.Object

org.deeplearning4j.nn.conf.memory.LayerMemoryReport.Builder

Enclosing class:

LayerMemoryReport

public static class LayerMemoryReport.Builder
extends Object

Constructor Summary

Constructors

Constructor	Description
Builder(String layerName, Class<?> layerType, InputType inputType, InputType outputType)

Method Summary

Modifier and Type	Method	Description
LayerMemoryReport	build()
LayerMemoryReport.Builder	cacheMemory(long cacheModeMemoryFixed, long cacheModeMemoryVariablePerEx)	Reports the cached/cacheable memory requirements.
LayerMemoryReport.Builder	cacheMemory(Map<CacheMode,Long> cacheModeMemoryFixed, Map<CacheMode,Long> cacheModeMemoryVariablePerEx)	Reports the cached/cacheable memory requirements.
LayerMemoryReport.Builder	standardMemory(long parameterSize, long updaterStateSize)	Report the standard memory
LayerMemoryReport.Builder	workingMemory(long fixedInference, long variableInferencePerEx, long fixedTrain, long variableTrainPerEx)	Report the working memory size, for both inference and training
LayerMemoryReport.Builder	workingMemory(long fixedInference, long variableInferencePerEx, Map<CacheMode,Long> fixedTrain, Map<CacheMode,Long> variableTrainPerEx)	Report the working memory requirements, for both inference and training.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

Builder

public Builder(String layerName,
               Class<?> layerType,
               InputType inputType,
               InputType outputType)

Parameters:

layerName - Name of the layer or graph vertex

layerType - Type of the layer or graph vertex

inputType - Input type to the layer/vertex

outputType - Output type from the layer/vertex

Method Detail

standardMemory

public LayerMemoryReport.Builder standardMemory(long parameterSize,
                                                long updaterStateSize)

Report the standard memory

Parameters:

parameterSize - Number of parameters

updaterStateSize - Size for the updater array

workingMemory

public LayerMemoryReport.Builder workingMemory(long fixedInference,
                                               long variableInferencePerEx,
                                               long fixedTrain,
                                               long variableTrainPerEx)

Report the working memory size, for both inference and training

Parameters:

fixedInference - Number of elements used for inference ( independent of minibatch size)

variableInferencePerEx - Number of elements used for inference, for each example

fixedTrain - Number of elements used for training (independent of minibatch size)

variableTrainPerEx - Number of elements used for training, for each example

workingMemory

public LayerMemoryReport.Builder workingMemory(long fixedInference,
                                               long variableInferencePerEx,
                                               Map<CacheMode,Long> fixedTrain,
                                               Map<CacheMode,Long> variableTrainPerEx)

Report the working memory requirements, for both inference and training. As noted in MemoryReport Working memory is memory That will be allocated in a ND4J workspace, or can be garbage collected at any points after the method returns.

Parameters:

fixedInference - Number of elements of working memory used for inference (independent of minibatch size)

variableInferencePerEx - Number of elements of working memory used for inference, for each example

fixedTrain - Number of elements of working memory used for training (independent of minibatch size), for each cache mode

variableTrainPerEx - Number of elements of working memory used for training, for each example, for each cache mode

cacheMemory

public LayerMemoryReport.Builder cacheMemory(long cacheModeMemoryFixed,
                                             long cacheModeMemoryVariablePerEx)

Reports the cached/cacheable memory requirements. This method assumes the caseload memory is the same for all cases, i.e., typically used with zeros (Layers that do not use caching)

Parameters:

cacheModeMemoryFixed - Number of elements of cache memory, independent of the mini batch size

cacheModeMemoryVariablePerEx - Number of elements of cache memory, for each example

cacheMemory

public LayerMemoryReport.Builder cacheMemory(Map<CacheMode,Long> cacheModeMemoryFixed,
                                             Map<CacheMode,Long> cacheModeMemoryVariablePerEx)

Reports the cached/cacheable memory requirements.

Parameters:

cacheModeMemoryFixed - Number of elements of cache memory, independent of the mini batch size

cacheModeMemoryVariablePerEx - Number of elements of cache memory, for each example

build

public LayerMemoryReport build()