Package ai.djl.nn.norm

Class GhostBatchNorm

java.lang.Object

ai.djl.nn.AbstractBaseBlock

ai.djl.nn.AbstractBlock

ai.djl.nn.norm.BatchNorm

ai.djl.nn.norm.GhostBatchNorm

All Implemented Interfaces:

Block

public class GhostBatchNorm
extends BatchNorm

GhostBatchNorm is similar to BatchNorm except that it splits a batch into a smaller sub-batches aka ghost batches, and normalize them individually to have a mean of 0 and variance of 1 and finally concatenate them again to a single batch. Each of the mini-batches contains a virtualBatchSize samples.

See Also:

• Ghost Normalization Paper

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static class	GhostBatchNorm.Builder	The Builder to construct a GhostBatchNorm.

Nested classes/interfaces inherited from class ai.djl.nn.norm.BatchNorm

BatchNorm.BaseBuilder<T extends BatchNorm.BaseBuilder<T>>

Field Summary

Fields inherited from class ai.djl.nn.AbstractBlock

children, parameters

Fields inherited from class ai.djl.nn.AbstractBaseBlock

inputNames, inputShapes, outputDataTypes, version

Constructor Summary

Constructors

Modifier	Constructor	Description
protected	GhostBatchNorm(GhostBatchNorm.Builder builder)

Method Summary

Modifier and Type	Method	Description
protected NDList	batchify(NDList[] subBatches)	Converts an array of NDList into an NDList using StackBatchifier and squeezes the first dimension created by it.
static GhostBatchNorm.Builder	builder()	Creates a builder to build a GhostBatchNorm.
protected NDList	forwardInternal(ParameterStore parameterStore, NDList inputs, boolean training, ai.djl.util.PairList<String,Object> params)	A helper for Block.forward(ParameterStore, NDList, boolean, PairList) after initialization.
protected NDList[]	split(NDList list)	Splits an NDList into the given size of sub-batch.
protected NDList	squeezeExtraDimensions(NDList batch)	Squeezes first axes of NDList.

Methods inherited from class ai.djl.nn.norm.BatchNorm

batchNorm, batchNorm, batchNorm, batchNorm, beforeInitialize, getOutputShapes, loadMetadata, prepare, saveMetadata

Methods inherited from class ai.djl.nn.AbstractBlock

addChildBlock, addChildBlock, addChildBlockSingleton, addParameter, getChildren, getDirectParameters

Methods inherited from class ai.djl.nn.AbstractBaseBlock

cast, clear, describeInput, forward, forward, forwardInternal, getInputShapes, getOutputDataTypes, getParameters, initialize, initializeChildBlocks, isInitialized, loadParameters, readInputShapes, saveInputShapes, saveParameters, setInitializer, setInitializer, setInitializer, toString

Methods inherited from class java.lang.Object

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

Methods inherited from interface ai.djl.nn.Block

forward, freezeParameters, freezeParameters, getOutputShapes

Constructor Details

GhostBatchNorm

protected GhostBatchNorm(GhostBatchNorm.Builder builder)

Method Details

forwardInternal

protected NDList forwardInternal(ParameterStore parameterStore,
 NDList inputs,
 boolean training,
 ai.djl.util.PairList<String,Object> params)

A helper for Block.forward(ParameterStore, NDList, boolean, PairList) after initialization.

Overrides:

forwardInternal in class BatchNorm

Parameters:

parameterStore - the parameter store

inputs - the input NDList

training - true for a training forward pass

params - optional parameters

Returns:

the output of the forward pass

split

protected NDList[] split(NDList list)

Splits an NDList into the given size of sub-batch.

This function unbatchifies the input NDList into mini-batches, each with the size of virtualBatchSize. If the batch size is divisible by the virtual batch size, all returned sub-batches will be the same size. If the batch size is not divisible by virtual batch size, all returned sub-batches will be the same size, except the last one.

Parameters:

list - the NDList that needs to be split

Returns:

an array of NDList that contains all the mini-batches

batchify

protected NDList batchify(NDList[] subBatches)

Converts an array of NDList into an NDList using StackBatchifier and squeezes the first dimension created by it. This makes the final NDArray same size as the splitted one.

Parameters:

subBatches - the input array of NDList

Returns:

the batchified NDList

squeezeExtraDimensions

protected NDList squeezeExtraDimensions(NDList batch)

Squeezes first axes of NDList.

Parameters:

batch - input array of NDList

Returns:

the squeezed NDList

builder

public static GhostBatchNorm.Builder builder()

Creates a builder to build a GhostBatchNorm.

Returns:

a new builder