org.nd4j.autodiff.util

Class SameDiffUtils

java.lang.Object

org.nd4j.autodiff.util.SameDiffUtils

public class SameDiffUtils
extends Object

Utilities for SameDiff training and inference

Constructor Summary

Constructors

Constructor and Description
SameDiffUtils()

Method Summary

Modifier and Type	Method and Description
static ExternalErrorsFunction	externalErrors(SameDiff sameDiff, Map<String,INDArray> externalGradients, SDVariable... inputs)
static ExternalErrorsFunction	externalErrors(SameDiff sameDiff, SDVariable[] inputs)
static List<INDArray>	getSingleOutput(List<Map<String,INDArray>> outputs, String output) Get a list of batch outputs for a single variable from a list of batch outputs for all variables
static SDVariable	reductionBroadcastableWithOrigShape(int origRank, int[] reduceDims, SDVariable toExpand) Add 1s as required to the array make an array possible to be broadcast with the original (pre-reduce) array.
static SDVariable	reductionBroadcastableWithOrigShape(SDVariable origInput, SDVariable axis, SDVariable toExpand)
static SDVariable	reductionShape(SDVariable shape, SDVariable axis, boolean keepDim)
static Map<String,INDArray>	stackOutputs(List<Map<String,INDArray>> outputs) Stack batch outputs, like an output from SameDiff.output(MultiDataSetIterator, String...)
static void	validateDifferentialFunctionSameDiff(SameDiff sameDiff, SDVariable function, DifferentialFunction op)

Methods inherited from class java.lang.Object

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

Constructor Detail

SameDiffUtils

public SameDiffUtils()

Method Detail

stackOutputs

public static Map<String,INDArray> stackOutputs(List<Map<String,INDArray>> outputs)

Stack batch outputs, like an output from SameDiff.output(MultiDataSetIterator, String...)

getSingleOutput

public static List<INDArray> getSingleOutput(List<Map<String,INDArray>> outputs,
                                             String output)

Get a list of batch outputs for a single variable from a list of batch outputs for all variables

externalErrors

public static ExternalErrorsFunction externalErrors(SameDiff sameDiff,
                                                    Map<String,INDArray> externalGradients,
                                                    SDVariable... inputs)

externalErrors

public static ExternalErrorsFunction externalErrors(SameDiff sameDiff,
                                                    SDVariable[] inputs)

reductionBroadcastableWithOrigShape

public static SDVariable reductionBroadcastableWithOrigShape(int origRank,
                                                             int[] reduceDims,
                                                             SDVariable toExpand)

Add 1s as required to the array make an array possible to be broadcast with the original (pre-reduce) array.

Example: if doing [a,b,c].sum(1), result is [a,c]. To 'undo' this in a way that can be auto-broadcast, we want to expand as required - i.e., [a,c] -> [a,1,c] which can be auto-broadcast with the original [a,b,c]. This is typically only used with reduction operations backprop.

Parameters:

origRank - Rank of the original array, before the reduction was executed

reduceDims - Dimensions that the original array was reduced from

toExpand - Array to add 1s to the shape to (such that it can be

Returns:

Reshaped array.

reductionBroadcastableWithOrigShape

public static SDVariable reductionBroadcastableWithOrigShape(SDVariable origInput,
                                                             SDVariable axis,
                                                             SDVariable toExpand)

reductionShape

public static SDVariable reductionShape(SDVariable shape,
                                        SDVariable axis,
                                        boolean keepDim)

validateDifferentialFunctionSameDiff

public static void validateDifferentialFunctionSameDiff(SameDiff sameDiff,
                                                        SDVariable function,
                                                        DifferentialFunction op)