apple.mlcompute

Class MLCTrainingGraph

java.lang.Object

org.moe.natj.general.NativeObject

org.moe.natj.objc.ObjCObject

apple.NSObject

apple.mlcompute.MLCGraph

apple.mlcompute.MLCTrainingGraph

All Implemented Interfaces:

NSObject

public class MLCTrainingGraph
extends MLCGraph

MLCTrainingGraph

A training graph created from one or more MLCGraph objects plus additional layers added directly to the training graph.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static interface	MLCTrainingGraph.Block_executeForwardWithBatchSizeOptionsCompletionHandler
static interface	MLCTrainingGraph.Block_executeForwardWithBatchSizeOptionsOutputsDataCompletionHandler
static interface	MLCTrainingGraph.Block_executeGradientWithBatchSizeOptionsCompletionHandler
static interface	MLCTrainingGraph.Block_executeGradientWithBatchSizeOptionsOutputsDataCompletionHandler
static interface	MLCTrainingGraph.Block_executeOptimizerUpdateWithOptionsCompletionHandler
static interface	MLCTrainingGraph.Block_executeWithInputsDataLossLabelsDataLossLabelWeightsDataBatchSizeOptionsCompletionHandler
static interface	MLCTrainingGraph.Block_executeWithInputsDataLossLabelsDataLossLabelWeightsDataOutputsDataBatchSizeOptionsCompletionHandler

Nested classes/interfaces inherited from class apple.NSObject

NSObject.Function_instanceMethodForSelector_ret, NSObject.Function_methodForSelector_ret

Constructor Summary

Constructors

Modifier	Constructor and Description
protected	MLCTrainingGraph(org.moe.natj.general.Pointer peer)

Method Summary

Modifier and Type	Method and Description
static boolean	accessInstanceVariablesDirectly()
boolean	addInputsLossLabels(NSDictionary<java.lang.String,? extends MLCTensor> inputs, NSDictionary<java.lang.String,? extends MLCTensor> lossLabels) Add the list of inputs to the training graph
boolean	addInputsLossLabelsLossLabelWeights(NSDictionary<java.lang.String,? extends MLCTensor> inputs, NSDictionary<java.lang.String,? extends MLCTensor> lossLabels, NSDictionary<java.lang.String,? extends MLCTensor> lossLabelWeights) Add the list of inputs to the training graph
boolean	addOutputs(NSDictionary<java.lang.String,? extends MLCTensor> outputs) Add the list of outputs to the training graph
static MLCTrainingGraph	alloc()
MLCTensor	allocateUserGradientForTensor(MLCTensor tensor) Allocate an entry for a user specified gradient for a tensor
static MLCTrainingGraph	allocWithZone(org.moe.natj.general.ptr.VoidPtr zone)
static boolean	automaticallyNotifiesObserversForKey(java.lang.String key)
boolean	bindOptimizerDataDeviceDataWithTensor(NSArray<? extends MLCTensorData> data, NSArray<? extends MLCTensorOptimizerDeviceData> deviceData, MLCTensor tensor) Associates the given optimizer data and device data buffers with the tensor.
static void	cancelPreviousPerformRequestsWithTarget(java.lang.Object aTarget)
static void	cancelPreviousPerformRequestsWithTargetSelectorObject(java.lang.Object aTarget, org.moe.natj.objc.SEL aSelector, java.lang.Object anArgument)
static NSArray<java.lang.String>	classFallbacksForKeyedArchiver()
static org.moe.natj.objc.Class	classForKeyedUnarchiver()
boolean	compileOptimizer(MLCOptimizer optimizer) Compile the optimizer to be used with a training graph.
boolean	compileWithOptionsDevice(long options, MLCDevice device) Compile the training graph for a device.
boolean	compileWithOptionsDeviceInputTensorsInputTensorsData(long options, MLCDevice device, NSDictionary<java.lang.String,? extends MLCTensor> inputTensors, NSDictionary<java.lang.String,? extends MLCTensorData> inputTensorsData) Compile the training graph for a device.
static java.lang.String	debugDescription_static()
static java.lang.String	description_static()
long	deviceMemorySize() [@property] The device memory size used by the training graph
boolean	executeForwardWithBatchSizeOptionsCompletionHandler(long batchSize, long options, MLCTrainingGraph.Block_executeForwardWithBatchSizeOptionsCompletionHandler completionHandler) Execute the forward pass of the training graph
boolean	executeForwardWithBatchSizeOptionsOutputsDataCompletionHandler(long batchSize, long options, NSDictionary<java.lang.String,? extends MLCTensorData> outputsData, MLCTrainingGraph.Block_executeForwardWithBatchSizeOptionsOutputsDataCompletionHandler completionHandler) Execute the forward pass for the training graph
boolean	executeGradientWithBatchSizeOptionsCompletionHandler(long batchSize, long options, MLCTrainingGraph.Block_executeGradientWithBatchSizeOptionsCompletionHandler completionHandler) Execute the gradient pass of the training graph
boolean	executeGradientWithBatchSizeOptionsOutputsDataCompletionHandler(long batchSize, long options, NSDictionary<java.lang.String,? extends MLCTensorData> outputsData, MLCTrainingGraph.Block_executeGradientWithBatchSizeOptionsOutputsDataCompletionHandler completionHandler) Execute the gradient pass of the training graph
boolean	executeOptimizerUpdateWithOptionsCompletionHandler(long options, MLCTrainingGraph.Block_executeOptimizerUpdateWithOptionsCompletionHandler completionHandler) Execute the optimizer update pass of the training graph
boolean	executeWithInputsDataLossLabelsDataLossLabelWeightsDataBatchSizeOptionsCompletionHandler(NSDictionary<java.lang.String,? extends MLCTensorData> inputsData, NSDictionary<java.lang.String,? extends MLCTensorData> lossLabelsData, NSDictionary<java.lang.String,? extends MLCTensorData> lossLabelWeightsData, long batchSize, long options, MLCTrainingGraph.Block_executeWithInputsDataLossLabelsDataLossLabelWeightsDataBatchSizeOptionsCompletionHandler completionHandler) Execute the training graph (forward, gradient and optimizer update) with given source and label data
boolean	executeWithInputsDataLossLabelsDataLossLabelWeightsDataOutputsDataBatchSizeOptionsCompletionHandler(NSDictionary<java.lang.String,? extends MLCTensorData> inputsData, NSDictionary<java.lang.String,? extends MLCTensorData> lossLabelsData, NSDictionary<java.lang.String,? extends MLCTensorData> lossLabelWeightsData, NSDictionary<java.lang.String,? extends MLCTensorData> outputsData, long batchSize, long options, MLCTrainingGraph.Block_executeWithInputsDataLossLabelsDataLossLabelWeightsDataOutputsDataBatchSizeOptionsCompletionHandler completionHandler) Execute the training graph (forward, gradient and optimizer update) with given source and label data
NSData	gradientDataForParameterLayer(MLCTensor parameter, MLCLayer layer) Get the gradient data for a trainable parameter associated with a layer
MLCTensor	gradientTensorForInput(MLCTensor input) Get the gradient tensor for an input tensor
static MLCTrainingGraph	graph()
static MLCTrainingGraph	graphWithGraphObjectsLossLayerOptimizer(NSArray<? extends MLCGraph> graphObjects, MLCLayer lossLayer, MLCOptimizer optimizer) Create a training graph
static long	hash_static()
MLCTrainingGraph	init()
static NSObject.Function_instanceMethodForSelector_ret	instanceMethodForSelector(org.moe.natj.objc.SEL aSelector)
static NSMethodSignature	instanceMethodSignatureForSelector(org.moe.natj.objc.SEL aSelector)
static boolean	instancesRespondToSelector(org.moe.natj.objc.SEL aSelector)
static boolean	isSubclassOfClass(org.moe.natj.objc.Class aClass)
static NSSet<java.lang.String>	keyPathsForValuesAffectingValueForKey(java.lang.String key)
boolean	linkWithGraphs(NSArray<? extends MLCTrainingGraph> graphs) Link mutiple training graphs
static MLCTrainingGraph	new_objc()
MLCOptimizer	optimizer() [@property] optimizer
static boolean	resolveClassMethod(org.moe.natj.objc.SEL sel)
static boolean	resolveInstanceMethod(org.moe.natj.objc.SEL sel)
NSArray<? extends MLCTensor>	resultGradientTensorsForLayer(MLCLayer layer) Get the result gradient tensors for a layer in the training graph
boolean	setTrainingTensorParameters(NSArray<? extends MLCTensorParameter> parameters) Set the input tensor parameters that also will be updated by the optimizer
static void	setVersion_static(long aVersion)
NSArray<? extends MLCTensor>	sourceGradientTensorsForLayer(MLCLayer layer) Get the source gradient tensors for a layer in the training graph
boolean	stopGradientForTensors(NSArray<? extends MLCTensor> tensors) Add the list of tensors whose contributions are not to be taken when computing gradients during gradient pass
static org.moe.natj.objc.Class	superclass_static()
void	synchronizeUpdates() Synchronize updates (weights/biases from convolution, fully connected and LSTM layers, tensor parameters) from device memory to host memory.
static long	version_static()

Methods inherited from class apple.mlcompute.MLCGraph

bindAndWriteDataForInputsToDeviceBatchSizeSynchronous, bindAndWriteDataForInputsToDeviceSynchronous, concatenateWithSourcesDimension, device, gatherWithDimensionSourceIndices, layers, nodeWithLayerSource, nodeWithLayerSources, nodeWithLayerSourcesDisableUpdate, nodeWithLayerSourcesLossLabels, reshapeWithShapeSource, resultTensorsForLayer, scatterWithDimensionSourceIndicesCopyFromReductionType, selectWithSourcesCondition, sourceTensorsForLayer, splitWithSourceSplitCountDimension, splitWithSourceSplitSectionLengthsDimension, summarizedDOTDescription, transposeWithDimensionsSource

Methods inherited from class apple.NSObject

accessibilityActivate, accessibilityActivationPoint, accessibilityAssistiveTechnologyFocusedIdentifiers, accessibilityAttributedHint, accessibilityAttributedLabel, accessibilityAttributedUserInputLabels, accessibilityAttributedValue, accessibilityContainerType, accessibilityCustomActions, accessibilityCustomRotors, accessibilityDecrement, accessibilityDragSourceDescriptors, accessibilityDropPointDescriptors, accessibilityElementAtIndex, accessibilityElementCount, accessibilityElementDidBecomeFocused, accessibilityElementDidLoseFocus, accessibilityElementIsFocused, accessibilityElements, accessibilityElementsHidden, accessibilityFrame, accessibilityHint, accessibilityIncrement, accessibilityLabel, accessibilityLanguage, accessibilityNavigationStyle, accessibilityPath, accessibilityPerformEscape, accessibilityPerformMagicTap, accessibilityRespondsToUserInteraction, accessibilityScroll, accessibilityTextualContext, accessibilityTraits, accessibilityUserInputLabels, accessibilityValue, accessibilityViewIsModal, addObserverForKeyPathOptionsContext, attemptRecoveryFromErrorOptionIndex, attemptRecoveryFromErrorOptionIndexDelegateDidRecoverSelectorContextInfo, autoContentAccessingProxy, awakeAfterUsingCoder, awakeFromNib, class_objc, classForCoder, classForKeyedArchiver, copy, dealloc, debugDescription, description, dictionaryWithValuesForKeys, didChangeValueForKey, didChangeValueForKeyWithSetMutationUsingObjects, didChangeValuesAtIndexesForKey, doesNotRecognizeSelector, fileManagerShouldProceedAfterError, fileManagerWillProcessPath, finalize_objc, forwardingTargetForSelector, forwardInvocation, hash, indexOfAccessibilityElement, isAccessibilityElement, isEqual, isKindOfClass, isMemberOfClass, isProxy, methodForSelector, methodSignatureForSelector, mutableArrayValueForKey, mutableArrayValueForKeyPath, mutableCopy, mutableOrderedSetValueForKey, mutableOrderedSetValueForKeyPath, mutableSetValueForKey, mutableSetValueForKeyPath, observationInfo, observeValueForKeyPathOfObjectChangeContext, performSelector, performSelectorInBackgroundWithObject, performSelectorOnMainThreadWithObjectWaitUntilDone, performSelectorOnMainThreadWithObjectWaitUntilDoneModes, performSelectorOnThreadWithObjectWaitUntilDone, performSelectorOnThreadWithObjectWaitUntilDoneModes, performSelectorWithObject, performSelectorWithObjectAfterDelay, performSelectorWithObjectAfterDelayInModes, performSelectorWithObjectWithObject, prepareForInterfaceBuilder, provideImageDataBytesPerRowOrigin_Size_UserInfo, removeObserverForKeyPath, removeObserverForKeyPathContext, replacementObjectForCoder, replacementObjectForKeyedArchiver, respondsToSelector, self, setAccessibilityActivationPoint, setAccessibilityAttributedHint, setAccessibilityAttributedLabel, setAccessibilityAttributedUserInputLabels, setAccessibilityAttributedValue, setAccessibilityContainerType, setAccessibilityCustomActions, setAccessibilityCustomRotors, setAccessibilityDragSourceDescriptors, setAccessibilityDropPointDescriptors, setAccessibilityElements, setAccessibilityElementsHidden, setAccessibilityFrame, setAccessibilityHint, setAccessibilityLabel, setAccessibilityLanguage, setAccessibilityNavigationStyle, setAccessibilityPath, setAccessibilityRespondsToUserInteraction, setAccessibilityTextualContext, setAccessibilityTraits, setAccessibilityUserInputLabels, setAccessibilityValue, setAccessibilityViewIsModal, setIsAccessibilityElement, setNilValueForKey, setObservationInfo, setShouldGroupAccessibilityChildren, setValueForKey, setValueForKeyPath, setValueForUndefinedKey, setValuesForKeysWithDictionary, shouldGroupAccessibilityChildren, superclass, validateValueForKeyError, validateValueForKeyPathError, valueForKey, valueForKeyPath, valueForUndefinedKey, willChangeValueForKey, willChangeValueForKeyWithSetMutationUsingObjects, willChangeValuesAtIndexesForKey

Methods inherited from class org.moe.natj.objc.ObjCObject

equals, toString

Methods inherited from class org.moe.natj.general.NativeObject

finalize, getPeer, getPeerPointer

Methods inherited from class java.lang.Object

clone, getClass, hashCode, notify, notifyAll, wait, wait, wait

Constructor Detail

MLCTrainingGraph

protected MLCTrainingGraph(org.moe.natj.general.Pointer peer)

Method Detail

accessInstanceVariablesDirectly

public static boolean accessInstanceVariablesDirectly()

addInputsLossLabels

public boolean addInputsLossLabels(NSDictionary<java.lang.String,? extends MLCTensor> inputs,
                                   NSDictionary<java.lang.String,? extends MLCTensor> lossLabels)

Add the list of inputs to the training graph

Parameters:

inputs - The inputs

lossLabels - The loss label inputs

Returns:

A boolean indicating success or failure

addInputsLossLabelsLossLabelWeights

public boolean addInputsLossLabelsLossLabelWeights(NSDictionary<java.lang.String,? extends MLCTensor> inputs,
                                                   NSDictionary<java.lang.String,? extends MLCTensor> lossLabels,
                                                   NSDictionary<java.lang.String,? extends MLCTensor> lossLabelWeights)

Add the list of inputs to the training graph

Each input, loss label or label weights tensor is identified by a NSString. When the training graph is executed, this NSString is used to identify which data object should be as input data for each tensor whose device memory needs to be updated before the graph is executed.

Parameters:

inputs - The inputs

lossLabels - The loss label inputs

lossLabelWeights - The loss label weights

Returns:

A boolean indicating success or failure

addOutputs

public boolean addOutputs(NSDictionary<java.lang.String,? extends MLCTensor> outputs)

Add the list of outputs to the training graph

Parameters:

outputs - The outputs

Returns:

A boolean indicating success or failure

alloc

public static MLCTrainingGraph alloc()

allocWithZone

public static MLCTrainingGraph allocWithZone(org.moe.natj.general.ptr.VoidPtr zone)

allocateUserGradientForTensor

public MLCTensor allocateUserGradientForTensor(MLCTensor tensor)

Allocate an entry for a user specified gradient for a tensor

Parameters:

tensor - A result tensor produced by a layer in the training graph that is input to some user specified code and will need to provide a user gradient during the gradient pass.

Returns:

A gradient tensor

automaticallyNotifiesObserversForKey

public static boolean automaticallyNotifiesObserversForKey(java.lang.String key)

bindOptimizerDataDeviceDataWithTensor

public boolean bindOptimizerDataDeviceDataWithTensor(NSArray<? extends MLCTensorData> data,
                                                     NSArray<? extends MLCTensorOptimizerDeviceData> deviceData,
                                                     MLCTensor tensor)

Associates the given optimizer data and device data buffers with the tensor. Returns true if the data is successfully associated with the tensor and copied to the device.

The caller must guarantee the lifetime of the underlying memory of \p data for the entirety of the tensor's lifetime. The \p deviceData buffers are allocated by MLCompute. This method must be called before executeOptimizerUpdateWithOptions or executeWithInputsData is called for the training graph. We recommend using this method instead of using [MLCTensor bindOptimizerData] especially if the optimizer update is being called multiple times for each batch.

Parameters:

data - The optimizer data to be associated with the tensor

deviceData - The optimizer device data to be associated with the tensor

tensor - The tensor

Returns:

A Boolean value indicating whether the data is successfully associated with the tensor .

cancelPreviousPerformRequestsWithTarget

public static void cancelPreviousPerformRequestsWithTarget(java.lang.Object aTarget)

cancelPreviousPerformRequestsWithTargetSelectorObject

public static void cancelPreviousPerformRequestsWithTargetSelectorObject(java.lang.Object aTarget,
                                                                         org.moe.natj.objc.SEL aSelector,
                                                                         java.lang.Object anArgument)

classFallbacksForKeyedArchiver

public static NSArray<java.lang.String> classFallbacksForKeyedArchiver()

classForKeyedUnarchiver

public static org.moe.natj.objc.Class classForKeyedUnarchiver()

compileOptimizer

public boolean compileOptimizer(MLCOptimizer optimizer)

Compile the optimizer to be used with a training graph.

Typically the optimizer to be used with a training graph is specifed when the training graph is created using graphWithGraphObjects:lossLayer:optimizer. The optimizer will be compiled in when compileWithOptions:device is called if an optimizer is specified with the training graph. In the case where the optimizer to be used is not known when the graph is created or compiled, this method can be used to associate and compile a training graph with an optimizer.

Parameters:

optimizer - The MLCOptimizer object

Returns:

A boolean indicating success or failure

compileWithOptionsDevice

public boolean compileWithOptionsDevice(long options,
                                        MLCDevice device)

Compile the training graph for a device.

Parameters:

options - The compiler options to use when compiling the training graph

device - The MLCDevice object

Returns:

A boolean indicating success or failure

compileWithOptionsDeviceInputTensorsInputTensorsData

public boolean compileWithOptionsDeviceInputTensorsInputTensorsData(long options,
                                                                    MLCDevice device,
                                                                    NSDictionary<java.lang.String,? extends MLCTensor> inputTensors,
                                                                    NSDictionary<java.lang.String,? extends MLCTensorData> inputTensorsData)

Compile the training graph for a device.

Specifying the list of constant tensors when we compile the graph allows MLCompute to perform additional optimizations at compile time.

Parameters:

options - The compiler options to use when compiling the training graph

device - The MLCDevice object

inputTensors - The list of input tensors that are constants

inputTensorsData - The tensor data to be used with these constant input tensors

Returns:

A boolean indicating success or failure

debugDescription_static

public static java.lang.String debugDescription_static()

description_static

public static java.lang.String description_static()

deviceMemorySize

public long deviceMemorySize()

[@property] The device memory size used by the training graph

Returns the total size in bytes of device memory used for all intermediate tensors for forward, gradient passes and optimizer update for all layers in the training graph. We recommend executing an iteration before checking the device memory size as the buffers needed get allocated when the corresponding pass such as gradient, optimizer update is executed.

Returns:

A NSUInteger value

executeForwardWithBatchSizeOptionsCompletionHandler

public boolean executeForwardWithBatchSizeOptionsCompletionHandler(long batchSize,
                                                                   long options,
                                                                   MLCTrainingGraph.Block_executeForwardWithBatchSizeOptionsCompletionHandler completionHandler)

Execute the forward pass of the training graph

Parameters:

batchSize - The batch size to use. For a graph where batch size changes between layers this value must be 0.

options - The execution options

completionHandler - The completion handler

Returns:

A boolean indicating success or failure

executeForwardWithBatchSizeOptionsOutputsDataCompletionHandler

public boolean executeForwardWithBatchSizeOptionsOutputsDataCompletionHandler(long batchSize,
                                                                              long options,
                                                                              NSDictionary<java.lang.String,? extends MLCTensorData> outputsData,
                                                                              MLCTrainingGraph.Block_executeForwardWithBatchSizeOptionsOutputsDataCompletionHandler completionHandler)

Execute the forward pass for the training graph

Parameters:

batchSize - The batch size to use. For a graph where batch size changes between layers this value must be 0.

options - The execution options

outputsData - The data objects to use for outputs

completionHandler - The completion handler

Returns:

A boolean indicating success or failure

executeGradientWithBatchSizeOptionsCompletionHandler

public boolean executeGradientWithBatchSizeOptionsCompletionHandler(long batchSize,
                                                                    long options,
                                                                    MLCTrainingGraph.Block_executeGradientWithBatchSizeOptionsCompletionHandler completionHandler)

Execute the gradient pass of the training graph

Parameters:

batchSize - The batch size to use. For a graph where batch size changes between layers this value must be 0.

options - The execution options

completionHandler - The completion handler

Returns:

A boolean indicating success or failure

executeGradientWithBatchSizeOptionsOutputsDataCompletionHandler

public boolean executeGradientWithBatchSizeOptionsOutputsDataCompletionHandler(long batchSize,
                                                                               long options,
                                                                               NSDictionary<java.lang.String,? extends MLCTensorData> outputsData,
                                                                               MLCTrainingGraph.Block_executeGradientWithBatchSizeOptionsOutputsDataCompletionHandler completionHandler)

Execute the gradient pass of the training graph

Parameters:

batchSize - The batch size to use. For a graph where batch size changes between layers this value must be 0.

options - The execution options

outputsData - The data objects to use for outputs

completionHandler - The completion handler

Returns:

A boolean indicating success or failure

executeOptimizerUpdateWithOptionsCompletionHandler

public boolean executeOptimizerUpdateWithOptionsCompletionHandler(long options,
                                                                  MLCTrainingGraph.Block_executeOptimizerUpdateWithOptionsCompletionHandler completionHandler)

Execute the optimizer update pass of the training graph

Parameters:

options - The execution options

completionHandler - The completion handler

Returns:

A boolean indicating success or failure

executeWithInputsDataLossLabelsDataLossLabelWeightsDataBatchSizeOptionsCompletionHandler

public boolean executeWithInputsDataLossLabelsDataLossLabelWeightsDataBatchSizeOptionsCompletionHandler(NSDictionary<java.lang.String,? extends MLCTensorData> inputsData,
                                                                                                        NSDictionary<java.lang.String,? extends MLCTensorData> lossLabelsData,
                                                                                                        NSDictionary<java.lang.String,? extends MLCTensorData> lossLabelWeightsData,
                                                                                                        long batchSize,
                                                                                                        long options,
                                                                                                        MLCTrainingGraph.Block_executeWithInputsDataLossLabelsDataLossLabelWeightsDataBatchSizeOptionsCompletionHandler completionHandler)

Execute the training graph (forward, gradient and optimizer update) with given source and label data

Execute the training graph with given source and label data. If an optimizer is specified, the optimizer update is applied. If MLCExecutionOptionsSynchronous is specified in 'options', this method returns after the graph has been executed. Otherwise, this method returns after the graph has been queued for execution. The completion handler is called after the graph has finished execution.

Parameters:

inputsData - The data objects to use for inputs

lossLabelsData - The data objects to use for loss labels

lossLabelWeightsData - The data objects to use for loss label weights

batchSize - The batch size to use. For a graph where batch size changes between layers this value must be 0.

options - The execution options

completionHandler - The completion handler

Returns:

A boolean indicating success or failure

executeWithInputsDataLossLabelsDataLossLabelWeightsDataOutputsDataBatchSizeOptionsCompletionHandler

public boolean executeWithInputsDataLossLabelsDataLossLabelWeightsDataOutputsDataBatchSizeOptionsCompletionHandler(NSDictionary<java.lang.String,? extends MLCTensorData> inputsData,
                                                                                                                   NSDictionary<java.lang.String,? extends MLCTensorData> lossLabelsData,
                                                                                                                   NSDictionary<java.lang.String,? extends MLCTensorData> lossLabelWeightsData,
                                                                                                                   NSDictionary<java.lang.String,? extends MLCTensorData> outputsData,
                                                                                                                   long batchSize,
                                                                                                                   long options,
                                                                                                                   MLCTrainingGraph.Block_executeWithInputsDataLossLabelsDataLossLabelWeightsDataOutputsDataBatchSizeOptionsCompletionHandler completionHandler)

Execute the training graph (forward, gradient and optimizer update) with given source and label data

Parameters:

inputsData - The data objects to use for inputs

lossLabelsData - The data objects to use for loss labels

lossLabelWeightsData - The data objects to use for loss label weights

outputsData - The data objects to use for outputs

batchSize - The batch size to use. For a graph where batch size changes between layers this value must be 0.

options - The execution options

completionHandler - The completion handler

Returns:

A boolean indicating success or failure

gradientDataForParameterLayer

public NSData gradientDataForParameterLayer(MLCTensor parameter,
                                            MLCLayer layer)

Get the gradient data for a trainable parameter associated with a layer

This can be used to get the gradient data for weights or biases parameters associated with a convolution, fully connected or convolution transpose layer

Parameters:

parameter - The updatable parameter associated with the layer

layer - A layer in the training graph. Must be one of the following: - MLCConvolutionLayer - MLCFullyConnectedLayer - MLCBatchNormalizationLayer - MLCInstanceNormalizationLayer - MLCGroupNormalizationLayer - MLCLayerNormalizationLayer - MLCEmbeddingLayer - MLCMultiheadAttentionLayer

Returns:

The gradient data. Will return nil if the layer is marked as not trainable or if training graph is not executed with separate calls to forward and gradient passes.

gradientTensorForInput

public MLCTensor gradientTensorForInput(MLCTensor input)

Get the gradient tensor for an input tensor

Parameters:

input - The input tensor

Returns:

The gradient tensor

graph

public static MLCTrainingGraph graph()

graphWithGraphObjectsLossLayerOptimizer

public static MLCTrainingGraph graphWithGraphObjectsLossLayerOptimizer(NSArray<? extends MLCGraph> graphObjects,
                                                                       MLCLayer lossLayer,
                                                                       MLCOptimizer optimizer)

Create a training graph

Parameters:

graphObjects - The layers from these graph objects will be added to the training graph

lossLayer - The loss layer to use. The loss layer can also be added to the training graph using nodeWithLayer:sources:lossLabels

optimizer - The optimizer to use

Returns:

A new training graph object

hash_static

public static long hash_static()

init

public MLCTrainingGraph init()

Overrides:

init in class MLCGraph

instanceMethodForSelector

public static NSObject.Function_instanceMethodForSelector_ret instanceMethodForSelector(org.moe.natj.objc.SEL aSelector)

instanceMethodSignatureForSelector

public static NSMethodSignature instanceMethodSignatureForSelector(org.moe.natj.objc.SEL aSelector)

instancesRespondToSelector

public static boolean instancesRespondToSelector(org.moe.natj.objc.SEL aSelector)

isSubclassOfClass

public static boolean isSubclassOfClass(org.moe.natj.objc.Class aClass)

keyPathsForValuesAffectingValueForKey

public static NSSet<java.lang.String> keyPathsForValuesAffectingValueForKey(java.lang.String key)

linkWithGraphs

public boolean linkWithGraphs(NSArray<? extends MLCTrainingGraph> graphs)

Link mutiple training graphs

This is used to link subsequent training graphs with first training sub-graph. This method should be used when we have tensors shared by one or more layers in multiple sub-graphs

Parameters:

graphs - The list of training graphs to link

Returns:

A boolean indicating success or failure

new_objc

public static MLCTrainingGraph new_objc()

optimizer

public MLCOptimizer optimizer()

[@property] optimizer

The optimizer to be used with the training graph

resolveClassMethod

public static boolean resolveClassMethod(org.moe.natj.objc.SEL sel)

resolveInstanceMethod

public static boolean resolveInstanceMethod(org.moe.natj.objc.SEL sel)

resultGradientTensorsForLayer

public NSArray<? extends MLCTensor> resultGradientTensorsForLayer(MLCLayer layer)

Get the result gradient tensors for a layer in the training graph

Parameters:

layer - A layer in the training graph

Returns:

A list of tensors

setTrainingTensorParameters

public boolean setTrainingTensorParameters(NSArray<? extends MLCTensorParameter> parameters)

Set the input tensor parameters that also will be updated by the optimizer

These represent the list of input tensors to be updated when we execute the optimizer update Weights, bias or beta, gamma tensors are not included in this list. MLCompute automatically adds them to the parameter list based on whether the layer is marked as updatable or not.

Parameters:

parameters - The list of input tensors to be updated by the optimizer

Returns:

A boolean indicating success or failure

setVersion_static

public static void setVersion_static(long aVersion)

sourceGradientTensorsForLayer

public NSArray<? extends MLCTensor> sourceGradientTensorsForLayer(MLCLayer layer)

Get the source gradient tensors for a layer in the training graph

Parameters:

layer - A layer in the training graph

Returns:

A list of tensors

stopGradientForTensors

public boolean stopGradientForTensors(NSArray<? extends MLCTensor> tensors)

Add the list of tensors whose contributions are not to be taken when computing gradients during gradient pass

Parameters:

tensors - The list of tensors

Returns:

A boolean indicating success or failure

superclass_static

public static org.moe.natj.objc.Class superclass_static()

synchronizeUpdates

public void synchronizeUpdates()

Synchronize updates (weights/biases from convolution, fully connected and LSTM layers, tensor parameters) from device memory to host memory.

version_static

public static long version_static()