caffe2 Namespace Reference

A global dictionary that holds information about what Caffe2 modules have been loaded in the current runtime, and also utility functions to load modules. More...

Namespaces
	enforce_detail
	Rich logging messages.

Data Structures
struct	AbsCPUFunctor
struct	AbsGradientCPUFunctor
struct	AbstractLengthsDef
class	AbstractLengthsGradientOp
class	AbstractLengthsOp
	Segment reduction op with optional fused embedding lookup. More...
class	AbstractLengthsWithMainInputAndForwardOutputGradientOp
class	AbstractLengthsWithMainInputGradientOp
struct	AbstractReduceBackDef
struct	AbstractReduceFrontDef
class	AbstractReduceFrontOrBackGradientOp
class	AbstractReduceFrontOrBackOp
	Simple non-segmented reduction over the first few dimensions of the tensor. More...
struct	AbstractSortedSegmentDef
class	AbstractSortedSegmentGradientOp
class	AbstractSortedSegmentOp
	Segment reduction op with optional fused embedding lookup. More...
struct	AbstractSortedSegmentRangeDef
class	AbstractSortedSegmentRangeGradientOp
class	AbstractSortedSegmentRangeOp
	Base implementation for segment reduction op that leverages continuity of the data. More...
struct	AbstractSparseLengthsDef
struct	AbstractSparseSortedSegmentDef
struct	AbstractSparseUnsortedSegmentDef
struct	AbstractUnsortedSegmentDef
class	AbstractUnsortedSegmentGradientOp
class	AbstractUnsortedSegmentOp
	Unsorted segment reduction op with optional fused embedding lookup. More...
class	AccumulateHistogramOp
class	AccumulateInputGradientOp
class	AccumulateOp
class	AccuracyOp
class	AdagradOp
class	AdamOp
class	AddPaddingOp
class	AffineChannelGradientOp
class	AffineChannelOp
class	AlgorithmsCache
class	AliasOp
	Alias op makes the output and the input share the same underlying storage. More...
struct	AlignedDeleter
struct	AllocAligned
class	AlternateLearningRate
struct	Analysis
class	APMeterOp
class	ArgMaxOp
class	ArgMinOp
class	ArgOpBase
class	ArgumentHelper
	A helper class to index into arguments. More...
class	AssertOp
class	AsyncDAGNet
class	AsyncNetBase
class	AsyncPollingNet
class	AsyncSchedulingNet
class	AsyncSimpleNet
class	AtomicIterOp
class	AveragedLoss
class	AveragedLossGradient
class	AveragePool
class	AvgExportedStat
class	BaseInputAccessor
class	BaseReducer
class	BaseReducerGradient
class	BatchBoxCoxOp
class	BatchBucketOneHotOp
class	BatchDenseToSparseOp
class	BatchGatherGradientOp
class	BatchGatherOp
class	BatchMatMulOp
class	BatchOneHotOp
class	BatchPermutationGradientOp
class	BatchPermutationOp
class	BatchSparseToDenseOp
class	BatchToSpaceOp
class	BBoxTransformOp
class	BernoulliJSDGradientOp
class	BernoulliJSDOp
class	BinaryElementwiseOp
	Performs a binary operation (e.g. More...
class	Blob
	Blob is a general container that hosts a typed pointer. More...
class	BlobDeserializerBase
	BlobDeserializerBase is an abstract class that deserializes a blob from a BlobProto or a TensorProto. More...
class	BlobSerializerBase
	BlobSerializerBase is an abstract class that serializes a blob to a string. More...
class	BlobsQueue
struct	BlobStatGetter
struct	BlobStatRegistry
class	BlockingCounter
class	BooleanMaskOp
class	BooleanUnmaskOp
class	BoxWithNMSLimitOp
class	BRGNCHWCToPackedInt8BGRAStylizerDeprocessOp
class	BufferedTokenizer
class	Caffe2FlagParser
class	Caffe2ModuleTestDynamicDummyOp
class	CastOp
class	CeilOp
class	ChannelBackpropStatsOp
class	ChannelShuffleGradientOp
class	ChannelShuffleOp
class	ChannelStatsOp
struct	CharRange
class	CheckCounterDoneOp
class	CheckpointOp
class	ClipGradientOp
class	ClipOp
class	ClipTensorByScalingOp
class	CloseBlobsQueueOp
class	CloseRebatchingQueueOp
class	Col2ImOp
class	CollectAndDistributeFpnRpnProposalsOp
class	CommonSubexpressionEliminationTransform
	Common Subexpression Elimination. More...
class	ConcatOp
class	ConditionalOp
class	ConstantFillOp
class	ConstantWarmupLearningRate
struct	ConvArgs
class	Converter
class	ConvGradientOp
class	ConvOp
class	ConvPoolOpBase
class	ConvToNNPackTransform
class	ConvTransposeGradientOp
class	ConvTransposeOp
class	ConvTransposeUnpoolBase
class	CopyFromGLOp
class	CopyFromOpenGLOp
class	CopyOnDeviceLikeOp
class	CopyOnDeviceLikeOp< CUDAContext, CUDAContext, CUDAContext >
class	CopyOp
class	CopyToOpenGLOp
struct	CosCPUFunctor
struct	CosGradientCPUFunctor
class	CosineEmbeddingCriterionGradientOp
class	CosineEmbeddingCriterionOp
class	CosineSimilarityGradientOp
class	CosineSimilarityOp
class	CountDownOp
class	Counter
class	CountUpOp
struct	CPUAllocator
class	CPUContext
	The CPU Context, representing the bare minimum of what a Context class in Caffe2 should implement. More...
struct	CPUEventWrapper
class	CpuId
	Identification of an Intel CPU. More...
class	CPUSparseLengthsReductionOp
struct	CpuUtilizationReportOp
class	CreateBlobsQueueOp
class	CreateCounterOp
class	CreateDBOp
class	CreateMapOp
class	CreateRebatchingQueueOp
class	CreateScopeOp
class	CreateTextFileReaderOp
class	CrossEntropyGradientOp
class	CrossEntropyOp
class	CUDAContext
struct	CudaDevicePropWrapper
struct	CudaEventWrapper
class	CUDARecurrentNetworkExecutor
class	CudaRTCFunction
class	CudnnConvGradientOp
class	CudnnConvOp
class	CudnnConvOpBase
class	CudnnConvTransposeGradientOp
class	CudnnConvTransposeOp
class	CudnnConvTransposeOpBase
class	cudnnFilterDescWrapper
class	CuDNNLRNGradientOp
class	CuDNNLRNOp
class	CuDNNReluGradientOp
class	CuDNNReluOp
class	CuDNNSoftmaxGradientOp
class	CuDNNSoftmaxOp
class	CudnnSpatialBNGradientOp
class	CudnnSpatialBNOp
class	CuDNNState
class	cudnnTensorDescWrapper
	cudnnTensorDescWrapper is the placeholder that wraps around a cudnnTensorDescriptor_t, allowing us to do descriptor change as-needed during runtime. More...
class	CuDNNTransposeOp
class	cudnnTypeWrapper
	cudnnTypeWrapper is a wrapper class that allows us to refer to the cudnn type in a template function. More...
class	cudnnTypeWrapper< double >
class	cudnnTypeWrapper< float >
class	cudnnTypeWrapper< float16 >
struct	CuDNNWorkspace
	CuDNNWorkspace is a wrapper around a raw cuda pointer that holds the cudnn scratch space. More...
class	CuDNNWrapper
	CuDNNWrapper is a class that wraps the cudnn handles and cudnn workspaces. More...
class	DAGNet
class	DAGNetBase
class	DBExistsOp
class	DecodedFrame
struct	DefaultCPUAllocator
class	DefaultEngine
class	DeformConvGradientOp
class	DeformConvOp
class	DeformConvOpBase
class	DequeueBlobsOp
class	DequeueRebatchingQueueOp
class	DetailedExportedStat
class	DeviceGuard
struct	DeviceTypeRegisterer
class	DiagonalFillOp
struct	DispatchHelper
struct	DispatchHelper< FixedValues< FirstVal, Values... >, ExtraArgs... >
struct	DispatchHelper< FixedValues<>, ExtraArgs... >
class	DivGradientOp
class	DoOp
class	DotProductGradientOp
class	DotProductOp
class	DotProductWithPaddingGradientOp
class	DotProductWithPaddingOp
class	DropoutGradientOp
class	DropoutOp
class	EigenConvOp
struct	EigenPowFunctor
class	ElementwiseLinearGradientOp
class	ElementwiseLinearOp
class	ElementwiseRTCOp
	A GPU operator that can generate limited elementwise operations. More...
class	EluGradientOp
class	EluOp
class	EnforceNotMet
class	EnqueueBlobsOp
class	EnqueueRebatchingQueueOp
class	EnsureDenseOp
	Pass inputs to outputs. More...
class	Event
struct	EventCreateFunctionRegisterer
struct	EventErrorMessageFunctionRegisterer
struct	EventFinishFunctionRegisterer
struct	EventQueryFunctionRegisterer
struct	EventRecordFunctionRegisterer
struct	EventResetFunctionRegisterer
struct	EventSetFinishedFunctionRegisterer
struct	EventWaitFunctionRegisterer
class	ExpandDimsOp
struct	ExpCPUFunctor
class	ExpLearningRate
class	ExportedStat
struct	ExportedStatValue
class	FeedBlobOp
class	FileReader
class	FileStoreHandler
class	FileStoreHandlerCreateOp
class	FillerOp
class	FindDuplicateElementsOp
class	FindOp
class	FixedDivisor
class	FixedDivisor< int32_t >
class	FixedLearningRate
struct	FixedType
struct	FixedValues
class	FlattenOp
class	FlattenToVecOp
class	FlexibleTopKGradientOp
class	FlexibleTopKOp
class	Float16ConstantFillOp
class	FloatToFused8BitRowwiseQuantizedOp
class	FloatToHalfOp
class	FloatToRowwiseQuantized8BitsOp
class	FloorOp
struct	ForEach
	ForEach is a unary functor that forwards each element of the input array into the elementwise Functor provided, and gathers the results of each call into the resulting array. More...
class	FP16MomentumSGDUpdateOp
class	FP32MomentumSGDUpdateOp
class	FreeOp
class	FtrlOp
struct	FtrlParams
class	FullyConnectedDecompGradientOp
class	FullyConnectedGradientOp
class	FullyConnectedOp
class	FullyConnectedOp_SPARSE
class	FullyConnectedOpDecomp
class	FullyConnectedOpPrune
class	FullyConnectedPruneGradientOp
class	FunHashGradientOp
class	FunHashOp
class	Fused8BitRowwiseQuantizedToFloatOp
class	GatherByKeyOp
class	GatherFused8BitRowwiseOp
class	GatherOp
class	GatherPaddingOp
class	GatherRangesOp
class	GatherRangesToDenseOp
class	GaussianFillOp
class	GenerateProposalsOp
struct	GenericTensorImplementation
class	GetAbsGradient
struct	GetAddPaddingGradient
class	GetAffineChannelGradient
class	GetAveragedLossGradient
class	GetBatchGatherGradient
class	GetBatchPermutationGradient
class	GetBatchToSpaceGradient
class	GetBernoulliJSDGradient
class	GetCastGradient
class	GetChannelShuffleGradient
class	GetCol2ImGradient
class	GetColwiseMaxGradient
class	GetConcatGradient
class	GetConvGradient
class	GetConvTransposeGradient
class	GetCosGradient
class	GetCosineSimilarityGradient
class	GetCrossEntropyGradient
class	GetDotProductGradient
class	GetDotProductWithPaddingGradient
class	GetDropoutGradient
struct	GetElementwiseLinearGradient
class	GetEluGradient
class	GetExpGradient
class	GetFCDecompGradient
class	GetFlattenGradient
class	GetFloatToHalfGradient
class	GetGroupSpatialSoftmaxGradient
class	GetGRUUnitGradient
class	GetHalfToFloatGradient
class	GetIm2ColGradient
class	GetInstanceNormGradient
class	GetL1DistanceGradient
class	GetLabelCrossEntropyGradient
class	GetLeakyReluGradient
class	GetLogGradient
class	GetLRNGradient
class	GetLSTMUnitGradient
struct	GetMakeTwoClassGradient
class	GetMatMulGradient
class	GetMaxGradient
class	GetMeanGradient
class	GetMinGradient
struct	GetNanCheckGradient
class	GetNCHW2NHWCGradient
struct	GetNegateGradientGradient
class	GetNegativeGradient
class	GetNHWC2NCHWGradient
class	GetNormalizeGradient
class	GetPackSegmentsGradient
class	GetPadImageGradient
class	GetPoolGradient
class	GetPReluGradient
class	GetPrependDimGradient
struct	GetRecurrentGradient
struct	GetRecurrentNetworkGradient
class	GetReduceBackMaxGradient
class	GetReduceBackMeanGradient
class	GetReduceBackSumGradient
class	GetReduceFrontMaxGradient
class	GetReduceFrontMeanGradient
class	GetReduceFrontSumGradient
class	GetReluGradient
struct	GetRemovePaddingGradient
class	GetReshapeGradient
class	GetResizeNearestGradient
class	GetReversePackedSegsGradient
class	GetRoIPoolGradient
class	GetRowwiseMaxGradient
class	GetSampleAsGradient
class	GetScaleGradient
class	GetSelectSmoothL1LossGradient
class	GetSeluGradient
class	GetSigmoidCrossEntropyLossGradient
struct	GetSigmoidCrossEntropyWithLogitsGradient
class	GetSigmoidFocalLossGradient
class	GetSigmoidGradient
class	GetSinGradient
class	GetSmoothL1LossGradient
class	GetSoftmaxFocalLossGradient
class	GetSoftmaxGradient
class	GetSoftplusGradient
class	GetSoftsignGradient
class	GetSpaceToBatchGradient
class	GetSpatialBNGradient
class	GetSplitGradient
class	GetSqrGradient
class	GetSqrtGradient
class	GetSquaredL2DistanceGradient
class	GetSquareRootDivideGradient
class	GetSumElementsGradient
class	GetSwishGradient
class	GetTanhGradient
class	GetTileGradient
class	GetTopKGradient
class	GetTransposeGradient
class	GetUnpackSegmentsGradient
class	GetUpsampleNearestGradient
struct	GetWeightedSigmoidCrossEntropyWithLogitsGradient
struct	GetZeroGradientOpGradient
class	GivenTensorFillOp
class	GLAveragePoolOp
class	GLConcatOp
class	GLContext
class	GLConvOp
class	GLFullyConnectedOp
class	GLMaxPoolOp
class	GLNet
class	GLNormalizePlanarYUVOp
class	GLNormPlanarYUVOp
class	GLPoolOp
class	GLPredictor
class	GLReluOp
class	GLReshapeOp
class	GLResizeNearestOp
class	GLSigmoidOp
class	GLSoftmaxOp
class	GLSpatialBNOp
class	GLSumOp
class	GLTensor
class	GluOp
class	GPUFallbackOp
	A templated class to allow one to wrap a CPU operator as a CUDA operator. More...
class	GradientMakerBase
struct	GradientNotImplementedYet
	A helper class to indicate that the gradient mechanism is not ready. More...
struct	GradientOpsMeta
	A struct that holds the gradient operators and related gradient maps. More...
struct	GradientWrapper
class	GroupSpatialSoftmaxGradientOp
class	GroupSpatialSoftmaxOp
class	GRUUnitGradientOp
class	GRUUnitOp
class	HalfToFloatOp
class	HasElementsOp
class	HasScopeOp
class	HillLearningRate
class	HSoftmaxGradientOp
class	HSoftmaxOp
class	HSoftmaxOpBase
class	HSoftmaxSearchOp
class	HuffmanTreeHierarchyOp
class	IfOp
class	Im2ColOp
class	ImageAllocator
class	ImageInputOp
struct	Index
struct	IndexBase
class	IndexCreateOp
class	IndexDeserializer
class	IndexFreezeOp
class	IndexGetOp
class	IndexHashOp
class	IndexLoadOp
class	IndexSerializer
class	IndexSizeOp
class	IndexStoreOp
class	InitRegisterer
class	InstanceNormGradientOp
class	InstanceNormOp
class	InvLearningRate
class	IsEmptyOp
class	IsMemberOfOp
class	IsMemberOfValueHolder
class	IterOp
class	KeySplitOp
class	KeyValueToMapOp
class	L1DistanceGradientOp
class	L1DistanceOp
class	LabelCrossEntropyGradientOp
class	LabelCrossEntropyOp
class	LambdaRankNdcgGradientOp
class	LambdaRankNdcgOp
class	LarsOp
class	LayerNormGradientOp
class	LayerNormOp
class	LeakyReluGradientOp
class	LeakyReluOp
class	LearningRateFunctor
class	LearningRateOp
class	LengthsGatherOp
struct	LengthsOpGetGradient
class	LengthsPartitionOp
class	LengthsRangeFillOp
class	LengthsTileOp
class	LengthsTopKGradientOp
class	LengthsTopKOp
class	LengthsToRangesOp
class	LengthsToSegmentIdsOp
class	LengthsToShapeOp
class	LengthsToWeightsOp
class	LinearWarmupLearningRate
class	LoadOp
class	LocallyConnectedGradientOp
class	LocallyConnectedOp
struct	LogCPUFunctor
class	LoggerVoidify
struct	LogitCPUFunctor
class	LogitGradientOp
class	LogMeanExpRangeReducer
class	LogMeanExpRangeReducer< T, CPUContext >
struct	LogMeanExpRangeReducerDef
class	LogMeanExpRangeReducerGradient
class	LogSumExpRangeReducer
class	LogSumExpRangeReducer< T, CPUContext >
struct	LogSumExpRangeReducerDef
class	LogSumExpRangeReducerGradient
class	LpNormGradientOp
class	LpNormOp
class	LpPool
class	LRNGradientOp
class	LRNOp
class	LRNOpBase
class	LSTMUnitGradientOp
class	LSTMUnitOp
struct	MakeAligned
class	MakeTwoClassGradientOp
class	MakeTwoClassOp
class	MapDeserializer
class	MapSerializer
class	MapToKeyValueOp
struct	MapTypeTraits
class	MarginRankingCriterionGradientOp
class	MarginRankingCriterionOp
class	MatMulOp
class	MaxGradientOp
class	MaxMinOpBase
class	MaxOp
class	MaxPool
class	MaxPoolGradientRTCOp
class	MaxPoolRTCOp
class	MaxPoolWithIndexGradientOp
class	MaxPoolWithIndexOp
class	MaxRangeReducer
class	MaxRangeReducer< T, CPUContext >
struct	MaxRangeReducerDef
class	MaxRangeReducerGradient
class	MaxReduceDimsGradientOp
class	MaxReduceDimsOp
class	MaxReducer
class	MaxReducer< T, CPUContext >
struct	MaxReducerDef
class	MaxReducerGradient
class	MaxReductionGradientOp
class	MaxReductionOp
class	MeanGradientOp
class	MeanOp
class	MeanRangeReducer
class	MeanRangeReducer< T, CPUContext >
struct	MeanRangeReducerDef
class	MeanRangeReducerGradient
class	MeanReducer
class	MeanReducer< T, CPUContext >
struct	MeanReducerDef
class	MeanReducerGradient
class	MemoryAllocationReporter
class	MergeDimOp
class	MergeIdListsOp
class	MessageLogger
class	MinGradientOp
class	MinOp
class	MKLContext
	The MKL Context, which is largely the same as the CPUContext. More...
class	ModOp
class	ModuleSchema
	A module schema that can be used to store specific information about different modules. More...
class	MomentumSGDOp
class	MomentumSGDUpdateOp
class	MPICommonWorldWrapper
	A simple wrapper over an MPI common world. More...
class	MPIDataTypeWrapper
struct	MPSCNNContext
class	MSRAFillOp
class	MultiClassAccuracyOp
class	MutexDeserializer
class	MutexSerializer
class	NanCheckOp
class	NCHW2NHWCOp
class	NegateGradientOp
struct	NegativeCPUFunctor
class	NetBase
class	NetObserverReporter
class	NetObserverReporterPrint
class	NGramFromCategoricalOp
class	NHWC2NCHWOp
class	NNApi
class	NNPACKConvOp
class	NoDefaultEngineOp
	A helper class to denote that an op does not have a default engine. More...
class	NoGradient
	A helper class to indicate that the operator does not need gradient computation. More...
class	NormalizeGradientOp
class	NormalizeL1Op
class	NormalizeOp
struct	NotFunctor
class	Observable
	Inherit to make your class observable. More...
class	ObserverBase
	Use this to implement a Observer using the Observer Pattern template. More...
class	ObserverConfig
class	OneHotOp
class	ONNXWhileOp
class	OpenGLAddOp
class	OpenGLConcatOp
class	OpenGLConvOp
class	OpenGLConvTransposeOp
class	OpenGLInstanceNormPReluOp
class	OpenGLMulOp
class	OpenGLPadImageOp
class	OpenGLPReluOp
class	OpenGLResizeNearestOp
class	OpenGLSigmoidOp
class	OpenGLSubOp
class	OpenGLTensorToTextureStylizerPreprocessOp
class	OpenGLTextureToTensorStylizerDeprocessOp
class	OpenGLTextureToTextureStylizerDeprocessOp
class	OpenGLTextureToTextureStylizerPreprocessOp
class	Operator
class	OperatorAttachingNetObserver
class	OperatorBase
class	OpSchema
	A class to record the schema of an op. More...
class	OpSchemaRegistry
	A registry to hold all the operator schemas. More...
struct	OpTask
	Data structure for a scheduled task in the task queue. More...
class	PackedInt8BGRANHWCToNCHWCStylizerPreprocessOp
class	PackRNNSequenceOpBase
class	PackSegmentsOp
class	PadEmptySamplesOp
class	PadImageGradientOp
class	PadImageOp
class	PairWiseLossGradientOp
class	PairWiseLossOp
class	Params
class	PartitionOp
class	PartitionOpBase
class	PatternNetTransform
	PatternNetTransform allows you to create transforms using a simple interface. More...
class	PercentileOp
class	PerfNetObserver
class	PerfOperatorObserver
class	PerplexityOp
class	PiecewiseLinearTransformOp
struct	PinnedCPUAllocator
	An allocator that does the CPU memory allocation with pinned memory. More...
struct	PlanExecutionTime
class	PolyLearningRate
class	PoolGradientOp
class	PoolOp
class	PowOp
class	Predictor
class	PrefetchOperator
class	PReluGradientOp
class	PReluOp
class	PrependDimOp
class	PrintOp
class	ProfileCounter
class	ProfileObserver
class	ProfileOperatorObserver
class	PSRoIPoolGradientOp
class	PSRoIPoolOp
class	QConvOp
struct	QConvState
class	QTensor
class	QTensorDeserializer
class	QTensorSerializer
class	QuantDecodeGradientOp
class	QuantDecodeOp
class	QuantDecompZstdOp
class	RangeFillOp
class	RangeOp
class	RebatchingQueue
class	RecurrentBaseOp
class	RecurrentGradientOp
class	RecurrentNetworkBlobFetcherOp
class	RecurrentNetworkExecutorBase
	RecurrentNetworkExecutor is a specialized runtime for recurrent neural networks (RNNs). More...
class	RecurrentNetworkGradientOp
class	RecurrentNetworkOp
class	RecurrentOp
class	RecurrentParamAccessOp
class	RedisStoreHandler
class	RedisStoreHandlerCreateOp
class	ReduceMeanOp
class	ReduceOpBase
class	ReduceSumOp
class	ReduceTailSumOp
class	Registerer
class	Registry
	A template class that allows one to register classes by keys. More...
class	ReluGradientOp
class	ReluOp
class	RemoveDataBlocksOp
class	RemovePaddingOp
class	ReplaceNaNOp
class	ResetCounterOp
class	ReshapeOp
class	ResizeLikeOp
class	ResizeNearestGradientOp
class	ResizeNearestOp
class	RetrieveCountOp
class	ReversePackedSegsOp
class	RMACRegionsOp
class	RmsPropOp
class	RNNApplyLinkOp
class	RNNCapableOperatorObserver
	Inherit to make your class observable. More...
struct	RNNNetOperator
	Struct for operator in a timestep and its dependenceis. More...
class	RoIAlignGradientOp
class	RoIAlignOp
class	RoIPoolFGradientOp
class	RoIPoolFOp
class	RoIPoolGradientOp
class	RoIPoolOp
class	RowMulOp
class	Rowwise8BitQuantizedToFloatOp
class	RowWiseArgMaxOp
class	RowWiseSparseAdagradOp
class	RowWiseSparseAdamOp
class	RunCountNetObserver
class	RunCountOperatorObserver
class	SafeDequeueBlobsOp
class	SafeEnqueueBlobsOp
struct	SameTypeAsInput
class	SampleAsGradientOp
class	SampleAsOp
struct	SampleInterval
class	SaveOp
class	ScaleOp
class	ScatterAssignOp
	Update slices of the tensor in-place by overriding. More...
class	ScatterWeightedSumOp
	Update slices of the tensor in-place with weighted sum. More...
class	SegmentIdsToLengthsOp
class	SegmentIdsToRangesOp
class	SegmentOneHotOp
struct	SegmentOpGetGradient
class	SelectGradientOpBase
class	SelectSmoothL1LossGradientOp
class	SelectSmoothL1LossOp
class	SeluGradientOp
class	SeluOp
class	SequenceMaskOp
class	ShapeOp
struct	SigmoidCPUFunctor
class	SigmoidCrossEntropyLossGradientOp
class	SigmoidCrossEntropyLossOp
class	SigmoidCrossEntropyWithLogitsGradientOp
class	SigmoidCrossEntropyWithLogitsOp
class	SigmoidFocalLossGradientOp
class	SigmoidFocalLossOp
struct	SigmoidGradientCPUFunctor
class	SignalHandler
struct	SignCPUFunctor
class	SimpleNet
class	SimpleQueue
struct	SinCPUFunctor
class	SingleOpTransform
	Single Op Transform Base class. More...
struct	SinGradientCPUFunctor
class	SinusoidPositionEncodingOp
class	SizeOp
class	SkipIndices
class	SkipIndices<>
class	SliceGradientOp
class	SliceOp
class	SmartTensorPrinter
class	SmoothL1LossGradientOp
class	SmoothL1LossOp
class	SNPEOp
class	SoftmaxFocalLossGradientOp
class	SoftmaxFocalLossOp
class	SoftmaxGradientOp
class	SoftmaxOp
class	SoftmaxWithLossGradientOp
class	SoftmaxWithLossOp
class	SoftplusGradientOp
class	SoftplusOp
struct	SoftsignCPUFunctor
struct	SoftsignGradientCPUFunctor
class	SpaceBatchOpBase
class	SpaceToBatchOp
class	SparseAdagradOp
class	SparseAdamOp
class	SparseFtrlOp
class	SparseFunHashGradientOp
class	SparseFunHashOp
class	SparseLengths8BitsRowwiseOp
class	SparseLengthsFused8BitRowwiseOp
class	SparseMatrixReshapeOp
class	SparseMomentumSGDUpdateOp
class	SparseNormalizeOp
class	SparseToDenseMaskBase
class	SparseToDenseMaskGradientOp
class	SparseToDenseMaskOp
class	SparseToDenseOp
class	SpatialBNGradientOp
class	SpatialBNOp
class	SpatialNarrowAsGradient
class	SpatialNarrowAsGradientOp
class	SpatialNarrowAsOp
class	SpatialSoftmaxWithLossGradientOp
class	SpatialSoftmaxWithLossOp
class	SplitOp
struct	SqrCPUFunctor
struct	SqrtCPUFunctor
class	SquaredL2DistanceGradientOp
class	SquaredL2DistanceOp
class	SquareRootDivideOp
class	SqueezeOp
struct	Stat
struct	StaticLinkingProtector
class	StatRegistry
	Holds a map of atomic counters keyed by name. More...
class	StatRegistryCreateOp
class	StatRegistryExportOp
class	StatRegistryUpdateOp
class	StatValue
class	StdDevExportedStat
class	StepLearningRate
class	StopGradientOp
struct	StopOnSignal
class	StoreAddOp
class	StoreGetOp
class	StoreHandler
struct	StoreHandlerTimeoutException
class	StoreSetOp
class	StoreWaitOp
class	StringDeserializer
	StringDeserializer is the deserializer for Strings. More...
class	StringJoinOp
struct	StringProvider
class	StringSerializer
	StringSerializer is the serializer for String. More...
class	StumpFuncOp
class	SumElementsGradientOp
class	SumElementsIntOp
class	SumElementsOp
class	SummarizeOp
class	SumOp
class	SumRangeReducer
class	SumRangeReducer< T, CPUContext >
struct	SumRangeReducerDef
class	SumRangeReducerGradient
class	SumReduceDimsGradientOp
class	SumReduceDimsOp
class	SumReduceLikeOp
class	SumReducer
class	SumReducer< T, CPUContext >
struct	SumReducerDef
class	SumReducerGradient
class	SumSqrElementsOp
struct	SwishCPUFunctor
class	SwishGradientOp
struct	TanhCPUFunctor
struct	TanhGradientCPUFunctor
struct	Task
class	TaskThreadPool
class	Tensor
	Tensor is the basic class in Caffe2 that stores a contiguous memory with its shape information. More...
class	TensorDeserializer
	TensorDeserializer is the deserializer for Tensors. More...
class	TensorPrinter
class	TensorProtosDBInput
class	TensorSerializer
	TensorSerializer is the serializer for Tensors. More...
struct	TensorTypes
struct	TensorTypes2
struct	TextFileReaderInstance
class	TextFileReaderReadOp
class	ThreadedRecurrentNetworkExecutor
class	ThreadLocalCUDAObjects
	A struct to host thread-local cuda objects. More...
class	ThreadPool
class	ThresholdedReluGradientOp
class	ThresholdedReluOp
struct	ThrowInTheTowelIfGradientIsCalled
	A helper class to indicate that the operator should have no gradient. More...
class	TileGradientOp
class	TileOp
class	TimeCounter
class	TimeObserver
class	TimeOperatorObserver
class	Timer
	A simple timer object for measuring time. More...
struct	TimerBeginOp
struct	TimerEndOp
struct	TimerGetAndEndOp
struct	TimerGetOp
class	TimerInstance
struct	Token
class	TokenizedString
class	Tokenizer
class	TopKGradientOp
class	TopKOp
class	Transform
	The Transform Base Object. More...
class	TransposeOp
class	TTContractionGradientOp
class	TTContractionOp
class	TTLinearGradientOp
class	TTLinearOp
class	TTPadGradientOp
class	TTPadOp
class	TypeMeta
	TypeMeta is a thin class that allows us to store the type of a container such as a blob, or the data type of a tensor, with a unique run-time id. More...
struct	TypeNameRegisterer
struct	TypeNameTraits
struct	TypeNameTraits< int32_t >
struct	TypeNameTraits< int64_t >
class	UnaryElementwiseWithArgsOp
class	UniformFillOp
class	UniqueOp
	Deduplicates input indices vector and optionally produces reverse remapping. More...
class	UniqueUniformFillOp
class	UnpackSegmentsOp
class	UnsafeCoalesceOp
class	UnsupportedOperatorFeature
class	UpsampleNearestGradientOp
class	UpsampleNearestOp
class	VariableLengthSequencePaddingOp
class	VideoDecoder
class	VideoInputOp
class	VideoIOContext
struct	VideoMeta
class	WallClockTimeOp
class	WeightedMultiSamplingOp
class	WeightedSampleDequeueBlobsOp
class	WeightedSampleOp
class	WeightedSigmoidCrossEntropyWithLogitsGradientOp
class	WeightedSigmoidCrossEntropyWithLogitsOp
class	WeightedSumGradientOp
class	WeightedSumOp
class	WeightedSumReducer
class	WeightedSumReducer< T, CPUContext >
struct	WeightedSumReducerDef
class	WeightedSumReducerGradient
class	WhereOp
class	WhileOp
struct	WithDefaultConstructor
	WithDefaultConstructor is a functor that can be used as the functor of an UnaryElementwiseWithArgsOp. More...
struct	WithoutBroadcast
class	Worker
class	WorkersPool
class	Workspace
	Workspace is a class that holds all the related objects created during runtime: (1) all blobs, and (2) all instantiated networks. More...
class	XavierFillOp
class	YellowFinOp
class	ZeroGradientOp
class	ZmqContext
class	ZmqMessage
class	ZmqSocket

Typedefs
using	MemoryDeleter = void(*)(void *)
typedef int64_t	TIndex
template<typename Key , typename Value >
using	CaffeMap = std::map< Key, Value >
typedef Tensor< CUDAContext >	TensorCUDA
typedef void(*	EventCreateFunction) (const DeviceOption &option, Event *)
typedef void(*	EventRecordFunction) (Event *, const void *, const char *)
typedef void(*	EventWaitFunction) (const Event *, void *)
typedef void(*	EventFinishFunction) (const Event *)
typedef EventStatus(*	EventQueryFunction) (const Event *)
typedef const std::string &(*	EventErrorMessageFunction) (const Event *)
typedef void(*	EventSetFinishedFunction) (const Event *, const char *)
typedef void(*	EventResetFunction) (Event *)
typedef ObserverBase< NetBase >	NetObserver
typedef std::function< std::unique_ptr< NetObserver >NetBase *)>	NetObserverCreator
typedef ObserverBase< OperatorBase >	OperatorObserver
typedef Registry< std::string, std::unique_ptr< OperatorBase >, const OperatorDef &, Workspace * > *(*	RegistryFunction) ()
using	EnginePrefType = std::vector< std::string >
using	PerOpEnginePrefType = CaffeMap< int, CaffeMap< std::string, EnginePrefType >>
using	GlobalEnginePrefType = CaffeMap< int, EnginePrefType >
typedef std::function< bool(int)>	ShouldContinue
using	ExportedStatList = std::vector< ExportedStatValue >
	Holds names and values of counters exported from a StatRegistry.
using	ExportedStatMap = std::unordered_map< std::string, int64_t >
typedef Tensor< CPUContext >	TensorCPU
typedef TypeMeta(*	TypeCall) (const void *)
typedef vector< TIndex >(*	TensorInfoCall) (const void *, bool *shares_data, size_t *capacity, DeviceOption *device)
typedef intptr_t	CaffeTypeId
typedef half_float::half	half
typedef half	DataType
template<typename T >
using	deleted_unique_ptr = std::unique_ptr< T, std::function< void(T *)>>
using	ParallelFor = std::function< void(size_t, std::function< void(size_t)>)>
using	NumericTypes = TensorTypes< int32_t, int64_t, float, double >
using	IntTypes = TensorTypes< int32_t, int64_t >
using	BoolTypes = TensorTypes< bool >
using	IntBoolTypes = TensorTypes< int32_t, int64_t, bool >
template<typename InputTypes , class Context , class Functor , class OutputType = SameTypeAsInput>
using	UnaryElementwiseOp = UnaryElementwiseWithArgsOp< InputTypes, Context, WithDefaultConstructor< Functor >, OutputType >
	UnaryElementwiseOp is a wrapper around UnaryElementwiseWithArgsOp, with the difference that it takes a functor with default constructor, e.g. More...
using	n = 2
using	MapType64To64 = MapTypeTraits< int64_t, int64_t >::MapType
using	MapType64To32 = MapTypeTraits< int64_t, int32_t >::MapType
using	MapType32To32 = MapTypeTraits< int32_t, int32_t >::MapType
using	MapType32To64 = MapTypeTraits< int32_t, int64_t >::MapType
template<typename ScalarFunctor , typename TypeMap = FixedType<std::string>>
using	StringElementwiseOp = UnaryElementwiseWithArgsOp< TensorTypes< std::string >, CPUContext, ForEach< ScalarFunctor >, TypeMap >
using	RebatchingQueuePtr = std::unique_ptr< RebatchingQueue >
template<typename T >
using	EArrXt = Eigen::Array< T, Eigen::Dynamic, 1 >
using	EArrXf = Eigen::ArrayXf
using	EArrXd = Eigen::ArrayXd
using	EArrXi = Eigen::ArrayXi
using	EArrXb = EArrXt< bool >
template<typename T >
using	EArrXXt = Eigen::Array< T, Eigen::Dynamic, Eigen::Dynamic >
using	EArrXXf = Eigen::ArrayXXf
template<typename T >
using	ERArrXXt = Eigen::Array< T, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor >
using	ERArrXXf = ERArrXXt< float >
template<typename T >
using	EVecXt = Eigen::Matrix< T, Eigen::Dynamic, 1 >
using	EVecXd = Eigen::VectorXd
using	EVecXf = Eigen::VectorXf
using	ERVecXd = Eigen::RowVectorXd
using	ERVecXf = Eigen::RowVectorXf
template<typename T >
using	EMatXt = Eigen::Matrix< T, Eigen::Dynamic, Eigen::Dynamic >
using	EMatXd = Eigen::MatrixXd
using	EMatXf = Eigen::MatrixXf
template<typename T >
using	ERMatXt = Eigen::Matrix< T, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor >
using	ERMatXd = ERMatXt< double >
using	ERMatXf = ERMatXt< float >
template<typename T >
using	EigenMatrixMap = Eigen::Map< Eigen::Matrix< T, Eigen::Dynamic, Eigen::Dynamic > >
template<typename T >
using	EigenArrayMap = Eigen::Map< Eigen::Array< T, Eigen::Dynamic, Eigen::Dynamic > >
template<typename T >
using	EigenVectorMap = Eigen::Map< Eigen::Matrix< T, Eigen::Dynamic, 1 > >
template<typename T >
using	EigenVectorArrayMap = Eigen::Map< Eigen::Array< T, Eigen::Dynamic, 1 > >
template<typename T >
using	ConstEigenMatrixMap = Eigen::Map< const Eigen::Matrix< T, Eigen::Dynamic, Eigen::Dynamic > >
template<typename T >
using	ConstEigenArrayMap = Eigen::Map< const Eigen::Array< T, Eigen::Dynamic, Eigen::Dynamic > >
template<typename T >
using	ConstEigenVectorMap = Eigen::Map< const Eigen::Matrix< T, Eigen::Dynamic, 1 > >
template<typename T >
using	ConstEigenVectorArrayMap = Eigen::Map< const Eigen::Array< T, Eigen::Dynamic, 1 > >

Enumerations
enum	CudaMemoryPoolType { NONE = 0, CUB = 1 }
enum	EventStatus { EVENT_INITIALIZED = 0, EVENT_SCHEDULED = 1, EVENT_SUCCESS = 2, EVENT_FAILED = 3 }
enum	StorageOrder { UNKNOWN = 0, NHWC = 1, NCHW = 2 }
enum	{ ALGO_FWD = 0, ALGO_WGRAD = 1, ALGO_DGRAD = 2 }
enum	PadMode { CONSTANT = 0, REFLECT = 1, EDGE = 2 }
enum	QuantDecodeRunTy { RUN_ALWAYS, RUN_ONCE }
enum	RecurrentParamOpMode { SET_PARAM, GET_PARAM }
enum	FLowAlgType { FarnebackOpticalFlow = 0, DensePyrLKOpticalFlow = 1, BroxOpticalFlow = 2, OpticalFlowDual_TVL1 = 3 }
enum	FlowDataType { Flow2C = 0, Flow3C = 1, FlowWithGray = 2, FlowWithRGB = 3 }
enum	SpecialFps { SAMPLE_NO_FRAME = 0, SAMPLE_ALL_FRAMES = -1, SAMPLE_TIMESTAMP_ONLY = -2 }
enum	VideoResType { USE_WIDTH_HEIGHT = 0, USE_MINIMAL_WIDTH_HEIGHT = 1, ORIGINAL_RES = 2 }
enum	DecodeType { DO_TMP_JITTER = 0, DO_UNIFORM_SMP = 1, USE_START_FRM = 2 }

Functions
void	ConvertToRawDataset (const string &input_db_name, const string &output_db_name)
void	ReadImage (std::ifstream *file, int *label, char *buffer)
void	WriteToDB (const string &filename, const int num_items, const int &offset, db::DB *db)
void	ConvertCIFAR ()
void	ConvertImageDataset (const string &input_folder, const string &list_filename, const string &output_db_name, const bool)
uint32_t	swap_endian (uint32_t val)
void	convert_dataset (const char *image_filename, const char *label_filename, const char *db_path, const int data_limit)
void	run ()
void	NoDelete (void *)
CPUAllocator *	GetCPUAllocator ()
void	SetCPUAllocator (CPUAllocator *alloc)
void	swap (Blob &lhs, Blob &rhs)
	CAFFE_DEFINE_TYPED_REGISTRY (BlobSerializerRegistry, CaffeTypeId, BlobSerializerBase, std::unique_ptr)
	CAFFE_DEFINE_REGISTRY (BlobDeserializerRegistry, BlobDeserializerBase)
	CAFFE_DECLARE_TYPED_REGISTRY (BlobSerializerRegistry, CaffeTypeId, BlobSerializerBase, std::unique_ptr)
unique_ptr< BlobSerializerBase >	CreateSerializer (CaffeTypeId id)
	CAFFE_DECLARE_REGISTRY (BlobDeserializerRegistry, BlobDeserializerBase)
unique_ptr< BlobDeserializerBase >	CreateDeserializer (const string &type)
bool	HasCudaRuntime ()
const std::map< string, string > &	GetBuildOptions ()
template<typename T , typename... Args>
std::enable_if<!std::is_array< T >::value, std::unique_ptr< T > >::type	make_unique (Args &&...args)
template<typename T >
std::enable_if< std::is_array< T >::value, std::unique_ptr< T > >::type	make_unique (const size_t n)
template<typename T , typename... Args>
std::enable_if< std::extent< T >::value!=0, std::unique_ptr< T > >::type	make_unique (Args &&...)=delete
template<typename Dst , typename Src >
Dst	dynamic_cast_if_rtti (Src ptr)
size_t	cudnnCompiledVersion ()
size_t	cudnnRuntimeVersion ()
void	CheckCuDNNVersions ()
cudnnTensorFormat_t	GetCudnnTensorFormat (const StorageOrder &order)
	A wrapper function to convert the Caffe storage order to cudnn storage order enum values.
int	NumCudaDevices ()
	Returns the number of devices.
void	SetDefaultGPUID (const int deviceid)
int	GetDefaultGPUID ()
int	CaffeCudaGetDevice ()
	Gets the current GPU id. More...
void	CaffeCudaSetDevice (const int id)
	Gets the current GPU id. More...
int	GetGPUIDForPointer (const void *ptr)
	Gets the GPU id that the current pointer is located at.
const cudaDeviceProp &	GetDeviceProperty (const int device)
	Gets the device property for the given device. More...
void	DeviceQuery (const int deviceid)
	Runs a device query function and prints out the results to LOG(INFO).
bool	GetCudaPeerAccessPattern (vector< vector< bool > > *pattern)
	Return a peer access pattern by returning a matrix (in the format of a nested vector) of boolean values specifying whether peer access is possible. More...
bool	TensorCoreAvailable ()
	Return the availability of TensorCores for math.
const char *	cublasGetErrorString (cublasStatus_t error)
	Return a human readable cublas error string.
const char *	curandGetErrorString (curandStatus_t error)
	Return a human readable curand error string.
int	CudaVersion ()
	A runtime function to report the cuda version that Caffe2 is built with.
bool	HasCudaGPU ()
	Check if the current running session has a cuda gpu present. More...
int	CAFFE_GET_BLOCKS (const int N)
	Compute the number of blocks needed to run N threads.
uint32_t	RandomNumberSeed ()
	A function to generate a random number seed that is unique in a best-effort basis, using an ever-incrementing seed and the current time.
CudaMemoryPoolType	GetCudaMemoryPoolType ()
	Gets the current memory pool type used by Caffe2. More...
	CAFFE_KNOWN_TYPE (db::DBReader)
	CAFFE_KNOWN_TYPE (db::Cursor)
void	EventCreateCPU (const DeviceOption &option, Event *event)
void	EventRecordCPU (Event *event, const void *, const char *err_msg)
void	EventFinishCPU (const Event *event)
void	EventWaitCPUCPU (const Event *event, void *)
EventStatus	EventQueryCPU (const Event *event)
const std::string &	EventErrorMessageCPU (const Event *event)
void	EventSetFinishedCPU (const Event *event, const char *err_msg)
void	EventResetCPU (Event *event)
	REGISTER_EVENT_CREATE_FUNCTION (CPU, EventCreateCPU)
	REGISTER_EVENT_RECORD_FUNCTION (CPU, EventRecordCPU)
	REGISTER_EVENT_WAIT_FUNCTION (CPU, CPU, EventWaitCPUCPU)
	REGISTER_EVENT_FINISH_FUNCTION (CPU, EventFinishCPU)
	REGISTER_EVENT_QUERY_FUNCTION (CPU, EventQueryCPU)
	REGISTER_EVENT_ERROR_MESSAGE_FUNCTION (CPU, EventErrorMessageCPU)
	REGISTER_EVENT_SET_FINISHED_FUNCTION (CPU, EventSetFinishedCPU)
	REGISTER_EVENT_RESET_FUNCTION (CPU, EventResetCPU)
bool	EventCanScheduleCPU (const Event *, const Event *)
void	EventCreateCUDA (const DeviceOption &option, Event *event)
void	EventRecordCUDA (Event *event, const void *context, const char *err_msg)
void	EventFinishCUDA (const Event *event)
void	EventWaitCUDACUDA (const Event *event, void *context)
void	EventWaitCPUCUDA (const Event *event, void *context)
void	EventWaitCUDACPU (const Event *event, void *context)
EventStatus	EventQueryCUDA (const Event *event)
const std::string &	EventErrorMessageCUDA (const Event *event)
void	EventSetFinishedCUDA (const Event *event, const char *err_msg)
void	EventResetCUDA (Event *event)
	REGISTER_EVENT_CREATE_FUNCTION (CUDA, EventCreateCUDA)
	REGISTER_EVENT_RECORD_FUNCTION (CUDA, EventRecordCUDA)
	REGISTER_EVENT_WAIT_FUNCTION (CUDA, CUDA, EventWaitCUDACUDA)
	REGISTER_EVENT_WAIT_FUNCTION (CPU, CUDA, EventWaitCPUCUDA)
	REGISTER_EVENT_WAIT_FUNCTION (CUDA, CPU, EventWaitCUDACPU)
	REGISTER_EVENT_FINISH_FUNCTION (CUDA, EventFinishCUDA)
	REGISTER_EVENT_QUERY_FUNCTION (CUDA, EventQueryCUDA)
	REGISTER_EVENT_ERROR_MESSAGE_FUNCTION (CUDA, EventErrorMessageCUDA)
	REGISTER_EVENT_SET_FINISHED_FUNCTION (CUDA, EventSetFinishedCUDA)
	REGISTER_EVENT_RESET_FUNCTION (CUDA, EventResetCUDA)
	REGISTER_EVENT_WAIT_FUNCTION (MKLDNN, CUDA, EventWaitCPUCUDA)
	REGISTER_EVENT_WAIT_FUNCTION (CUDA, MKLDNN, EventWaitCUDACPU)
	CAFFE_DEFINE_REGISTRY (Caffe2FlagsRegistry, Caffe2FlagParser, const string &)
void	SetUsageMessage (const string &str)
	Sets the usage message when a commandline tool is called with "--help".
const char *	UsageMessage ()
	Returns the usage message for the commandline tool set by SetUsageMessage.
bool	ParseCaffeCommandLineFlags (int *pargc, char ***pargv)
	Parses the commandline flags. More...
bool	CommandLineFlagsHasBeenParsed ()
	Checks if the commandline flags has already been passed.
	CAFFE_DECLARE_REGISTRY (Caffe2FlagsRegistry, Caffe2FlagParser, const string &)
OperatorDef *	AddOp (NetDef *netdef_ptr, string op_type, std::vector< string > inputs, std::vector< string > outputs)
bool	MatchStrings (string p, string s)
	This allows for the use of * and | to match operator types, engines, or any other property that is represented by strings. More...
bool	MatchArguments (const OperatorDef &p_op, const OperatorDef &g_op)
	This ensures that each named arg that exists in the pattern exists in g_op, is equal in value.
bool	GlobalInit (int *pargc, char ***argv)
	Initialize the global environment of caffe2. More...
bool	Caffe2CheckIntrinsicsFeatures (int *, char ***)
	REGISTER_CAFFE2_INIT_FUNCTION (Caffe2CheckIntrinsicsFeatures,&Caffe2CheckIntrinsicsFeatures,"Check intrinsics compatibility between the CPU feature and the binary.")
std::string	StripBasename (const std::string &full_path)
size_t	ReplaceAll (string &s, const char *from, const char *to)
void	SetStackTraceFetcher (std::function< string(void)> fetcher)
void	SetOperatorLogger (std::function< void(const OperatorDef &)> tracer)
std::function< void(const OperatorDef &)>	GetOperatorLogger ()
bool	InitCaffeLogging (int *argc, char **argv)
void	ShowLogInfoToStderr ()
	A utility to allow one to show log info to stderr after the program starts. More...
constexpr bool	IsUsingGoogleLogging ()
void	MakeStringInternal (std::stringstream &)
template<typename T >
void	MakeStringInternal (std::stringstream &ss, const T &t)
template<typename T , typename... Args>
void	MakeStringInternal (std::stringstream &ss, const T &t, const Args &...args)
template<typename... Args>
string	MakeString (const Args &...args)
template<>
string	MakeString (const string &str)
string	MakeString (const char *c_str)
template<class Container >
string	Join (const string &delimiter, const Container &v)
template<class T >
void	LogMessageFatal (const char *file, int line, const T &message)
template<typename T >
T &	CheckNotNullCommon (const char *file, int line, const char *names, T &t)
template<typename T >
T *	CheckNotNull (const char *file, int line, const char *names, T *t)
template<typename T >
T &	CheckNotNull (const char *file, int line, const char *names, T &t)
template<class First , class Second >
std::ostream &	operator<< (std::ostream &out, const std::pair< First, Second > &p)
template<class Iter >
void	PrintSequence (std::ostream &ss, Iter begin, Iter end)
const CaffeMap< string, const ModuleSchema * > &	CurrentModules ()
	Current Modules present in the Caffe2 runtime. More...
bool	HasModule (const string &name)
	Checks whether a module is already present in the current binary.
void	LoadModule (const string &name, const string &filename="")
	Load a module. More...
	CAFFE_DEFINE_REGISTRY (NetRegistry, NetBase, const std::shared_ptr< const NetDef > &, Workspace *)
void	AddGlobalNetObserverCreator (NetObserverCreator creator)
unique_ptr< NetBase >	CreateNet (const NetDef &net_def, Workspace *ws)
	Creates a network, accessing / creating blobs in the given workspace. More...
unique_ptr< NetBase >	CreateNet (const std::shared_ptr< const NetDef > &net_def, Workspace *ws)
	CAFFE_DECLARE_REGISTRY (NetRegistry, NetBase, const std::shared_ptr< const NetDef > &, Workspace *)
	CAFFE_DEFINE_SHARED_REGISTRY (ThreadPoolRegistry, TaskThreadPool, const DeviceOption &)
	CAFFE_REGISTER_CREATOR (ThreadPoolRegistry, CPU, AsyncNetCPUThreadPoolCreator)
std::shared_ptr< TaskThreadPool >	GetAsyncNetCPUThreadPool (int numa_node_id)
	CAFFE_DECLARE_SHARED_REGISTRY (ThreadPoolRegistry, TaskThreadPool, const DeviceOption &)
	REGISTER_NET (async_dag, AsyncDAGNet)
std::shared_ptr< TaskThreadPool >	GetAsyncNetGPUThreadPool (int gpu_id)
	CAFFE_REGISTER_CREATOR (ThreadPoolRegistry, CUDA, AsyncNetGPUThreadPoolCreator)
	REGISTER_NET (async_polling, AsyncPollingNet)
	REGISTER_NET (async_scheduling, AsyncSchedulingNet)
	REGISTER_NET (dag, DAGNet)
	REGISTER_NET (simple, SimpleNet)
	REGISTER_NET (async_simple, AsyncSimpleNet)
bool	IsNUMAEnabled ()
void	NUMABind (int numa_node_id)
int	GetNUMANode (const void *ptr)
int	GetNumNUMANodes ()
void	NUMAMove (void *ptr, size_t size, int numa_node_id)
int	GetCurrentNUMANode ()
const std::string	OpRegistryKey (const std::string &op_type, const std::string &engine)
void	SetPerOpEnginePref (const PerOpEnginePrefType &per_op_engine_pref)
void	SetGlobalEnginePref (const GlobalEnginePrefType &global_engine_pref)
void	SetEnginePref (const PerOpEnginePrefType &per_op_engine_pref, const GlobalEnginePrefType &global_engine_pref)
void	SetOpEnginePref (const std::string &op_type, const CaffeMap< int, EnginePrefType > &op_pref)
unique_ptr< OperatorBase >	CreateOperator (const OperatorDef &operator_def, Workspace *ws, int net_position)
std::map< int32_t, OperatorRegistry * > *	gDeviceTypeRegistry ()
	CAFFE_DEFINE_REGISTRY (CPUOperatorRegistry, OperatorBase, const OperatorDef &, Workspace *)
	CAFFE_REGISTER_DEVICE_TYPE (DeviceType::CPU, CPUOperatorRegistry)
	CAFFE_DEFINE_REGISTRY (CUDAOperatorRegistry, OperatorBase, const OperatorDef &, Workspace *)
	CAFFE_REGISTER_DEVICE_TYPE (DeviceType::CUDA, CUDAOperatorRegistry)
	CAFFE_DEFINE_REGISTRY (GradientRegistry, GradientMakerBase, const OperatorDef &, const vector< GradientWrapper > &)
GradientOpsMeta	GetGradientForOp (const OperatorDef &def, const vector< GradientWrapper > &g_output)
	Gets the GradientOpsMeta for the given operator def.
TensorShape	GetTensorShapeOfBlob (const Blob *b)
TensorShapes	InferBlobShapesAndTypesFromWorkspace (Workspace *ws, const vector< std::unique_ptr< NetDef >> &nets)
TensorShapes	InferBlobShapesAndTypesFromMap (const CaffeMap< std::string, std::vector< TIndex >> &blob_dimensions, const vector< std::unique_ptr< NetDef >> &nets)
std::map< string, std::pair< DeviceOption, DeviceOption > >	ValidateTensorDevices (OperatorBase &op, const OperatorDef &op_def)
std::set< std::string >	GetRegisteredOperators ()
	CAFFE2_DEFINE_TENSOR_TYPES_DISPATCHER (TensorTypes, DoRunWithType, DoRunWithOtherType) CAFFE2_DEFINE_TENSOR_TYPES_DISPATCHER(TensorTypes2
	CAFFE_DECLARE_REGISTRY (CPUOperatorRegistry, OperatorBase, const OperatorDef &, Workspace *)
	CAFFE_DECLARE_REGISTRY (CUDAOperatorRegistry, OperatorBase, const OperatorDef &, Workspace *)
	CAFFE_DECLARE_REGISTRY (GradientRegistry, GradientMakerBase, const OperatorDef &, const vector< GradientWrapper > &)
std::ostream &	operator<< (std::ostream &out, const OpSchema &schema)
template<typename T_I = int>
TensorShape	CreateTensorShape (vector< T_I > dims,::caffe2::TensorProto_DataType dt)
vector< TIndex >	GetDimsVector (const TensorShape &shape)
std::pair< std::vector< DeviceOption >, std::vector< DeviceOption > >	InferOpInputOutputDevice (const OperatorDef &op)
template<uint64_t OpsPerPoint>
OpSchema::Cost	PointwiseCostInference (const OperatorDef &, const vector< TensorShape > &inputs)
bool	RunPlanOnWorkspace (Workspace *ws, const PlanDef &plan, ShouldContinue shouldContinue)
	CAFFE_KNOWN_TYPE (QTensor< CPUContext >)
template<typename KeyType >
void	PrintOffendingKey (const KeyType &key)
template<>
void	PrintOffendingKey (const string &key)
template<typename F >
detail::ScopeGuardImplDecay< F >	MakeGuard (F &&f) noexcept(noexcept(detail::ScopeGuardImplDecay< F >(static_cast< F && >(f))))
	ScopeGuard is a general implementation of the "Initialization is Resource Acquisition" idiom. More...
ExportedStatMap	toMap (const ExportedStatList &stats)
	CAFFE_KNOWN_TYPE (Tensor< CPUContext >)
TypeCall	GetTypeCallFunction (CaffeTypeId id)
void	RegisterTypeCallFunction (CaffeTypeId id, TypeCall c)
TensorInfoCall	GetTensorInfoFunction (CaffeTypeId id)
void	RegisterTensorInfoFunction (CaffeTypeId id, TensorInfoCall c)
vector< TIndex >	ToVectorTIndex (const std::vector< int > &src)
	A utility function to convert vector<int> to vector<TIndex>.
TIndex	size_from_dim_ (int k, const vector< TIndex > &dims)
	Return product of all dimensions starting from K.
TIndex	size_to_dim_ (int k, const vector< TIndex > &dims)
TIndex	size_between_dim_ (int k, int l, const vector< TIndex > &dims)
int	canonical_axis_index_ (int axis_index, int ndims)
template<class Context >
TypeMeta	GetTensorType (const void *c)
template<class Context >
vector< TIndex >	GetTensorInfo (const void *c, bool *shares_data, size_t *capacity, DeviceOption *device)
	CAFFE_DEFINE_REGISTRY (TransformRegistry, Transform)
unique_ptr< Transform >	CreateTransform (string key)
NetDef	ApplyTransform (const string &key, const NetDef &netdef)
double	average_net_run_duration (const NetDef &netdef, const NetDef &init_netdef, const int warmup_runs, const int main_runs)
NetDef	ApplyTransformIfFaster (const string &key, const NetDef &netdef, const NetDef &init_netdef, const int warmup_runs, const int main_runs, const double improvement_threshold)
	CAFFE_DECLARE_REGISTRY (TransformRegistry, Transform)
std::map< CaffeTypeId, string > &	gTypeNames ()
std::set< string > &	gRegisteredTypeNames ()
std::mutex &	gCaffe2TypeRegistrationMutex ()
string	Demangle (const char *name)
string	GetExceptionString (const std::exception &e)
	CAFFE_KNOWN_TYPE (float)
	CAFFE_KNOWN_TYPE (int)
	CAFFE_KNOWN_TYPE (std::string)
	CAFFE_KNOWN_TYPE (bool)
	CAFFE_KNOWN_TYPE (uint8_t)
	CAFFE_KNOWN_TYPE (int8_t)
	CAFFE_KNOWN_TYPE (uint16_t)
	CAFFE_KNOWN_TYPE (int16_t)
	CAFFE_KNOWN_TYPE (int64_t)
	CAFFE_KNOWN_TYPE (float16)
	CAFFE_KNOWN_TYPE (double)
	CAFFE_KNOWN_TYPE (char)
	CAFFE_KNOWN_TYPE (std::unique_ptr< std::mutex >)
	CAFFE_KNOWN_TYPE (std::unique_ptr< std::atomic< bool >>)
	CAFFE_KNOWN_TYPE (std::vector< int64_t >)
	CAFFE_KNOWN_TYPE (std::vector< unsigned long >)
	CAFFE_KNOWN_TYPE (bool *)
	CAFFE_KNOWN_TYPE (char *)
	CAFFE_KNOWN_TYPE (int *)
TensorProto::DataType	TypeMetaToDataType (const TypeMeta &meta)
const TypeMeta &	DataTypeToTypeMeta (const TensorProto::DataType &dt)
StorageOrder	StringToStorageOrder (const string &str)
constexpr char	NameScopeSeparator ()
struct	CAFFE2_ALIGNED (2) __f16
template<typename T >
bool	fp16_type ()
template<>
bool	fp16_type< float16 > ()
std::string	GetUniqueName ()
	REGISTER_CPU_OPERATOR (CreateDB, CreateDBOp< CPUContext >)
	OPERATOR_SCHEMA (CreateDB).NumInputs(0).NumOutputs(1)
	NO_GRADIENT (CreateDB)
	REGISTER_CUDA_OPERATOR (CreateDB, CreateDBOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (FileStoreHandlerCreate, FileStoreHandlerCreateOp< CPUContext >)
	NumInputs (0).NumOutputs(1).SetDoc(R"DOC( Creates a unique_ptr<StoreHandler> that uses the filesystem as backing store (typically a filesystem shared between many nodes
such as NFS This store handler is not built to be fast Its recommended use is for integration tests and prototypes where extra dependencies are cumbersome Use an ephemeral path to ensure multiple processes or runs don t interfere DOC	Arg ("path","base path used by the FileStoreHandler").Arg("prefix"
such as NFS This store handler is not built to be fast Its recommended use is for integration tests and prototypes where extra dependencies are cumbersome Use an ephemeral path to ensure multiple processes or runs don t interfere DOC prefix for all keys used by this store	Output (0,"handler","unique_ptr<StoreHandler>")
	NO_GRADIENT (FileStoreHandlerCreateOp)
	REGISTER_CUDA_OPERATOR (FileStoreHandlerCreate, FileStoreHandlerCreateOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (RedisStoreHandlerCreate, RedisStoreHandlerCreateOp< CPUContext >)
host name of Redis server	Arg ("port","port number of Redis server").Arg("prefix"
	NO_GRADIENT (RedisStoreHandlerCreateOp)
	REGISTER_CUDA_OPERATOR (RedisStoreHandlerCreate, RedisStoreHandlerCreateOp< CUDAContext >)
	CAFFE_KNOWN_TYPE (std::unique_ptr< StoreHandler >)
	REGISTER_CPU_OPERATOR (StoreSet, StoreSetOp)
	NumInputs (2).NumOutputs(0).SetDoc(R"DOC( Set a blob in a store. The key is the input blob's name and the value is the data in that blob. The key can be overridden by specifying the 'blob_name' argument. )DOC").Arg("blob_name" = Ai * Bi
alternative key for the	blob (optional)") .Input(0
alternative key for the unique_ptr< StoreHandler >	Input (1,"data","data blob")
	REGISTER_CPU_OPERATOR (StoreGet, StoreGetOp)
	NumInputs (1).NumOutputs(1).SetDoc(R"DOC( Get a blob from a store. The key is the output blob's name. The key can be overridden by specifying the 'blob_name' argument. )DOC").Arg("blob_name"
alternative key for the unique_ptr< StoreHandler >	Output (0,"data","data blob")
	REGISTER_CPU_OPERATOR (StoreAdd, StoreAddOp)
the store initializes it to and then performs the add operation The operation returns the resulting counter value DOC	Arg ("blob_name","key of the counter (required)").Arg("add_value"
the store initializes it to and then performs the add operation The operation returns the resulting counter value DOC value that is	added (optional, default:1)") .Input(0
the store initializes it to and then performs the add operation The operation returns the resulting counter value DOC value that is unique_ptr< StoreHandler >	Output (0,"value","the current value of the counter")
	REGISTER_CPU_OPERATOR (StoreWait, StoreWaitOp)
	NumInputs (1, 2).NumOutputs(0).SetDoc(R"DOC( Wait for the specified blob names to be set. The blob names can be passed either as an input blob with blob names or as an argument. )DOC").Arg("blob_names"
names of the blobs to wait	for (optional)") .Input(0
names of the blobs to wait unique_ptr< StoreHandler >	Input (1,"names","names of the blobs to wait for (optional)")
	REGISTER_CPU_OPERATOR (FC_Decomp, FullyConnectedOpDecomp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (FCGradient_Decomp, FullyConnectedDecompGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (FC_Decomp).NumInputs(4).NumOutputs(1)
	OPERATOR_SCHEMA (FCGradient_Decomp).NumInputs(4).NumOutputs(3
	REGISTER_GRADIENT (FC_Decomp, GetFCDecompGradient)
	REGISTER_CUDA_OPERATOR (FC_Decomp, FullyConnectedOpDecomp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (FCGradient_Decomp, FullyConnectedDecompGradientOp< float, CUDAContext >)
	REGISTER_CPU_OPERATOR (TTContraction, TTContractionOp< float, CPUContext >)
	REGISTER_CUDA_OPERATOR (TTContraction, TTContractionOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (TTContractionGradient, TTContractionGradientOp< float, CUDAContext >)
	REGISTER_CPU_OPERATOR (ImageInput, ImageInputOp< CPUContext >)
	NumInputs (0, 1).NumOutputs(2
INT_MAX	TensorInferenceFunction ([](const OperatorDef &def, const vector< TensorShape > &){vector< TensorShape > out(2);ArgumentHelper helper(def);int batch_size=helper.GetSingleArgument< int >("batch_size", 0);int crop=helper.GetSingleArgument< int >("crop",-1);int color=helper.GetSingleArgument< int >("color", 1);CHECK_GT(crop, 0);out[0]=CreateTensorShape(vector< int >{batch_size, crop, crop, color?3:1}, TensorProto::FLOAT);out[1]=CreateTensorShape(vector< int >{1, batch_size}, TensorProto::INT32);return out;}).SetDoc(R"DOC( Imports and processes images from a database. For each run of the operator
INT_MAX batch_size images will be processed GPUs can optionally be used for part of the processing The following transformations are applied to the image A bounding box is applied to the initial	image (optional)-The image is rescaled either up or down(with the scale argument) or just up(with the minsize argument)-The image is randomly cropped(crop size is passed as an argument but the location of the crop is random except if is_test is passed in which case the image in cropped at the center)-The image is normalized.Each of its color channels can have separate normalization values The dimension of the output image will always be cropxcrop) DOC") .Arg("batch_size"
INT_MAX batch_size images will be processed GPUs can optionally be used for part of the processing The following transformations are applied to the image A bounding box is applied to the initial Number of images to output for each run of the	operator" ".Must be 1 or greater") .Arg ("color","Number of color channels (1 or 3). Defaults to 1").Arg("color_jitter"
INT_MAX batch_size images will be processed GPUs can optionally be used for part of the processing The following transformations are applied to the image A bounding box is applied to the initial Number of images to output for each run of the Whether or not to do color jitter Defaults to	Arg ("img_saturation","Image saturation scale used in color jittering. ""Defaults to 0.4").Arg("img_brightness"
INT_MAX batch_size images will be processed GPUs can optionally be used for part of the processing The following transformations are applied to the image A bounding box is applied to the initial Number of images to output for each run of the Whether or not to do color jitter Defaults to Image brightness scale used in color jittering Defaults to	Arg ("img_contrast","Image contrast scale used in color jittering. ""Defaults to 0.4").Arg("color_lighting"
INT_MAX batch_size images will be processed GPUs can optionally be used for part of the processing The following transformations are applied to the image A bounding box is applied to the initial Number of images to output for each run of the Whether or not to do color jitter Defaults to Image brightness scale used in color jittering Defaults to Whether or not to do color lighting Defaults to	Arg ("color_lighting_std","Std of normal distribution where color lighting"" scaling factor is sampled. Defaults to 0.1").Arg("scale_jitter_type"
INT_MAX batch_size images will be processed GPUs can optionally be used for part of the processing The following transformations are applied to the image A bounding box is applied to the initial Number of images to output for each run of the Whether or not to do color jitter Defaults to Image brightness scale used in color jittering Defaults to Whether or not to do color lighting Defaults to Scale the size of the smallest dimension of the image to this Scale and minsize are mutually exclusive Must be larger than crop	Arg ("minsize","Scale the size of the smallest dimension of the image to"" this only if the size is initially smaller. Scale and minsize are"" mutually exclusive. Must be larger than crop.").Arg("warp"
the other dimension is proportionally scaled Defaults to	Arg ("crop","Size to crop the image to. Must be provided").Arg("mirror"
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to	Arg ("mean","Mean by which to normalize color channels."" Defaults to 0.").Arg("mean_per_channel"
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to Vector of means per color	channel (1 or 3 elements).Defaults to mean argument.Channel order BGR") .Arg("std"
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to Vector of means per color Standard deviation by which to normalize color channels Defaults to	Arg ("std_per_channel","Vector of standard dev. per color channel "" (1 or 3 elements). Defaults to std argument. Channel order is BGR").Arg("bounding_ymin"
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to Vector of means per color Standard deviation by which to normalize color channels Defaults to Bounding box coordinate Defaults	to (none)") .Arg("bounding_xmin"
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to Vector of means per color Standard deviation by which to normalize color channels Defaults to Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults if the input is in Caffe format Defaults to	Arg ("use_gpu_transform","1 if GPU acceleration should be used."" Defaults to 0. Can only be 1 in a CUDAContext").Arg("decode_threads"
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to Vector of means per color Standard deviation by which to normalize color channels Defaults to Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults if the input is in Caffe format Defaults to Number of CPU decode transform threads Defaults to	Arg ("output_type","If gpu_transform, can set to FLOAT or FLOAT16.").Arg("db"
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to Vector of means per color Standard deviation by which to normalize color channels Defaults to Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults if the input is in Caffe format Defaults to Number of CPU decode transform threads Defaults to Name of the	database (if not passed as input)") .Arg("db_type"
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to Vector of means per color Standard deviation by which to normalize color channels Defaults to Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults if the input is in Caffe format Defaults to Number of CPU decode transform threads Defaults to Name of the Type of The sizes of any outputs besides the data and	label (should have a number of elements equal to the number of additional" "outputs)") .Arg("random_scale"
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to Vector of means per color Standard deviation by which to normalize color channels Defaults to Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults if the input is in Caffe format Defaults to Number of CPU decode transform threads Defaults to Name of the Type of The sizes of any outputs besides the data and shortest side desired for image resize Defaults to[-1,-1] or no random resize desired	Input (0,"reader","The input reader (a db::DBReader)").Output(0
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to Vector of means per color Standard deviation by which to normalize color channels Defaults to Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults if the input is in Caffe format Defaults to Number of CPU decode transform threads Defaults to Name of the Type of The sizes of any outputs besides the data and shortest side desired for image resize Defaults to[-1,-1] or no random resize desired Tensor containing the images	Output (1,"label","Tensor containing the labels").Output(2
	NO_GRADIENT (ImageInput)
template<class Context >
bool	RandomSizedCropping (cv::Mat *img, const int crop, std::mt19937 *randgen)
template<class Context >
void	Saturation (float *img, const int img_size, const float alpha_rand, std::mt19937 *randgen)
template<class Context >
void	Brightness (float *img, const int img_size, const float alpha_rand, std::mt19937 *randgen)
template<class Context >
void	Contrast (float *img, const int img_size, const float alpha_rand, std::mt19937 *randgen)
template<class Context >
void	ColorJitter (float *img, const int img_size, const float saturation, const float brightness, const float contrast, std::mt19937 *randgen)
template<class Context >
void	ColorLighting (float *img, const int img_size, const float alpha_std, const std::vector< std::vector< float >> &eigvecs, const std::vector< float > &eigvals, std::mt19937 *randgen)
template<class Context >
void	ColorNormalization (float *img, const int img_size, const int channels, const std::vector< float > &mean, const std::vector< float > &std)
template<class Context >
void	TransformImage (const cv::Mat &scaled_img, const int channels, float *image_data, const bool color_jitter, const float saturation, const float brightness, const float contrast, const bool color_lighting, const float color_lighting_std, const std::vector< std::vector< float >> &color_lighting_eigvecs, const std::vector< float > &color_lighting_eigvals, const int crop, const bool mirror, const std::vector< float > &mean, const std::vector< float > &std, std::mt19937 *randgen, std::bernoulli_distribution *mirror_this_image, bool is_test=false)
template<class Context >
void	CropTransposeImage (const cv::Mat &scaled_img, const int channels, uint8_t *cropped_data, const int crop, const bool mirror, std::mt19937 *randgen, std::bernoulli_distribution *mirror_this_image, bool is_test=false)
	REGISTER_CUDA_OPERATOR (ImageInput, ImageInputOp< CUDAContext >)
template<typename T_IN , typename T_OUT , class Context >
bool	TransformOnGPU (Tensor< Context > &X, Tensor< Context > *Y, Tensor< Context > &mean, Tensor< Context > &std, Context *context)
	REGISTER_EVENT_CREATE_FUNCTION (MKLDNN, EventCreateCPU)
	REGISTER_EVENT_RECORD_FUNCTION (MKLDNN, EventRecordCPU)
	REGISTER_EVENT_WAIT_FUNCTION (MKLDNN, MKLDNN, EventWaitCPUCPU)
	REGISTER_EVENT_WAIT_FUNCTION (MKLDNN, CPU, EventWaitCPUCPU)
	REGISTER_EVENT_WAIT_FUNCTION (CPU, MKLDNN, EventWaitCPUCPU)
	REGISTER_EVENT_FINISH_FUNCTION (MKLDNN, EventFinishCPU)
	REGISTER_EVENT_QUERY_FUNCTION (MKLDNN, EventQueryCPU)
	REGISTER_EVENT_ERROR_MESSAGE_FUNCTION (MKLDNN, EventErrorMessageCPU)
	REGISTER_EVENT_SET_FINISHED_FUNCTION (MKLDNN, EventSetFinishedCPU)
	REGISTER_EVENT_RESET_FUNCTION (MKLDNN, EventResetCPU)
	CAFFE_DECLARE_REGISTRY (MKLOperatorRegistry, OperatorBase, const OperatorDef &, Workspace *)
	CAFFE_KNOWN_TYPE (GLTensor< GLfloat >)
	CAFFE_KNOWN_TYPE (GLTensor< GLhalf >)
	CAFFE_KNOWN_TYPE (GLTensor< half >)
	CAFFE_KNOWN_TYPE (Tensor< GLContext >)
void	EventCreateOPENGL (const DeviceOption &, Event *)
void	EventRecordOPENGL (Event *, const void *, const char *)
void	EventWaitOPENGLOPENGL (const Event *, void *)
void	EventFinishOPENGL (const Event *)
void	EventResetOPENGL (Event *)
	REGISTER_EVENT_CREATE_FUNCTION (OPENGL, EventCreateOPENGL)
	REGISTER_EVENT_RECORD_FUNCTION (OPENGL, EventRecordOPENGL)
	REGISTER_EVENT_WAIT_FUNCTION (OPENGL, OPENGL, EventWaitOPENGLOPENGL)
	REGISTER_EVENT_FINISH_FUNCTION (OPENGL, EventFinishOPENGL)
	REGISTER_EVENT_RESET_FUNCTION (OPENGL, EventResetOPENGL)
template<typename T = half>
void	getTensorCPU (const GLTensor< T > &g_, TensorCPU &g)
	REGISTER_NET (opengl, GLNet)
	CAFFE_DEFINE_REGISTRY (GLOperatorRegistry, OperatorBase, const OperatorDef &, Workspace *)
	CAFFE_REGISTER_DEVICE_TYPE (DeviceType::OPENGL, GLOperatorRegistry)
	CAFFE_DECLARE_REGISTRY (GLOperatorRegistry, OperatorBase, const OperatorDef &, Workspace *)
void	dumpDefForOpenGL (const NetDef &d)
NetDef	rewritePredictNetForOpenGL (const NetDef &predictNet, bool runFusion, std::unordered_set< std::string > cpuOps)
bool	tryConvertToOpenGL (const NetDef &predictNet, NetDef *glPredictNet, bool runFusion, std::unordered_set< std::string > cpuOps)
	REGISTER_GL_OPERATOR (Relu, GLReluOp< half >)
	REGISTER_GL_OPERATOR (Sigmoid, GLSigmoidOp< DataType >)
	REGISTER_GL_OPERATOR (Concat, GLConcatOp< DataType >)
	REGISTER_GL_OPERATOR (Conv, GLConvOp< DataType >)
	REGISTER_GL_OPERATOR (CopyFromGL, CopyFromGLOp< DataType >)
	REGISTER_GL_OPERATOR (Sum, GLSumOp< DataType >)
	REGISTER_GL_OPERATOR (Add, GLSumOp< DataType >)
	REGISTER_GL_OPERATOR (FC, GLFullyConnectedOp< DataType >)
	REGISTER_GL_OPERATOR (NormalizePlanarYUV, GLNormalizePlanarYUVOp< DataType >)
	REGISTER_GL_OPERATOR (AveragePool, GLAveragePoolOp< DataType >)
	REGISTER_GL_OPERATOR (MaxPool, GLMaxPoolOp< DataType >)
	REGISTER_GL_OPERATOR (Reshape, GLReshapeOp< DataType >)
	REGISTER_GL_OPERATOR (ResizeNearest, GLResizeNearestOp< DataType >)
	REGISTER_GL_OPERATOR (Softmax, GLSoftmaxOp< DataType >)
	REGISTER_GL_OPERATOR (SpatialBN, GLSpatialBNOp< DataType >)
void	benchmarkModel (std::string init_net_pb, std::string predict_net_pb, std::string input_name, std::vector< int > input_dims, std::string net_name="benchmark_net", std::unordered_set< std::string > cpu_ops=std::unordered_set< std::string >({}))
template<typename T = float>
void	PopulateCPUBlob (Workspace *ws, bool random, std::string name, std::vector< int > dims, int val=1, int dist_shift=0, float variance=1)
template<typename T = half>
void	compareNetResult (Workspace &ws, NetDef &cpu_net, NetDef &gpu_net, string cpu_blob="ref_Y", string gpu_blob="gpu_Y", double tol=0.01, bool relative=false)
template<typename T = half>
void	compareNetResult4D (Workspace &ws, NetDef &cpu_net, NetDef &gpu_net, string cpu_blob="ref_Y", string gpu_blob="gpu_Y", double tol=0.05)
bool	tryConvertToMPSCNN (const NetDef &initNet, const NetDef &predictNet, NetDef *mpscnnPredictNet)
NetDef	annotateDefWithReadCounts (const NetDef &net)
NetDef	rewriteForMetal (const NetDef &net)
NetDef	runMPSCNNFusion (const NetDef &net)
void	dumpDef (const NetDef &d)
void	mpscnnRecordExecutionFinish ()
MPSCNNContext &	getMPSCNNContext ()
bool	tryConvertToMPSCNNIntermediateCopies (const NetDef &initNet, const NetDef &predictNet, NetDef *mpscnnPredictNet)
NetDef	setSpecialArgs (const NetDef &def)
void	testMPSCNN ()
void	compareModels (const NetDef &initNet, NetDef predictNet)
void	verifyRewrite (const NetDef &initNet, const NetDef &net, std::vector< int > inputDims)
	CAFFE_KNOWN_TYPE (GLImage< float >)
	CAFFE_KNOWN_TYPE (GLImage< uint8_t >)
	CAFFE_KNOWN_TYPE (GLImageVector< float >)
	CAFFE_KNOWN_TYPE (GLImageVector< uint8_t >)
template<class T >
void	shareInputGLImage (Workspace *ws, const std::string &name, GLImageVector< T > *input)
template<class T >
const GLImageVector< T > *	extractOutputGLImage (Workspace *ws, const std::string &name)
const NetDef	create_gl_run_net (const NetDef &init_net, const NetDef &run_net, bool use_texture_input)
NetDef	rewritePredictNetForOpenGL (const NetDef &predictNet, bool useTextureInput, bool useTiling, bool runFusion)
bool	tryConvertToOpenGL (const NetDef &initNet, const NetDef &predictNet, NetDef *glPredictNet, bool useTextureInput, bool useTiling, bool runFusion)
	REGISTER_CPU_OPERATOR (OpenGLAdd, OpenGLAddOp< float16_t >)
	OPERATOR_SCHEMA (OpenGLAdd).NumInputs(2).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLConcat, OpenGLConcatOp< float16_t >)
	OPERATOR_SCHEMA (OpenGLConcat).NumInputs(2
	NumOutputs (1, 2)
	REGISTER_CPU_OPERATOR (OpenGLConv, OpenGLConvOp< float16_t, false, false >)
	OPERATOR_SCHEMA (OpenGLConv).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLConvPRelu, OpenGLConvOp< float16_t, true, false >)
	OPERATOR_SCHEMA (OpenGLConvPRelu).NumInputs(4).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLConvRelu, OpenGLConvOp< float16_t, false, true >)
	OPERATOR_SCHEMA (OpenGLConvRelu).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLConvTranspose, OpenGLConvTransposeOp< float16_t, false, false >)
	OPERATOR_SCHEMA (OpenGLConvTranspose).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLConvTransposePRelu, OpenGLConvTransposeOp< float16_t, true, false >)
	OPERATOR_SCHEMA (OpenGLConvTransposePRelu).NumInputs(4).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLConvTransposeRelu, OpenGLConvTransposeOp< float16_t, false, true >)
	OPERATOR_SCHEMA (OpenGLConvTransposeRelu).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (CopyToOpenGL, CopyToOpenGLOp< float16_t >)
	OPERATOR_SCHEMA (CopyToOpenGL).NumInputs(1).NumOutputs(1).AllowInplace(
	REGISTER_CPU_OPERATOR (CopyFromOpenGL, CopyFromOpenGLOp< float16_t >)
	OPERATOR_SCHEMA (CopyFromOpenGL).NumInputs(1).NumOutputs(1).AllowInplace(
	REGISTER_CPU_OPERATOR (OpenGLInstanceNorm, OpenGLInstanceNormPReluOp< float16_t, false >)
	OPERATOR_SCHEMA (OpenGLInstanceNorm).NumInputs(3
	NumOutputs (1, 3).AllowInplace(
	REGISTER_CPU_OPERATOR (OpenGLInstanceNormPRelu, OpenGLInstanceNormPReluOp< float16_t, true >)
	OPERATOR_SCHEMA (OpenGLInstanceNormPRelu).NumInputs(3
	REGISTER_CPU_OPERATOR (OpenGLMul, OpenGLMulOp< float16_t >)
	REGISTER_CPU_OPERATOR (OpenGLNormalizePlanarYUV, GLNormPlanarYUVOp< float16_t >)
	OPERATOR_SCHEMA (OpenGLNormalizePlanarYUV).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLPadImage, OpenGLPadImageOp< float16_t >)
	OPERATOR_SCHEMA (OpenGLPadImage).NumInputs(1).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLPRelu, OpenGLPReluOp< float16_t, GLPRelu::PRelu >)
	IdenticalTypeAndShape ()
	REGISTER_CPU_OPERATOR (OpenGLRelu, OpenGLPReluOp< float16_t, GLPRelu::Relu >)
	REGISTER_CPU_OPERATOR (OpenGLResizeNearest, OpenGLResizeNearestOp< float16_t >)
	OPERATOR_SCHEMA (OpenGLResizeNearest).NumInputs(1).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLSigmoid, OpenGLSigmoidOp< float16_t, Sigmoid >)
	REGISTER_CPU_OPERATOR (OpenGLTanh, OpenGLSigmoidOp< float16_t, Tanh >)
	REGISTER_CPU_OPERATOR (OpenGLTensorToTextureStylizerPreprocess, OpenGLTensorToTextureStylizerPreprocessOp)
	OPERATOR_SCHEMA (OpenGLTensorToTextureStylizerPreprocess).NumInputs(2).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLTextureToTextureStylizerPreprocess, OpenGLTextureToTextureStylizerPreprocessOp< RGBA >)
	OPERATOR_SCHEMA (OpenGLTextureToTextureStylizerPreprocess).NumInputs(2).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLTextureToTensorStylizerDeprocess, OpenGLTextureToTensorStylizerDeprocessOp)
	OPERATOR_SCHEMA (OpenGLTextureToTensorStylizerDeprocess).NumInputs(2).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLTextureToTextureStylizerDeprocess, OpenGLTextureToTextureStylizerDeprocessOp< RGBA >)
	OPERATOR_SCHEMA (OpenGLTextureToTextureStylizerDeprocess).NumInputs(2).NumOutputs(1)
	REGISTER_CPU_OPERATOR (OpenGLSub, OpenGLSubOp< float16_t >)
	OPERATOR_SCHEMA (OpenGLSub).NumInputs(2).NumOutputs(1)
void	testOpenGL ()
void	compareModelsForOpenGL (std::string name, const NetDef &initNet, NetDef predictNet, int width, int height, int channel, std::string input_type, std::string input_order)
void	compareBatchedToTiledModels (std::string name, const NetDef &initNet, NetDef predictNet, int width, int height, int channel, std::string input_type, std::string input_order)
int	runModelBenchmarks (caffe2::NetDef &init_net, caffe2::NetDef &predict_net, int warm_up_runs, int main_runs, int channel, int height, int width, std::string input_type, std::string input_order, std::string engine, bool run_individual=false, bool use_texture_input=false, bool use_tiling=false, bool run_fusion=true)
std::string &	gSNPELocation ()
	REGISTER_CPU_OPERATOR (SNPE, SNPEOp)
void	uniformQuantize2b1b (const TensorCPU &X, const std::vector< std::unique_ptr< TensorCPU >> &XQ, float offset, float inter_center_distance)
void	qconv (const ConvArgs &args, const TensorCPU &X, const TensorCPU &W, const TensorCPU *b, TensorCPU *Y)
void	qpad_zero (const ConvArgs &args, const TensorCPU &X, TensorCPU *Y)
void	signQuantize (const TensorCPU &X, TensorCPU *XQ)
void	filterNormalization11 (const TensorCPU &WQ, TensorCPU *WQN)
void	filterNormalizationL1 (const TensorCPU &W, TensorCPU *WL1)
void	qim2col (const ConvArgs &args, const TensorCPU &XQ, const TensorCPU &WQ, TensorCPU *XQcol)
std::unique_ptr< QConvState >	create2b1bConvState (Workspace *ws, const TensorCPU &W, const TensorCPU *b)
void	run2b1bConvGeneric (QConvState *state, const ConvArgs &args, const TensorCPU &X, TensorCPU *Y)
void	run2b1bUnification (QConvState *state, size_t N, size_t C, const float *WQNVdata, const float *YQs0Vdata, const float *YQs1Vdata, size_t YQstride, float *Ydata, size_t Ystride, const float *bias)
	REGISTER_CPU_OPERATOR (QConv, QConvOp)
size_t	divRoundUp (size_t x, size_t d)
bool	run2b1bConvNeon (QConvState *state, const ConvArgs &args, const TensorCPU &X, TensorCPU *Y)
	CAFFE_KNOWN_TYPE (MPICommonWorldWrapper)
std::mutex &	MPIMutex ()
MPI_Comm	GlobalMPIComm ()
	Gets the global MPI communicator used by Caffe2. More...
void	SetGlobalMPIComm (MPI_Comm new_comm)
	Sets the global MPI communicator. More...
int	MPICommSize (MPI_Comm comm)
	A helper function to return the size of the given communicator.
int	MPICommRank (MPI_Comm comm)
	A helper function to return the rank of the given communicator.
void	MPISetupPeers (const int replicas, const string &role, const string &job_path)
	A function used to perform peer setup so one does not need to use mpirun / mpiexec to run the binary. More...
void	CheckInitializedMPI ()
	REGISTER_CPU_OPERATOR (Abs, UnaryElementwiseOp< TensorTypes< float >, CPUContext, AbsCPUFunctor >)
	REGISTER_CPU_OPERATOR (AbsGradient, BinaryElementwiseOp< TensorTypes< float >, CPUContext, WithoutBroadcast< AbsGradientCPUFunctor >>)
element wise DOC	Input (0,"input","Input tensor").Output(0
element wise DOC The absolute value of the input tensor computed element wise	InheritOnnxSchema ("Abs")
	OPERATOR_SCHEMA (AbsGradient).NumInputs(2).NumOutputs(1).IdenticalTypeAndShape()
	REGISTER_GRADIENT (Abs, GetAbsGradient)
	REGISTER_CPU_OPERATOR (Accumulate, AccumulateOp< float, CPUContext >)
we first initialize the output tensor to all and then do accumulation Any further calls to the	operator, given that no one else fiddles with the output in the interim, will do simple accumulations.Accumulation is done using Axpby operation as shown:Y=1 *X+gamma *Y where X is the input tensor, Y is the output tensor and gamma is the multiplier argument.) DOC") .Arg ("gamma","(float, default 1.0) Accumulation multiplier").Input(0
we first initialize the output tensor to all and then do accumulation Any further calls to the The input tensor that has to be accumulated to the output tensor If the output size is not the same as input the output tensor is first reshaped and initialized to and only accumulation is done	Output (0,"output","Accumulated output tensor")
	SHOULD_NOT_DO_GRADIENT (Accumulate)
	REGISTER_CPU_OPERATOR (Accuracy, AccuracyOp< float, CPUContext >)
NumInputs(2).NumOutputs(1).ScalarType(TensorProto	SHOULD_NOT_DO_GRADIENT (Accuracy)
	REGISTER_CPU_OPERATOR (RowWiseArgMax, RowWiseArgMaxOp< CPUContext >)
this	operator returns a 2D (N X 1) output tensor with with the index of the maximum value in each row.If there are duplicate max values in a row the index of the first occurence is returned.) DOC") .Input(0
this N X D input tensor	Output (0,"Z","2D (N X 1) output tensor")
	NO_GRADIENT (RowWiseArgMax)
	REGISTER_CPU_OPERATOR (ArgMax, ArgMaxOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (ArgMin, ArgMinOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (Assert, AssertOp< CPUContext >)
or long longs and checks if all values are true when coerced into a boolean In other for non bool types this asserts that all values in the tensor are non zero DOC	Arg ("error_msg","An error message to print when the assert fails.", false)
	REGISTER_CPU_OPERATOR (BatchGather, BatchGatherOp< CPUContext >)
	REGISTER_CPU_OPERATOR (BatchGatherGradient, BatchGatherGradientOp< CPUContext >)
output_dims	push_back (data_dims[0])
output_dims	insert (output_dims.end(), indices_dims.begin(), indices_dims.end())
output_dims	insert (output_dims.end(), data_dims.begin()+2, data_dims.end())
	SetDoc (R"DOC( Batch gather operation, first dimension in DATA is the batch size. Given DATA tensor of rank r >= 2, and INDICES tensor of rank q >= 1, gather entries of the outer-most dimension of DATA indexed by INDICES, and concatenate them in an output tensor of rank (q - 1) + (r - 1). Example: DATA = [ [1.0, 1.2, 2.4, 4.5], [2.3, 3.4, 3.6, 2.3], [4.5, 5.7, 1.2, 4.5], ] INDICES = [ [0, 2], ] OUTPUT = [ [1.0, 2.4], [2.3, 3.6], [4.5, 1.2], ] )DOC").Input(0
Tensor of rank of any rank q	Output (0,"OUTPUT","Tensor of rank (q - 1) + (r - 1).")
	OPERATOR_SCHEMA (BatchGatherGradient).NumInputs(3).NumOutputs(1)
	REGISTER_GRADIENT (BatchGather, GetBatchGatherGradient)
	REGISTER_CPU_OPERATOR (BatchMatMul, BatchMatMulOp< CPUContext >)
vector< TensorShape >	TensorInferenceForBatchMatMul (const OperatorDef &def, const vector< TensorShape > &in)
OpSchema::Cost	CostInferenceForBatchMatMul (const OperatorDef &def, const vector< TensorShape > &in)
where A has	shape (dim0, dim1,...M, K)
where A has B has	shape (dim0, dim1,...K, N)
	REGISTER_CPU_OPERATOR (BatchSparseToDense, BatchSparseToDenseOp< float, CPUContext >)
	NumInputs (3, 4).NumOutputs(1).SetDoc(R"DOC( Convert sparse matrix representation into dense matrix. A sparse matrix is represented by `lengths` vector
indices and values vector Each element in lengths	vector (lengths[`i`]) represents the number of indices in this batch(batch`i`).With in each batch
	REGISTER_CPU_OPERATOR (BooleanMask, BooleanMaskOp< CPUContext >)
	REGISTER_CPU_OPERATOR (BooleanMaskLengths, BooleanMaskLengthsOp< CPUContext >)
	SetDoc (R"DOC( Given a data tensor and a 1D boolean mask tensor, returns a tensor containing only the elements corresponding to positions where the mask is true. )DOC").Input(0
original data tensor	Input (1,"mask","A tensor of bools of same shape as `data`.").Output(0
original data tensor A tensor of same type as data	Output (1,"masked_indices","A tensor for indices.")
return the segment lengths of a corresponding segmented tensor after BooleanMask is applied DOC	Input (0,"lengths","A 1D int32 tensor representing segment lengths.").Input(1
return the segment lengths of a corresponding segmented tensor after BooleanMask is applied DOC A bool tensor of values to keep	Output (0,"masked_lengths","Segment lengths of a masked tensor.")
	NO_GRADIENT (BooleanMaskLengths)
template<typename Functor >
void	MaskWithFunctor (size_t N, size_t M, int B, const float *in, Functor fn, float fill_val, float *out)
template<typename Functor >
void	RepeatedMaskWithFunctor (size_t N, size_t M, int D, const float *in, Functor fn, float fill_val, float *out)
	REGISTER_CPU_OPERATOR (SequenceMask, SequenceMaskOp< CPUContext >)
	REGISTER_CPU_OPERATOR (BooleanUnmask, BooleanUnmaskOp< CPUContext >)
	NumInputs ([](int n){return n > 0 &&n%2==0;}).NumOutputs(1).SetDoc(R"DOC( Given a series of mask and values
reconstruct values together according to masks A comprehensive False False True Reconstruct We Note that for all mask there must be at least one True If for a field there are multiple True we will accept the first value For False False False True DOC	Output (0,"unmasked_data","The final reconstructed unmasked data")
	REGISTER_CPU_OPERATOR (Cast, CastOp< CPUContext >)
out	push_back (in[0])
out[0]	set_data_type (cast::GetCastDataType(helper,"to"))
	SetDoc (R"DOC( The operator casts the elements of a given input tensor to a data type specified by the 'to' argument and returns an output tensor of the same size in the converted type. The 'to' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message. If the 'to' argument is not provided or is not one of the enumerated types in DataType, Caffe2 throws an Enforce error. NOTE: Casting to and from strings is not supported yet. )DOC").Arg("to"
The data type to which the elements of the input tensor are cast Strictly must be one of the types from DataType enum in TensorProto	Input (0,"input","Input tensor to be cast.").Output(0
	REGISTER_GRADIENT (Cast, GetCastGradient)
	REGISTER_CPU_OPERATOR (Ceil, CeilOp< float, CPUContext >)
	SetDoc (R"DOC( Ceil takes one input data (Tensor<T>) and produces one output data (Tensor<T>) where the ceil function, y = ceil(x), is applied to the tensor elementwise. Currently supports only float32. )DOC").Input(0
ND input tensor	Output (0,"Y","ND input tensor")
	GRADIENT_NOT_IMPLEMENTED_YET (Ceil)
	REGISTER_CPU_OPERATOR (ChannelBackpropStats, ChannelBackpropStatsOp< CPUContext >)
	NumInputs (4).NumOutputs(2).SetDoc(R"DOC( Given an input tensor in NCHW format
the gradient for the output of SpatialBN and the per channel mean and inverse std var vectors for the computes the per channel bias and scale gradient to be used during the backward pass for subsequent spatial batch normalization gradient calculation the results of this op are subsequently reduced over multiple devices to obtain statistics over a larger batch size in cases where the batch size for a single model copy is too low to yield the full benefit of batch normalization The resulting bias and scale can then be plugged back into SpatialBNGradient to get results over the larger batch size DOC	Input (0,"X","The input 4-dimensional tensor of shape NCHW").Input(1
the gradient for the output of SpatialBN and the per channel mean and inverse std var vectors for the computes the per channel bias and scale gradient to be used during the backward pass for subsequent spatial batch normalization gradient calculation the results of this op are subsequently reduced over multiple devices to obtain statistics over a larger batch size in cases where the batch size for a single model copy is too low to yield the full benefit of batch normalization The resulting bias and scale can then be plugged back into SpatialBNGradient to get results over the larger batch size DOC The mean saved from the forward pass as a dimensional tensor of size C	Input (2,"inv_std","The saved inverse standard deviation as a 1-dimensional tensor ""of size C.").Input(3
the gradient for the output of SpatialBN and the per channel mean and inverse std var vectors for the computes the per channel bias and scale gradient to be used during the backward pass for subsequent spatial batch normalization gradient calculation the results of this op are subsequently reduced over multiple devices to obtain statistics over a larger batch size in cases where the batch size for a single model copy is too low to yield the full benefit of batch normalization The resulting bias and scale can then be plugged back into SpatialBNGradient to get results over the larger batch size DOC The mean saved from the forward pass as a dimensional tensor of size C Gradient for the output layer of here used as input because we are on the backward pass	Output (0,"scale_grad","Gradient for the scale vector").Output(1
	SHOULD_NOT_DO_GRADIENT (ChannelBackpropStats)
	REGISTER_CPU_OPERATOR (ChannelShuffle, ChannelShuffleOp< CPUContext >)
	REGISTER_CPU_OPERATOR (ChannelShuffleGradient, ChannelShuffleGradientOp< CPUContext >)
	REGISTER_GRADIENT (ChannelShuffle, GetChannelShuffleGradient)
	REGISTER_CPU_OPERATOR (ChannelStats, ChannelStatsOp< CPUContext >)
computes the sum of all elements per channel and the sum of all elements squared per channel These values can be reduced across multiple batches and used to obtain the mean and variance across the full set of batches Using the new mean and variance as input to SpatialBN has the effect of changing the batch size over which SpatialBN is applied DOC The output dimensional tensor of size C containing the sum of elements of X per channel	Output (1,"sumsq","The output 1-dimensional tensor of size C containing the sum of ""elements squared per channel.")
	SHOULD_NOT_DO_GRADIENT (ChannelStats)
	REGISTER_CPU_OPERATOR (Clip, ClipOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (ClipGradient, ClipGradientOp< float, CPUContext >)
Key value handler for	rendezvous (optional).") .Output(0
Key value handler for A common world for collective operations	Arg ("size","(int) size of the common world.").Arg("rank"
Existing common world to clone	Output (0,"comm_world","A common world for collective operations.")
	SetDoc ("Closes all connections managed by a common world.").Input(0
	NumInputsOutputs ([](int in, int out){return in >=2 &&out==(in-1);}).EnforceInplace([](int in
	InputsCanCrossDevices ().IdenticalTypeAndShapeOfInput(0).SetDoc(R"DOC( Does a broadcast operation from the root node to every other node. The tensor on each node should have been pre-created with the same shape and data type. )DOC").Input(0
The common world	Input (1,"X","A tensor to be broadcasted.").Output(0
The common world In place as input	Arg ("root","(int, default 0) the root to run broadcast from.")
The common world	Input (1,"X","A tensor to be reduced.").Output(0
The common world The reduced result on not set for other nodes	Arg ("root","(int, default 0) the root to run reduce into.")
	IdenticalTypeAndShapeOfInput (0).InputsCanCrossDevices().SetDoc(R"DOC( Does an allreduce operation among the nodes. Currently only Sum is supported. )DOC").Input(0
The common world	Input (1,"X","A tensor to be allreduced.").Output(0
The common world	Input (1,"X","A tensor to be reduce-scattered.").Output(0
	NumInputs (2, INT_MAX).NumOutputs(1).InputsCanCrossDevices().SetDoc(R"DOC( Does an allgather operation among the nodes. )DOC").Input(0
The common world	Input (1,"X","A tensor to be allgathered.").Output(0
	NumInputs ({2, 4}).NumOutputs(0).SetDoc(R"DOC( Sends the tensor to another node. )DOC").Input(0
The common world An int CPUtensor of size specifying the rank If this overrides the to argument of the op	Input (3,"tag","An int CPUtensor of size 1 specifying the tag to ""send the tensor with. This overrides the 'tag' ""argument of the op.").Arg("dst"
The common world An int CPUtensor of size specifying the rank If this overrides the to argument of the op The rank to send the tensor to	Arg ("tag","(int) a tag to send the tensor with.").Arg("raw_buffer"
	AllowInplace ({{2, 1},{3, 2}}).SetDoc(R"DOC( Receives the tensor from another node. )DOC").Input(0
The common world	Input (1,"Y","In-place output. If raw_buffer is specified, ""Y should have pre-allocated data and type..").Input(2
The common world An int CPUtensor of size specifying the rank If this overrides the from argument of the op The received tensor	Output (1,"src","The sender that sent the message as a CPUTensor ""of size 1 and of type int.").Output(2
The common world An int CPUtensor of size specifying the rank If this overrides the from argument of the op The received tensor The tag that the message is sent with as a CPUTensor of size and of type int	Arg ("src","(int) he rank to receive the tensor from.").Arg("tag"
The common world An int CPUtensor of size specifying the rank If this overrides the from argument of the op The received tensor The tag that the message is sent with as a CPUTensor of size and of type int int a tag to receive the tensor with	Arg ("raw_buffer","(bool) if set, only send the content and assume that the receiver ""has already known the tensor's shape and information.")
	SHOULD_NOT_DO_GRADIENT (CreateCommonWorld)
	SHOULD_NOT_DO_GRADIENT (CloneCommonWorld)
	SHOULD_NOT_DO_GRADIENT (DestroyCommonWorld)
	SHOULD_NOT_DO_GRADIENT (Broadcast)
	SHOULD_NOT_DO_GRADIENT (Reduce)
	SHOULD_NOT_DO_GRADIENT (Allgather)
	SHOULD_NOT_DO_GRADIENT (Allreduce)
	SHOULD_NOT_DO_GRADIENT (ReduceScatter)
	SHOULD_NOT_DO_GRADIENT (Barrier)
	SHOULD_NOT_DO_GRADIENT (SendTensor)
	SHOULD_NOT_DO_GRADIENT (ReceiveTensor)
	REGISTER_CPU_OPERATOR (CreateCommonWorld, NoDefaultEngineOp< CPUContext >)
	REGISTER_CPU_OPERATOR (CloneCommonWorld, NoDefaultEngineOp< CPUContext >)
	REGISTER_CPU_OPERATOR (DestroyCommonWorld, NoDefaultEngineOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Broadcast, NoDefaultEngineOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Reduce, NoDefaultEngineOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Allgather, NoDefaultEngineOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Allreduce, NoDefaultEngineOp< CPUContext >)
	REGISTER_CPU_OPERATOR (ReduceScatter, NoDefaultEngineOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Barrier, NoDefaultEngineOp< CPUContext >)
	REGISTER_CPU_OPERATOR (SendTensor, NoDefaultEngineOp< CPUContext >)
	REGISTER_CPU_OPERATOR (ReceiveTensor, NoDefaultEngineOp< CPUContext >)
	REGISTER_CUDA_OPERATOR (CreateCommonWorld, NoDefaultEngineOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (CloneCommonWorld, NoDefaultEngineOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (Broadcast, NoDefaultEngineOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (Reduce, NoDefaultEngineOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (Allgather, NoDefaultEngineOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (Allreduce, NoDefaultEngineOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (SendTensor, NoDefaultEngineOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (ReceiveTensor, NoDefaultEngineOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (Split, SplitOp< CPUContext >)
INT_MAX	Input (0,"input","The tensor to split").Input(1
INT_MAX Optional list of output	lengths (see also arg 'split')") .Arg("axis"
INT_MAX Optional list of output Which axis to split on	Arg ("split","length of each output").Arg("order"
INT_MAX Optional list of output Which axis to split on Either NHWC or will split on C defaults to NCHW	DeviceInferenceFunction (splitOpDevInfer).SetDoc(R"DOC( Split a tensor into a list of tensors
	REGISTER_CPU_OPERATOR (Concat, ConcatOp< CPUContext >)
	NumInputs (1, INT_MAX).NumOutputs(2).Arg("axis"
Which axis to concat on	Arg ("order","Either NHWC or NCHW, will concat on C axis, defaults to NCHW").Arg("add_axis"
Which axis to concat on Pass to add the axis specified in arg axis to all input tensors	TensorInferenceFunction ([](const OperatorDef &def, const vector< TensorShape > &in){ArgumentHelper helper(def);const int axis=helper.HasArgument("axis")?helper.GetSingleArgument< int >("axis",-1):GetDimFromOrderString(helper.GetSingleArgument< string >("order","NCHW"));bool add_axis=helper.GetSingleArgument< int >("add_axis", 0)!=0;const int canonical_axis=canonical_axis_index_(axis, in[0].dims_size());CAFFE_ENFORCE_GT(in.size(), 0);vector< int > split_shape(1, in.size());vector< int > out_shape(in[0].dims().begin(), in[0].dims().end());if(add_axis){for(int i=1;i< in.size();++i){CAFFE_ENFORCE_EQ(in[0].dims().size(), in[i].dims().size(),"All inputs of Concat should have same dims when add_axis = 1. ""Got different sizes for inputs 0 and ", i);for(int j=0;j< in[0].dims().size();++j){CAFFE_ENFORCE_EQ(in[0].dims(j), in[i].dims(j),"All inputs of Concat should have same dims when add_axis = 1. ""Got different dims for inputs 0 and ", i,". At dim: ", j);}}out_shape.insert(out_shape.begin()+canonical_axis, in.size());}else{for(int i=1;i< in.size();++i){CAFFE_ENFORCE_EQ(in[0].dims().size(), in[i].dims().size(),"All inputs of Concat should have same dims except ""canonical_axis dim that is equal to ", canonical_axis,"Got different sizes for inputs 0 and ", i);for(int j=0;j< in[0].dims().size();++j){if(j==canonical_axis){continue;}CAFFE_ENFORCE_EQ(in[0].dims(j), in[i].dims(j),"All inputs of Concat should have same dims except ""canonical_axis dim that is equal to ", canonical_axis,"Got different dims for inputs 0 and ", i,". At dim: ", j);}}for(int i=1;i< in.size();++i){out_shape[canonical_axis]+=in[i].dims(canonical_axis);}}if(def.output_size()==1){return vector< TensorShape >{CreateTensorShape(out_shape, in[0].data_type())};}return vector< TensorShape >{CreateTensorShape(out_shape, in[0].data_type()), CreateTensorShape(split_shape, TensorProto::INT32)};}).CostInferenceFunction(CostInferenceForConcat).DeviceInferenceFunction(concatOpDevInfer).SetDoc("Concatenate a list of tensors into a single tensor").Output(0
Which axis to concat on Pass to add the axis specified in arg axis to all input tensors Concatenated tensor	Output (1,"split_info","The dimensions of the inputs.").InheritOnnxSchema("Concat")
	REGISTER_CPU_OPERATOR (DepthSplit, SplitOp< CPUContext >)
	REGISTER_CPU_OPERATOR (DepthConcat, ConcatOp< CPUContext >)
INT_MAX	SetDoc ("Backward compatible operator name for Split.")
	REGISTER_GRADIENT (Split, GetSplitGradient)
	REGISTER_GRADIENT (DepthSplit, GetSplitGradient)
	REGISTER_GRADIENT (Concat, GetConcatGradient)
	REGISTER_GRADIENT (DepthConcat, GetConcatGradient)
	REGISTER_CUDA_OPERATOR (Split, SplitOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (Concat, ConcatOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (DepthSplit, SplitOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (DepthConcat, ConcatOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (Conditional, ConditionalOp< CPUContext >)
	NumInputs (3).NumOutputs(1).SetDoc(R"DOC( Given a 1-D tensor of boolean values
apply conditional	operator along the first dimension of DataT and DataF and return DataO.Note, DataT and DataF must have the exact same shape and type.) DOC") .Input (0,"Condition","Boolean tensor to select DataT or DataF").Input(1
apply conditional Data to use when True	Input (2,"DataF","Data to use when False").Output(0
	NO_GRADIENT (Conditional)
	REGISTER_CPU_OPERATOR (ConvGradient, ConvGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (ConvGradient).NumInputs(2
	REGISTER_CPU_OPERATOR (Conv1DGradient, ConvGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (Conv1DGradient).NumInputs(2
	REGISTER_CPU_OPERATOR (Conv2DGradient, ConvGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (Conv2DGradient).NumInputs(2
	REGISTER_CPU_OPERATOR (Conv3DGradient, ConvGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (Conv3DGradient).NumInputs(2
	REGISTER_GRADIENT (Conv, GetConvGradient)
	REGISTER_GRADIENT (Conv1D, GetConvGradient)
	REGISTER_GRADIENT (Conv2D, GetConvGradient)
	REGISTER_GRADIENT (Conv3D, GetConvGradient)
std::function< void(OpSchema &)>	ConvDocGenerator (const char *dim)
	REGISTER_CPU_OPERATOR (Conv, ConvOp< float, CPUContext >)
NumInputs(2, 3).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (Conv1D, ConvOp< float, CPUContext >)
NumInputs(2, 3).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (Conv2D, ConvOp< float, CPUContext >)
NumInputs(2, 3).NumOutputs(1).CostInferenceFunction(OpSchema	REGISTER_CPU_OPERATOR (Conv3D, ConvOp< float, CPUContext >)
	REGISTER_CUDNN_OPERATOR (Conv, CudnnConvOp)
	REGISTER_CUDNN_OPERATOR (ConvGradient, CudnnConvGradientOp)
	REGISTER_CUDNN_OPERATOR (Conv1D, CudnnConvOp)
	REGISTER_CUDNN_OPERATOR (Conv1DGradient, CudnnConvGradientOp)
	REGISTER_CUDNN_OPERATOR (Conv2D, CudnnConvOp)
	REGISTER_CUDNN_OPERATOR (Conv2DGradient, CudnnConvGradientOp)
	REGISTER_CUDNN_OPERATOR (Conv3D, CudnnConvOp)
	REGISTER_CUDNN_OPERATOR (Conv3DGradient, CudnnConvGradientOp)
	REGISTER_CPU_OPERATOR_WITH_ENGINE (Conv, EIGEN, EigenConvOp< float >)
	REGISTER_CPU_OPERATOR_WITH_ENGINE (Conv1D, EIGEN, EigenConvOp< float >)
	REGISTER_CPU_OPERATOR_WITH_ENGINE (Conv2D, EIGEN, EigenConvOp< float >)
	REGISTER_CPU_OPERATOR_WITH_ENGINE (Conv3D, EIGEN, EigenConvOp< float >)
	REGISTER_CUDA_OPERATOR (Conv, ConvOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (ConvGradient, ConvGradientOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (Conv1D, ConvOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (Conv1DGradient, ConvGradientOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (Conv2D, ConvOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (Conv2DGradient, ConvGradientOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (Conv3D, ConvOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (Conv3DGradient, ConvGradientOp< float, CUDAContext >)
template<>
void	createSharedBuffer< CPUContext > (Workspace *ws)
template<>
void	runWithSharedBuffer (Workspace *ws, std::function< void(Tensor< CPUContext > *buffer)> f)
template<typename Context >
void	createSharedBuffer (Workspace *ws)
	Creates a mutex and shared buffer in the workspace. More...
template<typename Context >
void	runWithSharedBuffer (Workspace *ws, std::function< void(Tensor< Context > *buffer)> f)
	Thread-safe, can be invoked from RunOnDevice() to serialize access to shared buffer.
template<>
void	createSharedBuffer< CUDAContext > (Workspace *ws)
template<>
void	runWithSharedBuffer (Workspace *ws, std::function< void(Tensor< CUDAContext > *buffer)> f)
	REGISTER_CPU_OPERATOR (ConvTransposeGradient, ConvTransposeGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (ConvTransposeGradient).NumInputs(3).NumOutputs(1
	REGISTER_GRADIENT (ConvTranspose, GetConvTransposeGradient)
	REGISTER_CPU_OPERATOR (ConvTranspose, ConvTransposeOp< float, CPUContext >)
	NumInputs (2, 3).NumOutputs(1).SetDoc(R"DOC( The transposed convolution consumes an input vector
this is done throughout the image data and the output is computed As a side note on the implementation which is why they are separate files DOC	Input (0,"X","Input data blob from previous layer; has size ""(N x C x H x W), where N is the batch size, C is the number of channels, and"" H and W are the height and width. Note that this is for the NCHW usage. On ""the other hand, the NHWC Op has a different set of dimension constraints.").Input(1
has	size (M x C x kH x kW)
has where C is the number of and kH and kW are the height and width of the kernel	Input (2,"bias","The 1D bias blob that is added through the convolution;""has size (C). Optional, if not passed, will treat it as all 0.").Output(0
has where C is the number of and kH and kW are the height and width of the kernel Output data blob that contains the result of the transposed convolution The output dimensions are functions of the kernel stride and pad lengths	InheritOnnxSchema ("ConvTranspose")
	REGISTER_CUDNN_OPERATOR (ConvTranspose, CudnnConvTransposeOp< float >)
	REGISTER_CUDNN_OPERATOR (ConvTransposeGradient, CudnnConvTransposeGradientOp< float >)
	REGISTER_CUDA_OPERATOR (ConvTranspose, ConvTransposeOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (ConvTransposeGradient, ConvTransposeGradientOp< float, CUDAContext >)
	REGISTER_CPU_OPERATOR (Cos, UnaryElementwiseOp< TensorTypes< float >, CPUContext, CosCPUFunctor >)
	REGISTER_CPU_OPERATOR (CosGradient, BinaryElementwiseOp< TensorTypes< float >, CPUContext, WithoutBroadcast< CosGradientCPUFunctor >>)
	OPERATOR_SCHEMA (CosGradient).NumInputs(2).NumOutputs(1).IdenticalTypeAndShape()
	REGISTER_GRADIENT (Cos, GetCosGradient)
	REGISTER_CPU_OPERATOR (CosineEmbeddingCriterion, CosineEmbeddingCriterionOp< CPUContext >)
	REGISTER_CPU_OPERATOR (CosineEmbeddingCriterionGradient, CosineEmbeddingCriterionGradientOp< CPUContext >)
	REGISTER_CPU_OPERATOR (CreateCounter, CreateCounterOp< int64_t, CPUContext >)
	REGISTER_CPU_OPERATOR (ResetCounter, ResetCounterOp< int64_t, CPUContext >)
	REGISTER_CPU_OPERATOR (CountDown, CountDownOp< int64_t, CPUContext >)
	REGISTER_CPU_OPERATOR (CheckCounterDone, CheckCounterDoneOp< int64_t, CPUContext >)
	REGISTER_CPU_OPERATOR (CountUp, CountUpOp< int64_t, CPUContext >)
	REGISTER_CPU_OPERATOR (RetrieveCount, RetrieveCountOp< int64_t, CPUContext >)
A blob pointing to an instance of a new counter	Arg ("init_count","Initial count for the counter, must be >= 0.")
	SetDoc (R"DOC( Resets a count-down counter with initial value specified by the 'init_count' argument. )DOC").Input(0
A blob pointing to an instance of a new counter	Output (0,"previous_value","(optional) Previous value of the counter.").Arg("init_count"
	REGISTER_CUDA_OPERATOR (CreateCounter, CreateCounterOp< int64_t, CUDAContext >)
	REGISTER_CUDA_OPERATOR (ResetCounter, ResetCounterOp< int64_t, CUDAContext >)
	REGISTER_CUDA_OPERATOR (CountDown, CountDownOp< int64_t, CUDAContext >)
	REGISTER_CUDA_OPERATOR (CheckCounterDone, CheckCounterDoneOp< int64_t, CUDAContext >)
	REGISTER_CUDA_OPERATOR (CountUp, CountUpOp< int64_t, CUDAContext >)
	REGISTER_CUDA_OPERATOR (RetrieveCount, RetrieveCountOp< int64_t, CUDAContext >)
	CAFFE_KNOWN_TYPE (detail::WorkspaceStack)
	REGISTER_CPU_OPERATOR (CreateScope, CreateScopeOp< CPUContext >)
	SHOULD_NOT_DO_GRADIENT (CreateScope)
	OPERATOR_SCHEMA (CreateScope).NumInputs(0).NumOutputs(1).SetDoc(R"DOC( 'CreateScope' operator initializes and outputs empty scope that is used by Do operator to store local blobs )DOC")
	REGISTER_CPU_OPERATOR (HasScope, HasScopeOp< CPUContext >)
	SHOULD_NOT_DO_GRADIENT (HasScope)
	OPERATOR_SCHEMA (HasScope).NumInputs(1).NumOutputs(1).SetDoc(R"DOC( Checks whether scope blob has any saved scopes left )DOC")
	REGISTER_CPU_OPERATOR (LabelCrossEntropy, LabelCrossEntropyOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (LabelCrossEntropyGradient, LabelCrossEntropyGradientOp< float, CPUContext >)
	SetDoc (R"DOC( Operator computes the cross entropy between the input and the label set. In practice, it is most commonly used at the end of models, after the SoftMax operator and before the AveragedLoss operator. Note that LabelCrossEntropy assumes that the label provided is either a 1D array of size N (batch size), or a 2D array of size N x 1 (batch size). Each entry in the label vector indicates which is the correct class; as such, each entry must be between 0 and D - 1, inclusive, where D is the total number of classes. The formula used is: Y[i] = -log(X[i][j]) where (i, j) is the classifier's prediction of the jth class (the correct one), and i is the batch size. Each log has a lower limit for numerical stability. )DOC").Input(0
X is a array of size N x where N is the batch size and D is the number of classes	Input (1,"label","Blob containing the labels used to compare the input").Output(0
	REGISTER_GRADIENT (LabelCrossEntropy, GetLabelCrossEntropyGradient)
	REGISTER_CPU_OPERATOR (MakeTwoClass, MakeTwoClassOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (MakeTwoClassGradient, MakeTwoClassGradientOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SigmoidCrossEntropyWithLogits, SigmoidCrossEntropyWithLogitsOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SigmoidCrossEntropyWithLogitsGradient, SigmoidCrossEntropyWithLogitsGradientOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (WeightedSigmoidCrossEntropyWithLogits, WeightedSigmoidCrossEntropyWithLogitsOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (WeightedSigmoidCrossEntropyWithLogitsGradient, WeightedSigmoidCrossEntropyWithLogitsGradientOp< float, CPUContext >)
out[0]	add_dims (in[0].dims(0))
out[0]	add_dims (2)
	SetDoc (R"DOC( Given a vector of probabilities, this operator transforms this into a 2-column matrix with complimentary probabilities for binary classification. In explicit terms, given the vector X, the output Y is vstack(1 - X, X). )DOC").Input(0
Input vector of probabilities	Output (0,"Y","2-column matrix with complimentary probabilities of X for ""binary classification")
	SetDoc (R"DOC( Given two matrices logits and targets, of same shape, (batch_size, num_classes), computes the sigmoid cross entropy between the two. Returns a tensor of shape (batch_size,) of losses for each example. )DOC").Input(0
matrix of logits for each example and class	Input (1,"targets","matrix of targets, same shape as logits.").Output(0
	SetDoc (R"DOC( Given three matrices: logits, targets, weights, all of the same shape, (batch_size, num_classes), computes the weighted sigmoid cross entropy between logits and targets. Specifically, at each position r,c, this computes weights[r, c] * crossentropy(sigmoid(logits[r, c]), targets[r, c]), and then averages over each row. Returns a tensor of shape (batch_size,) of losses for each example. )DOC").Input(0
matrix of logits for each example and class matrix of same shape as logits	Output (0,"xentropy","Vector with the total xentropy for each example.")
	REGISTER_GRADIENT (MakeTwoClass, GetMakeTwoClassGradient)
	REGISTER_GRADIENT (SigmoidCrossEntropyWithLogits, GetSigmoidCrossEntropyWithLogitsGradient)
	REGISTER_GRADIENT (WeightedSigmoidCrossEntropyWithLogits, GetWeightedSigmoidCrossEntropyWithLogitsGradient)
	REGISTER_CPU_OPERATOR (CrossEntropy, CrossEntropyOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (CrossEntropyGradient, CrossEntropyGradientOp< float, CPUContext >)
	SetDoc (R"DOC( Operator computes the cross entropy between the input and the label set. In practice, it is most commonly used at the end of models, after the SoftMax operator and before the AveragedLoss operator. Note that CrossEntropy assumes that the soft labels provided is a 2D array of size N x D (batch size x number of classes). Each entry in the 2D label corresponds to the soft label for the input, where each element represents the correct probability of the class being selected. As such, each element must be between 0 and 1, and all elements in an entry must sum to 1. The formula used is: Y[i] = sum_j (label[i][j] * log(X[i][j])) where (i, j) is the classifier's prediction of the jth class (the correct one), and i is the batch size. Each log has a lower limit for numerical stability. )DOC").Input(0
	REGISTER_GRADIENT (CrossEntropy, GetCrossEntropyGradient)
	CAFFE_KNOWN_TYPE (std::unique_ptr< dataset_ops::TreeCursor >)
	CAFFE_KNOWN_TYPE (dataset_ops::TensorVectorPtr< CPUContext >)
	CAFFE_KNOWN_TYPE (dataset_ops::SharedTensorVectorPtr)
	OPERATOR_SCHEMA (DeformConvGradient).NumInputs(4
	NumOutputs (2, 4)
OpSchema::Cost	CostInferenceForDotProduct (const OperatorDef &def, const vector< TensorShape > &in)
	REGISTER_CPU_OPERATOR (SquaredL2Distance, SquaredL2DistanceOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SquaredL2DistanceGradient, SquaredL2DistanceGradientOp< float, CPUContext >)
	SetDoc (R"DOC( Given two input float tensors X, Y, and produces one output float tensor of the L2 difference between X and Y that is computed as ||(X - Y)^2 / 2||. )DOC").Input(0
or input tensor	Input (1,"Y","1D or 2D input tensor (must have the same shape as X)").Output(0
	OPERATOR_SCHEMA (SquaredL2DistanceGradient).NumInputs(3).NumOutputs(2)
	REGISTER_GRADIENT (SquaredL2Distance, GetSquaredL2DistanceGradient)
	REGISTER_CPU_OPERATOR (L1Distance, L1DistanceOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (L1DistanceGradient, L1DistanceGradientOp< float, CPUContext >)
	SetDoc (R"DOC( Given two input float tensors X, Y, and produces one output float tensor of the L1 difference between X and Y, computed as L1(x,y) = sum over |x-y| )DOC").Input(0
	OPERATOR_SCHEMA (L1DistanceGradient).NumInputs(3).NumOutputs(2)
	REGISTER_GRADIENT (L1Distance, GetL1DistanceGradient)
	REGISTER_CPU_OPERATOR (DotProduct, DotProductOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (DotProductGradient, DotProductGradientOp< float, CPUContext >)
	SetDoc (R"DOC( Given two input float tensors X, Y, and produces one output float tensor of the dot product between X and Y. )DOC").Input(0
or input tensor output tensor	CostInferenceFunction (OpSchema::CostInferenceFunctionType(CostInferenceForDotProduct))
	OPERATOR_SCHEMA (DotProductGradient).NumInputs(3).NumOutputs(2)
	REGISTER_GRADIENT (DotProduct, GetDotProductGradient)
	REGISTER_CPU_OPERATOR (CosineSimilarity, CosineSimilarityOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (CosineSimilarityGradient, CosineSimilarityGradientOp< float, CPUContext >)
	SetDoc (R"DOC( Given two input float tensors X, Y, and produces one output float tensor of the cosine similarity between X and Y. )DOC").Input(0
	OPERATOR_SCHEMA (CosineSimilarityGradient).NumInputs(3).NumOutputs(2)
	REGISTER_GRADIENT (CosineSimilarity, GetCosineSimilarityGradient)
	REGISTER_CPU_OPERATOR (DotProductWithPadding, DotProductWithPaddingOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (DotProductWithPaddingGradient, DotProductWithPaddingGradientOp< float, CPUContext >)
Y with different shapes and produces one output float tensor of the dot product between X and Y We currently support two kinds of strategies to achieve this Before doing normal dot_product pad the smaller	tensor (using pad_value) to the same shape as the other one.2) replicate the smaller tensor to the same shape as the other one.Note the first dimension of X
Y with different shapes and produces one output float tensor of the dot product between X and Y We currently support two kinds of strategies to achieve this Before doing normal dot_product pad the smaller Y must be equal Only the second dimension of X or Y can be padded DOC	Input (0,"X","1D or 2D input tensor").Input(1
Y with different shapes and produces one output float tensor of the dot product between X and Y We currently support two kinds of strategies to achieve this Before doing normal dot_product pad the smaller Y must be equal Only the second dimension of X or Y can be padded DOC or input tensor	Output (0,"Z","1D output tensor").IdenticalTypeAndShapeOfInputDim(0
Y with different shapes and produces one output float tensor of the dot product between X and Y We currently support two kinds of strategies to achieve this Before doing normal dot_product pad the smaller Y must be equal Only the second dimension of X or Y can be padded DOC or input tensor	Arg ("pad_value","the padding value for tensors with smaller dimension").Arg("replicate"
	OPERATOR_SCHEMA (DotProductWithPaddingGradient).NumInputs(3).NumOutputs(2)
	REGISTER_GRADIENT (DotProductWithPadding, GetDotProductWithPaddingGradient)
	REGISTER_CPU_OPERATOR (Do, DoOp< CPUContext >)
INT_MAX	SetDoc (R"DOC( 'Do' control operator, executes a subnet in a separate workspace. Last blobs in the input and output lists should be the same blob created with CreateScope op. Arguments 'inner_blobs' and 'outer_blobs_idx' provide a mapping between selected inner blob names and corresponding outer blob indices. )DOC").Arg("net"
INT_MAX Subnet with blob bindings	Arg ("inner_blobs","List of inner net blob names to bind to outer workspace").Arg("outer_blobs_idx"
INT_MAX Subnet with blob bindings Indices of corresponding outer workspace in	operator outputs (skipping workspace blobs)") .Arg( "saved_fwd_blobs"
INT_MAX Subnet with blob bindings Indices of corresponding outer workspace in List of blobs from the forward Do	operator workspace needed" "in backward pass, used in gradient Do operator") .Arg ("reuse_workspace","Whether to reuse workspace or create a new one in a given scope").AllowInplace([](int in
	REGISTER_CUDA_OPERATOR (Do, DoOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (Dropout, DropoutOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (DropoutGrad, DropoutGradientOp< float, CPUContext >)
	AllowInplace ({{0, 0}}).TensorInferenceFunction([](const OperatorDef &def
	if (output_mask)
	SetDoc (R"DOC( Dropout takes one input data (Tensor<float>) and produces two Tensor outputs, output (Tensor<float>) and mask (Tensor<bool>). Depending on whether it is in test mode or not, the output Y will either be a random dropout, or a simple copy of the input. Note that our implementation of Dropout does scaling in the training phase, so during testing nothing needs to be done. )DOC").Arg("ratio"
default the ratio of random dropout	ArgIsTest ("(int) if nonzero, run dropout in test mode where ""the output is simply Y = X.").Input(0
default the ratio of random dropout The input data as Tensor	Output (0,"output","The output.").Output(1
default the ratio of random dropout The input data as Tensor The output mask If is_test is this output is not filled	InheritOnnxSchema ("Dropout")
	REGISTER_GRADIENT (Dropout, GetDropoutGradient)
	EIGEN_FUNCTOR (Add, EIGEN_ADD, NumericTypes, SameTypeAsInput)
	EIGEN_FUNCTOR (Div, EIGEN_DIV, NumericTypes, SameTypeAsInput)
void	ElementWiseDivide (CPUContext &, const int n, float *dXdata, float *dYdata, const float *dZdata, const float *Ydata, const float *Zdata)
	REGISTER_CPU_OPERATOR (DivGradient, DivGradientOp< CPUContext >)
	REGISTER_CPU_OPERATOR (ElementwiseLinear, ElementwiseLinearOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (ElementwiseLinearGradient, ElementwiseLinearGradientOp< float, CPUContext >)
w of size D and b of size the op computes Y of	size (N X D) where Y_
	REGISTER_GRADIENT (ElementwiseLinear, GetElementwiseLinearGradient)
	EIGEN_FUNCTOR (Mul, EIGEN_MUL, NumericTypes, SameTypeAsInput)
	NAIVE_FUNCTOR (LT, NAIVE_LT, NumericTypes, FixedType< bool >)
	NAIVE_FUNCTOR (LE, NAIVE_LE, NumericTypes, FixedType< bool >)
	NAIVE_FUNCTOR (GT, NAIVE_GT, NumericTypes, FixedType< bool >)
	NAIVE_FUNCTOR (GE, NAIVE_GE, NumericTypes, FixedType< bool >)
	NAIVE_FUNCTOR (EQ, NAIVE_EQ, IntBoolTypes, FixedType< bool >)
	NAIVE_FUNCTOR (And, NAIVE_AND, BoolTypes, FixedType< bool >)
	NAIVE_FUNCTOR (Or, NAIVE_OR, BoolTypes, FixedType< bool >)
	NAIVE_FUNCTOR (Xor, NAIVE_XOR, BoolTypes, FixedType< bool >)
	REGISTER_CPU_OPERATOR (Not, UnaryElementwiseOp< BoolTypes, CPUContext, NotFunctor >)
	REGISTER_CPU_OPERATOR (SumReduceLike, SumReduceLikeOp< CPUContext >)
template<typename Context >
std::tuple< size_t, size_t, size_t >	calculate_broadcast_sizes (const Tensor< Context > &A, const Tensor< Context > &B, int axis)
std::function< void(OpSchema &)>	MathDocGenerator (const char *name)
	CostInferenceFunction (PointwiseCostInference< 1 >).IdenticalTypeAndShapeOfInput(0).FillUsing(MathDocGenerator("addition")).InheritOnnxSchema("Add")
	OPERATOR_SCHEMA (DivGradient).NumInputs(3).NumOutputs(2).AllowInplace(
and the dimensions of the second input is the contiguous subset of the dimensions of the first For the following tensor shapes are	shape (B)
and the dimensions of the second input is the contiguous subset of the dimensions of the first For the following tensor shapes are i e B is a scalar	shape (A)
	EIGEN_FUNCTOR (Sub, EIGEN_SUB, NumericTypes, SameTypeAsInput)
	REGISTER_CPU_OPERATOR (Sum, SumOp< CPUContext >)
	CostInferenceFunction (CostInferenceForSum).InputsCanCrossDevices().IdenticalTypeAndShapeOfInput(0).SetDoc(R"DOC( Element-wise sum of each of the input tensors. The first input tensor can be used in-place as the output tensor
in which case the sum will be done in place and results will be accumulated in input0 All inputs and outputs must have the same shape and data type DOC	Input (0,"data_0","First of the input tensors. Can be inplace.").Output(0
in which case the sum will be done in place and results will be accumulated in input0 All inputs and outputs must have the same shape and data type DOC Output tensor Same dimension as inputs	InheritOnnxSchema ("Sum")
	REGISTER_CPU_OPERATOR (Elu, EluOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (EluGradient, EluGradientOp< float, CPUContext >)
is applied to the tensor elementwise DOC	Input (0,"X","1D input tensor").Output(0
is applied to the tensor elementwise DOC input tensor	InheritOnnxSchema ("Elu")
	SetDoc (R"DOC( EluGradient takes both Y and dY and uses this to update dX according to the chain rule and derivatives of the rectified linear function. )DOC")
	REGISTER_GRADIENT (Elu, GetEluGradient)
	REGISTER_CPU_OPERATOR (Exp, UnaryElementwiseOp< TensorTypes< float >, CPUContext, ExpCPUFunctor >)
element wise This operation can be done in an in place fashion by providing the same input and output blobs DOC The exponential of the input tensor computed element wise	InheritOnnxSchema ("Exp")
	REGISTER_GRADIENT (Exp, GetExpGradient)
	REGISTER_CPU_OPERATOR (ExpandDims, ExpandDimsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Squeeze, SqueezeOp< CPUContext >)
	TensorInferenceFunction ([](const OperatorDef &def, const vector< TensorShape > &in){ArgumentHelper helper(def);auto dims=helper.template GetRepeatedArgument< int >("dims");auto originalSize=dims.size();CAFFE_ENFORCE(originalSize > 0,"Parameter `dims` must be provided.");std::sort(dims.begin(), dims.end());dims.erase(std::unique(dims.begin(), dims.end()), dims.end());if(dims.size()< originalSize){LOG(WARNING)<< "Parameter `dims` has repeated dimensions.";}CAFFE_ENFORCE(dims.front() >=0,"Dimension ids must be non-negative.");CAFFE_ENFORCE_GE(in[0].dims_size()+dims.size(), dims.back()+1,"Input needs at least ",(1+dims.back()-dims.size())," dimensions given `dims`.");vector< TensorShape > out(1);int cur_pos=0;int idx=0;for(const auto new_dim:dims){for(int i=cur_pos;i< new_dim;i++){out[0].add_dims(in[0].dims(idx++));}out[0].add_dims(1);cur_pos=new_dim+1;}for(;idx< in[0].dims_size();idx++){out[0].add_dims(in[0].dims(idx));}out[0].set_data_type(in[0].data_type());return out;}).SetDoc(R"DOC( Insert single-dimensional entries to the shape of a tensor. Takes one required argument `dims`
	REGISTER_CUDA_OPERATOR (Squeeze, SqueezeOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (ExpandDims, ExpandDimsOp< CUDAContext >)
std::vector< TensorShape >	FCShapeInference (const OperatorDef &def, const vector< TensorShape > &in, bool pretransposed_weight)
OpSchema::Cost	CostInferenceForFC (const OperatorDef &def, const vector< TensorShape > &in)
	REGISTER_CPU_OPERATOR (FeedBlob, FeedBlobOp< CPUContext >)
	SHOULD_NOT_DO_GRADIENT (FeedBlob)
	NumInputs (0, 0).NumOutputs(1
	SetDoc (R"DOC( FeedBlobs the content of the blobs. The input and output blobs should be one-to-one inplace.)DOC").Arg("value"
	REGISTER_CPU_OPERATOR (UniformFill, UniformFillOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (UniformIntFill, UniformFillOp< int, CPUContext >)
	REGISTER_CPU_OPERATOR (UniqueUniformFill, UniqueUniformFillOp< CPUContext >)
	REGISTER_CPU_OPERATOR (ConstantFill, ConstantFillOp< CPUContext >)
	REGISTER_CPU_OPERATOR (DiagonalFill, DiagonalFillOp< CPUContext >)
	REGISTER_CPU_OPERATOR (GaussianFill, GaussianFillOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (XavierFill, XavierFillOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (MSRAFill, MSRAFillOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (RangeFill, RangeFillOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (LengthsRangeFill, LengthsRangeFillOp< CPUContext >)
	TensorInferenceFunction (FillerTensorInference<>).SetDoc(R"DOC( The operator fills the elements of the output tensor with a const ant value specified by the 'value' argument. The data type is specified by the 'dtype' argument. The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message. If the 'dtype' argument is not provided
template<int VALUE_TYPE = TensorProto_DataType_FLOAT>
std::vector< TensorShape >	FillerTensorInference (const OperatorDef &def, const vector< TensorShape > &in)
	REGISTER_CUDA_OPERATOR (LengthsRangeFill, GPUFallbackOp< LengthsRangeFillOp< CPUContext >>)
	Index (integers)") .Input(1
Needles query	Output (0,"query_indices","Indices of the needles in index or 'missing value'").Arg("missing_value"
Needles query Placeholder for items that are not found	SetDoc (R"DOC( Finds elements of second input from first input, outputting the last (max) index for each query. If query not find, inserts missing_value. See IndexGet() for a version that modifies the index when values are not found. )DOC")
	REGISTER_CPU_OPERATOR (Flatten, FlattenOp< CPUContext >)
vector< TensorShape >	out (1)
	for (auto d:in[0].dims())
out[0]	set_data_type (in[0].data_type())
out[0]	add_dims (outer)
out[0]	add_dims (inner)
	SetDoc (R"DOC( Flattens the input tensor into a 2D matrix. If input tensor has shape (d_0, d_1, ... d_n) then the output will have shape (d_0 X d_1 ... d_(axis-1), d_axis X d_(axis+1) ... X dn) )DOC").Input(0
A tensor of with input dimensions up to axis flattened to the outer dimension of the output and remaining input dimensions flattened into the inner dimension of the output	Arg ("axis","(Default to 1) Indicate up to which input dimensions ""(exclusive) should be flattened to the outer dimension of the output").InheritOnnxSchema("Flatten")
	REGISTER_GRADIENT (Flatten, GetFlattenGradient)
	REGISTER_CPU_OPERATOR (FlexibleTopK, FlexibleTopKOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (FlexibleTopKGradient, FlexibleTopKGradientOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (Floor, FloorOp< float, CPUContext >)
	SetDoc (R"DOC( Floor takes one input data (Tensor<T>) and produces one output data (Tensor<T>) where the floor function, y = floor(x), is applied to the tensor elementwise. Currently supports only float32. )DOC").Input(0
	GRADIENT_NOT_IMPLEMENTED_YET (Floor)
	REGISTER_CPU_OPERATOR (Free, FreeOp< CPUContext >)
	SHOULD_NOT_DO_GRADIENT (Free)
INT_MAX	SameNumberOfOutput ().EnforceOneToOneInplace().SetDoc(R"DOC( Frees the content of the blobs. The input and output blobs should be one-to-one inplace.)DOC")
	REGISTER_CUDA_OPERATOR (Free, FreeOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (FC, FullyConnectedOp< CPUContext >)
	REGISTER_CPU_OPERATOR (FCGradient, FullyConnectedGradientOp< CPUContext >)
	REGISTER_CPU_OPERATOR (FCTransposed, FullyConnectedOp< CPUContext, DefaultEngine, false >)
	REGISTER_CPU_OPERATOR (FCTransposedGradient, FullyConnectedGradientOp< CPUContext, DefaultEngine, false >)
	REGISTER_CUDA_OPERATOR (FC, FullyConnectedOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (FCGradient, FullyConnectedGradientOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (FCTransposed, FullyConnectedOp< CUDAContext, DefaultEngine, false >)
	REGISTER_CUDA_OPERATOR (FCTransposedGradient, FullyConnectedGradientOp< CUDAContext, DefaultEngine, false >)
	REGISTER_CPU_OPERATOR (FloatToFused8BitRowwiseQuantized, FloatToFused8BitRowwiseQuantizedOp< CPUContext >)
and then scaling each element to an bit number between and To later de quantize the	scale (range/255) and offset(bias) are stored alongside the data.More precisely
and then scaling each element to an bit number between and To later de quantize the the first bytes of each row in the output matrix are a bit float storing the the next bytes store the bias as a bit and all remaining bytes in the row encode single quantized values DOC	Input (0,"input","Float32 input data").Output(0
	NO_GRADIENT (FloatToFused8BitRowwiseQuantized)
	REGISTER_CPU_OPERATOR (Fused8BitRowwiseQuantizedToFloat, Fused8BitRowwiseQuantizedToFloatOp< CPUContext >)
followed by the bias as a bit float in the next and the quantized values in the preceding bytes of the row The output is a matrix containing only the but de quantized De quantization is performed by multiplying each value by its row s scale and bias parameters The de quantized values will thus not be exactly equal to the un quantized floating point values DOC	Input (0,"scale_bias_quantized_input","Fused scale, bias and quantized data").Output(0
	NO_GRADIENT (Fused8BitRowwiseQuantizedToFloat)
but operating on bit rowwise quantized matrices with fused	storage (where each row stores quantized values, and then the scale and offset).DATA needs to have rank 2 and INDICES needs to have rank 1.) DOC") .Input( 0
but operating on bit rowwise quantized matrices with fused uint8 tensor with rank obtained with	operator FloatToFused8BitRowwiseQuantized") .Input (1,"INDICES","Integer vector containing indices of the first dimension of DATA for""the rows that are being gathered").Output(0
but operating on bit rowwise quantized matrices with fused uint8 tensor with rank obtained with output	TensorInferenceFunction ([](const OperatorDef &def, const vector< TensorShape > &in){vector< TensorShape > out(1);for(auto d:in[1].dims()){out[0].add_dims(d);}for(int i=1;i< in[0].dims_size();++i){out[0].add_dims(in[0].dims(i));}out[0].set_data_type(in[0].data_type());return out;})
	REGISTER_CPU_OPERATOR (GatherFused8BitRowwise, GatherFused8BitRowwiseOp< CPUContext >)
	REGISTER_CPU_OPERATOR (GivenTensorFill, GivenTensorFillOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (GivenTensorDoubleFill, GivenTensorFillOp< double, CPUContext >)
	REGISTER_CPU_OPERATOR (GivenTensorBoolFill, GivenTensorFillOp< bool, CPUContext >)
	REGISTER_CPU_OPERATOR (GivenTensorIntFill, GivenTensorFillOp< int, CPUContext >)
	REGISTER_CPU_OPERATOR (GivenTensorInt64Fill, GivenTensorFillOp< int64_t, CPUContext >)
	REGISTER_CPU_OPERATOR (GivenTensorStringFill, GivenTensorFillOp< std::string, CPUContext >)
	NO_GRADIENT (GivenTensorFill)
	NO_GRADIENT (GivenTensorDoubleFill)
	NO_GRADIENT (GivenTensorBoolFill)
	NO_GRADIENT (GivenTensorIntFill)
	NO_GRADIENT (GivenTensorInt64Fill)
	NO_GRADIENT (GivenTensorStringFill)
	TensorInferenceFunction (FillerTensorInference< TensorProto_DataType_DOUBLE >)
	TensorInferenceFunction (FillerTensorInference< TensorProto_DataType_BOOL >)
	TensorInferenceFunction (FillerTensorInference< TensorProto_DataType_INT32 >)
	TensorInferenceFunction (FillerTensorInference< TensorProto_DataType_INT64 >)
	TensorInferenceFunction (FillerTensorInference< TensorProto_DataType_STRING >)
	REGISTER_CPU_OPERATOR (GRUUnit, GRUUnitOp< float, CPUContext >)
in a sequence length aware fashion given	the (fused) inputs X(TxNxD)
in a sequence length aware fashion given the previous hidden	state (NxD)
in a sequence length aware fashion given the previous hidden and the sequence	lengths (N)
in a sequence length aware fashion given the previous hidden and the sequence computes the GRU avoiding computation if the input is	invalid (as in, the value at X[t][n] >=seqLengths[n].) DOC") .Arg( "drop_states"
in a sequence length aware fashion given the previous hidden and the sequence computes the GRU avoiding computation if the input is Bool to determine if hidden state is zeroes or passed along for timesteps past the given sequence_length	Arg ("sequence_lengths","When false, the sequence lengths input is left out, ""and all following inputs are shifted left by one.").Output(0
	REGISTER_CPU_OPERATOR (GRUUnitGradient, GRUUnitGradientOp< float, CPUContext >)
	NumInputs (5, 6).NumOutputs(2).Arg("sequence_lengths"
	REGISTER_GRADIENT (GRUUnit, GetGRUUnitGradient)
out	push_back (X)
out[0]	set_data_type (TensorProto_DataType_FLOAT16)
out[0]	set_data_type (TensorProto_DataType_FLOAT)
The value for the elements of the output tensor	Arg ("shape","The shape of the output tensor.").Output(0
	REGISTER_GRADIENT (FloatToHalf, GetFloatToHalfGradient)
	REGISTER_GRADIENT (HalfToFloat, GetHalfToFloatGradient)
	NO_GRADIENT (Float16ConstantFill)
std::vector< TensorShape >	Float16FillerTensorInference (const OperatorDef &def, const vector< TensorShape > &in)
	REGISTER_CPU_OPERATOR (If, IfOp< CPUContext >)
INT_MAX	SetDoc (R"DOC( 'If' control operator, first input is a scalar boolean blob that stores condition value. Accepts 'then_net' (required) and 'else_net' (optional) arguments for 'then' and 'else' subnets respectively. Subnets are executed in the same workspace as 'If'. )DOC").Arg("then_net"
INT_MAX Net executed when condition is true	Arg ("else_net","Net executed when condition is false (optional)").Input(0
INT_MAX Net executed when condition is true Scalar boolean condition	AllowInplace ([](int in, int out) -> bool{return true;})
	REGISTER_CUDA_OPERATOR (If, IfOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (Im2Col, Im2ColOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (Col2Im, Col2ImOp< float, CPUContext >)
	REGISTER_GRADIENT (Im2Col, GetIm2ColGradient)
	REGISTER_GRADIENT (Col2Im, GetCol2ImGradient)
	switch (order)
	CAFFE_ENFORCE (H >=dkernel_h)
	CAFFE_ENFORCE (W >=dkernel_w)
	Input (0,"X","4-tensor in NCHW or NHWC.").Output(0
	OPERATOR_SCHEMA (Col2Im).NumInputs(2).NumOutputs(1)
	REGISTER_CUDA_OPERATOR (Im2Col, Im2ColOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (Col2Im, Col2ImOp< float, CUDAContext >)
	REGISTER_CPU_OPERATOR (IntIndexCreate, IndexCreateOp< int32_t >)
	REGISTER_CPU_OPERATOR (LongIndexCreate, IndexCreateOp< int64_t >)
	REGISTER_CPU_OPERATOR (StringIndexCreate, IndexCreateOp< std::string >)
	REGISTER_CPU_OPERATOR (IndexGet, IndexGetOp)
	REGISTER_CPU_OPERATOR (IndexLoad, IndexLoadOp)
	REGISTER_CPU_OPERATOR (IndexStore, IndexStoreOp)
	REGISTER_CPU_OPERATOR (IndexFreeze, IndexFreezeOp)
	REGISTER_CPU_OPERATOR (IndexSize, IndexSizeOp)
Max number of including the zero entry	Output (0,"handler","Pointer to an Index instance.")
Max number of including the zero entry	Output (0,"handle","Pointer to an Index instance.")
return an Int tensor of same shape containing the indices for each of the keys If the index is unknown entries are given index new entries are added into the index If an insert is necessary but max_elements has been fail DOC	Input (0,"handle","Pointer to an Index instance.").Input(1
return an Int tensor of same shape containing the indices for each of the keys If the index is unknown entries are given index new entries are added into the index If an insert is necessary but max_elements has been fail DOC Tensor of keys to be looked up	Output (0,"indices","Indices for each of the keys.")
disallowing creation of new index entries Should not be called concurrently with IndexGet DOC The input handle	EnforceInplace ({{0, 0}})
Pointer to an Index instance	Input (1,"items","1-D tensor with elements starting with index 1.").Output(0
Pointer to an Index instance	Output (0,"items","Scalar int64 tensor with number of entries.")
	NO_GRADIENT (IndexGetOp)
	NO_GRADIENT (IntIndexCreate)
	NO_GRADIENT (LongIndexCreate)
	NO_GRADIENT (StringIndexCreate)
	SHOULD_NOT_DO_GRADIENT (IndexFreeze)
	SHOULD_NOT_DO_GRADIENT (IndexLoad)
	SHOULD_NOT_DO_GRADIENT (IndexStore)
	SHOULD_NOT_DO_GRADIENT (IndexSize)
	CAFFE_KNOWN_TYPE (std::unique_ptr< caffe2::IndexBase >)
	REGISTER_BLOB_SERIALIZER ((TypeMeta::Id< std::unique_ptr< caffe2::IndexBase >>()), IndexSerializer)
	REGISTER_BLOB_DESERIALIZER (std::unique_ptr< caffe2::IndexBase >, IndexDeserializer)
	REGISTER_CPU_OPERATOR (InstanceNormGradient, InstanceNormGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (InstanceNormGradient).NumInputs(4
	NumOutputs (3)
	REGISTER_GRADIENT (InstanceNorm, GetInstanceNormGradient)
	REGISTER_CPU_OPERATOR (InstanceNorm, InstanceNormOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (BernoulliJSD, BernoulliJSDOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (BernoulliJSDGradient, BernoulliJSDGradientOp< float, CPUContext >)
array of probabilities for prediction	Input (0,"T","array of probabilities for target").Output(0
	OPERATOR_SCHEMA (BernoulliJSDGradient).NumInputs(3).NumOutputs(1)
	REGISTER_GRADIENT (BernoulliJSD, GetBernoulliJSDGradient)
	REGISTER_CPU_OPERATOR (KeySplit, KeySplitOp< int64_t, CPUContext >)
	NO_GRADIENT (KeySplitOp)
	OPERATOR_SCHEMA (KeySplit).NumInputs(1).NumOutputs(1
	REGISTER_CPU_OPERATOR (LayerNorm, LayerNormOp< CPUContext >)
	OPERATOR_SCHEMA (LayerNormGradient).NumInputs(5).NumOutputs(1)
	REGISTER_CPU_OPERATOR (LayerNormGradient, LayerNormGradientOp< CPUContext >)
	REGISTER_GRADIENT (LayerNorm, GetLayerNormGradient)
std::vector< int >	input_dims (input_dims_long.begin(), input_dims_long.end())
ArgumentHelper	helper (def)
std::vector< int >	stat_dims (input_dims.begin(), input_dims.begin()+canonical_axis)
stat_dims	push_back (1)
	SetDoc (R"DOC( Computes layer normalization as described in https://arxiv.org/pdf/1607.06450.pdf. Given an input vector x \in [a_0, a_1, ...,a_{k-1}, a_k, ..., a_{n-1}], this op treats dimensions a_k through a_{n-1} as feature vectors. For each feature vector, the op contains the mean and standard deviation. Then, it returns the normalized values (with respect to the feature vector). Note that this op does not contain the scale an bias terms described in the paper. Simply follow this op with an FC op to add those. Concretely, this op implements: h = \frac{1}{\sigma}(a - \mu) where \mu = \frac{1}{H}\sum_{i=1}^{H} a_i and \sigma = \sqrt{\frac{1}{H}\sum_{i=1}^{H}(a_i - \mu)^2} where H is the number of hidden units (i.e. product of dimensions from 'axis' to the end.) )DOC").Arg("axis"
Describes axis of the inputs Defaults to one because the axis most likely describes the batch size	Arg ("epsilon","(float) default to 0.001. Small value to be added to the stdev when"" dividing out by that value. This prevents division by zero.").Input(0
Describes axis of the inputs Defaults to one because the axis most likely describes the batch size Input tensor which layer normalization will be applied to	Output (0,"output","Normalized values").Output(1
Describes axis of the inputs Defaults to one because the axis most likely describes the batch size Input tensor which layer normalization will be applied to Mean values for each feature vector	Output (2,"stddev","Standard deviations for each feature vector")
	REGISTER_CPU_OPERATOR (LeakyRelu, LeakyReluOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (LeakyReluGradient, LeakyReluGradientOp< float, CPUContext >)
Coefficient of default value is and produces one output	data (Tensor< T >) where the function`f(x)
	Arg ("alpha","Coefficient of leakage").InheritOnnxSchema("LeakyRelu")
	REGISTER_GRADIENT (LeakyRelu, GetLeakyReluGradient)
	REGISTER_CPU_OPERATOR (SparseLengthsSumFused8BitRowwise, SparseLengthsFused8BitRowwiseOp< CPUContext >)
but operating on bit rowwise quantized matrices with fused	storage (where each row stores quantized values, and then 4-byte scale and 4-byte bias).) DOC") .Input( 0
but operating on bit rowwise quantized matrices with fused uint8 tensor obtained with	operator FloatToFused8BitRowwiseQuantized") .Input (1,"INDICES","Integer vector containing indices of the first ""dimension of DATA for the slices that are being aggregated").Input(2
but operating on bit rowwise quantized matrices with fused uint8 tensor obtained with Vector with the same sum of elements as the first dimension of DATA	Output (0,"output","output")
	NO_GRADIENT (SparseLengthsSumFused8BitRowwise)
	REGISTER_CPU_OPERATOR (SparseLengthsWeightedSumFused8BitRowwise, SparseLengthsFused8BitRowwiseOp< CPUContext, true >)
but operating on bit rowwise quantized matrices with fused uint8 tensor obtained with Vector with the same sum of elements as the first dimension of DATA	Input (3,"WEIGHTS","Vector of weights to scale rows of DATA with before reduction").Output(0
	NO_GRADIENT (SparseLengthsWeightedSumFused8BitRowwise)
	REGISTER_CPU_OPERATOR (SparseLengthsMeanFused8BitRowwise, SparseLengthsFused8BitRowwiseOp< CPUContext, false, true >)
	NO_GRADIENT (SparseLengthsMeanFused8BitRowwise)
	REGISTER_CPU_OPERATOR_STR ("SparseLengthsSum", CPUSparseLengthsReductionOp< float, TensorTypes< float, float16 >, 0, 0 >)
	REGISTER_CPU_OPERATOR_STR ("SparseLengthsWeightedSum", CPUSparseLengthsReductionOp< float, TensorTypes< float, float16 >, 1, 0 >)
	REGISTER_CPU_OPERATOR_STR ("SparseLengthsMean", CPUSparseLengthsReductionOp< float, TensorTypes< float, float16 >, 0, 1 >)
for each weights are accessed by where L is the length of given row This is basically a fused	operator of LengthsRangeFill+Gather+SparseWeightedSum) DOC") .Input (0,"DATA","uint8 tensor obtained with ""operator FloatToRowwiseQuantized8Bits").Input(1
for each weights are accessed by where L is the length of given row This is basically a fused Scalar multipliers for the input slices Must be a vector with the length matching the length of DATA	Input (2,"INDICES","Integer vector containing indices of the first ""dimension of DATA for the slices that are being aggregated").Input(3
	REGISTER_CPU_OPERATOR_STR ("SparseLengthsPositionalWeightedSum", CPUSparseLengthsReductionOp< float, TensorTypes< float, float16 >, 1, 0, 1 >)
	REGISTER_CPU_OPERATOR (Rowwise8BitQuantizedToFloat, Rowwise8BitQuantizedToFloatOp< CPUContext >)
	REGISTER_CPU_OPERATOR (FloatToRowwiseQuantized8Bits, FloatToRowwiseQuantized8BitsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (SparseLengthsSum8BitsRowwise, SparseLengths8BitsRowwiseOp< CPUContext >)
	REGISTER_CPU_OPERATOR (SparseLengthsWeightedSum8BitsRowwise, SparseLengths8BitsRowwiseOp< CPUContext, 1 >)
	REGISTER_CPU_OPERATOR (SparseLengthsMean8BitsRowwise, SparseLengths8BitsRowwiseOp< CPUContext, 0, 1 >)
	REGISTER_CPU_OPERATOR (SparseLengthsWeightedMean8BitsRowwise, SparseLengths8BitsRowwiseOp< CPUContext, 1, 1 >)
	NumInputs (5).NumOutputs(1).SetDoc(R"DOC( Variation of SparseLengthsWeightedSum operator
reshape it into matrix of	size (m_1, m_2 x...x m_n) and apply row-wise quantization.After this
	NO_GRADIENT (Rowwise8BitQuantizedToFloat)
	NO_GRADIENT (FloatToRowwiseQuantized8Bits)
	NO_GRADIENT (SparseLengthsSum8BitsRowwise)
	NO_GRADIENT (SparseLengthsWeightedSum8BitsRowwise)
	NO_GRADIENT (SparseLengthsMean8BitsRowwise)
	NO_GRADIENT (SparseLengthsWeightedMean8BitsRowwise)
	REGISTER_CPU_OPERATOR (LengthsTile, LengthsTileOp< CPUContext >)
	REGISTER_CUDA_OPERATOR (LengthsTile, LengthsTileOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (LengthsTopK, LengthsTopKOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (LengthsTopKGradient, LengthsTopKGradientOp< float, CPUContext >)
where segments are defined by their and concatenate them in an output tensor of the output value will be padded and the corresponding output indices will be padded by DOC	Input (0,"DATA","Tensor of rank 1. First dimension must be equal to the sum of ""lengths").Input(1
where segments are defined by their and concatenate them in an output tensor of the output value will be padded and the corresponding output indices will be padded by DOC Tensor of int32 lengths of rank	Output (0,"TopKValue","Output top k elements for each segment, with""shape=(SIZE(lengths), k)").Output(1
where segments are defined by their and concatenate them in an output tensor of the output value will be padded and the corresponding output indices will be padded by DOC Tensor of int32 lengths of rank Output indices in DATA corresponding to value in TopKValue	Arg ("k","the number of top values to return for each segment, if the number ""of values is smaller than k, the values would be padded with 0 and ""indices would be padded with -1.")
	OPERATOR_SCHEMA (LengthsTopKGradient).NumInputs(3).NumOutputs(1)
	REGISTER_GRADIENT (LengthsTopK, GetLengthsTopKGradient)
	REGISTER_CPU_OPERATOR (DBExists, DBExistsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Load, LoadOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Save, SaveOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Checkpoint, CheckpointOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Snapshot, CheckpointOp< CPUContext >)
A scalar bool Tensor	Arg ("absolute_path","(int, default 0) if set, use the db path directly and do not prepend ""the current root folder of the workspace.").Arg("db_name"
A scalar bool Tensor string the path to the db to load	Arg ("db_type","(string) the type of the db.")
	NumInputs (0, INT_MAX).NumOutputs(0
INT_MAX	SetDoc (R"DOC( The Load operator loads a set of serialized blobs from a db or multiple dbs. It takes [0, infinity) number of inputs and [0, infinity) number of outputs, using the db keys to match the db entries with the outputs. If at least one input is passed, then it is assumed that that input blobs are a set of DBReaders to load from. Otherwise the db or dbs argument is used to load blobs from one single db or multiple dbs respectively. db_type argument is used to specify the type of the input db/dbs. )DOC").Arg("absolute_path"
INT_MAX default if use the db path directly and do not prepend the current root folder of the workspace	Arg ("add_prefix","(string, default=\"\") blobs will be prefixed with this when loading.""Useful for avoiding collisions with blobs existing in the workspace.""The output blob names specified to this op should include this prefix.").Arg("strip_prefix"
template<typename... Ts>
string	FormatString (const string &pattern, Ts...values)
	REGISTER_CUDA_OPERATOR (Load, LoadOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (Save, SaveOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (Checkpoint, CheckpointOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (LRN, LRNOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (LRNGradient, LRNGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (LRN).NumInputs(1).NumOutputs(1
	InheritOnnxSchema ("LRN")
	OPERATOR_SCHEMA (LRNGradient).NumInputs(3).NumOutputs(1)
	REGISTER_GRADIENT (LRN, GetLRNGradient)
	REGISTER_CPU_OPERATOR (LC, LocallyConnectedOp< float, CPUContext >)
NumInputs(2, 3).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (LC1D, LocallyConnectedOp< float, CPUContext >)
NumInputs(2, 3).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (LC2D, LocallyConnectedOp< float, CPUContext >)
NumInputs(2, 3).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (LC3D, LocallyConnectedOp< float, CPUContext >)
NumInputs(2, 3).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (LCGradient, LocallyConnectedGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (LCGradient).NumInputs(2
	REGISTER_CPU_OPERATOR (LC1DGradient, LocallyConnectedGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (LC1DGradient).NumInputs(2
	REGISTER_CPU_OPERATOR (LC2DGradient, LocallyConnectedGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (LC2DGradient).NumInputs(2
	REGISTER_CPU_OPERATOR (LC3DGradient, LocallyConnectedGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (LC3DGradient).NumInputs(2
	REGISTER_GRADIENT (LC, GetLocallyConnectedGradient)
	REGISTER_GRADIENT (LC1D, GetLocallyConnectedGradient)
	REGISTER_GRADIENT (LC2D, GetLocallyConnectedGradient)
	REGISTER_GRADIENT (LC3D, GetLocallyConnectedGradient)
	REGISTER_CUDA_OPERATOR (LC, LocallyConnectedOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (LCGradient, LocallyConnectedGradientOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (LC1D, LocallyConnectedOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (LC1DGradient, LocallyConnectedGradientOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (LC2D, LocallyConnectedOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (LC2DGradient, LocallyConnectedGradientOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (LC3D, LocallyConnectedOp< float, CUDAContext >)
	REGISTER_CUDA_OPERATOR (LC3DGradient, LocallyConnectedGradientOp< float, CUDAContext >)
	REGISTER_CPU_OPERATOR (Log, UnaryElementwiseOp< TensorTypes< float >, CPUContext, LogCPUFunctor >)
element wise This operation can be done in an in place fashion by providing the same input and output blobs DOC The natural log of the input tensor computed element wise	InheritOnnxSchema ("Log")
	REGISTER_GRADIENT (Log, GetLogGradient)
	REGISTER_CPU_OPERATOR (Logit, UnaryElementwiseWithArgsOp< TensorTypes< float >, CPUContext, LogitCPUFunctor >)
	REGISTER_CPU_OPERATOR (LogitGradient, LogitGradientOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (AveragedLoss, AveragedLoss< float, CPUContext >)
	REGISTER_CPU_OPERATOR (AveragedLossGradient, AveragedLossGradient< float, CPUContext >)
NumInputs(1).NumOutputs(1).ScalarType(TensorProto	OPERATOR_SCHEMA (AveragedLossGradient).NumInputs(2).NumOutputs(1)
	REGISTER_GRADIENT (AveragedLoss, GetAveragedLossGradient)
	REGISTER_CPU_OPERATOR (LpPool, PoolOp< float, CPUContext, LpPool >)
	REGISTER_CPU_OPERATOR (LpPoolGradient, PoolGradientOp< float, CPUContext, LpPool >)
stride and pad lengths defined by the ConvPoolOpBase	operator.L-p pooling consisting of taking the L-p norm of a subset of the input tensor according to the kernel size and downsampling the data into the output blob Y for further processing.) DOC") .Input (0,"X","Input data tensor from the previous operator; dimensions ""depend on whether the NCHW or NHWC operators are being used. For example, ""in the former, the input has size (N x C x H x W), where N is the batch ""size, C is the number of channels, and H and W are the height and the width ""of the data. The corresponding permutation of dimensions is used in the ""latter case. ").Output(0
	OPERATOR_SCHEMA (LpPoolGradient).NumInputs(3).NumOutputs(1)
	REGISTER_GRADIENT (LpPool, GetPoolGradient)
	REGISTER_CPU_OPERATOR (LSTMUnit, LSTMUnitOp< CPUContext >)
	NumInputs (4, 5).NumOutputs(2).SetDoc(R"DOC( LSTMUnit computes the activations of a standard LSTM (without peephole connections)
in a sequence length aware fashion given the previous cell and the sequence computes the LSTM avoiding computation if the input is	invalid (as in, the value at X{t][n] >=seqLengths[n].) DOC") .Arg("forget_bias"
	REGISTER_CPU_OPERATOR (LSTMUnitGradient, LSTMUnitGradientOp< CPUContext >)
	NumInputs (8, 9).NumOutputs(3).Arg("sequence_lengths"
	REGISTER_GRADIENT (LSTMUnit, GetLSTMUnitGradient)
	CAFFE_KNOWN_TYPE (MapType64To64)
	REGISTER_CPU_OPERATOR (MarginRankingCriterion, MarginRankingCriterionOp< CPUContext >)
	REGISTER_CPU_OPERATOR (MarginRankingCriterionGradient, MarginRankingCriterionGradientOp< CPUContext >)
	X2 (Tensor< float >)
and label	Y (Tensor< int >) to produce the loss(Tensor< float >) where the loss function
and label	loss (X1, X2, Y)
	REGISTER_CPU_OPERATOR (Sqr, UnaryElementwiseOp< TensorTypes< float >, CPUContext, SqrCPUFunctor >)
Input tensor	Output (0,"output","Squared elements of the input")
	REGISTER_GRADIENT (Sqr, GetSqrGradient)
	REGISTER_CPU_OPERATOR (Sign, UnaryElementwiseOp< TensorTypes< float >, CPUContext, SignCPUFunctor >)
	REGISTER_CPU_OPERATOR (MatMul, MatMulOp< float, CPUContext >)
	if (trans_a)
	if (trans_b)
out[0]	add_dims (M)
out[0]	add_dims (N)
	SetDoc (R"DOC( Matrix multiplication Y = A * B, where A has size (M x K), B has size (K x N), and Y will have a size (M x N). )DOC").Input(0
matrix of	size (M x K)") .Input(1
matrix of matrix of	size (K x N)") .Output(0
matrix of matrix of matrix of Exclusive axis that divides the first and second dimension of matrix default to	Arg ("axis_b","Exclusive axis that divides the first and second dimension \ of matrix B, default to 1").Arg("trans_a"
	REGISTER_GRADIENT (MatMul, GetMatMulGradient)
	REGISTER_CUDA_OPERATOR (MatMul, MatMulOp< float, CUDAContext >)
	REGISTER_CPU_OPERATOR (Mean, MeanOp< CPUContext >)
	REGISTER_CPU_OPERATOR (MeanGradient, MeanGradientOp< CPUContext >)
	SetDoc (R"DOC( Element-wise mean of each of the input tensors. The first input tensor can be used in-place as the output tensor, in which case the mean will be done in place and results will be accumulated in input0. All inputs and outputs must have the same shape and data type. )DOC").Input(0
First of the input tensors Can be inplace	Output (0,"mean","Output tensor. Same dimension as inputs.")
	REGISTER_GRADIENT (Mean, GetMeanGradient)
	REGISTER_CPU_OPERATOR (MaxGradient, MaxGradientOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (MinGradient, MinGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (MaxGradient).NumInputs(3
INT_MAX	NumOutputs (1, INT_MAX)
	OPERATOR_SCHEMA (MinGradient).NumInputs(3
	REGISTER_GRADIENT (Max, GetMaxGradient)
	REGISTER_GRADIENT (Min, GetMinGradient)
	REGISTER_CPU_OPERATOR (Max, MaxOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (Min, MinOp< float, CPUContext >)
	SetDoc (R"DOC( Element-wise max of each of the input tensors. The first input tensor can be used in-place as the output tensor, in which case the max will be done in place and results will be accumulated in input0. All inputs and outputs must have the same shape and data type. )DOC").Input(0
First of the input tensors Can be inplace	Output (0,"max","Output tensor. Same dimension as inputs.").InheritOnnxSchema("Max")
	SetDoc (R"DOC( Element-wise min of each of the input tensors. The first input tensor can be used in-place as the output tensor, in which case the min will be done in place and results will be accumulated in input0. All inputs and outputs must have the same shape and data type. )DOC").Input(0
First of the input tensors Can be inplace	Output (0,"min","Output tensor. Same dimension as inputs.").InheritOnnxSchema("Min")
	REGISTER_CPU_OPERATOR (MultiClassAccuracy, MultiClassAccuracyOp< float, CPUContext >)
D float	tensor (N, D,) of predicted scores of each class for" "each data.N is the number of instances
D float i batch size D is number of possible classes labels	Input (1,"labels","1-D int tensor (N,) of labels for each instance.").Output(0
D float i batch size D is number of possible classes labels D float	tensor (D,) of accuracy for each class.If a class has no" "instance in the batch
D float i batch size D is number of possible classes labels D float its accuracy score is set to zero	Output (1,"amounts","1-D int tensor (D,) of number of instances for each class in the batch.")
	SHOULD_NOT_DO_GRADIENT (MultiClassAccuracy)
	REGISTER_CPU_OPERATOR (NegateGradient, NegateGradientOp< CPUContext >)
	SetDoc (R"DOC( NegagteGradient operator in forward pass simply copies input to the output, and in backward pass, flips the sign of the output gradient )DOC")
	REGISTER_GRADIENT (NegateGradient, GetNegateGradientGradient)
	REGISTER_CPU_OPERATOR (Negative, UnaryElementwiseOp< TensorTypes< float, double, int, long >, CPUContext, NegativeCPUFunctor >)
input tensor	Output (0,"Y","1D input tensor").InheritOnnxSchema("Neg")
	REGISTER_GRADIENT (Negative, GetNegativeGradient)
	REGISTER_CPU_OPERATOR (NGramFromCategorical, NGramFromCategoricalOp< float, int64_t, CPUContext >)
	NO_GRADIENT (NGramFromCategorical)
	OPERATOR_SCHEMA (NGramFromCategorical).NumInputs(1).NumOutputs(1)
	REGISTER_CPU_OPERATOR (NormalizeL1, NormalizeL1Op< float, CPUContext >)
axis to normalize	SetDoc (R"DOC( Given a matrix, apply L1-normalization along the specified axis. )DOC")
	REGISTER_CPU_OPERATOR (Normalize, NormalizeOp< float, CPUContext >)
axis to normalize	SetDoc (R"DOC( Given a matrix, apply L2-normalization along the specified dimension. )DOC").IdenticalTypeAndShape()
	REGISTER_CPU_OPERATOR (NormalizeGradient, NormalizeGradientOp< float, CPUContext >)
	REGISTER_GRADIENT (Normalize, GetNormalizeGradient)
vector< TensorShape >	TensorInferenceForBatchOneHot (const OperatorDef &, const vector< TensorShape > &in)
OpSchema::Cost	CostInferenceForBatchOneHot (const OperatorDef &def, const vector< TensorShape > &in)
	REGISTER_CPU_OPERATOR (BatchBucketOneHot, BatchBucketOneHotOp< CPUContext >)
	REGISTER_CPU_OPERATOR (BatchOneHot, BatchOneHotOp< CPUContext >)
	REGISTER_CPU_OPERATOR (OneHot, OneHotOp< CPUContext >)
	REGISTER_CPU_OPERATOR (SegmentOneHot, SegmentOneHotOp)
	REGISTER_CPU_OPERATOR (ONNXWhile, ONNXWhileOp< CPUContext >)
INT_MAX	SetDoc (R"DOC( *** EXPERIMENTAL. This operator is a work-in-progress. No assumption should be made about the stability or correctness of this op. *** Generic Looping construct confirming to the ONNX Loop operator spec. This loop has multiple termination conditions: 1. Trip count. Iteration count specified at runtime. Set by specifying the input M. Optional. Set to empty string to omit. Note that a static trip count (specified at graph construction time) can be specified by passing in a constant node for input M. 2. Loop termination condition. This is an input to the op that determines whether to run the first interation and also a loop-carried dependency for the body graph. The body graph must yield a value for the condition variable, whether this input is provided or not. This table summarizes the operating modes of this operator with equivalent C-style code: Operator inputs defined as (max_trip_count, condition_var). Omitted optional inputs are represented as empty string. Concretely, in this caffe2 op an input is marked as omitted by setting its 'has_{name}' argument to False. input ("", ""): for (int i=0; ; ++i) { cond = ... // Note this value is ignored, but is required in the body } input ("", cond) // Note this is analogous to a while loop bool cond = ...; for (int i=0; cond; ++i) { cond = ...; } input ("", 1) // Note this is analogous to a do-while loop bool cond = true for (int i=0; cond; ++i) { cond = ...; } input (trip_count, "") // Note this is analogous to a for loop int trip_count = ... for (int i=0; i < trip_count; ++i) { cond = ...; // ignored } input (trip_count, cond) int trip_count = ...; bool cond = ...; for (int i=0; i < trip_count && cond; ++i) { cond = ...; } )DOC").Arg("loop_net"
INT_MAX Net executed on each iteration	Input (0,"condition","Scalar boolean condition").AllowInplace([](int in
	REGISTER_CPU_OPERATOR (NHWC2NCHW, NHWC2NCHWOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (NCHW2NHWC, NCHW2NHWCOp< float, CPUContext >)
out[0]	add_dims (in[0].dims(3))
out[0]	add_dims (in[0].dims(1))
out[0]	add_dims (in[0].dims(2))
	SetDoc (R"DOC( The operator switches the order of data in a tensor from NHWC- sample index N, height H, width H and channels C, to the NCHW order. )DOC").Input(0
The input	data (Tensor< float >) in the NHWC order.") .Output( 0
The input The output	tensor (Tensor< float >) in the NCHW order.")
	OPERATOR_SCHEMA (NCHW2NHWC).NumInputs(1).NumOutputs(1).SetDoc(R"DOC( The operator switches the order of data in a tensor from NCHW- sample index N
channels height H and width to the NHWC order DOC	Input (0,"data","The input data (Tensor<float>) in the NCHW order.").Output(0
	REGISTER_GRADIENT (NHWC2NCHW, GetNHWC2NCHWGradient)
	REGISTER_GRADIENT (NCHW2NHWC, GetNCHW2NHWCGradient)
	REGISTER_CPU_OPERATOR (PackSegments, PackSegmentsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (UnpackSegments, UnpackSegmentsOp< CPUContext >)
	SetDoc ("Map N dim tensor to N+1 dim based on length blob. Sequences that \ are shorter than the longest sequence are padded with zeros.").Input(0
d int long tensor contains the length in each of the output	Input (1,"tensor","N dim Tensor.").Output(0
d int long tensor contains the length in each of the output N dim Tensor where	dim (1) is the max length" "
d int long tensor contains the length in each of the output N dim Tensor where	dim (0) is the batch size.") .Output( 1
d int long tensor contains the length in each of the output N dim Tensor where dim boolean false where packed_tensor is true otherwise	Arg ("pad_minf","Padding number in the packed segments. Use true to pad \ -infinity, otherwise pad zeros").Arg("return_presence_mask"
d int long tensor contains the length in each of the input	Input (1,"tensor","N+1 dim Tensor.").Output(0
	REGISTER_GRADIENT (PackSegments, GetPackSegmentsGradient)
	REGISTER_GRADIENT (UnpackSegments, GetUnpackSegmentsGradient)
PadMode	StringToPadMode (const string &mode)
	REGISTER_CPU_OPERATOR (PadImage, PadImageOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (PadImageGradient, PadImageGradientOp< float, CPUContext >)
CPUContext::PadTensorInference	SetDoc (R"DOC( PadImage pads values around the boundary of an image according to the pad values and stride sizes defined by the ConvPoolOpBase operator. )DOC").Input(0
dimensions depend on whether the NCHW or NHWC operators are being used For in the the input has	size (N x C x H x W)
dimensions depend on whether the NCHW or NHWC operators are being used For in the the input has where N is the batch C is the number of and H and W are the height and the width of the data The corresponding permutation of dimensions is used in the latter case	Output (0,"Y","Output data tensor from padding the H and W dimensions on ""the tensor. Dimensions will vary based on various pad and stride ""sizes.")
	OPERATOR_SCHEMA (PadImageGradient).NumInputs(1).NumOutputs(1)
	REGISTER_GRADIENT (PadImage, GetPadImageGradient)
	REGISTER_CPU_OPERATOR (Percentile, PercentileOp< CPUContext >)
given a sample set of raw labeled with their corresponding percentiles from the same distribution In this	operator takes as input a tensor of floats to find the percentile values for, a 2D tensor of floats, where the first column of the tensor represents sampled values, and the second column represents the percentile labels, and a tensor of integers lengths.This lengths tensor is used because the operator works on multiple sets of raw values at the same time.For example, for an input:original_values=[[3, 5, 3], [5, 1, 6]], lengths=[2, 1, 1], value_to_pct=[[3, 0.2],[5, 0.5],[1, 0.3],[3.0.6]] Our operator expects that each column i of the input tensor is sampled from distribution i.Lengths tells us that the first two elements in value_to_pct are sampled from distribution 1, the next is from distribution two, and the last is from distribution 3.We expect the output of our operator to give us[[0.2, 1.0, 0.6],[0.5, 0.3, 1.0]].To calculate the percentile of an element, we check to see if its value is already mapped to a percentile in value_to_pct.If so, we return that value.If not, we linearly interpolate between the two closest values in value_to_pct.If the value is larger than all values in value_to_pct, we return 1.If it's smaller than all the values, we return 0.) DOC") .Input (0,"original_values","Input 2D tensor of floats, representing the original, raw data to calculate percentiles for.").Input(1
given a sample set of raw labeled with their corresponding percentiles from the same distribution In this Sorted with columns Each element in the first column is a float representing the raw value of a sample Its corresponding element in the next column represents the percentile it maps to	Input (2,"lengths","1D tensor, representing the length of each distribution. We expect that the sum of elements of this tensor"" is equal to the total length of value_to_pct.").Output(0
	NO_GRADIENT (Percentile)
	REGISTER_CPU_OPERATOR (Perplexity, PerplexityOp< float, CPUContext >)
	OPERATOR_SCHEMA (Perplexity).NumInputs(1).NumOutputs(1).SetDoc(R"DOC( Perplexity calculates how well a probability distribution predicts a sample. Perplexity takes a 1-D tensor containing a batch of probabilities. Each value in the tensor belongs to a different sample and represents the probability of the model predicting the true label for that sample. The operator returns a single (float) perplexity value for the batch. )DOC").Input(0
The input data as Tensor It contains a batch of true label or target probabilities	Output (0,"output","The output- a single (float) perplexity value for the ""batch")
	SHOULD_NOT_DO_GRADIENT (Perplexity)
	REGISTER_CPU_OPERATOR (PiecewiseLinearTransform, PiecewiseLinearTransformOp< float, CPUContext >)
	NumInputs (1, 4).NumOutputs(1).SetDoc(R"DOC( PiecewiseLinearTransform takes inputs -- predictions
a D or D slopes and intercepts The output tensor has the same shape of input predictions and contains the predictions transformed by the piecewise linear functions Each column of predictions has its own piecewise linear transformation functions Therefore the size of piecewise function parameters are pieces x except for binary predictions where only the positive prediction needs them Note that in each low bound is excluded while high bound is included Also the piecewise linear function must be continuous Notes If the input is binary	predictions (Nx2 or Nx1 tensor)
a D or D slopes and intercepts The output tensor has the same shape of input predictions and contains the predictions transformed by the piecewise linear functions Each column of predictions has its own piecewise linear transformation functions Therefore the size of piecewise function parameters are pieces x except for binary predictions where only the positive prediction needs them Note that in each low bound is excluded while high bound is included Also the piecewise linear function must be continuous Notes If the input is binary set the binary arg to true so that one group of piecewise linear functions is	needed (see details below).-The transform parameters(bounds
a D or D slopes and intercepts The output tensor has the same shape of input predictions and contains the predictions transformed by the piecewise linear functions Each column of predictions has its own piecewise linear transformation functions Therefore the size of piecewise function parameters are pieces x except for binary predictions where only the positive prediction needs them Note that in each low bound is excluded while high bound is included Also the piecewise linear function must be continuous Notes If the input is binary set the binary arg to true so that one group of piecewise linear functions is intercepts can be passed either through args or through input blobs If we have multiple groups of piecewise linear each group has the same number of pieces If a prediction is out of the it is capped to the smallest or largest bound DOC	Arg ("bounds","1-D vector of size (prediction_dimensions x (pieces+1)) contain the ""upper bounds of each piece of linear function. One special case is ""the first bound is the lower bound of whole piecewise function and we ""treat it the same as the left most functions. (bounds, slopes, ""intercepts) can be passed through either arg or input blobs.").Arg("slopes"
a D or D slopes and intercepts The output tensor has the same shape of input predictions and contains the predictions transformed by the piecewise linear functions Each column of predictions has its own piecewise linear transformation functions Therefore the size of piecewise function parameters are pieces x except for binary predictions where only the positive prediction needs them Note that in each low bound is excluded while high bound is included Also the piecewise linear function must be continuous Notes If the input is binary set the binary arg to true so that one group of piecewise linear functions is intercepts can be passed either through args or through input blobs If we have multiple groups of piecewise linear each group has the same number of pieces If a prediction is out of the it is capped to the smallest or largest bound DOC D vector of	size (prediction_dimensions x pieces) containing the" "slopes of linear function") .Arg( "intercepts"
	REGISTER_CPU_OPERATOR (AveragePoolGradient, PoolGradientOp< float, CPUContext, AveragePool< float >>)
	OPERATOR_SCHEMA (AveragePoolGradient).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (AveragePool1DGradient, PoolGradientOp< float, CPUContext, AveragePool< float >>)
	OPERATOR_SCHEMA (AveragePool1DGradient).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (AveragePool2DGradient, PoolGradientOp< float, CPUContext, AveragePool< float >>)
	OPERATOR_SCHEMA (AveragePool2DGradient).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (AveragePool3DGradient, PoolGradientOp< float, CPUContext, AveragePool< float >>)
	OPERATOR_SCHEMA (AveragePool3DGradient).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (MaxPoolGradient, PoolGradientOp< float, CPUContext, MaxPool< float >>)
	OPERATOR_SCHEMA (MaxPoolGradient).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (MaxPool1DGradient, PoolGradientOp< float, CPUContext, MaxPool< float >>)
	OPERATOR_SCHEMA (MaxPool1DGradient).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (MaxPool2DGradient, PoolGradientOp< float, CPUContext, MaxPool< float >>)
	OPERATOR_SCHEMA (MaxPool2DGradient).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (MaxPool3DGradient, PoolGradientOp< float, CPUContext, MaxPool< float >>)
	OPERATOR_SCHEMA (MaxPool3DGradient).NumInputs(3).NumOutputs(1)
	REGISTER_GRADIENT (AveragePool, GetPoolGradient)
	REGISTER_GRADIENT (AveragePool1D, GetPoolGradient)
	REGISTER_GRADIENT (AveragePool2D, GetPoolGradient)
	REGISTER_GRADIENT (AveragePool3D, GetPoolGradient)
	REGISTER_GRADIENT (MaxPool, GetPoolGradient)
	REGISTER_GRADIENT (MaxPool1D, GetPoolGradient)
	REGISTER_GRADIENT (MaxPool2D, GetPoolGradient)
	REGISTER_GRADIENT (MaxPool3D, GetPoolGradient)
std::function< void(OpSchema &)>	AveragePoolDocGenerator (const char *dim)
std::function< void(OpSchema &)>	MaxPoolDocGenerator (const char *dim)
	REGISTER_CPU_OPERATOR (AveragePool, PoolOp< float, CPUContext, AveragePool< float >>)
NumInputs(1).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (AveragePool1D, PoolOp< float, CPUContext, AveragePool< float >>)
NumInputs(1).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (AveragePool2D, PoolOp< float, CPUContext, AveragePool< float >>)
NumInputs(1).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (AveragePool3D, PoolOp< float, CPUContext, AveragePool< float >>)
NumInputs(1).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (MaxPool, PoolOp< float, CPUContext, MaxPool< float >>)
NumInputs(1).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (MaxPool1D, PoolOp< float, CPUContext, MaxPool< float >>)
NumInputs(1).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (MaxPool2D, PoolOp< float, CPUContext, MaxPool< float >>)
NumInputs(1).NumOutputs(1).TensorInferenceFunction(ConvPoolOpBase< CPUContext >	REGISTER_CPU_OPERATOR (MaxPool3D, PoolOp< float, CPUContext, MaxPool< float >>)
	REGISTER_CPU_OPERATOR (Pow, PowOp< TensorTypes< float >, CPUContext, EigenPowFunctor, SameTypeAsInput >).NumInputs(1
	NumOutputs (1).Arg("exponent"
The exponent of the power function	AllowInplace ({{0, 0},{1, 0}}).IdenticalTypeAndShapeOfInput(0).SetDoc(R"DOC( Pow takes input data (Tensor<T>) and an argument exponent
	REGISTER_CPU_OPERATOR (PRelu, PReluOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (PReluGradient, PReluGradientOp< float, CPUContext >)
	OPERATOR_SCHEMA (PReluGradient).NumInputs(4).NumOutputs(2).SetDoc(R"DOC( PReluGradient takes both Y and dY and uses this to update dX and dW according to the chain rule and derivatives of the rectified linear function. )DOC")
	REGISTER_GRADIENT (PRelu, GetPReluGradient)
	REGISTER_CPU_OPERATOR (PrependDim, PrependDimOp< CPUContext >)
	REGISTER_CPU_OPERATOR (MergeDim, MergeDimOp< CPUContext >)
	SetDoc (R"DOC( Reshape the tensor by prepending a dimension of fixed size and dividing the size of the next dimension by that amount. )DOC").Arg("dim_size"
Size of the dimension to prepend	Input (0,"data","An input tensor.").Output(0
	SetDoc (R"DOC( Merge first two dimensions in a single dimension with size dim(0) * dim(1). )DOC").Input(0
An input tensor	Output (0,"reshaped","Reshaped tensor.")
	REGISTER_GRADIENT (PrependDim, GetPrependDimGradient)
	REGISTER_CUDA_OPERATOR (PrependDim, PrependDimOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (MergeDim, MergeDimOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (QuantDecode, QuantDecodeOp< QuantDecodeRunTy::RUN_ALWAYS >)
	REGISTER_CPU_OPERATOR (QuantDecodeGradient, QuantDecodeGradientOp)
vector< TIndex >	ConvertFromInputIndex (TIndex index, vector< TIndex > &dims)
TIndex	ConvertToOutputIndex (const vector< int > &axes, const vector< TIndex > &nd_idx, vector< TIndex > &dims)
template<typename T >
T	Add (T x, T y)
template<typename T , class Context >
void	ComputeOp (const T *X_data, const TIndex X_size, vector< TIndex > &dims, T *Y_data, vector< int > &axes, int keepdims, T(*binary_op)(T, T))
	REGISTER_CPU_OPERATOR (ReduceSum, ReduceSumOp< float, CPUContext >)
then the resulted tensor have the reduced dimension pruned DOC	Arg ("axes","A list of integers, along which to reduce.").Arg("keepdims"
then the resulted tensor have the reduced dimension pruned DOC Keep the reduced	dimension (s) or not
then the resulted tensor have the reduced dimension pruned DOC Keep the reduced default keeps the reduced An input tensor	Output (0,"reduced","Reduced output tensor.")
	GRADIENT_NOT_IMPLEMENTED_YET (ReduceSum)
	REGISTER_CPU_OPERATOR (ReduceMean, ReduceMeanOp< float, CPUContext >)
	GRADIENT_NOT_IMPLEMENTED_YET (ReduceMean)
	REGISTER_CPU_OPERATOR (ReduceFrontSum, SumReduceDimsOp< CPUContext, true, false >)
	REGISTER_CPU_OPERATOR (ReduceFrontSumGradient, SumReduceDimsGradientOp< CPUContext, true, false >)
	REGISTER_GRADIENT (ReduceFrontSum, GetReduceFrontSumGradient)
	REGISTER_CPU_OPERATOR (ReduceBackSum, SumReduceDimsOp< CPUContext, false, false >)
	REGISTER_CPU_OPERATOR (ReduceBackSumGradient, SumReduceDimsGradientOp< CPUContext, false, false >)
	REGISTER_GRADIENT (ReduceBackSum, GetReduceBackSumGradient)
Number of dimensions to reduce	SetDoc (R"DOC( Reduces the input tensor along the first dimension of the input tensor by applying 'Sum'. When lengths is given, sum is only computed with subsets of elements correspondingly. )DOC").Input(0
Number of dimensions to reduce T< D1..., Dn > Input data	Input (1,"lengths","Num of elements in each sample, should have size D2 x D3 x ... x Dn.").TensorInferenceFunction([](const OperatorDef &def
	OPERATOR_SCHEMA (ReduceFrontSumGradient).NumInputs(2
Number of dimensions to reduce	SetDoc (R"DOC( Reduces the input tensor along the last dimension of the input tensor by applying 'Sum'. When lengths is given, sum is only computed with subsets of elements correspondingly. )DOC").Input(0
Number of dimensions to reduce T< D1..., Dn > Input data	Input (1,"lengths","Num of elements in each sample, should have size D1 x D2 x ... x D(n-1).").TensorInferenceFunction([](const OperatorDef &def
	OPERATOR_SCHEMA (ReduceBackSumGradient).NumInputs(2
	REGISTER_CPU_OPERATOR (ReduceFrontMean, SumReduceDimsOp< CPUContext, true, true >)
	REGISTER_CPU_OPERATOR (ReduceFrontMeanGradient, SumReduceDimsGradientOp< CPUContext, true, true >)
	REGISTER_GRADIENT (ReduceFrontMean, GetReduceFrontMeanGradient)
Number of dimensions to reduce	SetDoc (R"DOC( Reduces the input tensor along the first dimension of the input tensor by applying 'Mean'. When lengths is given, mean is only computed with subsets of elements correspondingly. )DOC").Input(0
	OPERATOR_SCHEMA (ReduceFrontMeanGradient).NumInputs(2
	REGISTER_CPU_OPERATOR (ReduceBackMean, SumReduceDimsOp< CPUContext, false, true >)
	REGISTER_CPU_OPERATOR (ReduceBackMeanGradient, SumReduceDimsGradientOp< CPUContext, false, true >)
	REGISTER_GRADIENT (ReduceBackMean, GetReduceBackMeanGradient)
Number of dimensions to reduce	SetDoc (R"DOC( Reduces the input tensor along the last dimension of the input tensor by applying 'Mean'. When lengths is given, mean is only computed with subsets of elements correspondingly. )DOC").Input(0
	OPERATOR_SCHEMA (ReduceBackMeanGradient).NumInputs(2
	REGISTER_CPU_OPERATOR (ReduceFrontMax, MaxReduceDimsOp< float, CPUContext, true >)
	REGISTER_CPU_OPERATOR (ReduceFrontMaxGradient, MaxReduceDimsGradientOp< float, CPUContext, true >)
	REGISTER_CPU_OPERATOR (ReduceBackMax, MaxReduceDimsOp< float, CPUContext, false >)
	REGISTER_CPU_OPERATOR (ReduceBackMaxGradient, MaxReduceDimsGradientOp< float, CPUContext, false >)
	REGISTER_GRADIENT (ReduceFrontMax, GetReduceFrontMaxGradient)
	REGISTER_GRADIENT (ReduceBackMax, GetReduceBackMaxGradient)
Number of dimensions to reduce	SetDoc (R"DOC( Reduces the input tensor along the first dimension of the input tensor by applying 'Max'. When lengths is given, max is only computed with subsets of elements correspondingly. )DOC").Input(0
Number of dimensions to reduce T< D1..., Dn > Input data	Input (1,"lengths","Num of elements in each sample, should have size D2 x D3 ... x Dn.").TensorInferenceFunction([](const OperatorDef &def
	OPERATOR_SCHEMA (ReduceFrontMaxGradient).NumInputs(3
Number of dimensions to reduce	SetDoc (R"DOC( Reduces the input tensor along the last dimension of the input tensor by applying 'Max'. When lengths is given, max is only computed with subsets of elements correspondingly. )DOC").Input(0
	OPERATOR_SCHEMA (ReduceBackMaxGradient).NumInputs(3
	REGISTER_CPU_OPERATOR (SumElements, SumElementsOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SumElementsInt, SumElementsIntOp< int, CPUContext >)
	REGISTER_CPU_OPERATOR (SumSqrElements, SumSqrElementsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (SumElementsGradient, SumElementsGradientOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (RowwiseMax, MaxReductionOp< float, CPUContext, true >)
	REGISTER_CPU_OPERATOR (RowwiseMaxGradient, MaxReductionGradientOp< float, CPUContext, true >)
	REGISTER_CPU_OPERATOR (ColwiseMaxGradient, MaxReductionGradientOp< float, CPUContext, false >)
	REGISTER_CPU_OPERATOR (ColwiseMax, MaxReductionOp< float, CPUContext, false >)
NumInputs(1).NumOutputs(1).ScalarType(TensorProto NumInputs(1).NumOutputs(1).ScalarType(TensorProto	SHOULD_NOT_DO_GRADIENT (SumElementsInt)
NumInputs(1).NumOutputs(1).ScalarType(TensorProto	OPERATOR_SCHEMA (SumElementsGradient).NumInputs(2).NumOutputs(1)
	REGISTER_GRADIENT (SumElements, GetSumElementsGradient)
A tenosr of dimensions batch_size x M x N to compute rowwise max	Output (0,"Y","batch_size x M rowwise-max results matrix.")
	OPERATOR_SCHEMA (RowwiseMaxGradient).NumInputs(3).NumOutputs(1)
	REGISTER_GRADIENT (RowwiseMax, GetRowwiseMaxGradient)
	OPERATOR_SCHEMA (ColwiseMaxGradient)
A tenosr of dimensions batch_size x M x N to compute colwise max	Output (0,"Y","batch_size x N column-max results matrix.")
	OPERATOR_SCHEMA (ColumnMaxGradient).NumInputs(3).NumOutputs(1)
	REGISTER_GRADIENT (ColwiseMax, GetColwiseMaxGradient)
	REGISTER_CPU_OPERATOR (Relu, ReluOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (ReluGradient, ReluGradientOp< float, CPUContext >)
	CostInferenceFunction (CostInferenceForRelu).IdenticalTypeAndShape().SetDoc(R"DOC( Relu takes one input data (Tensor<T>) and produces one output data (Tensor<T>) where the rectified linear function
is applied to the tensor elementwise DOC input tensor	InheritOnnxSchema ("Relu")
	SetDoc (R"DOC( ReluGradient takes both Y and dY and uses this to update dX according to the chain rule and derivatives of the rectified linear function. )DOC")
	REGISTER_GRADIENT (Relu, GetReluGradient)
	REGISTER_GRADIENT (ReluFp16, GetReluGradient)
	REGISTER_CPU_OPERATOR (ReplaceNaN, ReplaceNaNOp< CPUContext >)
	SHOULD_NOT_DO_GRADIENT (ReplaceNaN)
	REGISTER_CPU_OPERATOR (Reshape, ReshapeOp< float, CPUContext >)
out[1]	set_data_type (TensorProto::INT64)
out[1]	add_dims (in[0].dims_size())
	if (!helper.HasArgument("shape"))
	CAFFE_ENFORCE_EQ (in.size(), 1,"New shape must not be specified by the input blob and the ""argument `shape` at the same time.")
	for (int i=0;i< actualNewShape.size();++i)
	if (unknownIdx!=-1)
	for (const auto d:actualNewShape)
an extra argument shape must be specified It outputs the reshaped tensor as well as the original shape At most one dimension of the new shape can be In this the value is inferred from the size of the tensor and the remaining dimensions A dimension could also in which case the actual dimension value is going to be copied from the input tensor DOC	Arg ("shape","New shape").Input(0
an extra argument shape must be specified It outputs the reshaped tensor as well as the original shape At most one dimension of the new shape can be In this the value is inferred from the size of the tensor and the remaining dimensions A dimension could also in which case the actual dimension value is going to be copied from the input tensor DOC An input tensor	Input (1,"new_shape","New shape.").Output(0
an extra argument shape must be specified It outputs the reshaped tensor as well as the original shape At most one dimension of the new shape can be In this the value is inferred from the size of the tensor and the remaining dimensions A dimension could also in which case the actual dimension value is going to be copied from the input tensor DOC An input tensor Reshaped data	Output (1,"old_shape","Original shape.").InheritOnnxSchema("Reshape")
	REGISTER_GRADIENT (Reshape, GetReshapeGradient)
	REGISTER_CUDA_OPERATOR (Reshape, ReshapeOp< float, CUDAContext >)
void	resizeNearest2x (int batch_size, int num_channels, int input_height, int input_width, const float *input, float *output)
	REGISTER_CPU_OPERATOR (ResizeNearest, ResizeNearestOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (ResizeNearestGradient, ResizeNearestGradientOp< float, CPUContext >)
Scale along width dimension	Arg ("height_scale","Scale along height dimension").SetDoc(R"DOC( Resizes the spatial dimensions of the input using nearest neighbor interpolation. The `width_scale` and `height_scale` arguments control the size of the output
Scale along width dimension which is given Input tensor	Output (0,"Y","Output tensor")
	REGISTER_GRADIENT (ResizeNearest, GetResizeNearestGradient)
	REGISTER_CPU_OPERATOR (ReversePackedSegs, ReversePackedSegsOp< CPUContext >)
leaving paddings unchanged This	operator is used to reverse input of a recurrent neural network to make it a BRNN.) DOC") .Input (0,"data","a 3-D (lengths, segments, embeddings,) tensor.").Input(1
leaving paddings unchanged This length of each segment	Output (0,"reversed data","a (lengths, segments, embeddings,) tensor with each segment reversed""and paddings unchanged.")
	REGISTER_GRADIENT (ReversePackedSegs, GetReversePackedSegsGradient)
	REGISTER_CPU_OPERATOR (RMACRegions, RMACRegionsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (RecurrentNetworkBlobFetcher, RecurrentNetworkBlobFetcherOp< CPUContext >)
Prefix string to prepend extracted blobs	Input (0,"ScratchWorkspaceBlob","Name of scratch workspace blob returned by recurrent network.").Output(0
	SHOULD_NOT_DO_GRADIENT (RecurrentNetworkBlobFetcher)
	REGISTER_CUDA_OPERATOR (RecurrentNetworkBlobFetcher, RecurrentNetworkBlobFetcherOp< CUDAContext >)
template<>
std::unique_ptr< RecurrentNetworkExecutorBase >	createRNNExecutor< CPUContext > (const NetDef &step_net_def, std::map< string, string > &recurrent_input_map, std::string timestep_blob, ArgumentHelper rnn_args)
	Implementation of RecurrentNetworkExecutor that uses thread pool for multithreaded execution of RNNs. More...
template<class Context >
std::unique_ptr< RecurrentNetworkExecutorBase >	createRNNExecutor (const NetDef &step_net_def, std::map< string, string > &recurrent_input_map, std::string timestep_blob, ArgumentHelper rnn_args)
template<>
std::unique_ptr< RecurrentNetworkExecutorBase >	createRNNExecutor< CUDAContext > (const NetDef &step_net_def, std::map< string, string > &recurrent_input_map, std::string timestep_blob, ArgumentHelper arg_helper)
	CAFFE_KNOWN_TYPE (detail::ScratchWorkspaces)
	REGISTER_CPU_OPERATOR (RecurrentNetwork, RecurrentNetworkOp< CPUContext >)
INT_MAX	SetDoc (R"DOC( Run the input network in a recurrent fashion. This can be used to implement fairly general recurrent neural networks (RNNs). The operator proceeds as follows. - First, initialized the states from the input recurrent states - For each timestep T, apply the links (that map offsets from input/output tensors into the inputs/outputs for the `step` network) - Finally, alias the recurrent states to the specified output blobs. This is a fairly special-case meta-operator, and so the implementation is somewhat complex. It trades of generality (and frankly usability) against performance and control (compared to e.g. TF dynamic_rnn, Theano scan, etc). See the usage examples for a flavor of how to use it. )DOC")
	REGISTER_CPU_OPERATOR (RecurrentNetworkGradient, RecurrentNetworkGradientOp< CPUContext >)
	OPERATOR_SCHEMA (RecurrentNetworkGradient)
	REGISTER_CPU_OPERATOR (rnn_internal_accumulate_gradient_input, AccumulateInputGradientOp< CPUContext >)
INT_MAX	EnforceInplace ({{2, 0}}).Private().SetDoc(R"DOC( Internal RNN operator. )DOC")
	REGISTER_CPU_OPERATOR (rnn_internal_apply_link, RNNApplyLinkOp< CPUContext >)
	Private ().SetDoc(R"DOC( Internal RNN operator. )DOC")
	REGISTER_GRADIENT (RecurrentNetwork, GetRecurrentNetworkGradient)
	REGISTER_CUDNN_OPERATOR (Recurrent, RecurrentOp< float >)
	OPERATOR_SCHEMA (Recurrent).NumInputs(4).NumOutputs(5).SetDoc(R"DOC( Recurrent wraps the CuDNN R5 RNN implementation. See the CuDNN R5 documentation for more information. In general
the implementation takes an	input (TxNxD) tensor
the implementation takes an the hidden state	input (NxD)
the implementation takes an the hidden state the cell and a weight	tensor (effectively an opaque blob, where the size and layout is dictated by CuDNN).The outputs are the output(again
the implementation takes an the hidden state the cell and a weight the final hidden cell	states (NxD).These can be reset(at sequence boundaries across minibatches) by multiplying by zero.The CuDNN arguments(hidden_size
	REGISTER_CUDNN_OPERATOR (RecurrentGradient, RecurrentGradientOp< float >)
	NumInputs (7).NumOutputs(6).AllowInplace(
	REGISTER_CUDNN_OPERATOR (RecurrentParamSet, RecurrentParamAccessOp< float, SET_PARAM >)
	SetDoc ("Set individual parameters of a recurrent net.").Arg("param_type"
R	DOC (Type of param to be set:"input_gate_w","forget_gate_w","cell_w","output_gate_w""input_gate_b","forget_gate_b","cell_b","output_gate_b") DOC") .Arg("input_type"
R recurrent or input	Arg ("layer","layer index (starting from 0)").Input(0
R recurrent or input R	DOC (Input blob.Needed for inferring the shapes.A dummy tensor matching the input shape is ok.) DOC") .Input(1
R recurrent or input R Blob holding all the parameters	Input (2,"param","Values for the specified parameter").Output(0
R recurrent or input R Blob holding all the parameters Blob holding all the	parameters (same as input(1))")
	REGISTER_CUDNN_OPERATOR (RecurrentParamGet, RecurrentParamAccessOp< float, GET_PARAM >)
R recurrent or input R Blob holding all the parameters	Output (0,"param","Blob holding the requested values")
	REGISTER_GRADIENT (Recurrent, GetRecurrentGradient)
	REGISTER_CPU_OPERATOR (RoIAlignGradient, RoIAlignGradientOp< float, CPUContext >)
See RoIPoolF	Input (1,"RoIs","See RoIPoolF.").Input(2
See RoIPoolF Gradient of forward	output (Y)") .Output(0
See RoIPoolF Gradient of forward Gradient of forward	input (X)")
	REGISTER_GRADIENT (RoIAlign, GetRoIAlignGradient)
	REGISTER_CPU_OPERATOR (RoIAlign, RoIAlignOp< float, CPUContext >)
Spatial scale of the input feature map X relative to the input image E if X has a stride of w r t the input image	Arg ("pooled_h","(int) default 1; Pooled output Y's height.").Arg("pooled_w"
Pooled output Y s width	Arg ("sampling_ratio","(int) default -1; number of sampling points in the interpolation grid ""used to compute the output value of each pooled output bin. If > 0, ""then exactly sampling_ratio x sampling_ratio grid points are used. If ""<= 0, then an adaptive number of grid points are used (computed as ""ceil(roi_width / pooled_w), and likewise for height).").Input(0
Pooled output Y s width feature map input of	shape (N, C, H, W).") .Input( 1
	REGISTER_CPU_OPERATOR (RoIPool, RoIPoolOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (RoIPoolGradient, RoIPoolGradientOp< float, CPUContext >)
	TensorInferenceFunction ([](const OperatorDef &def, const vector< TensorShape > &in){ArgumentHelper helper(def);const StorageOrder order=StringToStorageOrder(helper.GetSingleArgument< string >("order","NCHW"));const TensorShape &X=in[0];const int num_channels=(order==StorageOrder::NCHW?X.dims(1):X.dims(3));const TensorShape &R=in[1];const int num_rois=R.dims(0);const int pooled_height=helper.GetSingleArgument< int >("pooled_h", 1);const int pooled_width=helper.GetSingleArgument< int >("pooled_w", 1);TensorShape Y=CreateTensorShape(vector< int >({num_rois, num_channels, pooled_height, pooled_width}), X.data_type());bool is_test=helper.GetSingleArgument< int >(OpSchema::Arg_IsTest, 0);if(!is_test){TensorShape argmaxes=Y;argmaxes.set_data_type(TensorProto_DataType_INT32);return vector< TensorShape >({Y, argmaxes});}else{return vector< TensorShape >({Y});}}).SetDoc(R"DOC( Carries out ROI Pooling for Faster-RCNN. Depending on the mode
there are multiple output	argmaxes (train mode) Output case) DOC") .Arg( "is_test"
there are multiple output If run in test mode and skip computation of argmaxes(used for" "gradient computation).Only one output tensor is produced." "(Default	OPERATOR_SCHEMA (RoIPoolGradient).NumInputs(4).NumOutputs(1)
	REGISTER_GRADIENT (RoIPool, GetRoIPoolGradient)
	REGISTER_CPU_OPERATOR (Scale, ScaleOp< CPUContext >)
	REGISTER_GRADIENT (Scale, GetScaleGradient)
	REGISTER_CUDA_OPERATOR (Scale, ScaleOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (SparseLengthsIndicesInGradientWeightedSumWithMainInputGradient, AbstractLengthsWithMainInputGradientOp< float, int, CPUContext, WeightedSumReducerDef::template ReducerGradient< float, CPUContext >, true, true >)
	REGISTER_CPU_OPERATOR (SparseLengthsIndicesInGradientWeightedSumGradient, AbstractLengthsGradientOp< float, int, CPUContext, WeightedSumReducerDef::template ReducerGradient< float, CPUContext >, true >)
	REGISTER_CPU_OPERATOR (SparseLengthsIndicesInGradientSumGradient, AbstractLengthsGradientOp< float, int, CPUContext, SumReducerDef::template ReducerGradient< float, CPUContext >, true >)
	OPERATOR_SCHEMA (LengthsIndicesInGradientSumGradient).NumInputs(3).NumOutputs(1)
	REGISTER_CPU_OPERATOR (LengthsIndicesInGradientSumGradient, AbstractLengthsGradientOp< float, int, CPUContext, SumReducerDef::template ReducerGradient< float, CPUContext >, true >)
	REGISTER_CPU_OPERATOR (Selu, SeluOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SeluGradient, SeluGradientOp< float, CPUContext >)
is applied to the tensor elementwise DOC	Arg ("alpha","(float) default to 1.6732~; affects the activation function itself. ""This should go with the weight initialization in the paper. "" See https://arxiv.org/abs/1706.02515 ").Arg("scale"
affects the activation function itself	Input (0,"X","input tensor").Output(0
affects the activation function itself input tensor	InheritOnnxSchema ("Selu")
	SetDoc (R"DOC( SeluGradient takes both Y and dY and uses this to update dX according to the chain rule and derivatives of the selu function. )DOC").Arg("alpha"
affects the activation function itself	Input (0,"Y","input tensor").Input(1
	REGISTER_GRADIENT (Selu, GetSeluGradient)
	REGISTER_CPU_OPERATOR (AddPadding, AddPaddingOp< CPUContext >)
	REGISTER_CPU_OPERATOR (RemovePadding, RemovePaddingOp< CPUContext >)
	REGISTER_CPU_OPERATOR (GatherPadding, GatherPaddingOp< CPUContext >)
	REGISTER_CPU_OPERATOR (PadEmptySamples, PadEmptySamplesOp< CPUContext >)
	REGISTER_GRADIENT (AddPadding, GetAddPaddingGradient)
	REGISTER_GRADIENT (RemovePadding, GetRemovePaddingGradient)
	SetDoc (R"DOC( Given a partitioned tensor T<N, D1..., Dn>, where the partitions are defined as ranges on its outer-most (slowest varying) dimension N, with given range lengths, return a tensor T<N + 2*padding_width, D1 ..., Dn> with paddings added to the start and end of each range. Optionally, different paddings can be provided for beginning and end. Paddings provided must be a tensor T<D1..., Dn>. If no padding is provided, add zero padding. If no lengths vector is provided, add padding only once, at the start and end of data. )DOC").Arg("padding_width"
Number of copies of padding to add around each range	Arg ("end_padding_width","(Optional) Specifies a different end-padding width.").Input(0
Number of copies of padding to add around each range T< N, D1..., Dn > Input data	Input (1,"lengths","(i64) Num of elements in each range. sum(lengths) = N.").Input(2
Number of copies of padding to add around each range T< N, D1..., Dn > Input data T< D1..., Dn > Padding data for range start	Input (3,"end_padding","T<D1..., Dn> (optional) Padding for range end. ""If not provided, start_padding is used as end_padding as well.").Output(0
Number of copies of padding to add around each range T< N, D1..., Dn > Input data T< D1..., Dn > Padding data for range start T< N+2 *padding_width, D1..., Dn > Padded data	Output (1,"lengths_out","(i64, optional) Lengths for each padded range.")
	SetDoc (R"DOC( Remove padding around the edges of each segment of the input data. This is the reverse opration of AddPadding, and uses the same arguments and conventions for input and output data format. )DOC").Arg("padding_width"
Outer size of padding to remove around each range T< N, D1..., Dn > Input data	Input (1,"lengths","(i64) Num of elements in each range. sum(lengths) = N. ""If not provided, considers all data as a single segment.").Output(0
Outer size of padding to remove around each range T< N, D1..., Dn > Input data T< N-2 *padding_width, D1..., Dn > Unpadded data	Output (1,"lengths_out","(i64, optional) Lengths for each unpadded range.")
	SetDoc (R"DOC( Gather the sum of start and end paddings in a padded input sequence. Used in order to compute the gradients of AddPadding w.r.t the padding tensors. )DOC").Arg("padding_width"
Outer size of padding present around each range T< N, D1..., Dn > Padded input data Sum of all start or of all paddings if end_padding_sum is not provided	Output (1,"end_padding_sum","T<D1..., Dn> Sum of all end paddings, if provided.")
INT_MAX	SetDoc (R"DOC( Pad empty field given lengths and index features, Input(0) is a blob pointing to the lengths of samples in one batch, [Input(1),... Input(num_fields)] a list of tensors containing the data for each field of the features. PadEmptySamples is thread safe. )DOC").Input(0
INT_MAX A blob containing a pointer to the lengths	Output (0,"out_lengths","Tensor containing lengths with empty sample padded.")
	REGISTER_CPU_OPERATOR (Shape, ShapeOp< CPUContext >)
out[0]	add_dims (in[0].dims().size())
out[0]	set_data_type (TensorProto::INT32)
	SetDoc ("Produce a 1D int64 tensor with the shape of the input tensor.")
	SHOULD_NOT_DO_GRADIENT (Shape)
	REGISTER_CUDA_OPERATOR (Shape, ShapeOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (Sigmoid, UnaryElementwiseOp< TensorTypes< float >, CPUContext, SigmoidCPUFunctor >)
	REGISTER_CPU_OPERATOR (SigmoidGradient, BinaryElementwiseOp< TensorTypes< float >, CPUContext, WithoutBroadcast< SigmoidGradientCPUFunctor >>)
is applied to the tensor elementwise DOC output tensor	InheritOnnxSchema ("Sigmoid")
	SetDoc (R"DOC( SigmoidGradient takes both Y and dY and uses this to update dX according to the chain rule and derivatives of the sigmoid function. )DOC")
	REGISTER_GRADIENT (Sigmoid, GetSigmoidGradient)
	REGISTER_CPU_OPERATOR (Sin, UnaryElementwiseOp< TensorTypes< float >, CPUContext, SinCPUFunctor >)
	REGISTER_CPU_OPERATOR (SinGradient, BinaryElementwiseOp< TensorTypes< float >, CPUContext, WithoutBroadcast< SinGradientCPUFunctor >>)
	OPERATOR_SCHEMA (SinGradient).NumInputs(2).NumOutputs(1).IdenticalTypeAndShape()
	REGISTER_GRADIENT (Sin, GetSinGradient)
	REGISTER_CPU_OPERATOR (SinusoidPositionEncoding, SinusoidPositionEncodingOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Slice, SliceOp< int, CPUContext >)
	REGISTER_CPU_OPERATOR (SliceGradient, SliceGradientOp< int, CPUContext >)
	NumInputs (1, 3).NumOutputs(1).SetDoc(R"DOC( Produces a slice of the input tensor. Currently
	REGISTER_CPU_OPERATOR (Softmax, SoftmaxOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SoftmaxGradient, SoftmaxGradientOp< float, CPUContext >)
it will be coerced into one For an arbitrary n dimensional tensor X in[a_0, a_1,..., a_{k-1}, a_k,..., a_{n-1}] and k is the axis then X will be coerced into a dimensional tensor with dimensions[a_0 *...*a_{k-1}, a_k *...*a_{n-1}] For the default case where this means the X tensor will be coerced into a tensor of where a_0 is often the batch size In this we must have or else the	operator will throw errors.) DOC") .Arg ("axis","(int) default to 1; describes the axis of the inputs when coerced ""to 2D; defaults to one because the 0th axis most likely describes ""the batch_size").Input(0
it will be coerced into one For an arbitrary n dimensional tensor X in[a_0, a_1,..., a_{k-1}, a_k,..., a_{n-1}] and k is the axis then X will be coerced into a dimensional tensor with dimensions[a_0 *...*a_{k-1}, a_k *...*a_{n-1}] For the default case where this means the X tensor will be coerced into a tensor of where a_0 is often the batch size In this we must have or else the The input tensor that s coerced into a matrix of	size (NxD)" "as described above.") .Output(0
it will be coerced into one For an arbitrary n dimensional tensor X in[a_0, a_1,..., a_{k-1}, a_k,..., a_{n-1}] and k is the axis then X will be coerced into a dimensional tensor with dimensions[a_0 *...*a_{k-1}, a_k *...*a_{n-1}] For the default case where this means the X tensor will be coerced into a tensor of where a_0 is often the batch size In this we must have or else the The input tensor that s coerced into a matrix of The softmax normalized output values with the same shape as input tensor	InheritOnnxSchema ("Softmax")
	OPERATOR_SCHEMA (SoftmaxGradient).NumInputs(2).NumOutputs(1)
	REGISTER_GRADIENT (Softmax, GetSoftmaxGradient)
	REGISTER_GRADIENT (SoftmaxFp16, GetSoftmaxGradient)
void	SoftmaxCPU (CPUContext &context, const int N, const int D, const float *Xdata, float *Ydata, float *scale, const float *sum_multiplier, bool logarithmic, float *rowmax)
	REGISTER_CPU_OPERATOR (SoftmaxWithLoss, SoftmaxWithLossOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SoftmaxWithLossGradient, SoftmaxWithLossGradientOp< float, CPUContext >)
vector< TensorShape >	out (2)
out[0]	set_data_type (logits.data_type())
out[0]	add_dims (batch_size)
out[0]	add_dims (num_classes)
	SetDoc (R"DOC( Combined Softmax and Cross-Entropy loss operator. The operator computes the softmax normalized values for each layer in the batch of the given input, after which cross-entropy loss is computed. This operator is numerically more stable than separate Softmax and CrossEntropy ops. The inputs are a 2-D tensor (Tensor<float>) of size (batch_size x input_feature_dimensions) and tensor of labels (ground truth). Output is tensor with the probability for each label for each example (N x D) and averaged loss (scalar). Use parameter label_prob=1 to enable inputting labels as a probability distribution. Optional third input blob can be used to weight the samples for the loss. )DOC").Input(0
Unscaled log probabilities	Input (1,"labels","Ground truth").Input(2
Unscaled log probabilities Optional blob to be used to weight the samples for the loss	Output (0,"softmax","Tensor with softmax cross entropy loss").Output(1
	OPERATOR_SCHEMA (SoftmaxWithLossGradient).NumOutputs(1)
	REGISTER_CPU_OPERATOR (Softplus, SoftplusOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SoftplusGradient, SoftplusGradientOp< float, CPUContext >)
is applied to the tensor elementwise DOC input tensor	InheritOnnxSchema ("Softplus")
	REGISTER_GRADIENT (Softplus, GetSoftplusGradient)
	REGISTER_CPU_OPERATOR (Softsign, UnaryElementwiseOp< TensorTypes< float >, CPUContext, SoftsignCPUFunctor >)
	REGISTER_CPU_OPERATOR (SoftsignGradient, BinaryElementwiseOp< TensorTypes< float >, CPUContext, WithoutBroadcast< SoftsignGradientCPUFunctor >>)
by providing the same input and output blobs DOC	Input (0,"input","1-D input tensor").Output(0
by providing the same input and output blobs DOC The	softsign (x/1+|x|) values of the input tensor" "computed element-wise") .InheritOnnxSchema("Softsign")
	SetDoc (R"DOC( Calculates the softsign gradient (sgn(x)/(1+|x|)^2) of the given input tensor element-wise. )DOC").Input(0
D input tensor	Input (1,"input","1-D input tensor").Output(0
D input tensor The softsign	gradient (sgn(x)/(1+|x|)^2) values of the input tensor" "computed element-wise")
	REGISTER_GRADIENT (Softsign, GetSoftsignGradient)
	REGISTER_CPU_OPERATOR (SpaceToBatch, SpaceToBatchOp< CPUContext >)
	OPERATOR_SCHEMA (SpaceToBatch).NumInputs(1).NumOutputs(1).SetDoc(R"DOC( SpaceToBatch for 4-D tensors of type T. Zero-pads and then rearranges (permutes) blocks of spatial data into batch. More specifically
	REGISTER_CPU_OPERATOR (BatchToSpace, BatchToSpaceOp< CPUContext >)
	OPERATOR_SCHEMA (BatchToSpace).NumInputs(1).NumOutputs(1).SetDoc(R"DOC( BatchToSpace for 4-D tensors of type T. Rearranges (permutes) data from batch into blocks of spatial data
	REGISTER_GRADIENT (SpaceToBatch, GetSpaceToBatchGradient)
	REGISTER_GRADIENT (BatchToSpace, GetBatchToSpaceGradient)
template<typename Context >
void	spaceToBatch (const Tensor< Context > &input, int pad_t, int pad_l, int block_size, Tensor< Context > *output, Context *)
template<typename Context >
void	batchToSpace (const Tensor< Context > &input, int pad_t, int pad_l, int block_size, Tensor< Context > *output, Context *)
	REGISTER_CPU_OPERATOR (SparseNormalize, SparseNormalizeOp< float, CPUContext >)
Parameters to be normalized	Input (1,"indices","Sparse indices").Input(2
Parameters to be normalized Gradient computed	Output (0,"output_param","Normalized parameters").EnforceOneToOneInplace().Arg("use_max_norm"
Parameters to be normalized Gradient computed A bool variable to control whether to use max norm or constant norm When constant norm is used so that all the embedding vectors are scaled to have a L2 norm equals to	A (see blow arugment norm=A).If use_max_norm
Parameters to be normalized Gradient computed A bool variable to control whether to use max norm or constant norm When constant norm is used so that all the embedding vectors are scaled to have a L2 norm equals to max norm is used so that embedding is scaled so that its l2 norm is no larger than A If an embedding s norm is less than A the embedding is left unchanged The default is True	Arg ("norm","L2 norm of the embedding. The default is 1.0.").SetDoc(R"DOC( Given a sparse matrix
	SHOULD_NOT_DO_GRADIENT (SparseNormalize)
	REGISTER_CPU_OPERATOR (SparseToDense, SparseToDenseOp< CPUContext >)
value represented as indices vector and values tensor into a compacted tensor where the first dimension is determined by the first dimension of the input if it is given or the max index Missing values are filled with zeros The op supports duplicated indices and performs summation over corresponding values This behavior is useful for converting GradientSlices into dense representation After running this	len (mask)]+shape(default_value)`(if`lengths`is not provided the" "first dimension is omitted)")
	REGISTER_CPU_OPERATOR (SpatialBNGradient, SpatialBNGradientOp< CPUContext >)
	NumInputs ({5, 7}).NumOutputs(3).AllowInplace(
	REGISTER_GRADIENT (SpatialBN, GetSpatialBNGradient)
	REGISTER_CPU_OPERATOR (SpatialBN, SpatialBNOp< CPUContext >)
	REGISTER_CUDA_OPERATOR (SpatialBN, CudnnSpatialBNOp)
	REGISTER_CUDA_OPERATOR (SpatialBNGradient, CudnnSpatialBNGradientOp)
	REGISTER_CUDNN_OPERATOR (SpatialBN, CudnnSpatialBNOp)
	REGISTER_CUDNN_OPERATOR (SpatialBNGradient, CudnnSpatialBNGradientOp)
	REGISTER_CPU_OPERATOR (SpatialSoftmaxWithLoss, SpatialSoftmaxWithLossOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SpatialSoftmaxWithLossGradient, SpatialSoftmaxWithLossGradientOp< float, CPUContext >)
	CAFFE_ENFORCE_EQ (logits.dims_size(), 4)
	CAFFE_ENFORCE_EQ (labels.dims_size(), 3)
	SetDoc (R"DOC( Combined Spatial Softmax and Cross-Entropy loss operator. Similar to SoftmaxWithLoss, this operator computes the spatial softmax normalized values for each layer in the batch of the given input, after which cross-entropy loss is computed. This operator is numerically more stable than separate Softmax and CrossEntropy ops. The inputs are a 2-D tensor (Tensor<float>) of size (batch_size x input_feature_dimensions) and tensor of labels (ground truth). Output is tensor with the probability for each label in a pixel for each example (N x D x W x H) and averaged loss (scalar). For spatial softmax, weighting is by x,y position of the input. )DOC").Input(0
	OPERATOR_SCHEMA (SpatialSoftmaxWithLossGradient).NumOutputs(1)
	REGISTER_CPU_OPERATOR (Sqrt, UnaryElementwiseOp< TensorTypes< float >, CPUContext, SqrtCPUFunctor >)
	REGISTER_GRADIENT (Sqrt, GetSqrtGradient)
	REGISTER_CPU_OPERATOR (SquareRootDivide, SquareRootDivideOp< CPUContext >)
	SetDoc (R"DOC( Given DATA tensor with first dimension N and SCALE vector of the same size N produces an output tensor with same dimensions as DATA. Which consists of DATA slices. i-th slice is divided by sqrt(SCALE[i]) elementwise. If SCALE[i] == 0 output slice is identical to the input one (no scaling) Example: Data = [ [2.0, 4.0], [9.0, 12.0] ] SCALE = [4, 9] OUTPUT = [ [1.0, 2.0], [3.0, 4.0] ] )DOC")
	REGISTER_GRADIENT (SquareRootDivide, GetSquareRootDivideGradient)
	REGISTER_CPU_OPERATOR (StatRegistryCreate, StatRegistryCreateOp)
	REGISTER_CPU_OPERATOR (StatRegistryUpdate, StatRegistryUpdateOp)
	REGISTER_CPU_OPERATOR (StatRegistryExport, StatRegistryExportOp)
	REGISTER_CPU_OPERATOR (TimerBegin, TimerBeginOp)
	REGISTER_CPU_OPERATOR (TimerEnd, TimerEndOp)
	REGISTER_CPU_OPERATOR (TimerGetAndEnd, TimerGetAndEndOp)
	REGISTER_CPU_OPERATOR (TimerGet, TimerGetOp)
	REGISTER_CPU_OPERATOR (CpuUtilizationReport, CpuUtilizationReportOp)
or the global with the values of counters for the given keys DOC	Input (0,"keys","1D string tensor with the key names to update.").Input(1
or the global with the values of counters for the given keys DOC int64 tensor with the values to update	Input (2,"handle","If provided, update the given StatRegistry. ""Otherwise, update the global singleton.")
If export values from given StatRegistry export values from the global singleton StatRegistry	Output (0,"keys","1D string tensor with exported key names").Output(1
If export values from given StatRegistry export values from the global singleton StatRegistry int64 tensor with exported values	Output (2,"timestamps","The unix timestamp at counter retrieval.").Arg("reset"
returning a pointer to it The timer is stopped by calling TimerEnd DOC	Arg ("counter_name","Name of the timer. If not provided, use output name.").Output(0
publishing a CAFFE_EVENT	Input (0,"timer","Pointer to timer, obtained from TimerBegin.")
Pointer to obtained from TimerBegin	Output (0,"nanos","nanoseconds in int64")
Delta in max CPU utilization in percentage as a float value	Arg ("stats_name","String name of the stat entry holding CPU utilization")
	CAFFE_KNOWN_TYPE (TimerInstance *)
	CAFFE_KNOWN_TYPE (std::unique_ptr< caffe2::StatRegistry >)
	REGISTER_CPU_OPERATOR (StopGradient, StopGradientOp< CPUContext >)
	NumInputs (1, 1).NumOutputs(1
	NO_GRADIENT (StopGradient)
	REGISTER_CUDA_OPERATOR (StopGradient, StopGradientOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (StumpFunc, StumpFuncOp< float, float, CPUContext >)
tensor of float	Output (0,"Y","tensor of float").SetDoc(R"DOC( Converts each input element into either high_ or low_value based on the given threshold. )DOC")
	NO_GRADIENT (StumpFunc)
	REGISTER_CPU_OPERATOR (Summarize, SummarizeOp< float, CPUContext >)
	SetDoc (R"DOC( Summarize computes four statistics of the input tensor (Tensor<float>)- min, max, mean and standard deviation. The output will be written to a 1-D tensor of size 4 if an output tensor is provided. Else, if the argument 'to_file' is greater than 0, the values are written to a log file in the root folder. )DOC").Arg("to_file"
default flag to indicate if the summarized statistics have to be written to a log file	Input (0,"data","The input data as Tensor<float>.").Output(0
	SHOULD_NOT_DO_GRADIENT (Summarize)
	REGISTER_CPU_OPERATOR (Swish, UnaryElementwiseOp< TensorTypes< float, double >, CPUContext, SwishCPUFunctor >)
	REGISTER_CPU_OPERATOR (SwishGradient, SwishGradientOp< CPUContext >)
	SetDoc (R"DOC( SwishGradient takes X, Y and dY and uses this to update dX according to the chain rule and derivatives of the swish function. )DOC")
	REGISTER_GRADIENT (Swish, GetSwishGradient)
	REGISTER_CPU_OPERATOR (Tanh, UnaryElementwiseOp< TensorTypes< float >, CPUContext, TanhCPUFunctor >)
	REGISTER_CPU_OPERATOR (TanhGradient, BinaryElementwiseOp< TensorTypes< float >, CPUContext, WithoutBroadcast< TanhGradientCPUFunctor >>)
by providing the same input and output blobs DOC The hyperbolic tangent values of the input tensor computed element wise	InheritOnnxSchema ("Tanh")
	OPERATOR_SCHEMA (TanhGradient).NumInputs(2).NumOutputs(1).AllowInplace(
	REGISTER_GRADIENT (Tanh, GetTanhGradient)
	REGISTER_CPU_OPERATOR (TensorProtosDBInput, TensorProtosDBInput< CPUContext >)
INT_MAX	SetDoc (R"DOC( TensorProtosDBInput is a simple input operator that basically reads things from a db where each key-value pair stores an index as key, and a TensorProtos object as value. These TensorProtos objects should have the same size, and they will be grouped into batches of the given size. The DB Reader is provided as input to the operator and it returns as many output tensors as the size of the TensorProtos object. Each output will simply be a tensor containing a batch of data with size specified by the 'batch_size' argument containing data from the corresponding index in the TensorProtos objects in the DB. )DOC").Arg("batch_size"
INT_MAX default the number of samples in a batch The default value of means that the	operator will attempt to insert the" "entire data in a single output blob.") .Input (0,"data","A pre-initialized DB reader. Typically, this is obtained ""by calling CreateDB operator with a db_name and a db_type. The ""resulting output blob is a DB Reader tensor").Output(0
INT_MAX default the number of samples in a batch The default value of means that the The output tensor in which the batches of data are returned The number of output tensors is equal to the size	of (number of TensorProto's in) the TensorProtos objects stored in the" "DB as values.Each output tensor will be of size specified by the" "'batch_size'argument of the operator")
	NO_GRADIENT (TensorProtosDBInput)
	REGISTER_CUDA_OPERATOR (TensorProtosDBInput, TensorProtosDBInput< CUDAContext >)
void	convert (TensorProto_DataType dst_type, const char *src_start, const char *src_end, void *dst)
	CAFFE_KNOWN_TYPE (std::unique_ptr< TextFileReaderInstance >)
	REGISTER_CPU_OPERATOR (CreateTextFileReader, CreateTextFileReaderOp)
	REGISTER_CPU_OPERATOR (TextFileReaderRead, TextFileReaderReadOp)
Path to the file	Arg ("num_passes","Number of passes over the file.").Arg("field_types"
Path to the file List with type of each field Type enum is found at core DataType	Output (0,"handler","Pointer to the created TextFileReaderInstance.")
INT_MAX	SetDoc ("Read a batch of rows from the given text file reader instance. ""Expects the number of fields to be equal to the number of outputs. ""Each output is a 1D tensor containing the values for the given field ""for each row. When end of file is reached, returns empty tensors.").Input(0
INT_MAX Pointer to an existing TextFileReaderInstance	Arg ("batch_size","Maximum number of rows to read.")
	NO_GRADIENT (CreateTextFileReader)
	NO_GRADIENT (TextFileReaderRead)
	REGISTER_CPU_OPERATOR (ThresholdedRelu, ThresholdedReluOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (ThresholdedReluGradient, ThresholdedReluGradientOp< float, CPUContext >)
	CostInferenceFunction (PointwiseCostInference< 2 >).IdenticalTypeAndShape().SetDoc(R"DOC( ThresholdedRelu takes one input data (Tensor) and produces one output data (Tensor) where the rectified linear function
	REGISTER_CPU_OPERATOR (Tile, TileOp< CPUContext >)
	REGISTER_CPU_OPERATOR (TileGradient, TileGradientOp< float, CPUContext >)
	if (in.size() > 1)
out[0]	set_dims (canonical_axis, out[0].dims().Get(canonical_axis)*tiles)
	SetDoc (R"DOC( Constructs a tensor by tiling a given tensor along a specified axis. This operation creates a new tensor by replicating the input tensor 'tiles' times along dimension 'axis'. The output tensor's 'axis'th dimension has input.dims(axis) * tiles elements, and the values of input are replicated 'tiles' times along the 'axis'th dimension. For example, tiling [[a b c d]] by tile=2, axis=0 produces [[a b c d], [a b c d]]. )DOC").Arg("tiles"
Number of replicas	Arg ("axis","Axis to replicate along").Input(0
Number of replicas The input tensor	Input (1,"tiles","(optional) Number of replicas (overrides argument)").Input(2
Number of replicas The input tensor optional Axis to replicate	along (overrides argument)") .Output( 0
Number of replicas The input tensor optional Axis to replicate Tensor that will contain input replicated along the given axis	InheritOnnxSchema ("Tile")
	OPERATOR_SCHEMA (TileGradient).NumInputs(1
	REGISTER_GRADIENT (Tile, GetTileGradient)
	REGISTER_CPU_OPERATOR (TopK, TopKOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (TopKGradient, TopKGradientOp< float, CPUContext >)
	TensorInferenceFunction ([](const OperatorDef &def, const vector< TensorShape > &in){vector< TensorShape > out={in[0], in[0]};ArgumentHelper helper(def);auto k=helper.GetSingleArgument("k",-1);auto dims_size=in[0].dims_size();out[0].set_dims(dims_size-1, k);out[1].set_dims(dims_size-1, k);out[1].set_data_type(TensorProto_DataType_INT32);if(def.output_size() > 2){TensorShape flatten_indices_shape;flatten_indices_shape.set_data_type(TensorProto_DataType_INT32);flatten_indices_shape.add_dims(std::accumulate(in[0].dims().begin(), in[0].dims().end()-1, 1, std::multiplies< long >())*k);out.push_back(flatten_indices_shape);}return out;}).SetDoc(R"DOC( Retrieve the top-K elements for the last dimension. Given an input tensor of shape [a_1
r and integer argument return two k which contains the values of the top k elements along the last dimension Index tensor of shape[a_1, a_2,..., a_n, k] which contains the indices of the top k	elements (original indices from the input tensor).Given two equivalent values
r and integer argument return two k which contains the values of the top k elements along the last dimension Index tensor of shape[a_1, a_2,..., a_n, k] which contains the indices of the top k this	operator uses the indices along the last dim-ension as a tiebreaker.That is, the element with the lower index will appear first.) DOC") .Input (0,"X","Tensor of shape [a_1, a_2, ..., a_n, r]").Output(0
r and integer argument return two k which contains the values of the top k elements along the last dimension Index tensor of shape[a_1, a_2,..., a_n, k] which contains the indices of the top k this Tensor of shape[a_1, a_2,..., a_n, k] containing top K values from the input tensor	Output (1,"Indices","Tensor of shape [a_1, a_2, ..., a_n, k] containing"" the corresponding input tensor indices for the top K values.").Output(2
r and integer argument return two k which contains the values of the top k elements along the last dimension Index tensor of shape[a_1, a_2,..., a_n, k] which contains the indices of the top k this Tensor of shape[a_1, a_2,..., a_n, k] containing top K values from the input tensor Flatten Tensor of shape[a_1 *a_2 *...*a_n *k] containing the indices into the flatten input	Arg ("k","Number of top elements to retrieve")
	OPERATOR_SCHEMA (TopKGradient).NumInputs(3).NumOutputs(1)
	REGISTER_GRADIENT (TopK, GetTopKGradient)
	REGISTER_CPU_OPERATOR (Transpose, TransposeOp< CPUContext >)
	if (axes.empty())
	CAFFE_ENFORCE (valid_axes,"Axes argument passed in had invalid values")
	CAFFE_ENFORCE (axes.size()==tensor_size,"Axes argument passed in had the incorrect size")
	for (auto axis=axes.begin();axis!=axes.end();++axis)
	SetDoc (R"DOC( Transpose the input tensor similar to numpy.transpose. For example, when axes=(1, 0, 2), given an input tensor of shape (1, 2, 3), the output shape will be (2, 1, 3). )DOC").Arg("axes"
A list of integers By reverse the otherwise permute the axes according to the values given Transposed output	InheritOnnxSchema ("Transpose")
	REGISTER_GRADIENT (Transpose, GetTransposeGradient)
	REGISTER_CUDNN_OPERATOR (Transpose, CuDNNTransposeOp)
	REGISTER_CPU_OPERATOR (WallClockTime, WallClockTimeOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Print, PrintOp< CPUContext >)
	REGISTER_CPU_OPERATOR (FlattenToVec, FlattenToVecOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Alias, AliasOp< CPUContext >)
	REGISTER_CPU_OPERATOR (ResizeLike, ResizeLikeOp< CPUContext >)
	REGISTER_CPU_OPERATOR (SumInt, SumOp< CPUContext >)
	REGISTER_CPU_OPERATOR (WeightedSum, WeightedSumOp< CPUContext >)
	REGISTER_CPU_OPERATOR (WeightedSumGradient, WeightedSumGradientOp< CPUContext >)
	REGISTER_CPU_OPERATOR (ScatterWeightedSum, ScatterWeightedSumOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (ScatterAssign, ScatterAssignOp< CPUContext >)
	REGISTER_CPU_OPERATOR (EnsureCPUOutput, CopyOp< CPUContext, CPUContext, CPUContext >)
	REGISTER_CPU_OPERATOR (CopyFromCPUInput, CopyOp< CPUContext, CPUContext, CPUContext >)
	REGISTER_CPU_OPERATOR (CopyOnDeviceLike, CopyOnDeviceLikeOp< CPUContext, CPUContext, CPUContext >)
	REGISTER_CPU_OPERATOR (Copy, CopyOp< CPUContext, CPUContext, CPUContext >)
	REGISTER_CPU_OPERATOR (LengthsToShape, LengthsToShapeOp< CPUContext >)
	REGISTER_CPU_OPERATOR (HasElements, HasElementsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (IsEmpty, IsEmptyOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Gather, GatherOp< CPUContext >)
	REGISTER_CPU_OPERATOR (GatherRanges, GatherRangesOp< CPUContext >)
	REGISTER_CPU_OPERATOR (LengthsGather, LengthsGatherOp< CPUContext >)
	REGISTER_CPU_OPERATOR (Unique, UniqueOp< CPUContext >)
	REGISTER_CPU_OPERATOR (LengthsToSegmentIds, LengthsToSegmentIdsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (LengthsToRanges, LengthsToRangesOp< CPUContext >)
	REGISTER_CPU_OPERATOR (SegmentIdsToLengths, SegmentIdsToLengthsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (SegmentIdsToRanges, SegmentIdsToRangesOp< CPUContext >)
	REGISTER_CPU_OPERATOR (LengthsToWeights, LengthsToWeightsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (EnsureDense, EnsureDenseOp< CPUContext >)
	REGISTER_CPU_OPERATOR (AccumulateHistogram, AccumulateHistogramOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (UnsafeCoalesce, UnsafeCoalesceOp< CPUContext >)
bool saves contents to the root folder of the current appending the tensor contents to a file named after the blob name logs to stderr	Input (0,"tensor","The tensor to print.")
	OPERATOR_SCHEMA (LengthsToShape).NumInputs(1).NumOutputs(1)
out[0]	add_dims (total)
	SetDoc (R"DOC( Flattens the input tensor into a 1D vector. )DOC").Input(0
out	push_back (in[1])
	SetDoc (R"DOC( Produces tensor containing data of first input and shape of second input. )DOC").Input(0
Tensor whose data will be copied into the output	Input (1,"shape_tensor","Tensor whose shape will be applied to output.").Output(0
	NumInputs ([](int n){return(n > 0 &&n%2==0);}).NumOutputs(1).AllowInplace(
weight tensor pairs Input should be in the form where X_i all have the same and weight_i are size tensors that specifies the weight of each vector Note that if one wants to do in place it could only be done with X_0 also as the but not other X_i DOC	Input (0,"data_0","First of the input tensors.").Input(0
weight tensor pairs Input should be in the form where X_i all have the same and weight_i are size tensors that specifies the weight of each vector Note that if one wants to do in place it could only be done with X_0 also as the but not other X_i DOC Weight of the first input in the sum	Output (0,"output","Result containing weighted elem-wise sum of inputs.")
	NumInputs ([](int n){return(n > 0 &&n%2==1);}).NumOutputs(1
	NumInputs ([](int n){return(n > 3 &&(n-3)%2==0);}).NumOutputs(1).EnforceInplace(
	SetDoc (R"DOC( Similar to WeightedSum, computes the weighted sum of several tensors, with the difference that inputs are sliced tensors. The first tensor has to be in-place and only slices of it on the first dimension as indexed by INDICES will be updated. Note: The op pretty much ignores the exact shapes of the input arguments and cares only about sizes. It's done for performance consideration to avoid unnecessary reshapes. Only first dimension of X_0 is important, let's call it N. If M is the total size of X_0 and K is the size of INDICES then X_i is assumed to be of shape K x (M / N) regardless of the real shape. Note: Each update in INDICES is applied independently which means that if duplicated elements are present in INDICES the corresponding slice of X_0 will be scaled multiple times. Manual collapsing of INDICES is required beforehand if necessary. Note: Updates are applied sequentially by inputs which might have undesired consequences if the input tensor is accessed concurrently by different op (e.g. when doing Hogwild). Other threads might see intermediate results even on individual slice level, e.g. X_0 scaled by weight_0 but without any updates applied. Currently only works on CPU because of access to INDICES. )DOC").Input(0
Tensor to be updated	Input (1,"Weight_0","Scalar weight for X_0, applied only to slices affected.").Input(2
Tensor to be updated D list of indices on the first dimension of X_0 that need to be updated	Input (3,"X_1","Update slices, with shape len(INDICES) + shape(X_0)[1:]").Input(4
Tensor to be updated D list of indices on the first dimension of X_0 that need to be updated Scalar weight for X_1 update	Output (0,"X_0","Has to be exactly the same tensor as the input 0").EnforceInplace(
	SetDoc (R"DOC( Update slices of the tensor in-place by overriding current value. Note: The op pretty much ignores the exact shapes of the input arguments and cares only about sizes. It's done for performance consideration to avoid unnecessary reshapes. Only first dimension of X_0 is important, let's call it N. If M is the total size of X_0 and K is the size of INDICES then X_i is assumed to be of shape K x (M / N) regardless of the real shape. Note: Each update in INDICES is applied independently which means that if duplicated elements are present in INDICES arbitrary one will win. Currently only works on CPU because of access to INDICES. )DOC").Input(0
Tensor to be updated	Input (1,"INDICES","1-D list of indices on the first dimension""of X_0 that need to be updated").Input(2
	REGISTER_CUDA_OPERATOR (Print, PrintOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (Flatten, FlattenOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (FlattenToVec, FlattenToVecOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (Alias, AliasOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (ResizeLike, ResizeLikeOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (Sum, SumOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (WeightedSum, WeightedSumOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (EnsureCPUOutput, CopyOp< CUDAContext, CPUContext, CUDAContext >)
	REGISTER_CUDA_OPERATOR (CopyFromCPUInput, CopyOp< CUDAContext, CUDAContext, CPUContext >)
	REGISTER_CUDA_OPERATOR (CopyGPUToCPU, CopyOp< CUDAContext, CPUContext, CUDAContext >)
	REGISTER_CUDA_OPERATOR (CopyCPUToGPU, CopyOp< CUDAContext, CUDAContext, CPUContext >)
	REGISTER_CUDA_OPERATOR (Copy, CopyOp< CUDAContext, CUDAContext, CUDAContext >)
	REGISTER_CUDA_OPERATOR (CopyOnDeviceLike, CopyOnDeviceLikeOp< CUDAContext, CUDAContext, CUDAContext >)
	REGISTER_CUDA_OPERATOR (UnsafeCoalesce, UnsafeCoalesceOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (VariableLengthSequencePadding, VariableLengthSequencePaddingOp< float, CPUContext >)
	SetDoc (R"DOC( Super special-case operator. Used to pad a tensor to mimic pytorch's pad_packed_sequence. Given an input tensor INPUT of size NxBxM and an input tensor LENS of size B, where N = maximum sequence length B = batch size M = hidden size set each element of INPUT to zero if it is is past the end of the corresponding sequence (i.e. if LENS[j] > i for an index (i,j,k)). )DOC")
	REGISTER_CPU_OPERATOR (WeightedMultiSampling, WeightedMultiSamplingOp< CPUContext >)
	if (in[0].dims(0)==0)
	if (args.HasArgument("num_samples"))
	SetDoc (R"DOC( The operator performs sampling based on the input sampling weights. All weights are cummulative probability thus sorted. The output is a 1-D tensor (Tensor<int>). If two inputs are given, the second input is used to provide shape of the output sample tensor. Otherwise, we use argument `num_samples` to determine the number of samples to generate. )DOC").Input(0
An optional D Tensor< float > Input cumulative sampling	probability (such as[0.2, 0.5, 0.8, 1.5])." "All weights must be non-negative numbers.Note that the last value of" "CDF is not necessary 1.If the last value is not 1
An optional D Tensor< float > Input cumulative sampling all values in sampling_cdf will be scaled by this number	Input (1,"shape_tensor (optional)","Tensor whose shape will be applied to output.").Output(0
An optional D Tensor< float > Input cumulative sampling all values in sampling_cdf will be scaled by this number The output tensor contains indices sampled from distribution given by the weight vector in the input tensor The output is a D Tensor< int > of size determined by argument num_samples or the second input tensor	Arg ("num_samples","number of samples to sample from the input data")
	SHOULD_NOT_DO_GRADIENT (WeightedMultiSample)
	REGISTER_CPU_OPERATOR (WeightedSample, WeightedSampleOp< float, CPUContext >)
	TensorInferenceFunction ([](const OperatorDef &def, const vector< TensorShape > &in){vector< TensorShape > out(2);int batch_size=in[0].dims(0);out[0]=CreateTensorShape(vector< int >{batch_size}, TensorProto::INT32);out[1]=CreateTensorShape(vector< int >{batch_size}, TensorProto::FLOAT);return out;}).SetDoc(R"DOC( The operator performs sampling based on the input sampling weights for each batch. All weights must be non-negative numbers. The input is a 2-D tensor (Tensor<float>) of size (batch_size x weights_dim). For each batch
an index is randomly sampled from the distribution given by the weights of the corresponding batch The output is a D	tensor (Tensor< int >) of size(batch_size x 1) and contains the index(es) of the sampled output.) DOC") .Input( 0
an index is randomly sampled from the distribution given by the weights of the corresponding batch The output is a D A D Tensor< float > of	size (batch_size x weights_dim)." "All weights must be non-negative numbers.") .Input( 1
an index is randomly sampled from the distribution given by the weights of the corresponding batch The output is a D A D Tensor< float > of An optional D Tensor< float > of The output tensor contains	index (es) sampled from distribution given" "by the weight vector(s) in the input tensor" "The output is a 1-D Tensor< int > of size(batch_size x 1)") .Output( 1
an index is randomly sampled from the distribution given by the weights of the corresponding batch The output is a D A D Tensor< float > of An optional D Tensor< float > of The output tensor contains The output tensor contains	value (s) selected by the sampled index(es)" "It is a 1-D Tensor< float > of size(batch_size x 1)")
	SHOULD_NOT_DO_GRADIENT (WeightedSample)
	REGISTER_CPU_OPERATOR (While, WhileOp< CPUContext >)
INT_MAX	SetDoc (R"DOC( 'While' control operator, first input is a scalar boolean blob that stores loop's condition value. Accepts 'loop_net' (required) and 'cond_net' (optional) arguments for loop's body and condition subnets respectively. If condition subnet is specified, it is executed before the first and after each iteration. Subnets are executed in the same workspace as 'While'. )DOC").Arg("loop_net"
INT_MAX Net executed on each iteration	Arg ("cond_net","Net to (re)compute condition value").Input(0
	REGISTER_CUDA_OPERATOR (While, WhileOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (ZeroGradient, ZeroGradientOp< CPUContext >)
	REGISTER_GRADIENT (ZeroGradient, GetZeroGradientOpGradient)
	REGISTER_CUDA_OPERATOR (ZeroGradient, ZeroGradientOp< CUDAContext >)
	EMBEDDING_SPECIALIZATION (int32_t, float, float, false)
	EMBEDDING_SPECIALIZATION (int64_t, float, float, false)
	EMBEDDING_SPECIALIZATION (int32_t, float16, float, false)
	EMBEDDING_SPECIALIZATION (int64_t, float16, float, false)
	EMBEDDING_SPECIALIZATION (int32_t, uint8_t, float, false)
	EMBEDDING_SPECIALIZATION (int64_t, uint8_t, float, false)
	EMBEDDING_SPECIALIZATION (int32_t, float, float, true)
	EMBEDDING_SPECIALIZATION (int64_t, float, float, true)
	EMBEDDING_SPECIALIZATION (int32_t, float16, float, true)
	EMBEDDING_SPECIALIZATION (int64_t, float16, float, true)
	EMBEDDING_SPECIALIZATION (int32_t, uint8_t, float, true)
	EMBEDDING_SPECIALIZATION (int64_t, uint8_t, float, true)
template<typename IndexType , typename InType , typename OutType , bool IS_WEIGHT_POSITIONAL = false>
void	EmbeddingLookup (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const InType *input, const IndexType *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, OutType *out)
	Embedding lookup with reduction. More...
void	EmbeddingLookup_int32_t_float_float_false__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float *input, const int32_t *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, float *out)
void	EmbeddingLookup_int32_t_float_float_true__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float *input, const int32_t *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, float *out)
void	EmbeddingLookup_int64_t_float_float_false__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float *input, const int64_t *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, float *out)
void	EmbeddingLookup_int64_t_float_float_true__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float *input, const int64_t *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, float *out)
void	EmbeddingLookup_int32_t_float16_float_false__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float16 *input, const int32_t *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, float *out)
void	EmbeddingLookup_int32_t_float16_float_true__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float16 *input, const int32_t *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, float *out)
void	EmbeddingLookup_int64_t_float16_float_false__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float16 *input, const int64_t *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, float *out)
void	EmbeddingLookup_int64_t_float16_float_true__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float16 *input, const int64_t *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, float *out)
void	EmbeddingLookup_int32_t_uint8_t_float_false__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const uint8_t *input, const int32_t *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, float *out)
void	EmbeddingLookup_int32_t_uint8_t_float_true__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const uint8_t *input, const int32_t *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, float *out)
void	EmbeddingLookup_int64_t_uint8_t_float_false__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const uint8_t *input, const int64_t *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, float *out)
void	EmbeddingLookup_int64_t_uint8_t_float_true__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const uint8_t *input, const int64_t *indices, const int *lengths, const float *weights, const float *scale_bias, bool normalize_by_lengths, float *out)
void	Fused8BitRowwiseEmbeddingLookup_int32_t_float_float_false__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float *input, const int32_t *indices, const int *lengths, const float *weights, bool normalize_by_lengths, float *out)
void	Fused8BitRowwiseEmbeddingLookup_int32_t_float_float_true__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float *input, const int32_t *indices, const int *lengths, const float *weights, bool normalize_by_lengths, float *out)
void	Fused8BitRowwiseEmbeddingLookup_int64_t_float_float_false__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float *input, const int64_t *indices, const int *lengths, const float *weights, bool normalize_by_lengths, float *out)
void	Fused8BitRowwiseEmbeddingLookup_int64_t_float_float_true__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float *input, const int64_t *indices, const int *lengths, const float *weights, bool normalize_by_lengths, float *out)
void	Fused8BitRowwiseEmbeddingLookup_int32_t_float16_float_false__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float16 *input, const int32_t *indices, const int *lengths, const float *weights, bool normalize_by_lengths, float *out)
void	Fused8BitRowwiseEmbeddingLookup_int32_t_float16_float_true__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float16 *input, const int32_t *indices, const int *lengths, const float *weights, bool normalize_by_lengths, float *out)
void	Fused8BitRowwiseEmbeddingLookup_int64_t_float16_float_false__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float16 *input, const int64_t *indices, const int *lengths, const float *weights, bool normalize_by_lengths, float *out)
void	Fused8BitRowwiseEmbeddingLookup_int64_t_float16_float_true__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const float16 *input, const int64_t *indices, const int *lengths, const float *weights, bool normalize_by_lengths, float *out)
void	Fused8BitRowwiseEmbeddingLookup_int32_t_uint8_t_float_false__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const uint8_t *input, const int32_t *indices, const int *lengths, const float *weights, bool normalize_by_lengths, float *out)
void	Fused8BitRowwiseEmbeddingLookup_int32_t_uint8_t_float_true__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const uint8_t *input, const int32_t *indices, const int *lengths, const float *weights, bool normalize_by_lengths, float *out)
void	Fused8BitRowwiseEmbeddingLookup_int64_t_uint8_t_float_false__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const uint8_t *input, const int64_t *indices, const int *lengths, const float *weights, bool normalize_by_lengths, float *out)
void	Fused8BitRowwiseEmbeddingLookup_int64_t_uint8_t_float_true__avx2_fma (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const uint8_t *input, const int64_t *indices, const int *lengths, const float *weights, bool normalize_by_lengths, float *out)
	FUSED_8BIT_ROWWISE_EMBEDDING_SPECIALIZATION (int32_t, uint8_t, float)
	FUSED_8BIT_ROWWISE_EMBEDDING_SPECIALIZATION (int64_t, uint8_t, float)
template<typename IndexType , typename InType , typename OutType , bool IS_WEIGHT_POSITIONAL = false>
void	Fused8BitRowwiseEmbeddingLookup (const TIndex block_size, const TIndex output_size, const TIndex index_size, const TIndex data_size, const InType *input, const IndexType *indices, const int *lengths, const float *weights, bool normalize_by_lengths, OutType *out)
	Embedding lookup with reduction. More...
template<>
void	TypedAxpy< float, float > (int N, const float a, const float *x, float *y)
void	TypedAxpy_float16_float__base (int N, const float a, const float16 *x, float *y)
template<>
void	TypedAxpy< float16, float > (int N, const float a, const float16 *x, float *y)
void	TypedAxpy_uint8_float__base (int N, const float a, const std::uint8_t *x, float *y)
template<>
void	TypedAxpy< std::uint8_t, float > (int N, const float a, const std::uint8_t *x, float *y)
template<typename IN , typename OUT >
void	TypedAxpy (int N, const OUT a, const IN *x, OUT *y)
void	TypedAxpy_float16_float__avx_f16c (int N, const float a, const float16 *x, float *y)
void	TypedAxpy_float16_float__avx2_fma (int N, const float a, const float16 *x, float *y)
void	TypedAxpy_uint8_float__avx2_fma (int N, const float a, const std::uint8_t *x, float *y)
	CAFFE_KNOWN_TYPE (std::shared_ptr< BlobsQueue >)
	REGISTER_CPU_OPERATOR (CreateBlobsQueue, CreateBlobsQueueOp< CPUContext >)
	REGISTER_CPU_OPERATOR (EnqueueBlobs, EnqueueBlobsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (DequeueBlobs, DequeueBlobsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (CloseBlobsQueue, CloseBlobsQueueOp< CPUContext >)
	REGISTER_CPU_OPERATOR (SafeEnqueueBlobs, SafeEnqueueBlobsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (SafeDequeueBlobs, SafeDequeueBlobsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (WeightedSampleDequeueBlobs, WeightedSampleDequeueBlobsOp< CPUContext >)
	OPERATOR_SCHEMA (CreateBlobsQueue).NumInputs(0).NumOutputs(1)
	NumInputsOutputs ([](int inputs, int outputs){return inputs >=2 &&outputs >=1 &&inputs==outputs+1;}).EnforceInplace([](int input
	NumInputsOutputs ([](int inputs, int outputs){return inputs==1 &&outputs >=1;}).SetDoc(R"DOC( Dequeue the blobs from queue. )DOC").Arg("timeout_secs"
Timeout in The shared pointer for the BlobsQueue	Output (0,"blob","The blob to store the dequeued data")
	OPERATOR_SCHEMA (CloseBlobsQueue).NumInputs(1).NumOutputs(0)
	NumInputsOutputs ([](int inputs, int outputs){return inputs >=2 &&outputs >=2 &&inputs==outputs;}).EnforceInplace([](int input
	SetDoc (R"DOC( Enqueue the blobs into queue. When the queue is closed and full, the output status will be set to true which can be used as exit criteria for execution step. The 1st input is the queue and the last output is the status. The rest are data blobs. )DOC").Input(0
	NumInputsOutputs ([](int inputs, int outputs){return inputs==1 &&outputs >=2;}).SetDoc(R"DOC( Dequeue the blobs from queue. When the queue is closed and empty
the output status will be set to true which can be used as exit criteria for execution step The input is the queue and the last output is the status The rest are data blobs DOC	Arg ("num_records","(default 1) If > 1, multiple records will be dequeued and tensors ""for each column will be concatenated. This requires all tensors in ""the records to be at least 1D, and to have the same inner dimensions.").Input(0
INT_MAX	SetDoc (R"DOC( Dequeue the blobs from multiple queues. When one of queues is closed and empty, the output status will be set to true which can be used as exit criteria for execution step. The 1st input is the queue and the last output is the status. The rest are data blobs. )DOC").Arg("weights"
INT_MAX Weights for sampling from multiple queues	Arg ("table_idx_blob","The index of the blob (among the output blob list) ""that will be used to store the index of the table chosen to read the ""current batch.")
	NO_GRADIENT (CreateBlobsQueue)
	NO_GRADIENT (EnqueueBlobs)
	NO_GRADIENT (DequeueBlobs)
	NO_GRADIENT (CloseBlobsQueue)
	NO_GRADIENT (SafeEnqueueBlobs)
	NO_GRADIENT (SafeDequeueBlobs)
	NO_GRADIENT (WeightedSampleDequeueBlobs)
	REGISTER_CUDA_OPERATOR (CreateBlobsQueue, CreateBlobsQueueOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (EnqueueBlobs, EnqueueBlobsOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (DequeueBlobs, DequeueBlobsOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (CloseBlobsQueue, CloseBlobsQueueOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (SafeEnqueueBlobs, SafeEnqueueBlobsOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (SafeDequeueBlobs, SafeDequeueBlobsOp< CUDAContext >)
	CAFFE_KNOWN_TYPE (RebatchingQueuePtr)
	REGISTER_CPU_OPERATOR (Adagrad, AdagradOp< float, CPUContext >)
	SetDoc (R"DOC( Computes the AdaGrad update for an input gradient and accumulated history. Concretely, given inputs (param, grad, moment, learning_rate), computes new_moment = moment + square(grad) new_grad = learning_rate * grad / (sqrt(new_moment) + epsilon) new_param = param + new_grad and returns (new_param, new_moment). )DOC").Input(0
Parameters to be updated	Input (1,"moment","Moment history").Input(2
Parameters to be updated Gradient computed	Input (3,"lr","learning rate").Output(0
Parameters to be updated Gradient computed Updated parameters	Output (1,"output_moment","Updated moment").Arg("epsilon"
Parameters to be updated Gradient computed Updated parameters Default	Arg ("decay","Default 1. If it is in (0, 1), the gradient square sum ""is decayed by this factor.")
	REGISTER_CPU_OPERATOR (SparseAdagrad, SparseAdagradOp< float, CPUContext >)
runs the dense AdaGrad update	on (param, grad, moment[indices], lr)
runs the dense AdaGrad update and	returns (new_param, new_moment) as in the dense case.) DOC") .Input(0
runs the dense AdaGrad update and Parameters to be updated Sparse indices	Input (3,"grad","Gradient computed").Input(4
runs the dense AdaGrad update and Parameters to be updated Sparse indices learning rate	Output (0,"output_param","Updated parameters").Output(1
runs the dense AdaGrad update and Parameters to be updated Sparse indices learning rate Updated moment	Arg ("epsilon","Default 1e-5")
	REGISTER_CPU_OPERATOR (RowWiseSparseAdagrad, RowWiseSparseAdagradOp< float, CPUContext >)
runs a modified sparse Adagrad update and	returns (new_param, new_momwnr)
runs a modified sparse Adagrad update and where moment is a tensor with length equal to the number of rows in and the new moment is calculated by adding the average squared sum of gradients across each row Note that indices must also be a tensor indexing into the rows of param DOC	Input (0,"param","Parameters to be updated").Input(1
runs a modified sparse Adagrad update and where moment is a tensor with length equal to the number of rows in and the new moment is calculated by adding the average squared sum of gradients across each row Note that indices must also be a tensor indexing into the rows of param DOC Moment history	Input (2,"indices","Sparse indices").Input(3
runs a modified sparse Adagrad update and where moment is a tensor with length equal to the number of rows in and the new moment is calculated by adding the average squared sum of gradients across each row Note that indices must also be a tensor indexing into the rows of param DOC Moment history Gradient computed	Input (4,"lr","learning rate").Output(0
runs a modified sparse Adagrad update and where moment is a tensor with length equal to the number of rows in and the new moment is calculated by adding the average squared sum of gradients across each row Note that indices must also be a tensor indexing into the rows of param DOC Moment history Gradient computed Updated parameters	Output (1,"output_moment_1","Updated moment").Arg("epsilon"
	SHOULD_NOT_DO_GRADIENT (Adagrad)
	SHOULD_NOT_DO_GRADIENT (SparseAdagrad)
	SHOULD_NOT_DO_GRADIENT (RowWiseSparseAdagrad)
template<typename Context >
void	adagrad_update (int N, const float *w, const float *g, const float *h, float *nw, float *nh, float epsilon, float decay, const float *lr, Context *)
	REGISTER_CPU_OPERATOR (Adam, AdamOp< float, CPUContext >)
	NumInputs (6).NumOutputs(3).AllowInplace(
	SetDoc (R"DOC( Computes the Adam update (https://arxiv.org/abs/1412.6980) for an input gradient and momentum parameters. Concretely, given inputs (param, m1, m2, grad, lr, iters), t = iters + 1 corrected_local_rate = lr * sqrt(1 - power(beta2, t)) / (1 - power(beta1, t)) m1_o = (beta1 * m1) + (1 - beta1) * grad m2_o = (beta2 * m2) + (1 - beta2) * np.square(grad) grad_o = corrected_local_rate * m1_o / \ (sqrt(m2_o) + epsilon) param_o = param + grad_o and returns (param_o, m1_o, m2_o) )DOC").Input(0
Parameters to be updated	Input (1,"moment_1","First moment history").Input(2
Parameters to be updated Second moment history learning rate	Input (5,"iter","iteration number").Output(0
Parameters to be updated Second moment history learning rate Updated parameters	Output (1,"output_moment_1","Updated first moment").Output(2
Parameters to be updated Second moment history learning rate Updated parameters Updated second moment	Arg ("beta1","Default 0.9").Arg("beta2"
	REGISTER_CPU_OPERATOR (SparseAdam, SparseAdamOp< float, CPUContext >)
	SetDoc (R"DOC( Computes the Adam Update for the sparse case. Given inputs (param, moment1, moment2, indices, grad, lr, iter), runs the dense Adam on (param, moment1[indices], momemnt2[indices], lr, iter) and returns (new_param, new_moment1, new_moment2) as in dense case )DOC").Input(0
Parameters to be updated Second moment history	Input (3,"indices","Sparse indices").Input(4
Parameters to be updated Second moment history Gradient computed	Input (5,"lr","learning rate").Input(6
Parameters to be updated Second moment history Gradient computed iteration number Updated first moment	Output (2,"output_moment_2","Updated second moment").Arg("beta1"
Parameters to be updated Second moment history Gradient computed iteration number Updated first moment Default	Arg ("beta2","Default 0.999").Arg("epsilon"
	REGISTER_CPU_OPERATOR (RowWiseSparseAdam, RowWiseSparseAdamOp< float, CPUContext >)
	SetDoc (R"DOC( Computes a modified Adam Update for the sparse case. Given inputs (param, moment1, moment2, indices, grad, lr, iter), runs the Adam update on (param, moment1[indices], moment2[indices], lr, iter) and returns (new_param, new_moment1, new_moment2), where moment2 is a 1D tensor with length equal to the number of rows in param: shape(moment2) == shape(param)[0]. Each element of moment2 is applied to an entire row of param, and the new moment2 values are calculated by averaging across the row. )DOC").Input(0
	SHOULD_NOT_DO_GRADIENT (Adam)
	SHOULD_NOT_DO_GRADIENT (SparseAdam)
	SHOULD_NOT_DO_GRADIENT (RowWiseSparseAdam)
template<typename Context >
void	adam_update (int N, const float *g, const float *m, const float *v, float *ng, float *nm, float *nv, float beta1, float beta2, float eps_hat, float correction, const float *lr, Context *)
template<typename Context >
void	adam_compute (int N, const float *w, const float *g, const float *m, const float *v, float *nw, float *nm, float *nv, float beta1, float beta2, float eps_hat, float correction, const float *lr, Context *)
	REGISTER_CPU_OPERATOR (ClipTensorByScaling, ClipTensorByScalingOp< CPUContext >)
	SetDoc (R"DOC( Clips the input tensor by scaling based on the input value and the threshold. The value is usually the (pre-computed) norm of the tensor. If the value is larger than the threshold, scaling would be performed in this way: tensor *= (threshold / value). An optional input called additional_threshold can be provided which will scale the original threshold before it is used. That is, the final threshold will become threshold * additional_threshold. This op could be used for gradient clipping. )DOC").Input(0
Tensor of floats to be clipped	Input (1,"val","Value to be compared against the threshold").Input(2
Tensor of floats to be clipped An optional additonal threshold to scale the orignal threshold	Arg ("threshold","Threshold to determine whether to scale down the tensor").Output(0
	SHOULD_NOT_DO_GRADIENT (ClipTensorByScaling)
template<class Context >
void	fp16_momentum_sgd_update (int N, const float16 *g, const float16 *m, float16 *ng, float16 *nm, const float *lr, float momentum, bool nesterov, float weight_decay, bool fp32_update, float16 *param, Context *)
template<class Context >
void	fp32_momentum_sgd_update (int N, const float *g, const float *m, float *ng, float *nm, const float *lr, float momentum, bool nesterov, float weight_decay, float *param, Context *)
template<class T >
T	sgn (const T x)
template<typename T >
void	ftrl_compute (const T w, const T n, const T z, const T g, T &nw, T &nn, T &nz, const FtrlParams< T > &params)
template<typename Context , typename T >
void	ftrl_update (int N, const T *w, const T *nz, const T *g, T *new_w, T *new_nz, const FtrlParams< T > &params, Context *)
	REGISTER_CPU_OPERATOR (Iter, IterOp< CPUContext >)
	REGISTER_CPU_OPERATOR (AtomicIter, AtomicIterOp< CPUContext >)
	REGISTER_BLOB_SERIALIZER ((TypeMeta::Id< std::unique_ptr< std::mutex >>()), MutexSerializer)
	REGISTER_BLOB_DESERIALIZER (std::unique_ptr< std::mutex >, MutexDeserializer)
	SetDoc (R"DOC( Stores a singe integer, that gets incremented on each call to Run(). Useful for tracking the iteration count during SGD, for example. )DOC")
	SetDoc (R"DOC( Similar to Iter, but takes a mutex as the first input to make sure that updates are carried out atomically. This can be used in e.g. Hogwild sgd algorithms. )DOC").Input(0
The mutex used to do atomic increment	Input (1,"iter","The iter counter as an int64_t TensorCPU.")
	NO_GRADIENT (Iter)
	NO_GRADIENT (AtomicIter)
void	IncrementIter (TensorCPU *output)
	REGISTER_CUDA_OPERATOR (Iter, IterOp< CUDAContext >)
	REGISTER_CUDA_OPERATOR (AtomicIter, AtomicIterOp< CUDAContext >)
	REGISTER_CPU_OPERATOR (Lars, LarsOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (LearningRate, LearningRateOp< float, CPUContext >)
	REGISTER_CUDA_OPERATOR (LearningRate, LearningRateOp< float, CUDAContext >)
	REGISTER_CPU_OPERATOR (MomentumSGD, MomentumSGDOp< float, CPUContext >)
	TensorInferenceFunction ([](const OperatorDef &, const vector< TensorShape > &in){vector< TensorShape > out(2);out[0]=in[0];out[1]=in[1];return out;}).SetDoc(R"DOC( Computes a momentum SGD update for an input gradient and momentum parameters. Concretely
given	inputs (grad, m, lr) and parameters(momentum
given adjusted_gradient m_new Output	is (grad, momentum) Note the difference to MomemtumSGDUpdate
given adjusted_gradient m_new Output which actually performs the parameter	update (and is thus faster).) DOC")
	SHOULD_NOT_DO_GRADIENT (MomentumSGD)
	REGISTER_CPU_OPERATOR (MomentumSGDUpdate, MomentumSGDUpdateOp< float, CPUContext >)
	TensorInferenceFunction ([](const OperatorDef &, const vector< TensorShape > &in){vector< TensorShape > out(3);out[0]=in[0];out[1]=in[1];out[2]=in[3];return out;}).SetDoc(R"DOC( Performs a momentum SGD update for an input gradient and momentum parameters. Concretely
given	inputs (grad, m, lr, param) and arguments(momentum
given param	return ((1+momentum)*m_new-momentum *m, m_new, param) Output is(grad
	SHOULD_NOT_DO_GRADIENT (MomentumSGDUpdate)
	REGISTER_CPU_OPERATOR (SparseMomentumSGDUpdate, SparseMomentumSGDUpdateOp< float, CPUContext >)
	EnforceInplace ({{1, 1},{3, 2}}).TensorInferenceFunction([](const OperatorDef &
	SetDoc (R"DOC( Performs a momentum SGD update analogous to MomentumSGDUpdate, but using a GradientSlice and indices into the full param and momentum tables. Both param and momentum should be in-place (corresponding inputs and outputs should be the same blobs). )DOC").Input(0
GradientSlice with gradients for updated indices	Input (1,"moment","Momentum blob, same shape as param.").Input(2
GradientSlice with gradients for updated indices Learning rate	Input (3,"param","Full parameter blob.").Input(4
GradientSlice with gradients for updated indices Learning rate	Indices (in first dimension of param) where updates are performed.") .Output(0
GradientSlice with gradients for updated indices Learning rate Adjusted gradient	Output (1,"output_moment","Updated momentum.").Output(2
GradientSlice with gradients for updated indices Learning rate Adjusted gradient Updated parameter	Arg ("momentum","Momentum hyperparameter.").Arg("nesterov"
	SHOULD_NOT_DO_GRADIENT (SparseMomentumSGDUpdate)
template<typename Context >
void	momentum_sgd_update (const int N, const float *g, const float *m, float *ng, float *nm, const float *lr, const float momentum, const bool nesterov, float *param, Context *)
template<>
void	rmsprop_update< CPUContext > (int N, const float *g, const float *ms, const float *mom, float *ng, float *nms, float *nmom, float decay, float momentum, float epsilon, const float *lr, CPUContext *)
	REGISTER_CPU_OPERATOR (RmsProp, RmsPropOp< float, CPUContext >)
	SetDoc (R"DOC( Computes the RMSProp update (http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf). Concretely, given inputs (grad, mean_squares, mom, lr), computes: mean_squares_o = mean_squares + (1 - decay) * (square(grad) - mean_squares) mom_o = momentum * mom + lr * grad / sqrt(epsilon + mean_squares_o) grad_o = mom_o Returns (grad_o, mean_squares_o, mom_o). )DOC")
	SHOULD_NOT_DO_GRADIENT (RmsProp)
template<typename Context >
void	rmsprop_update (int N, const float *g, const float *ms, const float *mom, float *ng, float *nms, float *nmom, float decay, float momentum, float epsilon, const float *lr, Context *context)
	REGISTER_CPU_OPERATOR (YellowFin, YellowFinOp< float, CPUContext >)
	NumInputs (10).NumOutputs(8).AllowInplace(
	SetDoc (R"DOC( Computes the YellowFin update (https://arxiv.org/abs/1706.03471) and performs momentum SGD optimization step. lr and mu are not being shared between parameters. curv_win, g_avg, g2_avg and scalars_memory are just auxiliary memory for computing moving averages (see the publication). Takes arguments beta: coefficient for moving averages, curv_win_width: timeframe when average squared gradient is being stored, epsilon: for numerical purposes, nesterov and zero_debias for debias of moving average. )DOC").Input(0
Parameters to be updated	Input (1,"moment","Momentum").Input(2
Parameters to be updated Learning rate	Input (3,"mu","Momentum coefficient").Input(4
Parameters to be updated Learning rate Memory for latest curvature ranges	Input (5,"g_avg","Moving average of gradient").Input(6
Parameters to be updated Learning rate Memory for latest curvature ranges Moving average of squared gradient	Input (7,"scalars_memory","Memory for stateful scalars").Input(8
Parameters to be updated Learning rate Memory for latest curvature ranges Moving average of squared gradient Gradient computed	Input (9,"iter","Iteration number").Output(0
Parameters to be updated Learning rate Memory for latest curvature ranges Moving average of squared gradient Gradient computed Parameters to be updated	Output (1,"output_moment","Momentum").Output(2
Parameters to be updated Learning rate Memory for latest curvature ranges Moving average of squared gradient Gradient computed Parameters to be updated Output learning rate	Output (3,"output_mu","Output momentum coefficient").Output(4
Parameters to be updated Learning rate Memory for latest curvature ranges Moving average of squared gradient Gradient computed Parameters to be updated Output learning rate Output memory for latest curvature ranges	Output (5,"output_g_avg","Output moving average of gradient").Output(6
Parameters to be updated Learning rate Memory for latest curvature ranges Moving average of squared gradient Gradient computed Parameters to be updated Output learning rate Output memory for latest curvature ranges Output moving average of squared gradient	Output (7,"output_scalars_memory","Output memory for stateful scalars").Arg("beta"
Parameters to be updated Learning rate Memory for latest curvature ranges Moving average of squared gradient Gradient computed Parameters to be updated Output learning rate Output memory for latest curvature ranges Output moving average of squared gradient Default	Arg ("curv_win_width","Default 20").Arg("epsilon"
Parameters to be updated Learning rate Memory for latest curvature ranges Moving average of squared gradient Gradient computed Parameters to be updated Output learning rate Output memory for latest curvature ranges Output moving average of squared gradient Default Default	Arg ("nesterov","Default false").Arg("zero_debias"
	SHOULD_NOT_DO_GRADIENT (YellowFin)
void	initNNPACK ()
	REGISTER_CPU_OPERATOR_WITH_ENGINE (Conv, NNPACK, NNPACKConvOp)
	REGISTER_CPU_OPERATOR (QuantDecompZstd, QuantDecompZstdOp)
INT_MAX	SetDoc (R"DOC( Decompress a set of tensors that are compressed using zstd. The data can be compressed using mutils.compress_data_list(), see quant_decomp_op_test.py for an example. The number of outputs depended on the input. )DOC").Input(0
INT_MAX Compressed data in	tensor (uint8_t)
	SHOULD_NOT_DO_GRADIENT (QuantDecompZstd)
bool	are_nodes_common (const Graph &g, int model_idx, int candidate_idx)
	REGISTER_TRANSFORM (CommonSubexpressionElimination, CommonSubexpressionEliminationTransform)
	REGISTER_TRANSFORM (ConvToNNPack, ConvToNNPackTransform)
bool	compare_ops (const OperatorDef &p_op, const OperatorDef &g_op, bool arg_match)
const CpuId &	GetCpuId ()
template<class Map , typename Key = typename Map::key_type, typename Value = typename Map::mapped_type>
Map::mapped_type	get_default (const Map &map, const Key &key, Value &&dflt)
template<typename T , typename T2 >
__device__ T	mixed_mult (T data1, T2 data2)
template<typename T , typename T2 >
__device__ T	mixed_add (T data1, T2 data2)
template<typename TIN , typename TOUT >
__device__ void	mixed_store (TIN *data_in, TOUT *data_out)
template<typename T >
__device__ void	mixed_store (T *data_in, T *data_out)
void	MurmurHash3_x86_32 (const void *key, int len, uint32_t seed, void *out)
void	MurmurHash3_x86_128 (const void *key, const int len, uint32_t seed, void *out)
void	MurmurHash3_x64_128 (const void *key, const int len, const uint32_t seed, void *out)
const ::std::string &	GetEmptyStringAlreadyInited ()
void	ShutdownProtobufLibrary ()
std::string	DeviceTypeName (const int32_t &d)
bool	IsSameDevice (const DeviceOption &lhs, const DeviceOption &rhs)
bool	ReadStringFromFile (const char *filename, string *str)
bool	WriteStringToFile (const string &str, const char *filename)
string	ProtoDebugString (const Message &proto)
bool	ParseProtoFromLargeString (const string &str, Message *proto)
bool	ReadProtoFromTextFile (const char *filename, Message *proto)
void	WriteProtoToTextFile (const Message &proto, const char *filename)
bool	ReadProtoFromBinaryFile (const char *filename, MessageLite *proto)
void	WriteProtoToBinaryFile (const MessageLite &proto, const char *filename)
bool	operator== (const NetDef &l, const NetDef &r)
std::ostream &	operator<< (std::ostream &output, const NetDef &n)
template<>
Argument	MakeArgument (const string &name, const MessageLite &value)
bool	HasOutput (const OperatorDef &op, const std::string &output)
bool	HasInput (const OperatorDef &op, const std::string &input)
const Argument &	GetArgument (const OperatorDef &def, const string &name)
bool	GetFlagArgument (const OperatorDef &def, const string &name, bool def_value)
Argument *	GetMutableArgument (const string &name, const bool create_if_missing, OperatorDef *def)
bool	ReadProtoFromBinaryFile (const string filename, MessageLite *proto)
void	WriteProtoToBinaryFile (const MessageLite &proto, const string &filename)
bool	ReadProtoFromTextFile (const string filename, Message *proto)
void	WriteProtoToTextFile (const Message &proto, const string &filename)
bool	ReadProtoFromFile (const char *filename, Message *proto)
bool	ReadProtoFromFile (const string &filename, Message *proto)
template<class IterableInputs = std::initializer_list<string>, class IterableOutputs = std::initializer_list<string>, class IterableArgs = std::initializer_list<Argument>>
OperatorDef	CreateOperatorDef (const string &type, const string &name, const IterableInputs &inputs, const IterableOutputs &outputs, const IterableArgs &args, const DeviceOption &device_option=DeviceOption(), const string &engine="")
template<class IterableInputs = std::initializer_list<string>, class IterableOutputs = std::initializer_list<string>>
OperatorDef	CreateOperatorDef (const string &type, const string &name, const IterableInputs &inputs, const IterableOutputs &outputs, const DeviceOption &device_option=DeviceOption(), const string &engine="")
template<typename T >
Argument	MakeArgument (const string &name, const T &value)
template<typename T >
void	AddArgument (const string &name, const T &value, OperatorDef *def)
bool	operator== (const DeviceOption &dl, const DeviceOption &dr)
std::vector< std::string >	split (char separator, const std::string &string)
size_t	editDistance (const std::string &s1, const std::string &s2, size_t max_distance)
int32_t	editDistanceHelper (const char *s1, size_t s1_len, const char *s2, size_t s2_len, std::vector< size_t > &current, std::vector< size_t > &previous, std::vector< size_t > &previous1, size_t max_distance)
int	Do256NOPs ()
template<typename T >
T	WaitForVariableChange (std::atomic< T > *var, T initial_value, std::condition_variable *cond, std::mutex *mutex)
void	OpticalFlowExtractor (const cv::Mat &prev_gray, const cv::Mat &curr_gray, const int flow_alg_type, cv::Mat &flow)
void	MergeOpticalFlow (cv::Mat &prev_flow, const cv::Mat &curr_flow)
void	MultiFrameOpticalFlowExtractor (const std::vector< cv::Mat > &grays, const int optical_flow_alg_type, cv::Mat &flow)
	REGISTER_CPU_OPERATOR (VideoInput, VideoInputOp< CPUContext >)
	TensorInferenceFunction ([](const OperatorDef &def, const vector< TensorShape > &){ArgumentHelper helper(def);int batch_size=helper.GetSingleArgument< int >("batch_size", 0);int clip_per_video=helper.GetSingleArgument< int >("clip_per_video", 1);int crop_height=helper.GetSingleArgument< int >("crop_height", helper.GetSingleArgument< int >("crop_size", 0));int crop_width=helper.GetSingleArgument< int >("crop_width", helper.GetSingleArgument< int >("crop_size", 0));int length_rgb=helper.GetSingleArgument< int >("length_rgb", 0);int channels_rgb=helper.GetSingleArgument< int >("channels_rgb", 3);int length_of=helper.GetSingleArgument< int >("length_of", 0);int channels_of=helper.GetSingleArgument< int >("channels_of", 2);bool get_rgb=helper.GetSingleArgument< bool >("get_rgb", true);bool get_optical_flow=helper.GetSingleArgument< bool >("get_optical_flow", false);bool do_multi_label=helper.GetSingleArgument< bool >("do_multi_label", false);bool get_video_id=helper.GetSingleArgument< bool >("get_video_id", false);int output_size=1;if(get_rgb){output_size++;}if(get_optical_flow){output_size++;}if(get_video_id){output_size++;}int index=0;vector< TensorShape > out(output_size);CHECK_GT(crop_height, 0);CHECK_GT(crop_width, 0);batch_size *=clip_per_video;if(get_rgb){out[index++]=CreateTensorShape(vector< int >{batch_size, channels_rgb, length_rgb, crop_height, crop_width}, TensorProto::FLOAT);}if(get_optical_flow){out[index++]=CreateTensorShape(vector< int >{batch_size, channels_of, length_of, crop_height, crop_width}, TensorProto::FLOAT);}if(!do_multi_label){out[index++]=CreateTensorShape(vector< int >{1, batch_size}, TensorProto::INT32);}else{int num_of_class=helper.GetSingleArgument< int >("num_of_class", 0);out[index++]=CreateTensorShape(vector< int >{batch_size, num_of_class}, TensorProto::INT32);}if(get_video_id){out[index]=CreateTensorShape(vector< int >{1, batch_size}, TensorProto::INT32);}return out;})
	NO_GRADIENT (VideoInput)
	REGISTER_CUDA_OPERATOR (VideoInput, VideoInputOp< CUDAContext >)
void	Saturation (float *clip, const int length, const int crop_height, const int crop_width, const float alpha_rand, std::mt19937 *randgen)
void	Brightness (float *clip, const int length, const int crop_height, const int crop_width, const float alpha_rand, std::mt19937 *randgen)
void	Contrast (float *clip, const int length, const int crop_height, const int crop_width, const float alpha_rand, std::mt19937 *randgen)
void	ColorJitter (float *clip, const int length, const int crop_height, const int crop_width, const float saturation, const float brightness, const float contrast, std::mt19937 *randgen)
void	ColorLighting (float *clip, const int length, const int crop_height, const int crop_width, const float alpha_std, const std::vector< std::vector< float >> &eigvecs, const std::vector< float > &eigvals, std::mt19937 *randgen)
void	ColorNormalization (float *clip, const int length, const int crop_height, const int crop_width, const int channels, const std::vector< float > &mean, const std::vector< float > &inv_std)
void	ClipTransformRGB (const unsigned char *buffer_rgb, const int multi_crop_count, const int crop_height, const int crop_width, const int length_rgb, const int channels_rgb, const int sampling_rate_rgb, const int height, const int width, const int h_off, const int w_off, const int *multi_crop_h_off, const int *multi_crop_w_off, const bool mirror_me, const bool color_jitter, const float saturation, const float brightness, const float contrast, const bool color_lighting, const float color_lighting_std, const std::vector< std::vector< float >> &color_lighting_eigvecs, const std::vector< float > &color_lighting_eigvals, const std::vector< float > &mean_rgb, const std::vector< float > &inv_std_rgb, std::mt19937 *randgen, float *transformed_clip)
void	ClipTransformOpticalFlow (const unsigned char *buffer_rgb, const int crop_height, const int crop_width, const int length_of, const int channels_of, const int sampling_rate_of, const int height, const int width, const cv::Rect &rect, const int channels_rgb, const bool mirror_me, const int flow_alg_type, const int flow_data_type, const int frame_gap_of, const bool do_flow_aggregation, const std::vector< float > &mean_of, const std::vector< float > &inv_std_of, float *transformed_clip)
void	FreeDecodedData (std::vector< std::unique_ptr< DecodedFrame >> &sampledFrames)
bool	DecodeMultipleClipsFromVideo (const char *video_buffer, const std::string &video_filename, const int encoded_size, const Params &params, const int start_frm, const int clip_per_video, const bool use_local_file, int &height, int &width, std::vector< unsigned char * > &buffer_rgb)
	REGISTER_CPU_OPERATOR (AffineChannel, AffineChannelOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (AffineChannelGradient, AffineChannelGradientOp< float, CPUContext >)
	SetDoc (R"DOC( Applies a separate affine transformation to each channel of the input. Useful for replacing spatial batch norm with its equivalent fixed transformation. )DOC").Input(0
feature map input of input of	shape (C)
the c th element is the scale factor of the affine transformation for the c th channel of the input	Input (2,"bias","1D input of shape (C); the c-th element is the bias of the affine ""transformation for the c-th channel of the input.").Output(0
	Input (0,"scale","See AffineChannel.").Input(1
	REGISTER_GRADIENT (AffineChannel, GetAffineChannelGradient)
	REGISTER_CPU_OPERATOR (BatchPermutation, BatchPermutationOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (BatchPermutationGradient, BatchPermutationGradientOp< float, CPUContext >)
NumInputs(2).NumOutputs(1).SetDoc(R"DOC( Permute the batch elements of the input tensor X according to the permutation specified in the input indices. Warning gradient comptuation is only correct if indices is a permutation DOC	Input (0,"X","Tensor of at least 1D shape (N, D0, D1, ...).").Input(1
NumInputs(2).NumOutputs(1).SetDoc(R"DOC( Permute the batch elements of the input tensor X according to the permutation specified in the input indices. Warning gradient comptuation is only correct if indices is a permutation DOC tensor of type int with	shape (N,) specifying a valid permutation" "of the indices in[0
NumInputs(2).NumOutputs(1).SetDoc(R"DOC( Permute the batch elements of the input tensor X according to the permutation specified in the input indices. Warning gradient comptuation is only correct if indices is a permutation DOC tensor of type int with	N (inclusive).") .Output( 0
NumInputs(2).NumOutputs(1).SetDoc(R"DOC( Permute the batch elements of the input tensor X according to the permutation specified in the input indices. Warning gradient comptuation is only correct if indices is a permutation DOC tensor of type int with Tensor with the same shape as X where	the (D0, D1,...) dimensional" "batch elements of X are permuted according to the input indices.")
See BatchPermutation	Input (1,"dY","Gradient of forward output 0 (Y).").Output(0
	REGISTER_GRADIENT (BatchPermutation, GetBatchPermutationGradient)
	REGISTER_CPU_OPERATOR (GroupSpatialSoftmax, GroupSpatialSoftmaxOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (GroupSpatialSoftmaxGradient, GroupSpatialSoftmaxGradientOp< float, CPUContext >)
number of classes in each softmax group	Input (0,"scores","4D tensor of softmax inputs (called 'scores' or 'logits') with shape ""(N, C, H, W), where C = num_anchors * num_classes defines num_anchors ""groups of contiguous num_classes softmax inputs.").Output(0
See GroupSpatialSoftmax	Input (1,"d_probabilities","Gradient of forward output 0 (probabilities).").Output(0
See GroupSpatialSoftmax Gradient of forward	input (scores).")
	REGISTER_GRADIENT (GroupSpatialSoftmax, GetGroupSpatialSoftmaxGradient)
	REGISTER_CPU_OPERATOR (PSRoIPool, PSRoIPoolOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (PSRoIPoolGradient, PSRoIPoolGradientOp< float, CPUContext >)
Spatial scale of the input feature map X relative to the input image E if X has a stride of w r t the input image	Arg ("group_size","(int) default 1; pooled_h = pooled_w = group_size where pooled_{h,w} ""is the pooled output Y's height and width, respectively.").Arg("output_dim"
number of channels in the pooled which might be the number of classes is used for classification or if used for class agnostic bounding box regression	Input (0,"X","4D position sensitive feature map input of shape (N, C, H, W), where ""C = group_size**2 * output_dim.").Input(1
	REGISTER_CPU_OPERATOR (RoIPoolF, RoIPoolFOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (RoIPoolFGradient, RoIPoolFGradientOp< float, CPUContext >)
Pooled output Y s width	Input (0,"X","4D feature map input of shape (N, C, H, W).").Input(1
	REGISTER_CPU_OPERATOR (SampleAs, SampleAsOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SampleAsGradient, SampleAsGradientOp< float, CPUContext >)
Tensor of at least	shape (N,...).") .Input( 1
See SampleAs	Input (1,"labels","See SampleAs.").Input(2
	REGISTER_GRADIENT (SampleAs, GetSampleAsGradient)
	REGISTER_CPU_OPERATOR (SelectSmoothL1Loss, SelectSmoothL1LossOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SelectSmoothL1LossGradient, SelectSmoothL1LossGradientOp< float, CPUContext >)
L2 to L1 transition point	Arg ("scale","(float) default 1.0; multiply the loss by this scale factor.").Input(0
L2 to L1 transition point tensor of bounding box regression predictions with	shape (N, 4 *num_bbox_classes *num_anchors, H, W).") .Input( 1
L2 to L1 transition point tensor of bounding box regression predictions with tensor of labels	shape (M, 4) for 4 contiguous channels starting" "at each of the M locations selected by the locations input.") .Input( 2
L2 to L1 transition point tensor of bounding box regression predictions with tensor of labels tensor of shape(M, 4) that identifies M 'select'locations" "encoded by the four colums the loss is divided by	max (1, normalizer).") .Output( 0
See SelectSmoothL1Loss	Input (1,"Y","See SelectSmoothL1Loss.").Input(2
See SelectSmoothL1Loss See SelectSmoothL1Loss	Input (3,"normalizer","See SelectSmoothL1Loss.").Input(4
See SelectSmoothL1Loss See SelectSmoothL1Loss Gradient of forward	output (loss).") .Output( 0
See SelectSmoothL1Loss See SelectSmoothL1Loss Gradient of forward Gradient of forward	input (Y_hat).")
	REGISTER_GRADIENT (SelectSmoothL1Loss, GetSelectSmoothL1LossGradient)
	REGISTER_CPU_OPERATOR (SigmoidCrossEntropyLoss, SigmoidCrossEntropyLossOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SigmoidCrossEntropyLossGradient, SigmoidCrossEntropyLossGradientOp< float, CPUContext >)
multiply the loss by this scale factor	Arg ("normalize","(int) default 1; if true, divide the loss by the number of targets > ""-1.").Input(0
multiply the loss by this scale factor Tensor of predicted	logits (shape must be at least 1D).") .Input( 1
multiply the loss by this scale factor Tensor of predicted Tensor of targets of type int and same shape as logits X	Output (0,"loss","Scalar loss.")
See SigmoidCrossEntropyLoss	Input (1,"targets","See SigmoidCrossEntropyLoss.").Input(2
	REGISTER_GRADIENT (SigmoidCrossEntropyLoss, GetSigmoidCrossEntropyLossGradient)
	REGISTER_CPU_OPERATOR (SigmoidFocalLoss, SigmoidFocalLossOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SigmoidFocalLossGradient, SigmoidFocalLossGradientOp< float, CPUContext >)
where N is the number of elements in the H and W are the height and and each of length num_classes For the binary form of Focal num_classes does not include the background category(So, for COCO, num_classes=80, not 81.) The binary form of focal loss is multiply the loss by this scale factor	Arg ("alpha","(float) default 0.25; Focal Loss's alpha hyper-parameter.").Arg("gamma"
Focal Loss s gamma hyper parameter	Arg ("num_classes","(int) default 80; number of classes (excluding background).").Input(0
Focal Loss s gamma hyper parameter tensor of sigmoid	inputs (called 'scores'or 'logits') with shape" "(N
See SigmoidFocalLoss	Input (1,"labels","See SigmoidFocalLoss.").Input(2
See SigmoidFocalLoss See SigmoidFocalLoss	Input (3,"d_loss","Gradient of forward output 0 (loss)").Output(0
See SigmoidFocalLoss See SigmoidFocalLoss Gradient of forward	input (logits)")
	REGISTER_GRADIENT (SigmoidFocalLoss, GetSigmoidFocalLossGradient)
	REGISTER_CPU_OPERATOR (SmoothL1Loss, SmoothL1LossOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SmoothL1LossGradient, SmoothL1LossGradientOp< float, CPUContext >)
NumInputs(4).NumOutputs(1).SetDoc(R"DOC( Smooth L1 Loss is a minor variation of Huber loss in which the point of transition between L2 loss and L1 loss is adjustable by a hyper-parameter beta L2 to L1 transition point Tensor of	predictions (at least 1D).") .Input( 1
NumInputs(4).NumOutputs(1).SetDoc(R"DOC( Smooth L1 Loss is a minor variation of Huber loss in which the point of transition between L2 loss and L1 loss is adjustable by a hyper-parameter beta L2 to L1 transition point Tensor of Tensor of labels with the same shape as Y_hat	Input (2,"alpha_in","Tensor of inside weights with the same shape as Y.").Input(3
See SmoothL1Loss	Input (1,"Y","See SmoothL1Loss.").Input(2
See SmoothL1Loss See SmoothL1Loss	Input (3,"alpha_out","See SmoothL1Loss.").Input(4
	REGISTER_GRADIENT (SmoothL1Loss, GetSmoothL1LossGradient)
	REGISTER_CPU_OPERATOR (SoftmaxFocalLoss, SoftmaxFocalLossOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (SoftmaxFocalLossGradient, SoftmaxFocalLossGradientOp< float, CPUContext >)
where N is the number of elements in the H and W are the height and and where t is the	target (ground truth) class
Focal Loss s gamma hyper parameter	Arg ("num_classes","(int) default 81; number of classes in each softmax group.").Input(0
the loss is normalized by Scalar loss	Output (1,"probabilities","4D tensor of softmax probabilities with shape (N, C, H, W), where ""C = num_anchors * num_classes, and softmax was applied to each of the ""num_anchors groups; within a group the num_classes values sum to 1.")
See SoftmaxFocalLoss	Input (1,"labels","See SoftmaxFocalLoss.").Input(2
See SoftmaxFocalLoss See SoftmaxFocalLoss	Input (3,"probabilities","Output 1 from SoftmaxFocalLoss; See SoftmaxFocalLoss.").Input(4
	REGISTER_GRADIENT (SoftmaxFocalLoss, GetSoftmaxFocalLossGradient)
	REGISTER_CPU_OPERATOR (SpatialNarrowAs, SpatialNarrowAsOp< CPUContext >)
	REGISTER_CPU_OPERATOR (SpatialNarrowAsGradient, SpatialNarrowAsGradientOp< CPUContext >)
or input of	shape (N, H0, W0) or(N
or input of W0	Input (1,"B","3D or 4D input of shape (N, H1, W1) or (N, C, H1, W1), where H1 <= H0 ""and W1 <= W0.").Output(0
or input of W0 Sub window of A containing	rows (inclusive) and columns" "[0
or input of W0 Sub window of A containing	W1 (inclusive).")
See SpatialNarrowAs	Input (1,"B","See SpatialNarrowAs.").Input(2
See SpatialNarrowAs Gradient of forward	output (C).") .Output( 0
See SpatialNarrowAs Gradient of forward Gradient of forward	input (A)")
	REGISTER_GRADIENT (SpatialNarrowAs, SpatialNarrowAsGradient)
	REGISTER_CPU_OPERATOR (UpsampleNearest, UpsampleNearestOp< float, CPUContext >)
	REGISTER_CPU_OPERATOR (UpsampleNearestGradient, UpsampleNearestGradientOp< float, CPUContext >)
integer upsampling factor feature map of	shape (N, C, scale *H, scale *W)
	REGISTER_GRADIENT (UpsampleNearest, GetUpsampleNearestGradient)
	REGISTER_CPU_OPERATOR (Caffe2ModuleTestDynamicDummy, Caffe2ModuleTestDynamicDummyOp)
	OPERATOR_SCHEMA (Caffe2ModuleTestDynamicDummy)
	REGISTER_CAFFE2_EARLY_INIT_FUNCTION (registerGlobalPerfNetObserverReporter,&registerGlobalPerfNetObserverReporter,"Caffe2 print net observer reporter")
	REGISTER_CAFFE2_EARLY_INIT_FUNCTION (registerGlobalPerfNetObserverCreator,&registerGlobalPerfNetObserverCreator,"Caffe2 net global observer creator")
	CAFFE2_MODULE (caffe2_rocksdb,"RocksDB implementation for caffe2::DB.")

Variables
DoRunWithOtherType2 typedef Registry< std::string, std::unique_ptr< OperatorBase >, const OperatorDef &, Workspace * >	OperatorRegistry
const int	kCIFARSize = 32
const int	kCIFARImageNBytes = kCIFARSize * kCIFARSize * 3
const int	kCIFAR10BatchSize = 10000
const int	kCIFAR10TestDataSize = 10000
const int	kCIFAR10TrainBatches = 5
const int	kCIFAR100TrainDataSize = 50000
const int	kCIFAR100TestDataSize = 10000
constexpr size_t	gCaffe2Alignment = 32
constexpr auto	kTensorBlobType = "Tensor"
constexpr auto	kChunkIdSeparator = "#%"
constexpr int	kDefaultChunkSize = -1
constexpr int	kNoChunking = 0
std::atomic< bool >	g_caffe2_has_cuda_linked {false}
constexpr int	CAFFE_CUDA_NUM_THREADS = 512
constexpr int	CAFFE_MAXIMUM_NUM_BLOCKS = 4096
constexpr int	MaxDeviceTypes = DeviceType::COMPILE_TIME_MAX_DEVICE_TYPES
std::function< void(const OperatorDef &)>	GetOperatorLogger ()
	DoRunWithType2
constexpr int	kCannotComputeNumOutputs = -1
constexpr auto	kQTensorBlobQType = "QTensor"
constexpr int	k_limit_default_ = 1000
	float16
constexpr auto	kBlobName = "blob_name"
constexpr auto	kAddValue = "add_value"
alternative key for the	handler
INT_MAX batch_size images will be processed GPUs can optionally be used for part of the processing The following transformations are applied to the image A bounding box is applied to the initial Number of images to output for each run of the Whether or not to do color jitter Defaults to Image brightness scale used in color jittering Defaults to Whether or not to do color lighting Defaults to	Type
INT_MAX batch_size images will be processed GPUs can optionally be used for part of the processing The following transformations are applied to the image A bounding box is applied to the initial Number of images to output for each run of the Whether or not to do color jitter Defaults to Image brightness scale used in color jittering Defaults to Whether or not to do color lighting Defaults to Scale the size of the smallest dimension of the image to this Scale and minsize are mutually exclusive Must be larger than crop	If
INT_MAX batch_size images will be processed GPUs can optionally be used for part of the processing The following transformations are applied to the image A bounding box is applied to the initial Number of images to output for each run of the Whether or not to do color jitter Defaults to Image brightness scale used in color jittering Defaults to Whether or not to do color lighting Defaults to Scale the size of the smallest dimension of the image to this Scale and minsize are mutually exclusive Must be larger than crop both dimensions of the image will be set to minsize or	scale
	otherwise
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to Vector of means per color Standard deviation by which to normalize color channels Defaults to Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults if the input is in Caffe format Defaults to Number of CPU decode transform threads Defaults to Name of the Type of The sizes of any outputs besides the data and shortest side desired for image resize Defaults to[-1,-1] or no random resize desired	data
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to Vector of means per color Standard deviation by which to normalize color channels Defaults to Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults if the input is in Caffe format Defaults to Number of CPU decode transform threads Defaults to Name of the Type of The sizes of any outputs besides the data and shortest side desired for image resize Defaults to[-1,-1] or no random resize desired Tensor containing the images additional	outputs
the other dimension is proportionally scaled Defaults to Whether or not to mirror the image Defaults to Vector of means per color Standard deviation by which to normalize color channels Defaults to Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults Bounding box coordinate Defaults if the input is in Caffe format Defaults to Number of CPU decode transform threads Defaults to Name of the Type of The sizes of any outputs besides the data and shortest side desired for image resize Defaults to[-1,-1] or no random resize desired Tensor containing the images additional Any outputs after the first will be Tensors read from the input	TensorProtos
const char *const	snpe_ffi_so = "libsnpe_ffi.so"
constexpr size_t	k2b1bXBits = 2
constexpr size_t	kL1CacheSizeBytes = 16 * 1024
constexpr size_t	kGEMMTileSize = 64
constexpr size_t	kGEMMTileDepthBytes = 16
element wise DOC	output
we add to	it
we first initialize the output tensor to all	zeros
we first initialize the output tensor to all and then do accumulation Any further calls to the	input
we first initialize the output tensor to all and then do accumulation Any further calls to the The input tensor that has to be accumulated to the output tensor If the output size is not the same as input	size = 1
we first initialize the output tensor to all and then do accumulation Any further calls to the The input tensor that has to be accumulated to the output tensor If the output size is not the same as input the output tensor is first reshaped and initialized to	zero
we first initialize the output tensor to all and then do accumulation Any further calls to the The input tensor that has to be accumulated to the output tensor If the output size is not the same as input the output tensor is first reshaped and initialized to and only	then
this	X = in[0]
	a_1
	ints
	longs
or long longs and checks if all values are true when coerced into a boolean In other	words
const vector< TensorShape > &	in
ArgumentHelper	helper (def)
vector< int >	output_dims = GetDimsVector(in[1])
const auto &	data_dims = GetDimsVector(in[0])
const auto &	indices_dims = GetDimsVector(in[1])
	out [0] = CreateTensorShape(output_dims, TensorProto::FLOAT)
	DATA
Tensor of rank	r
indices	vector
indices and values vector Each element in lengths indices should not have duplicate number For	example
	The
original data tensor	masked_data
return the segment lengths of a corresponding segmented tensor after BooleanMask is applied DOC	mask
const float	minf = -1.0f * std::numeric_limits<float>::infinity()
reconstruct values together according to masks A comprehensive	False
reconstruct values together according to masks A comprehensive	True
reconstruct values together according to masks A comprehensive False	values1 = 1.0
reconstruct values together according to masks A comprehensive False	mask2 = False
reconstruct values together according to masks A comprehensive False False	values2
reconstruct values together according to masks A comprehensive False False True	values3 = 4.0
reconstruct values together according to masks A comprehensive False False True Reconstruct	by
reconstruct values together according to masks A comprehensive False False True Reconstruct	mask3
reconstruct values together according to masks A comprehensive False False True Reconstruct We	get
reconstruct values together according to masks A comprehensive False False True Reconstruct We Note that for all mask	positions
reconstruct values together according to masks A comprehensive False False True Reconstruct We Note that for all mask there must be at least one True If for a field there are multiple True	s
The data type to which the elements of the input tensor are cast Strictly must be one of the types from DataType enum in TensorProto Output tensor with the same shape as input with type specified by the to	argument
the gradient for the output of SpatialBN and the per channel mean and inverse std var vectors for the computes the per channel bias and scale gradient to be used during the backward pass for subsequent spatial batch normalization gradient calculation	Typically
the gradient for the output of SpatialBN and the per channel mean and inverse std var vectors for the computes the per channel bias and scale gradient to be used during the backward pass for subsequent spatial batch normalization gradient calculation the results of this op are subsequently reduced over multiple devices to obtain statistics over a larger batch size in cases where the batch size for a single model copy is too low to yield the full benefit of batch normalization The resulting bias and scale can then be plugged back into SpatialBNGradient to get results over the larger batch size DOC	mean
the gradient for the output of SpatialBN and the per channel mean and inverse std var vectors for the computes the per channel bias and scale gradient to be used during the backward pass for subsequent spatial batch normalization gradient calculation the results of this op are subsequently reduced over multiple devices to obtain statistics over a larger batch size in cases where the batch size for a single model copy is too low to yield the full benefit of batch normalization The resulting bias and scale can then be plugged back into SpatialBNGradient to get results over the larger batch size DOC The mean saved from the forward pass as a dimensional tensor of size C	output_grad
the gradient for the output of SpatialBN and the per channel mean and inverse std var vectors for the computes the per channel bias and scale gradient to be used during the backward pass for subsequent spatial batch normalization gradient calculation the results of this op are subsequently reduced over multiple devices to obtain statistics over a larger batch size in cases where the batch size for a single model copy is too low to yield the full benefit of batch normalization The resulting bias and scale can then be plugged back into SpatialBNGradient to get results over the larger batch size DOC The mean saved from the forward pass as a dimensional tensor of size C Gradient for the output layer of	SpatialBN
the gradient for the output of SpatialBN and the per channel mean and inverse std var vectors for the computes the per channel bias and scale gradient to be used during the backward pass for subsequent spatial batch normalization gradient calculation the results of this op are subsequently reduced over multiple devices to obtain statistics over a larger batch size in cases where the batch size for a single model copy is too low to yield the full benefit of batch normalization The resulting bias and scale can then be plugged back into SpatialBNGradient to get results over the larger batch size DOC The mean saved from the forward pass as a dimensional tensor of size C Gradient for the output layer of here used as input because we are on the backward pass	bias_grad
computes the sum of all elements per channel and the sum of all elements squared per channel These values can be reduced across multiple batches and used to obtain the mean and variance across the full set of batches Using the new mean and variance as input to SpatialBN has the effect of changing the batch size over which SpatialBN is applied DOC	sum
	kv_handler
Key value handler for	comm_world
Key value handler for A common world for collective operations int rank of this node in the common	world
	existing_comm_world
	common_world
The common world to be	destroyed
The common world	Y
The common world The reduced result on	root
The common world The allreduced	tensor
The common world The allreduced same on all	nodes
The common world	dst
The common world An int CPUtensor of size specifying the rank If	given
The common world An int CPUtensor of size specifying the rank If this overrides the to argument of the op The rank to send the tensor to bool if	set
The common world An int CPUtensor of size specifying the rank If this overrides the to argument of the op The rank to send the tensor to bool if only send the content and assume that the receiver has already known the tensor s shape and	information
The common world	src
The common world An int CPUtensor of size specifying the rank If this overrides the from argument of the op The received tensor	tag
INT_MAX	split
INT_MAX Optional list of output Which axis to split on Either NHWC or	NCWH
INT_MAX Optional list of output Which axis to split on Either NHWC or will split on C	axis
Which axis to concat on Pass to add the axis specified in arg axis to all input tensors	concat_result
apply conditional	DataT
apply conditional Data to use when True	DataO
apply conditional Data to use when True Output data after applying	ConditionalOp
const char *	kConvDoc
the filter	blob
the filter and the bias and computes the output Note that other	parameters
the filter and the bias and computes the output Note that other such as the stride and kernel or the pads sizes in each direction are not necessary for input because they are provided by the ConvTransposeUnpoolOpBase	operator.Various dimension checks are done implicitly, and the sizes are specified in the Input docs for this operator.As is expected, the filter is deconvolved with a subset of the image and the bias is added
this is done throughout the image data and the output is computed As a side note on the implementation	layout
this is done throughout the image data and the output is computed As a side note on the implementation which is why they are separate files DOC	filter
this is done throughout the image data and the output is computed As a side note on the implementation which is why they are separate files DOC The filter blob that will be used in the transposed	convolution
has where C is the number of	channels
element wise DOC The cosine of the input tensor computed element	wise
	counter
A blob pointing to an instance of a new counter Resets counter to this	value
Input blob from the previous	layer
Input blob from the previous which is almost always the result of a softmax	operation
X is a array of size N x	D
X is a array of size N x where N is the batch size and D is the number of classes Output blob after the cross entropy	computation
	logits = in[0]
matrix of logits for each example and class	xentropy
matrix of logits for each example and class	weights
or input tensor	Z
Y with different shapes and produces one output float tensor of the dot product between X and Y We currently support two kinds of strategies to achieve this Before doing normal dot_product pad the smaller Y must be equal Only the second dimension of X or Y can be padded DOC or input tensor whether to replicate the smaller tensor or	not
INT_MAX Subnet with blob bindings Indices of corresponding outer workspace	blobs
INT_MAX Subnet with blob bindings Indices of corresponding outer workspace in	order
INT_MAX Subnet with blob bindings Indices of corresponding outer workspace in List of blobs from the forward Do int out	bool { return true
ArgumentHelper	argsHelper (def)
auto	output_mask = !argsHelper.GetSingleArgument<bool>("is_test", 0)
	float
default the ratio of random dropout The input data as Tensor The output mask If is_test is	nonzero
describes the axis of the	inputs
defaults to one because the axis most likely describes the	batch_size
const char *	kBroadcastDoc
and the dimensions of the second input is the contiguous subset of the dimensions of the first For the following tensor shapes are	supported
element wise This operation can be done in an in place fashion	too
the data type of value is used The output tensor shape is specified by the shape argument If the number of input	is
the data type of value is used The output tensor shape is specified by the shape argument If the number of input the shape will be identical to that of the input at run time with optional additional dimensions appended at the end as specified by extra_shape argument In that case the shape argument should not be set If input_as_shape is set to	true
	index = 0
	query
TIndex	outer = 1
TIndex	inner = 1
A tensor of	rank
and then scaling each element to an bit number between and To later de quantize	values
followed by the bias as a bit float in the next	bytes
followed by the bias as a bit float in the next and the quantized values in the preceding bytes of the row The output is a matrix containing only the but de quantized De quantization is performed by multiplying each value by its row s scale and bias parameters The de quantized values will thus not be exactly equal to the	original
followed by the bias as a bit float in the next and the quantized values in the preceding bytes of the row The output is a matrix containing only the but de quantized De quantization is performed by multiplying each value by its row s scale and bias parameters The de quantized values will thus not be exactly equal to the un quantized floating point values DOC	float_input
but operating on bit rowwise quantized matrices with fused uint8 tensor with rank obtained with	OUTPUT
in a sequence length aware fashion	Concretely
in a sequence length aware fashion given the previous hidden and the sequence computes the GRU	activations
in a sequence length aware fashion given the previous hidden and the sequence computes the GRU avoiding computation if the input is Bool to determine if hidden state is zeroes or passed along for timesteps past the given sequence_length	hidden
in a sequence length aware fashion given the previous hidden and the sequence computes the GRU avoiding computation if the input is Bool to determine if hidden state is zeroes or passed along for timesteps past the given sequence_length The new GRU hidden state calculated by this	op
When	false
When the sequence lengths input is left and all following inputs are shifted left by	one
INT_MAX Net executed when condition is true	condition
auto	pad = helper.GetSingleArgument<int>("pad", 0)
auto	kernel_h
auto	kernel_w
auto	dilation_h
auto	dilation_w
auto	stride_h
auto	stride_w
int	N = 0
int	C = 0
int	H = 0
int	W = 0
const int	dkernel_h = dilation_h * (kernel_h - 1) + 1
const int	dkernel_w = dilation_w * (kernel_w - 1) + 1
const int	out_h = (H + 2 * pad - dkernel_h) / stride_h + 1
const int	out_w = (W + 2 * pad - dkernel_w) / stride_w + 1
Max number of	elements
return an Int tensor of same shape containing the indices for each of the keys If the index is	frozen
return an Int tensor of same shape containing the indices for each of the keys If the index is unknown entries are given index	Otherwise
return an Int tensor of same shape containing the indices for each of the keys If the index is unknown entries are given index new entries are added into the index If an insert is necessary but max_elements has been	reached
return an Int tensor of same shape containing the indices for each of the keys If the index is unknown entries are given index new entries are added into the index If an insert is necessary but max_elements has been fail DOC	keys
disallowing creation of new index entries Should not be called concurrently with IndexGet DOC	handle
Pointer to an Index instance The input handle If skips the first entry of the tensor This allows to load tensors that are aligned with an	embedding
Pointer to an Index instance The input handle If skips the first entry of the tensor This allows to load tensors that are aligned with an where the first entry corresponds to the default index	entry
the first element of the output tensor will be element of index DOC	items
array of probabilities for prediction	L
array of probabilities for prediction array of JSD	losses
	INT_MAX
auto	input_dims_long = GetDimsVector(in[0])
const auto	canonical_axis
int default	to
Coefficient of	leakage
but operating on bit rowwise quantized matrices with fused uint8 tensor obtained with	LENGTHS
	where
for each	row
for each weights are accessed by	indices [0..L-1]
for each weights are accessed by where L is the length of given row This is basically a fused	WEIGHT
	m_2
	m_n
reshape it into matrix of we compute	scale_i
reshape it into matrix of we compute where min_i and max_i minimum and maximum elements of i th and quantize each element	r_
reshape it into matrix of we compute where min_i and max_i minimum and maximum elements of i th and quantize each element each row r_i of which stores a pair	s_i
reshape it into matrix of we compute where min_i and max_i minimum and maximum elements of i th and quantize each element each row r_i of which stores a pair	b_i
and LENGTHS tensor of duplicate each entry of the outer most dimension of DATA according to and concatenate them in an output tensor of rank r	Example
where segments are defined by their and concatenate them in an output tensor of	shape =(SIZE(LENGTHs), k). In case there's less than k values in a segment
where segments are defined by their and concatenate them in an output tensor of the output value will be padded and the corresponding output indices will be padded by DOC Tensor of int32 lengths of rank	TopKIndices
	exists
INT_MAX	int
INT_MAX default if use the db path directly and do not prepend the current root folder of the workspace	string
INT_MAX default if use the db path directly and do not prepend the current root folder of the workspace	default
but allows one to save to db every few	iterations
stride	sizes
stride and pad lengths defined by the ConvPoolOpBase Output data tensor from L p pooling across the input tensor Dimensions will vary based on various	kernel
stride and pad lengths defined by the ConvPoolOpBase Output data tensor from L p pooling across the input tensor Dimensions will vary based on various	stride
and label is applied to the tensor elementwise If	y
ArgumentHelper	arg_helper (def)
int	axis_a = arg_helper.GetSingleArgument<int>("axis_a", 1)
int	axis_b = arg_helper.GetSingleArgument<int>("axis_b", 1)
int	trans_a = arg_helper.GetSingleArgument<bool>("trans_a", false)
int	trans_b = arg_helper.GetSingleArgument<bool>("trans_b", false)
int	canonical_axis_a = canonical_axis_index_(axis_a, in[0].dims().size())
int	canonical_axis_b = canonical_axis_index_(axis_b, in[0].dims().size())
int	M = size_to_dim_(canonical_axis_a, GetDimsVector(in[0]))
	A
matrix of	B
	data_0
	prediction
D float i	e
D float i batch size D is number of possible classes labels	accuracies
axis to	normalize
	lengths
d int long tensor contains the length in each of the output	packed_tensor
d int long tensor contains the length in each of the output N dim Tensor where	presence_mask
d int long tensor contains the length in each of the output N dim Tensor where dim boolean false where packed_tensor is	padded
d int long tensor contains the length in each of the input N dim	Tensor
CPUContext::PadTensorInference Input data tensor from the previous	operator
dimensions depend on whether the NCHW or NHWC operators are being used For in the	former
given a sample set of raw labeled with their corresponding percentiles from the same distribution In	particular
given a sample set of raw labeled with their corresponding percentiles from the same distribution In this	value_to_pct
given a sample set of raw labeled with their corresponding percentiles from the same distribution In this Sorted with columns Each element in the first column is a float representing the raw value of a sample Its corresponding element in the next column represents the percentile it maps to	percentile_values
given a sample set of raw labeled with their corresponding percentiles from the same distribution In this Sorted with columns Each element in the first column is a float representing the raw value of a sample Its corresponding element in the next column represents the percentile it maps to tensor of	floats
given a sample set of raw labeled with their corresponding percentiles from the same distribution In this Sorted with columns Each element in the first column is a float representing the raw value of a sample Its corresponding element in the next column represents the percentile it maps to tensor of with the same dimensions as the flattened input tensor Each element of this corresponds to the percentile calculated for	original_values [i]
	probabilities
a D or D slopes and intercepts The output tensor has the same shape of input predictions and contains the predictions transformed by the piecewise linear functions Each column of predictions has its own piecewise linear transformation functions Therefore the size of piecewise function parameters are pieces x	prediction_dimensions
a D or D slopes and intercepts The output tensor has the same shape of input predictions and contains the predictions transformed by the piecewise linear functions Each column of predictions has its own piecewise linear transformation functions Therefore the size of piecewise function parameters are pieces x except for binary predictions where only the positive prediction needs them Note that in each	piece
a D or D slopes and intercepts The output tensor has the same shape of input predictions and contains the predictions transformed by the piecewise linear functions Each column of predictions has its own piecewise linear transformation functions Therefore the size of piecewise function parameters are pieces x except for binary predictions where only the positive prediction needs them Note that in each low bound is excluded while high bound is included Also the piecewise linear function must be continuous Notes If the input is binary set the binary arg to true so that one group of piecewise linear functions is	slopes
a D or D slopes and intercepts The output tensor has the same shape of input predictions and contains the predictions transformed by the piecewise linear functions Each column of predictions has its own piecewise linear transformation functions Therefore the size of piecewise function parameters are pieces x except for binary predictions where only the positive prediction needs them Note that in each low bound is excluded while high bound is included Also the piecewise linear function must be continuous Notes If the input is binary set the binary arg to true so that one group of piecewise linear functions is intercepts can be passed either through args or through input blobs If we have multiple groups of piecewise linear	functions
a D or D slopes and intercepts The output tensor has the same shape of input predictions and contains the predictions transformed by the piecewise linear functions Each column of predictions has its own piecewise linear transformation functions Therefore the size of piecewise function parameters are pieces x except for binary predictions where only the positive prediction needs them Note that in each low bound is excluded while high bound is included Also the piecewise linear function must be continuous Notes If the input is binary set the binary arg to true so that one group of piecewise linear functions is intercepts can be passed either through args or through input blobs If we have multiple groups of piecewise linear each group has the same number of pieces If a prediction is out of the	bounds
const char *	kAveragePoolDoc
const char *	kMaxPoolDoc
and produces one output is applied to the data tensor elementwise DOC	Slope
Size of the dimension to prepend	reshaped
Number of dimensions to reduce	data_in
the value to replace	NaN
auto	actualNewShape = helper.GetRepeatedArgument<int64_t>("shape")
int64_t	totalSize = 1
int	unknownIdx = -1
	else
an extra argument shape must be specified It outputs the reshaped tensor as well as the original shape At most one dimension of the new shape can be In this	case
an extra argument shape must be specified It outputs the reshaped tensor as well as the original shape At most one dimension of the new shape can be In this the value is inferred from the size of the tensor and the remaining dimensions A dimension could also	be
	segments
	embeddings
Prefix string to prepend extracted blobs	blob_names
Prefix string to prepend extracted blobs tensor of strings containing extracted blob	names
the implementation takes an the hidden state the cell and a weight	TxNxD
the implementation takes an the hidden state the cell and a weight the final hidden cell	bidirectional
the implementation takes an the hidden state the cell and a weight the final hidden cell	num_layers
the implementation takes an the hidden state the cell and a weight the final hidden cell	rnn_mode
the implementation takes an the hidden state the cell and a weight the final hidden cell input_mode are passed directly through to CuDNN	DOC
R recurrent or input R	all_params
See RoIPoolF	dY
See RoIPoolF Gradient of forward	dX
Spatial scale of the input feature map X relative to the input image E	g
Pooled output Y s width feature map input of	RoIs
there are multiple output	cases
default the scale to	apply
affects the activation function itself This should go with the weight initialization in the paper See	https
Number of copies of padding to add around each range T< N, D1..., Dn > Input data	start_padding
Number of copies of padding to add around each range T< N, D1..., Dn > Input data T< D1..., Dn > Padding data for range start	data_out
Outer size of padding present around each range T< N, D1..., Dn > Padded input data	padding_sum
Outer size of padding present around each range T< N, D1..., Dn > Padded input data Sum of all start	paddings
	rather
it will be coerced into one For an arbitrary n dimensional tensor X in[a_0, a_1,..., a_{k-1}, a_k,..., a_{n-1}] and k is the axis	provided
it will be coerced into one For an arbitrary n dimensional tensor X in[a_0, a_1,..., a_{k-1}, a_k,..., a_{n-1}] and k is the axis then X will be coerced into a dimensional tensor with dimensions[a_0 *...*a_{k-1}, a_k *...*a_{n-1}] For the default case where this means the X tensor will be coerced into a tensor of	dimensions [a_0, a_1 *...*a_{n-1}]
it will be coerced into one For an arbitrary n dimensional tensor X in[a_0, a_1,..., a_{k-1}, a_k,..., a_{n-1}] and k is the axis then X will be coerced into a dimensional tensor with dimensions[a_0 *...*a_{k-1}, a_k *...*a_{n-1}] For the default case where this means the X tensor will be coerced into a tensor of where a_0 is often the batch size In this	situation
it will be coerced into one For an arbitrary n dimensional tensor X in[a_0, a_1,..., a_{k-1}, a_k,..., a_{n-1}] and k is the axis then X will be coerced into a dimensional tensor with dimensions[a_0 *...*a_{k-1}, a_k *...*a_{n-1}] For the default case where this means the X tensor will be coerced into a tensor of where a_0 is often the batch size In this we must have	a_0
auto	labels = in[1]
const int	num_classes
Unscaled log probabilities	weight_tensor
Unscaled log probabilities Optional blob to be used to weight the samples for the loss	loss
this op outputs a copy of the input tensor where values from the height and width dimensions are moved to the batch dimension After the zero	padding
followed by cropping This is the reverse transformation of SpaceToBatch More	specifically
	param
Parameters to be normalized	grad
Parameters to be normalized Gradient computed A bool variable to control whether to use max norm or constant norm When	use_max_norm = false
Parameters to be normalized Gradient computed A bool variable to control whether to use max norm or constant norm When constant norm is used so that all the embedding vectors are scaled to have a L2 norm equals to max norm is used so that embedding is scaled so that its l2 norm is no larger than A If an embedding s norm is less than A	originally
Unscaled log probabilities Optional blob to be used to weight the samples for the loss With spatial weighting is by	x
A Blob pointing to the newly created	StatRegistry
If export values from given StatRegistry export values from the global singleton StatRegistry int64 tensor with exported values default true Whether to atomically reset the counters	afterwards
returning a pointer to it The timer is stopped by calling TimerEnd DOC	timer
returning a pointer to it The timer is stopped by calling TimerEnd DOC Pointer to to be passed to	TimerEnd
stops the timer publishing a CAFFE_EVENT DOC	nanos
stops the timer publishing a CAFFE_EVENT DOC nanoseconds in	int64
	utilization
Delta in max CPU utilization	observed
default flag to indicate if the summarized statistics have to be written to a log file D	max
default flag to indicate if the summarized statistics have to be written to a log file D mean and standard	deviation
constexpr char	kSummaryzeOpExtension [] = ".summary"
auto	tiles = helper.GetSingleArgument<int32_t>("tiles", 1)
Number of replicas The input tensor optional Axis to replicate	tiled_data
	a_2
	a_n
r and integer argument	k
r and integer argument return two k which contains the values of the top k elements along the last dimension Index tensor of shape[a_1, a_2,..., a_n, k] which contains the indices of the top k this	Values
vector< int >	axes = helper.GetRepeatedArgument<int>("axes")
auto	valid_axes
A list of integers By reverse the otherwise permute the axes according to the values given	transposed
	time
The time in	nanoseconds
bool	if
bool saves contents to the root folder of the current	workspace
int	total = 1
A tensor of sharing its	storage
weight tensor pairs Input should be in the form	X_0
weight tensor pairs Input should be in the form	weight_0
weight tensor pairs Input should be in the form	X_1
weight tensor pairs Input should be in the form	weight_1
Tensor to be updated	INDICES
Tensor to be updated D list of indices on the first dimension of X_0 that need to be updated	Weight_1
Tensor to be updated	SLICES
Tensor to be updated Update	slices
const char	kPrintFileExtension [] = ".log"
const ArgumentHelper	args (def)
	sampling_cdf
An optional D Tensor< float > Input cumulative sampling all values in sampling_cdf will be scaled by this number	sampled_indexes
an index is randomly sampled from the distribution given by the weights of the corresponding batch The output is a D	sampling_weights
an index is randomly sampled from the distribution given by the weights of the corresponding batch The output is a D A D Tensor< float > of	sampling_values
an index is randomly sampled from the distribution given by the weights of the corresponding batch The output is a D A D Tensor< float > of An optional D Tensor< float > of The output tensor contains	sampled_values
Timeout in	secs
Timeout in	queue
The shared pointer for the	BlobsQueue
the output status will be set to true which can be used as exit criteria for execution step The input is the queue and the last output is the status The rest are data blobs DOC The shared pointer for the BlobsQueue	status
the output status will be set to true which can be used as exit criteria for execution step The input is the queue and the last output is the status The rest are data blobs DOC The shared pointer for the BlobsQueue Is set to depending on the success of	dequeue
Parameters to be updated Gradient computed	output_param
	moment
	lr
runs the dense AdaGrad update and Parameters to be updated Sparse indices learning rate	output_moment_1
runs a modified sparse Adagrad update and where moment is a tensor with length equal to the number of rows in and the new moment is calculated by adding the average squared sum of gradients across each row Note that indices must also be a tensor indexing into the rows of param DOC Moment history Gradient computed Updated parameters	Default
Parameters to be updated	moment_2
Parameters to be updated Second moment history learning rate Updated parameters	output_moment_2
Parameters to be updated Second moment history Gradient computed	iter
	input_tensor
Tensor of floats to be clipped	additional_threshold
Tensor of floats to be clipped An optional additonal threshold to scale the orignal threshold	clipped
	mutex
given	nesterov
given	computes
given	adjusted_gradient
given param	momentum
given param parameter Note the difference to	MomentumSGD
GradientSlice with gradients for updated indices Learning rate Adjusted gradient Updated parameter boolean Whether to use Nesterov Accelerated	Gradient
Parameters to be updated Learning rate	curv_win
Parameters to be updated Learning rate Memory for latest curvature ranges	g2_avg
Parameters to be updated Learning rate Memory for latest curvature ranges Moving average of squared gradient Gradient computed Parameters to be updated	output_lr
Parameters to be updated Learning rate Memory for latest curvature ranges Moving average of squared gradient Gradient computed Parameters to be updated Output learning rate	output_curv_win
Parameters to be updated Learning rate Memory for latest curvature ranges Moving average of squared gradient Gradient computed Parameters to be updated Output learning rate Output memory for latest curvature ranges	output_g2_avg
INT_MAX	compressed
constexpr size_t	kDefaultMinWorkSize = 80
constexpr size_t	kCacheLineSize = 64
constexpr size_t	kGEMMLOWPCacheLineSize = 64
const int	kMaxBusyWaitNOPs = 32 * 1000 * 1000
where N is the number of elements in the	batch
where N is the number of elements in the H and W are the height and	width
where N is the number of elements in the H and W are the height and and each of length num_classes The softmax is applied to each group independently	See
number of classes in each softmax group tensor of softmax probabilities with where and softmax was applied to each of the num_anchors	groups
	scores
See GroupSpatialSoftmax	d_scores
L2 to L1 transition point	Y_hat
L2 to L1 transition point tensor of bounding box regression predictions with tensor of labels	locations
See SelectSmoothL1Loss See SelectSmoothL1Loss	d_loss
See SelectSmoothL1Loss See SelectSmoothL1Loss Gradient of forward	d_Y_hat
where indicates that the corresponding sample should be ignored	and
multiply the loss by this scale factor Tensor of predicted	targets
where N is the number of elements in the H and W are the height and and each of length num_classes For the binary form of Focal	Loss
See SigmoidFocalLoss	normalizer
See SigmoidFocalLoss See SigmoidFocalLoss	d_logits
NumInputs(4).NumOutputs(1).SetDoc(R"DOC( Smooth L1 Loss is a minor variation of Huber loss in which the point of transition between L2 loss and L1 loss is adjustable by a hyper-parameter beta L2 to L1 transition point Tensor of Tensor of labels with the same shape as Y_hat	alpha_out
See SmoothL1Loss	alpha_in
where N is the number of elements in the H and W are the height and and where	p_i = exp(s_i) / sum_j exp(s_j)
or input of	H0
See SpatialNarrowAs	dC
See SpatialNarrowAs Gradient of forward	dA

Detailed Description

A global dictionary that holds information about what Caffe2 modules have been loaded in the current runtime, and also utility functions to load modules.

Copyright (c) 2016-present, Facebook, Inc.

Copyright (c) 2016, NVIDIA CORPORATION, All rights reserved.

Copyright 2016 Facebook.

Simple registry implementation in Caffe2 that uses static variables to register object creators during program initialization time.

Author

Tudor Bosman (tudor.nosp@m.b@fb.nosp@m..com)

Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:

1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Typedef Documentation

template<typename InputTypes , class Context , class Functor , class OutputType = SameTypeAsInput>

using caffe2::UnaryElementwiseOp = typedef UnaryElementwiseWithArgsOp< InputTypes, Context, WithDefaultConstructor<Functor>, OutputType>

UnaryElementwiseOp is a wrapper around UnaryElementwiseWithArgsOp, with the difference that it takes a functor with default constructor, e.g.

that does not need to take into consideration any arguments during operator creation.

Definition at line 93 of file elementwise_op.h.

Function Documentation

int caffe2::CaffeCudaGetDevice

(

)

Gets the current GPU id.

This is a simple wrapper around cudaGetDevice().

Definition at line 109 of file common_gpu.cc.

void caffe2::CaffeCudaSetDevice

(

const int

id

)

Gets the current GPU id.

This is a simple wrapper around cudaGetDevice().

Definition at line 122 of file common_gpu.cc.

unique_ptr< NetBase > caffe2::CreateNet	(	const NetDef &	net_def,
		Workspace *	ws
	)

Creates a network, accessing / creating blobs in the given workspace.

Note that this is different from Workspace::CreateNet. The latter adds the created net object to the workspace's net map, while this function returns a standalone net object.

Definition at line 103 of file net.cc.

template<>

std::unique_ptr<RecurrentNetworkExecutorBase> caffe2::createRNNExecutor< CPUContext >	(	const NetDef &	step_net_def,
		std::map< string, string > &	recurrent_input_map,
		std::string	timestep_blob,
		ArgumentHelper	rnn_args
	)

Implementation of RecurrentNetworkExecutor that uses thread pool for multithreaded execution of RNNs.

Used with CPU.

Definition at line 13 of file recurrent_network_executor.cc.

template<typename Context >

void caffe2::createSharedBuffer

(

Workspace *

ws

)

Creates a mutex and shared buffer in the workspace.

Not thread-safe, must be called from the constructor.

const CaffeMap< string, const ModuleSchema * > & caffe2::CurrentModules

(

)

Current Modules present in the Caffe2 runtime.

Returns: map: a map of modules and (optionally) their description. The key is the module name, and the value is the description for that module. The module name is recommended to be the part that constitutes the trunk of the dynamic library: for example, a module called libcaffe2_db_rocksdb.so should have the name "caffe2_db_rocksdb". The reason we do not use "lib" is because it's somewhat redundant, and the reason we do not include ".so" is for cross-platform compatibility on platforms like mac os.

Definition at line 27 of file module.cc.

template<typename IndexType , typename InType , typename OutType , bool IS_WEIGHT_POSITIONAL = false>

void caffe2::EmbeddingLookup	(	const TIndex	block_size,
		const TIndex	output_size,
		const TIndex	index_size,
		const TIndex	data_size,
		const InType *	input,
		const IndexType *	indices,
		const int *	lengths,
		const float *	weights,
		const float *	scale_bias,
		bool	normalize_by_lengths,
		OutType *	out
	)

Embedding lookup with reduction.

input of size data_size * block_size indices of size index_size lengths of size output_size weights nullptr or array of size index_size out of size output_size * block_size sum(lengths[i]) == index_size

Behavior is roughly equivalent to pseudocode:

pos = 0 for (i = 0..index_size-1) for (k = 0..block_size-1) out[i*block_size + k] = 0 for (j = 0..lengths[i]-1) for (k = 0..block_size-1) out[i*block_size + k] += input[indices[pos]*block_size + k] * (weights ? weights[IS_WEIGHT_POSITIONAL ? j : pos] : 1.0) pos += 1 if (normalize_weights && lengths[i] > 0) for (k = 0..block_size-1) out[i*block_size + k] /= lengths[i]

template<typename IndexType , typename InType , typename OutType , bool IS_WEIGHT_POSITIONAL = false>

void caffe2::Fused8BitRowwiseEmbeddingLookup	(	const TIndex	block_size,
		const TIndex	output_size,
		const TIndex	index_size,
		const TIndex	data_size,
		const InType *	input,
		const IndexType *	indices,
		const int *	lengths,
		const float *	weights,
		bool	normalize_by_lengths,
		OutType *	out
	)

Embedding lookup with reduction.

input of size data_size * (block_size + 8B) indices of size index_size lengths of size output_size weights nullptr or array of size index_size out of size output_size * block_size sum(lengths[i]) == index_size

Note that block_size should be the number of quantized values per row in the data, i.e. excluding the scale and bias. The total (fused) block size is assumed to be this block_size, plus 4 bytes for scale and 4 bytes for bias.

Behavior is roughly equivalent to pseudocode:

pos = 0 fused_block_size = block_size + 8B // quantized values and scale and bias for (i = 0..index_size-1) for (k = 0..block_size-1) out[i*block_size + k] = 0 for (j = 0..lengths[i]-1) for (k = 0..block_size-1) out[i*block_size + k] += input[indices[pos]*(fused_block_size) + k] * (weights ? weights[IS_WEIGHT_POSITIONAL ? j : pos] : 1.0) pos += 1 if (normalize_weights && lengths[i] > 0) for (k = 0..block_size-1) out[i*block_size + k] /= lengths[i]

CudaMemoryPoolType caffe2::GetCudaMemoryPoolType

(

)

Gets the current memory pool type used by Caffe2.

The memory pool is set up during caffe2's global initialization time.

bool caffe2::GetCudaPeerAccessPattern

(

vector< vector< bool > > *

pattern

)

Return a peer access pattern by returning a matrix (in the format of a nested vector) of boolean values specifying whether peer access is possible.

This function returns false if anything wrong happens during the query of the GPU access pattern.

Definition at line 218 of file common_gpu.cc.

const cudaDeviceProp & caffe2::GetDeviceProperty

(

const int

device

)

Gets the device property for the given device.

This function is thread safe.

Definition at line 166 of file common_gpu.cc.

bool caffe2::GlobalInit	(	int *	pargc,
		char ***	argv
	)

Initialize the global environment of caffe2.

Caffe2 uses a registration pattern for initialization functions. Custom initialization functions should take the signature bool (func)(int, char***) where the pointers to argc and argv are passed in. Caffe2 then runs the initialization in three phases: (1) Functions registered with REGISTER_CAFFE2_EARLY_INIT_FUNCTION. Note that since it is possible the logger is not initialized yet, any logging in such early init functions may not be printed correctly. (2) Parses Caffe-specific commandline flags, and initializes caffe logging. (3) Functions registered with REGISTER_CAFFE2_INIT_FUNCTION. If there is something wrong at each stage, the function returns false. If the global initialization has already been run, the function returns false as well.

Definition at line 18 of file init.cc.

MPI_Comm caffe2::GlobalMPIComm

(

)

Gets the global MPI communicator used by Caffe2.

In default, this is MPI_COMM_WORLD unless you call SetGlobalMPIComm().

Definition at line 20 of file mpi_common.cc.

bool caffe2::HasCudaGPU ( )

inline

Check if the current running session has a cuda gpu present.

Note that this is different from having caffe2 built with cuda. Building Caffe2 with cuda only guarantees that this function exists. If there are no cuda gpus present in the machine, or there are hardware configuration problems like an insufficient driver, this function will still return false, meaning that there is no usable GPU present.

In the open source build, it is possible that Caffe2's GPU code is dynamically loaded, and as a result a library could be only linked to the CPU code, but want to test if cuda is later available or not. In this case, one should use HasCudaRuntime() from common.h.

Definition at line 99 of file common_gpu.h.

caffe2::IdenticalTypeAndShape

(

)

Initial value:

= alpha * (exp(x) - 1.) for x <
0`, `f(x) = x for x >= 0`.

void caffe2::LoadModule	(	const string &	name,
		const string &	filename = ""
	)

Load a module.

Inputs: name: a module name or a path name. It is recommended that you use the name of the module, and leave the full path option to only experimental modules. filename: (optional) a filename that serves as a hint to load the module.

Definition at line 52 of file module.cc.

template<typename F >

detail::ScopeGuardImplDecay<F> caffe2::MakeGuard ( F &&  f )

noexcept

ScopeGuard is a general implementation of the "Initialization is Resource Acquisition" idiom.

Basically, it guarantees that a function is executed upon leaving the currrent scope unless otherwise told.

The MakeGuard() function is used to create a new ScopeGuard object. It can be instantiated with a lambda function, a std::function<void()>, a functor, or a void(*)() function pointer.

Usage example: Add a friend to memory iff it is also added to the db.

void User::addFriend(User& newFriend) { // add the friend to memory friends_.push_back(&newFriend);

// If the db insertion that follows fails, we should // remove it from memory. auto guard = MakeGuard([&] { friends_.pop_back(); });

// this will throw an exception upon error, which // makes the ScopeGuard execute UserCont::pop_back() // once the Guard's destructor is called. db_->addFriend(GetName(), newFriend.GetName());

// an exception was not thrown, so don't execute // the Guard. guard.dismiss(); }

Examine ScopeGuardTest.cpp for some more sample usage.

Stolen from: Andrei's and Petru Marginean's CUJ article: http://drdobbs.com/184403758 and the loki library: http://loki-lib.sourceforge.net/index.php?n=Idioms.ScopeGuardPointer and triendl.kj article: http://www.codeproject.com/KB/cpp/scope_guard.aspx

Definition at line 153 of file scope_guard.h.

bool caffe2::MatchStrings	(	string	p,
		string	s
	)

This allows for the use of * and | to match operator types, engines, or any other property that is represented by strings.

For example, if we wanted to match an operator to Conv or FC, we can give: "Conv|FC" as the type() of that op.

Definition at line 214 of file graph.cc.

void caffe2::MPISetupPeers	(	const int	replicas,
		const string &	role,
		const string &	job_path
	)

A function used to perform peer setup so one does not need to use mpirun / mpiexec to run the binary.

Note that if you use mpirun or mpiexec to set up the common world, do not use this function - MPI_Init would have already set that up.

This also assumes that you have a common path (like NFS) that multiple instances can read from.

Inputs: replicas (int): the number of replicas that mpi will run with. role (string): the role of this process, "server" or "client". job_path (string): a file name that the server will write its port into and the clients will read the server's port from.

Definition at line 94 of file mpi_common.cc.

bool caffe2::ParseCaffeCommandLineFlags	(	int *	pargc,
		char ***	pargv
	)

Parses the commandline flags.

This command parses all the commandline arguments passed in via pargc and argv. Once it is finished, partc and argv will contain the remaining commandline args that caffe2 does not deal with. Note that following convention, argv[0] contains the binary name and is not parsed.

Definition at line 55 of file flags.cc.

void caffe2::SetGlobalMPIComm

(

MPI_Comm

new_comm

)

Sets the global MPI communicator.

Caffe2 takes over the ownership of the passed in communicator.

Definition at line 24 of file mpi_common.cc.

void caffe2::ShowLogInfoToStderr

(

)

A utility to allow one to show log info to stderr after the program starts.

This is similar to calling GLOG's –logtostderr, or setting caffe2_log_level to smaller than INFO. You are recommended to only use this in a few sparse cases, such as when you want to write a tutorial or something. Normally, use the commandline flags to set the log level.

Definition at line 196 of file logging.cc.

Variable Documentation

it will be coerced into one For an arbitrary n dimensional tensor X in [a_0, a_1, ..., a_{k-1}, a_k, ..., a_{n-1}] and k is the axis then X will be coerced into a dimensional tensor with dimensions [a_0 * ... * a_{k-1}, a_k * ... * a_{n-1}] For the default case where this means the X tensor will be coerced into a tensor of where a_0 is often the batch size In this we must have caffe2::a_0

Initial value:

= N and a_1 * ... * a_{n-1} = D.
Each of these dimensions must be matched correctly

Definition at line 98 of file softmax_op.cc.

where indicates that the corresponding sample should be ignored caffe2::and

Initial value:

{0, 1} correspond to the binary classes 0 and 1. By
default the loss is divided by the number of targets > -1 and then multiplied by
the `scale` op argument. The divisive normalization may be disable by setting
the op argument `normalize` to 0 (the multiplication by `scale` still takes
effect).

This op fuses sigmoid and cross entropy for numerical stability in both forward
and gradient computation.
)DOC")
    .Arg(
        "scale",
        "(float) default 1.0

Definition at line 34 of file sigmoid_cross_entropy_loss_op.cc.

Number of replicas The input tensor caffe2::axis

Initial value:

=1
  shape(A) = (2, 3, 2, 5)

Definition at line 29 of file concat_split_op.cc.

auto caffe2::batch_size

Initial value:

=
              size_to_dim_(canonical_axis, GetDimsVector(logits))

Definition at line 102 of file elementwise_linear_op.cc.

const auto caffe2::canonical_axis

Initial value:

=
          canonical_axis_index_(axis, in[0].dims().size())

Definition at line 169 of file layer_norm_op.cc.

Number of replicas caffe2::data

Initial value:

= slope * x for x < 0`,
`f(x) = x for x >= 0`.

Definition at line 101 of file image_input_op.cc.

int caffe2::default

Initial value:

=\"\") characters in the provided blob "
        " names that match strip_prefix will be removed prior to loading."
        " Also, characters that precede strip_prefix will be removed. Useful "
        " for removing device scope from blob names.")
    .Arg("db", "(string) the path to the db to load.")
    .Arg(
        "dbs",
        "(list of strings) the paths to the dbs to load. This is used for loading"
        " blobs from multiple databases. If it is set, argument in \"db\" will be"
        " ignored.")
    .Arg("db_type", "(string) the type of the db.")
    .Arg(
        "keep_device",
        "(int, default 0) if nonzero, the blobs are loaded into the device that "
        "is specified in the serialized BlobProto. Otherwise, the device will be "
        "set as the one that the Load operator is being run under.")
    .Arg(
        "load_all",
        "(int, default 0) if nonzero, will load all blobs pointed to by the db "
        "to the workspace overwriting/creating blobs as needed.")
    .Arg(
        "allow_incomplete",
        "(bool, default false) if true, will allow not loading all the output "
        "blobs specified in the outputs")
    .Arg(
        "source_blob_names",
        "(list of strings) if set, used instead of output "
        "blob names, to specify which blobs in the db shall be loaded. Must be "
        "the same length as number of output blobs.")

Definition at line 48 of file load_save_op.cc.

auto caffe2::dilation_h

Initial value:

= helper.GetSingleArgument<int>(
              "dilation_h", helper.GetSingleArgument<int>("dilation", 1))

Definition at line 40 of file im2col_op.cc.

auto caffe2::dilation_w

Initial value:

= helper.GetSingleArgument<int>(
              "dilation_w", helper.GetSingleArgument<int>("dilation", 1))

Definition at line 42 of file im2col_op.cc.

given param caffe2::else

Initial value:

{
            CAFFE_ENFORCE_EQ(
                totalSize,
                size,
                "Argument `shape` does not agree with the input data.",
                " (",
                totalSize,
                " != ",
                size,
                ")")

Definition at line 88 of file reshape_op.cc.

const vector< TensorShape > & caffe2::in

Initial value:

{
      vector<TensorShape> out(1)

Definition at line 12 of file batch_gather_ops.cc.

Focal Loss s gamma hyper parameter tensor of softmax caffe2::inputs

Initial value:

= X_{nd} * w_d + b_d
  )DOC")
    .Input(0, "X", "2D input tensor of size (N X D) data")
    .Input(1, "w", "1D scaling factors of size D")
    .Input(2, "b", "1D biases of size D")
    .Output(0, "Y", "2D output tensor")
    .Arg(
        "axis",
        "default to 1

Definition at line 92 of file elementwise_linear_op.cc.

const char* caffe2::kAveragePoolDoc

Initial value:

= R"DOC(
consumes an input blob X and applies average pooling across the
the blob according to kernel sizes, stride sizes, and pad lengths defined by the
ConvPoolOpBase operator. Average pooling consisting of averaging all values of a
subset of the input tensor according to the kernel size and downsampling the
data into the output blob Y for further processing.
)DOC"

Definition at line 730 of file pool_op.cc.

const char* caffe2::kBroadcastDoc

Initial value:

= R"DOC(
If necessary the right-hand-side argument will be broadcasted to match the
shape of left-hand-side argument. When broadcasting is specified, the second
tensor can either be of size 1 (a scalar value), or having its shape as a
contiguous subset of the first tensor's shape. The starting of the mutually
equal shape is specified by the argument "axis", and if it is not set, suffix
matching is assumed. 1-dim expansion doesn't work yet.

For example, the following tensor shapes are supported (with broadcast=1):

  shape(A) = (2, 3, 4, 5), shape(B) = (,), i.e. B is a scalar
  shape(A) = (2, 3, 4, 5), shape(B) = (5,)
  shape(A) = (2, 3, 4, 5), shape(B) = (4, 5)
  shape(A) = (2, 3, 4, 5), shape(B) = (3, 4), with axis=1
  shape(A) = (2, 3, 4, 5), shape(B) = (2), with axis=0

Argument `broadcast=1` needs to be passed to enable broadcasting.
)DOC"

Definition at line 7 of file elementwise_op_schema.cc.

const char* caffe2::kConvDoc

Initial value:

= R"DOC(
Note that other parameters, such as the stride and
kernel size, or the pads' sizes in each direction are not necessary for input
because they are provided by the ConvPoolOpBase operator. Various dimension
checks are done implicitly, and the sizes are specified in the Input docs for
this operator. As is expected, the filter is convolved with a subset of the
image and the bias is added; this is done throughout the image data and the
output is computed. As a side note on the implementation layout:
conv_op_impl.h is the templated implementation of the conv_op.h file, which is
why they are separate files.
)DOC"

Definition at line 7 of file conv_op.cc.

auto caffe2::kernel_h

Initial value:

= helper.GetSingleArgument<int>(
              "kernel_h", helper.GetSingleArgument<int>("kernel", 0))

Definition at line 36 of file im2col_op.cc.

auto caffe2::kernel_w

Initial value:

= helper.GetSingleArgument<int>(
              "kernel_w", helper.GetSingleArgument<int>("kernel", 0))

Definition at line 38 of file im2col_op.cc.

const char* caffe2::kMaxPoolDoc

Initial value:

= R"DOC(
consumes an input blob X and applies max pooling across the
the blob according to kernel sizes, stride sizes, and pad lengths defined by the
ConvPoolOpBase operator. Max pooling consisting of taking the maximum value of a
subset of the input tensor according to the kernel size and downsampling the
data into the output blob Y for further processing.
)DOC"

Definition at line 738 of file pool_op.cc.

auto caffe2::num_classes

Initial value:

=
              size_from_dim_(canonical_axis, GetDimsVector(logits))

Definition at line 28 of file softmax_with_loss_op.cc.

auto caffe2::order

Initial value:

= StringToStorageOrder(
              helper.GetSingleArgument<string>("order", "NCHW"))

Definition at line 17 of file do_op.cc.

See SoftmaxFocalLoss See SoftmaxFocalLoss Gradient of forward caffe2::output

Initial value:

= net.BooleanUnmask([mask1, values1, mask2, values2], ["output"])

We get:
  output = 1.0

Definition at line 45 of file abs_op.cc.

Tensor of rank caffe2::r

Initial value:

= 2.")
    .Input(1, "INDICES", "Tensor of int32/int64 indices

Definition at line 29 of file batch_gather_ops.cc.

reshape it into matrix of we compute where min_i and max_i minimum and maximum elements of i th and quantize each element caffe2::r_

Initial value:

{ij} as
0 <= round(r_ij - bias_i) / scale_i) < 256. Instead of input tensor
we obtain uint8 tensor and auxiliary information as scale and bias to
restore input tensor (with losses).
)DOC")
    .Input(0, "input", "input")
    .Output(0, "quantized_input", "quantized_input")
    .Output(
        1,
        "scale_bias",
        "Matrix of floats

Definition at line 176 of file lengths_reducer_rowwise_8bit_ops.cc.

A tensor of caffe2::rank

Initial value:

= axis.")
    .Output(
        0,
        "output",
        "A 2D tensor with the contents of the input tensor

Definition at line 32 of file flatten_op.cc.

reshape it into matrix of we compute caffe2::scale_i

Initial value:

= (min_i - max_i) / 255 and  bias_i = min_i for
i-th row r_i of reshaped matrix

Definition at line 174 of file lengths_reducer_rowwise_8bit_ops.cc.

auto caffe2::stride_h

Initial value:

= helper.GetSingleArgument<int>(
              "stride_h", helper.GetSingleArgument<int>("stride", 1))

Definition at line 44 of file im2col_op.cc.

auto caffe2::stride_w

Initial value:

= helper.GetSingleArgument<int>(
              "stride_w", helper.GetSingleArgument<int>("stride", 1))

Definition at line 46 of file im2col_op.cc.

auto caffe2::valid_axes

Initial value:

=
            std::all_of(axes.begin(), axes.end(), [&tensor_size](int& axis) {
              return axis >= 0 && axis < tensor_size;
            })

Definition at line 35 of file transpose_op.cc.

reconstruct values together according to masks A comprehensive False False True Reconstruct We Note that for all mask there must be at least one True If for a field there are multiple True we will accept the first value For False False False True caffe2::values2

Initial value:

= 2.0
  mask3   = False

Definition at line 69 of file boolean_unmask_ops.cc.

caffe2::y

Initial value:

== 1 then it assumed the first input should be ranked higher
(have a larger value) than the second input

Definition at line 80 of file margin_ranking_criterion_op.cc.