Lite算子支持

Linux Ascend 端侧 推理应用 初级 中级 高级

本文列举MindSpore Lite支持的算子。

操作名	CPU FP16	CPU FP32	CPU Int8	CPU UInt8	GPU FP16	GPU FP32	NPU	支持的TensorFlow Lite算子	支持的Caffe Lite算子	支持的Onnx Lite算子	支持的TensorFlow 算子
Abs	✅	✅	✅	✅	✅	✅	✅	Abs		Abs	Abs
Add	✅	✅	✅	✅	✅	✅	✅	Add		Add, Int8Add	Add, AddV2
AddGrad		✅
AddN		✅						AddN
Assert		✅									Assert
Argmax	✅	✅	✅	✅	✅	✅	✅	Argmax	ArgMax	ArgMax	Argmax
Argmin	✅	✅	✅	✅	✅	✅		Argmin			ArgMin
AvgPool	✅	✅	✅	✅	✅	✅	✅	MeanPooling	Pooling	AveragePool, GlobalAveragePool, Int8AveragePool	AvgPool
AvgPoolGrad		✅
BatchNorm	✅	✅	✅	✅	✅	✅			BatchNorm	BatchNormalization
BatchNormGrad		✅
BatchToSpace		✅	✅	✅	✅	✅		BatchToSpace, BatchToSpaceND			BatchToSpace, BatchToSpaceND
BiasAdd	✅	✅	✅	✅	✅	✅				BiasAdd	BiasAdd
BiasAddGrad		✅
Broadcast		✅						BroadcastTo		Expand
Cast	✅	✅	✅	✅	✅	✅	✅	Cast, QUANTIZE, DEQUANTIZE		Cast	Cast
Ceil	✅	✅	✅	✅	✅	✅	✅	Ceil		Ceil	Ceil
Concat	✅	✅	✅	✅	✅	✅	✅	Concat	Concat	Concat	ConcatV2
ConstantOfShape		✅								ConstantOfShape
Conv2d	✅	✅	✅	✅	✅	✅	✅	Conv2D	Convolution	Conv, Int8Conv, ConvRelu, Int8ConvRelu	Conv2D
Conv2dGrad		✅
Conv2dTranspose	✅	✅	✅	✅	✅	✅	✅	DeConv2D	Deconvolution	ConvTranspose	Conv2DBackpropInput
Conv2dTransposeGrad		✅
Cos	✅	✅	✅	✅	✅	✅	✅	Cos		Cos	Cos
Crop	✅	✅	✅	✅					Crop
CropAndResize		✅									CropAndResize
CustomExtractFeatures		✅						ExtractFeatures
CustomNormalize		✅						Normalize
CustomPredict		✅						Predict
DeDepthwiseConv2D		✅	✅	✅					Deconvolution
DepthToSpace		✅	✅	✅	✅	✅		DepthToSpace		DepthToSpace
DepthwiseConv2dNative	✅	✅	✅	✅	✅	✅	✅	DepthwiseConv2D	Convolution		DepthwiseConv2dNative
DetectionPostProcess		✅	✅	✅				Custom
Div	✅	✅	✅	✅	✅	✅	✅	Div, RealDiv		Div	Div, RealDiv
DivGrad		✅
Eltwise	✅	✅	✅	✅	✅	✅	✅		Eltwise	Sum, Max[3]
Elu		✅							Elu	Elu, NonMaxSuppression	NonMaxSuppressionV3
EluGrad		✅
Equal	✅	✅	✅	✅	✅	✅	✅	Equal		Equal	Equal
Exp		✅			✅	✅		Exp	Exp	Exp	Exp
ExpandDims	✅	✅	✅	✅				ExpandDims			ExpandDims
Fill		✅						Fill			Fill
Flatten	✅	✅							Flatten
Floor	✅	✅	✅	✅	✅	✅	✅	flOOR		Floor	Floor
FloorDiv	✅	✅			✅	✅	✅	FloorDiv			FloorDiv
FloorMod	✅	✅			✅	✅	✅	FloorMod			FloorMod
FullConnection	✅	✅	✅	✅	✅	✅	✅	FullyConnected	InnerProduct
FusedBatchNorm	✅	✅	✅	✅			✅	FusedBatchNorm			FusedBatchNorm, FusedBatchNormV3
GatherNd		✅	✅	✅	✅	✅		GatherND			GatherNd
Gather	✅	✅	✅	✅	✅	✅	✅	Gather		Gather	GatherV2
Greater	✅	✅	✅	✅	✅	✅	✅	Greater		Greater	Greater
GreaterEqual	✅	✅	✅	✅	✅	✅	✅	GreaterEqual			GreaterEqual
GRU	✅	✅
HardTanh	✅	✅
HashtableLookup		✅						HashtableLookup
HSigmoid		✅		✅
Hswish	✅	✅	✅	✅	✅	✅	✅	HardSwish
HswishGrad		✅
InstanceNorm	✅	✅						InstanceNorm
InvertPermutation		✅									InvertPermutation
L2Norm		✅	✅					L2_NORMALIZATION
LayerNorm		✅	✅
LeakyReLU	✅	✅	✅	✅	✅	✅	✅	LeakyRelu		LeakyRelu	LeakyRelu
LeakyReLUGrad		✅
Less	✅	✅	✅	✅	✅	✅	✅	Less		Less	Less
LessEqual	✅	✅	✅	✅	✅	✅	✅	LessEqual			LessEqual
LRN		✅						LocalResponseNorm		Lrn, LRN
Log	✅	✅	✅	✅	✅	✅	✅	Log		Log	Log
LogGrad	✅	✅
LogicalAnd	✅	✅			✅	✅	✅	LogicalAnd		And	LogicalAnd
LogicalNot	✅	✅	✅	✅	✅	✅	✅	LogicalNot		Not	LogicalNot
LogicalOr	✅	✅			✅	✅	✅	LogicalOr		Or	LogicalOr
LshProjection		✅						LshProjection
LSTM	✅	✅								LSTM
MatMul	✅	✅	✅	✅	✅	✅	✅			MatMul	MatMul, BatchMatMul
MatMulGrad		✅
Maximum	✅	✅			✅	✅	✅	Maximum			Maximum
MaximumGrad		✅
MaxPool	✅	✅	✅	✅	✅	✅	✅	MaxPooling	Pooling	MaxPool, GlobalMaxPool	MaxPool
MaxPoolGrad		✅
Merge	✅	✅									Merge
Minimum	✅	✅			✅	✅	✅	Minimum		Min	Minimum
MinimumGrad		✅
Mul	✅	✅	✅	✅	✅	✅	✅	Mul		Mul	Mul
MulGrad		✅
Neg	✅	✅			✅	✅	✅	Neg		Neg
NegGrad		✅
NotEqual	✅	✅	✅	✅	✅	✅	✅	NotEqual			NotEqual
OneHot		✅			✅	✅		OneHot		OneHot	OneHot
Pad	✅	✅	✅	✅	✅	✅	✅	Pad, MirrorPad		Pad	MirrorPad, Pad
Pow		✅	✅	✅	✅	✅		Pow	Power	Pow[2]	Pow
PowGrad		✅
PReLU		✅			✅	✅		PRELU	PReLU	PRelu
RandomStandardNormal		✅									RandomStandardNormal
RandomUniform		✅									RandomUniform
Range		✅						Range			Range, RaggedRange
Rank		✅						Rank			Rank
Reciprocal	✅	✅	✅				✅
ReduceAll		✅									All
ReduceASum		✅			✅	✅			Reduction
ReduceMax	✅	✅	✅	✅	✅	✅		ReduceMax		ReduceMax	Max
ReduceMean	✅	✅	✅	✅	✅	✅		Mean	Reduction	ReduceMean	Mean
ReduceMin		✅	✅	✅	✅	✅		ReduceMin		ReduceMin	Min
ReduceProd		✅	✅	✅	✅	✅		ReduceProd		ReduceProd	Prod
ReduceSum		✅	✅	✅	✅	✅		Sum	Reduction	ReduceSum	Sum
ReduceSumSquare		✅	✅	✅					Reduction	ReduceSumSquare
ReLU	✅	✅	✅	✅	✅	✅	✅	Relu	ReLU	Relu	Relu
ReLUGrad	✅	✅
ReLU6	✅	✅	✅	✅	✅	✅	✅	Relu6	ReLU6	Clip[1]	Relu6
ReLU6Grad	✅	✅
Reshape	✅	✅	✅	✅	✅	✅	✅	Reshape	Reshape	Reshape,Flatten	Reshape
Resize		✅	✅	✅	✅	✅	✅	ResizeBilinear, NearestNeighbor	Interp		ResizeBilinear, ResizeBicubic, ResizeNearestNeighbor
ResizeGrad		✅
Reverse		✅						reverse			ReverseV2
ReverseSequence		✅						ReverseSequence			ReverseSequence
Round	✅	✅	✅	✅	✅	✅	✅	Round		Round	Round
Rsqrt	✅	✅	✅	✅	✅	✅	✅	Rsqrt			Rsqrt
Select		✅									Select
Selu											Selu
Scale	✅	✅	✅	✅	✅	✅	✅		Scale
ScatterNd		✅						ScatterNd
Shape	✅	✅	✅	✅	✅	✅	✅	Shape		Shape	Shape
Sigmoid	✅	✅	✅	✅	✅	✅	✅	Logistic	Sigmoid	Sigmoid	Sigmoid
SigmoidGrad	✅	✅
Sin	✅	✅	✅	✅	✅	✅	✅	Sin		Sin	Sin
Size		✅									Size
Slice	✅	✅	✅	✅	✅	✅	✅	Slice	Slice	Slice	Slice
SkipGram		✅						SKipGram
Softmax	✅	✅	✅	✅	✅	✅	✅	Softmax	Softmax	Softmax	Softmax
SoftmaxGrad		✅
SpaceToBatch		✅	✅	✅	✅	✅		SpaceToBatch
SpaceToBatchND		✅	✅	✅	✅	✅		SpaceToBatchND			SpaceToBatchND
SpaceToDepth		✅			✅	✅		SpaceToDepth		SpaceToDepth
SparseToDense		✅			✅	✅		SpareToDense
Split	✅	✅	✅	✅			✅	Split, SplitV		Split	Split, SplitV
Sqrt	✅	✅	✅	✅	✅	✅	✅	Sqrt		Sqrt	Sqrt
Square	✅	✅	✅	✅	✅	✅	✅	Square			Square
SquaredDifference	✅	✅			✅	✅	✅	SquaredDifference			SquaredDifference
Squeeze	✅	✅	✅	✅	✅	✅		Squeeze		Squeeze	Squeeze
StridedSlice	✅	✅	✅	✅	✅	✅	✅	StridedSlice			StridedSlice
Stack	✅	✅			✅	✅		Stack			Pack
Sub	✅	✅	✅	✅	✅	✅	✅	Sub		Sub	Sub
SubGrad		✅
Swish	✅	✅
Switch	✅	✅									Switch
Tanh	✅	✅			✅	✅	✅	Tanh	TanH	Tanh, Sign	Tanh
TanhGrad		✅
TensorListFromTensor	✅	✅									TensorListFromTensor
TensorListGetItem	✅	✅									TensorListGetItem
TensorListReserve	✅	✅									TensorListReserve
TensorListSetItem	✅	✅									TensorListSetItem
TensorListStack	✅	✅									TensorListStack
Tile	✅	✅						Tile	Tile	Tile	Tile
TopK		✅	✅	✅				TopKV2		TopK	TopKV2
Transpose	✅	✅	✅		✅	✅	✅	Transpose	Permute	Transpose	Transpose
Unique		✅						Unique
Unsqueeze	✅	✅	✅	✅			✅			Unsqueeze
Unstack		✅						Unstack
Where		✅						Where			Where
ZerosLike		✅						ZerosLike			ZerosLike
转换工具支持的其他算子[4]										Loop, Dropout, If	Dropout, Enter, Exit, If, IsFinite, LinSpace, LoopCond, NextIteration, StatelessIf, StatelessWhile, While

[1] Clip：仅支持将clip(0, 6)转换为Relu6。

[2] Pow：仅支持指数为单个常数。

[3] Sum与Max：仅支持输入个数为2。

[4] 转换工具支持，但不需要具体实现的算子，一般这类算子在转化工具中被优化而消失，如被融合掉或者使用其他算子代替。