Codegen算子支持

Linux Ascend 端侧 推理应用 初级 中级 高级

本文列举MindSpore Lite Codegen支持的算子。

操作名	CPU FP32	CPU Int8	CMSIS Int8	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
Add	✅	✅	✅	Add		Add, Int8Add	Add, AddV2
AddN	✅			AddN
AvgPool	✅	✅	✅	MeanPooling	Pooling	AveragePool, GlobalAveragePool, Int8AveragePool
BatchNorm		✅	✅		BatchNorm	BatchNormalization
BiasAdd	✅					BiasAdd	BiasAdd
Cast	✅	✅	✅	Cast, QUANTIZE, DEQUANTIZE		Cast	Cast
Ceil	✅			Ceil		Ceil
Concat	✅	✅	✅	Concat	Concat	Concat	ConcatV2
Conv2d	✅	✅	✅	Conv2D	Convolution	Conv, Int8Conv, ConvRelu, Int8ConvRelu	Conv2D
Cos	✅			Cos		Cos
DetectionPostProcess		✅		Custom
Div	✅	✅	✅	Div, RealDiv		Div	Div, RealDiv
Eltwise	✅				Eltwise	Sum, Max[3]
Equal	✅			Equal		Equal	Equal
ExpandDims	✅			ExpandDims			ExpandDims
Floor	✅			flOOR		Floor
FloorDiv	✅			FloorDiv
FloorMod	✅			FloorMod
FullConnection	✅	✅	✅	FullyConnected	InnerProduct
Greater	✅			Greater		Greater	Greater
GreaterEqual	✅			GreaterEqual			GreaterEqual
Less	✅			Less		Less	Less
LessEqual	✅			LessEqual			LessEqual
Log	✅			Log		Log
LogicalAnd	✅			LogicalAnd		And	LogicalAnd
LogicalNot	✅			LogicalNot		Not
LogicalOr	✅			LogicalOr		Or
MatMul	✅	✅				MatMul	MatMul
Maximum	✅			Maximum			Maximum
MaxPool	✅	✅	✅	MaxPooling	Pooling	MaxPool, GlobalMaxPool
Minimum	✅			Minimum		Min	Minimum
Mul	✅	✅	✅	Mul		Mul	Mul
Neg	✅			Neg		Neg
NotEqual	✅			NotEqual			NotEqual
ReLU	✅	✅	✅	Relu	ReLU	Relu	Relu
ReLU6	✅	✅	✅	Relu6	ReLU6	Clip[1]	Relu6
Reshape	✅	✅	✅	Reshape	Reshape	Reshape,Flatten	Reshape
Resize		✅		ResizeBilinear, NearestNeighbor	Interp
Round	✅			Round		Round	Round
Rsqrt	✅			Rsqrt
Sigmoid	✅	✅	✅	Logistic	Sigmoid	Sigmoid	Sigmoid
Sin	✅			Sin		Sin
Softmax	✅	✅	✅	Softmax	Softmax	Softmax
Sqrt	✅			Sqrt		Sqrt
Square	✅			Square
SquaredDifference	✅			SquaredDifference
Squeeze	✅			Squeeze		Squeeze	Squeeze
Sub	✅	✅	✅	Sub		Sub	Sub

[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。