MindSpore Lite Kit Operator List

This document provides the CPU backend operators supported by MindSpore Lite Kit and their mapping to ONNX Opset 18 operators. When using the model conversion tool to convert an ONNX model to an MS model for deployment, you can refer to this list for the supported ONNX operators, thereby ensuring the success of model conversion.

NOTE

All the following operators do not support int64 input tensors.

MindSpore Lite Operator	Description	ONNX Operator	Operator Specifications
Abs	Computes the absolute value element-wise.	Abs	Do not support uint8 tensors. Do not support empty input tensor quantization parameters.
Activation	Computes the activation function.	Celu Clip Elu Gelu HSigmoid LeakyRelu PRelu Relu Sigmoid SoftMax SoftPlus Tanh	-
AddFusion	Computes element-wise addition.	Add	-
ArgMaxFusion	Computes the maximum value along a specified dimension.	ArgMax	Do not support uint8 tensors. Do not support empty input tensor quantization parameters.
AvgPoolFusion	Performs average pooling.	AveragePool GlobalAveragePool GlobalMaxPool MaxPool	-
BatchNorm	Performs batch normalization.	BatchNormalization	-
BroadcastTo	Expands the dimensions of a tensor.	Expand	-
Cast	Converts the data type.	Cast	Do not support the following conversions: fp32 to int8, fp32 to uint32, int32 to int8, int32 to uint32, int32 to uint8, int8 to bool, and int8 to uint8.
Ceil	Rounds up to an integer.	Ceil	-
Clip	Limits the element range.	Clip	-
Concat	Concatenates tensors.	Concat	-
Conv2DFusion	Performs 2D convolution.	Conv	-
Cos	Computes the cosine element-wise.	Cos	-
CumSum	Accumulates the elements.	CumSum	-
DepthToSpace	Rearranges data from depth into blocks of spatial data.	DepthToSpace	Do not support uint8 tensors. Do not support input tensors with unknown dimensions.
DivFusion	Computes element-wise division.	Div	Do not support division by 0.
Eltwise	Computes element-wise summation.	Sum	-
Equal	Checks whether the inputs are equal.	Equal	Do not support uint8 input tensors, or output of bool tensors for input int8 tensors.
Erf	Computes the error function.	Erf	-
ExpFusion	Computes the exponential element-wise.	Exp	-
Flatten	Flattens data by dimension.	Flatten	Do not support uint8 tensors.
Floor	Rounds down to the nearest integer.	Floor	-
FusedBatchNorm	Performs fused batch normalization.	BatchNormalization	-
Gather	Gathers elements from a tensor along a specified axis.	Gather	Do not support uint8 tensors. Do not support QuantType_QUANT_NONE.
GatherD	Gathers elements from a tensor using dynamic indices.	GatherElements	-
GatherNd	Gathers elements from a tensor using N-dimensional indices.	GatherND	-
InstanceNorm	Performs instance normalization.	InstanceNormalization	-
Log	Computes the natural logarithm element-wise.	Log	Do not support negative input tensors.
LogicalNot	Computes the element-wise logical NOT.	Not	-
LogSoftmax	Applies Softmax followed by Log to the input tensor.	LogSoftmax	Do not support inf input tensors.
LRN	Applies local response normalization to reduce overfitting.	LRN	-
MatMulFusion	Performs matrix multiplication on two input tensors by computing the inner product using the input tensors and a set of learned weights, followed by adding a bias.	Gemm MatMul	-
Maximum	Computes the element-wise maximum.	Max	-
MaxPoolFusion	Performs max pooling operation.	GlobalMaxPool MaxPool	-
Minimum	Computes the element-wise minimum.	Min	-
Mod	Computes the element-wise modulus.	Mod	-
MulFusion	Computes the element-wise multiplication.	Mul	-
Neg	Computes the element-wise negation.	Neg	-
PadFusion	Adds padding to a tensor.	Pad	Do not support int32 tensors.
PowFusion	Computes the element-wise power.	Pow	Support only exponents that are a single constant.
PReLUFusion	Performs parametric ReLU activation.	PRelu	-
Range	Generates a sequence of numbers with a specified start, end, and step.	Range	-
Reciprocal	Computes the element-wise reciprocal.	Reciprocal	-
Reshape	Reshapes a tensor.	Reshape	-
Round	Rounds to the nearest integer.	Round	-
ScatterNdUpdate	Updates specified elements of a tensor using given indices and values.	ScatterND	-
Shape	Obtains the shape of a tensor.	Shape	-
Sin	Computes the element-wise sine.	Sin	-
Size	Obtains the size of a tensor.	Size	-
SliceFusion	Slices a tensor.	Slice	-
Softmax	Applies the Softmax function to normalize a tensor along a specified axis.	Softmax	-
SpaceToDepth	Rearranges blocks of spatial data into depth.	SpaceToDepth	-
Sqrt	Computes the element-wise square root.	Sqrt	-
Squeeze	Removes dimensions of size 1.	Squeeze	-
StridedSlice	Slices a tensor with stride.	Slice	-
SubFusion	Computes the element-wise subtraction.	Sub	-
TileFusion	Constructs a tensor by replicating input tensor multiple times.	Tile	Do not support int8 tensors.
TopKFusion	Obtains the top K elements from the input tensor.	TopK	-
Transpose	Transposes a tensor.	Transpose	-
Tril	Obtains the lower triangular part of a matrix.	Trilu (upper=0)	-
Triu	Obtains the upper triangular part of a matrix.	Trilu (upper=1)	-
Unsqueeze	Inserts a new dimension into the input tensor.	Unsqueeze	-
Where	Selects elements based on specified conditions.	Where	-