Supported ONNX Ops

Operator Support

ONNX converters currently support the following operators and parameters:

Operator

Opset Version Supported

Notes

Abs

1,6,13

Add

1,6,7,13,14

And

1,7

AveragePool

1,7,10,11,19,22

ArgMax

1,11,12,13

Only default value(0) is supported for select_last_index

ArgMin

1,11,12,13

Only default value(0) is supported for select_last_index

Asin

7

Atan

7

BatchNormalization

1,6,7,9,14,15

Momentum is not supported. Only default value (0) for training_mode is supported. All inputs after the first must be static. Only the first output is generated.

Ceil

1,6,13

Clip

1,6,11,12

Concat

1,4,11,13

Constant

1,9,11,12,13

sparse tensor is not currently supported.

ConstantOfShape

9

Conv

1,11,22

“kernel_shape” which differs from that of weights is not supported.

ConvTranspose

1,11,22

Cos

7,22

CumSum

11,14

value of axis attribute should be in the range of (-input_rank,

input_rank )

Cast

1,6,9,13

DepthToSpace

1,11,13

DequantizeLinear

10,13,19,21

Div

1,6,7,13,14

Dropout

1,6,7,10

Only test mode (i.e. noop) is supported. Mask output is not generated.

Einsum

12

Elu

1,6,22

Equal

1,7,11,13

Erf

9,13

Only certain cases which can be matched to GeLU will be supported

Exp

1,6,13

Expand

8,13

Static shape tensor only supported

Flatten

1,9,11,13

Floor

1,6

Gather

1,11,13

GatherElements

11,13

GatherND

11,12,13

Gelu

20

Value of approximate attribute should be none or tanh

Gemm

1,6,7,9,11,13

Gemm is only supported in the case where it is used to emulate FC; transA=0, transB=1, broadcast=1.

GlobalAveragePool

1,22

Only Supports 3D,4D & 5D inputs

GlobalMaxPool

1,22

Only Supports 3D,4D & 5D inputs

Greater

1,7,9,13

GreaterOrEqual

12

GridSample

16,20,22

Only Supports 4D & 5D inputs

GroupNormalization

18,21

GRU

1,7,14,22

Custom activations, user-defined activations and clip are not supported.

HardSigmoid

1,6

HardSwish

14,22

Identity

1,13,14,16

InstanceNormalization

1,6,22

If

1,11,13,16,19,21,23

THEN and ELSE branches can produce more than one output but the number of outputs produced by the THEN branch must be equal to the number of outputs produced by the ELSE branch. The output shape and data type of the corresponding outputs of the THEN and ELSE branches must be equal (i.e., shape and data type of 1st output of THEN branch must be equal to those of the ELSE branch, and so on.

IsNaN

9, 13, 20

IsInf

10, 20

LayerNormalization

17

LRN

1,13

LeakyRelu

1,6

Less

1,7,9,13

LessOrEqual

12

Log

1,6,13

LogSoftmax

1,11

LpNormalization

1,22

Only support attribute p=2

LpPool

1,2,11,18,22

LSTM

1,7,14,22

Custom activations, user-defined activations, clip and input-forget parameters are not supported. Only default value(0) is supported for layout

Matmul

1,9,13

Max

1,6,8,12,13

MaxPool

1,8,10,11,12,22

Only support Dilations = 1. Only support single output, doesn’t support Indices output.

MaxRoiPool

1,22

Min

1,6,8,12,13

Mod

10,13

Mul

1,6,7,13,14

Neg

1,6,13

NonMaxSuppression

10,11

NonZero

9,13

Not

1

OneHot

9,11

Or

1,7

Pad

1,2,11,13,18,19,21

Constant, Reflect & Edge padding mode is supported.

Pow

1,7,12,13,15

PRelu

1,6,7,9

Slope must be static.

QLinearConv

10

Translate to Conv.

QLinearMatMul

10,21

Translate to MatMul.

QuantizeLinear

10,13,19,21

RandomUniformLike

1,22

Range

11

Reciprocal

1,6

Translate to Div.

Reducel2

1,11,13,18

ReduceLogSumExp

1,11,13,18

Translate to Log + ReduceSum + Exp

ReduceMax

1,11,12,13,18,20

ReduceMean

1,11,13,18

ReduceMin

1,11,12,13,18,20

ReduceProd

1,11,13,18

ReduceSum

1,11,13

ReduceSumSquare

1,11,13,18

Relu

1,6

Reshape

1,5,13,14

Only default value(0) is supported for allow_zero

Resize

10,11,13,18,19

Only 3D,4D & 5D inputs are supported.

RoiAlign

10,16,22

Only output_haf_pixel is supported for coordinate transformation mode

Round

11,22

RNN

1, 7

Custom and user-defined activations (activation_alpha and beta) are not supported.

Scatter

9,11

ScatterElements

11,13,16

ScatterND

11,13,16

Shape

1,13,15

Shape is only supported at conversion time. The input to Shape and all operators which manipulate the resulting output shape (eg slice, concat, etc) are also computed at conversion time.

Sign

13

Sigmoid

1,6,13

Sin

7,22

Slice

1,10,11,13

Size

1,13,15,19,21,23

Size is supported only at conversion time. The input to Size and all operators which manipulate the resulting output shape (eg slize, concat, etc) are also computed at conversion time.

Softmax

1,11,13

Softplus

1,22

Split

1,2,11,13,18

Squeeze

1,11,13

Sqrt

1,6,13

STFT

17

Sub

1,6,7,13,14

Sum

1,6,8,13

Tanh

1,6,13

Tile

1.6,13

TopK

1,10,11

Transpose

1,13

Unsqueeze

1,11,13

Upsample

1,7,9

Only bilinear and nearest neighbor modes are supported.

FC

1

Both “axis” and “axis_w” must be 1.

ScaledTanh

1,6

Scaledtanh is removed in ONNX release v1.5.0

ThresholdedRelu

10,22

SpaceToDepth

1,13

Where

9,16

Xor

1,7