Supported ONNX operators

When you import a model, each Open Neural Network Exchange (ONNX) operator in the model graph becomes a Sentis layer. A Sentis layer has the same name as the ONNX operator, unless the table shows the operator maps to a different layer. Refer to How Sentis optimizes a model for more information.

Supported ONNX operators

The table below shows which ONNX operators Sentis supports, and which data types Sentis supports for each backend type.

Name Supported data types with BackendType.CPU Supported data types with BackendType.GPUCompute Supported data types with BackendType.GPUPixel Notes
Abs float, int float, int float, int
Acos float float float
Acosh float float float
Add float, int float, int float, int
And int int int
ArgMax float, int float, int float, int
ArgMin float, int float, int float, int
Asin float float float
Asinh float float float
Atan float float float
Atanh float float float
AveragePool float float (1D and 2D only) float (1D and 2D only) The ceil_mode and count_include_pad parameters aren't supported.
BatchNormalization float float float The momentum, spatial and training_mode parameters aren't supported.
Bernoulli float, int float, int float, int
Cast float, int, short float, int, short float, int, short
CastLike float, int, short float, int, short float, int, short
Ceil float float float
Celu float float float
Clip float, int float, int float, int
Compress float, int Not supported Not supported
Concat float, int float, int float, int
Constant - - - The sparse_value parameter isn't supported.
ConstantOfShape float, int float, int float, int
Conv float float (1D, 2D and 3D)* float (1D, 2D and 3D)
ConvTranspose float float (1D, 2D and 3D) float (1D, 2D and 3D) The dilations, group and output_shape parameters aren't supported.
Cos float float float
Cosh float float float
CumSum float, int float, int float, int
DepthToSpace float float float
Div float, int float, int float, int
Dropout - - - The operator maps to the Sentis layer Identity
Einsum float float (1 or 2 inputs only) Not supported
Elu float float float
Equal float, int float, int float, int
Erf float float float
Exp float float float
Expand float, int float, int float, int
Flatten float, int float, int float, int
Floor float float float
Gather float, int float, int float, int
GatherElements float, int float, int float, int
GatherND float, int float, int float, int
Gemm float float float
GlobalAveragePool float float float
GlobalMaxPool float float float
Greater float, int float, int float, int
GreaterOrEqual float, int float, int float, int
GridSample float float float
Hardmax float float float
HardSigmoid float float float
HardSwish float float float
Identity float, int float, int float, int
InstanceNormalization float float float
IsInf float float float (Infs not supported)
IsNaN float float float (NaNs not supported)
LayerNormalization float float float
LeakyRelu float float float
Less float, int float, int float, int
LessOrEqual float, int float, int float, int
Log float float float
LogSoftmax float float float
LRN float Not supported Not supported
LSTM float float Not supported
MatMul float float* float
Max float, int float, int float, int
MaxPool float float (1D and 2D only) float (1D and 2D only) The ceil_mode, dilations and storage_order parameters aren't supported.
Mean float float float The operator maps to the Sentis layers Add and ScalarMad.
Min float, int float, int float, int
Mod float, int float, int float, int
Mul float, int float, int float, int
Multinomial float Not supported Not supported
Neg float, int float, int float, int
NonMaxSuppression float float Not supported
NonZero float, int Not supported Not supported
Not int int int
OneHot float, int float, int float, int
Or int int int
Pad float, int float, int float, int
Pow float, int float, int float, int
PRelu float float float
RandomNormal float float float
RandomNormalLike float float float
RandomUniform float float float
RandomUniformLike float float float
Range float, int float, int float, int
Reciprocal float float float
ReduceL1 float, int float*, int* float, int
ReduceL2 float float* float
ReduceLogSum float float* float
ReduceLogSumExp float float float
ReduceMax float, int float*, int* float, int
ReduceMean float float* float
ReduceMin float, int float*, int* float, int
ReduceProd float, int float*, int* float, int
ReduceSum float, int float*, int* float, int
ReduceSumSquare float, int float*, int* float, int
Relu float float float
Reshape float, int float, int float, int
Resize float float float The cubic_coeff_a, exclude_outside, extrapolation_value and roi parameters aren't supported.
RoiAlign float float float The coordinate_transformation_mode parameter isn't supported.
Round float float float
Scatter (deprecated) float, int float, int float, int The operator maps to the Sentis layer ScatterElements.
ScatterElements float, int float, int (no ScatterReductionMode) float, int
ScatterND float, int float, int float, int
Selu float float float
Shape - - - The operator returns a CPU tensor without downloading the input tensor.
Shrink float float float
Sigmoid float float float
Sign float, int float, int float, int
Sin float float float
Sinh float float float
Size - - - The operator returns a CPU tensor without downloading the input tensor.
Slice float, int float, int float, int
Softmax float float float
Softplus float float float
Softsign float float float
SpaceToDepth float float float
Split float, int float, int float, int
Sqrt float float float
Squeeze float, int float, int float, int
Sub float, int float, int float, int
Sum float, int float, int float, int The operator maps to the Sentis layer Add.
Tan float float float
Tanh float float float
ThresholdedRelu float float float
Tile float, int float, int float, int
TopK float float Not supported
Transpose float, int float, int float, int
Trilu float, int float, int float, int
Unsqueeze float, int float, int float, int
Upsample (deprecated) float float float The operator maps to the Sentis layer Resize.
Where float, int float, int float, int
Xor int int int

* Sentis uses DirectML to accelerate these operators on supported hardware.

Sentis-only layers

Sentis might create the following layers when it optimizes the model.

Name Supported data types with BackendType.CPU Supported data types with BackendType.GPUCompute Supported data types with BackendType.GPUPixel
Dense float float* float
DequantizeUint8 byte byte byte
Gelu float float float
GeluFast float float float
MatMul2D float float* float
MoveDim float, int float, int float, int
Narrow float, int float, int float, int
RandomChoice float, int float, int float, int
Relu6 float float float
RMSNormalization float float float
ScalarMad float, int float, int float, int
Select float, int float, int float, int
SliceSet float, int float, int float, int
Square float, int float, int float, int
Swish float float float
ScaleBias float float float

* Sentis uses DirectML to accelerate these operators on supported hardware.

Unsupported operators

The following ONNX operators aren't supported in the current version of Sentis.

  • BitShift
  • ConcatFromSequence
  • ConvInteger
  • DequantizeLinear
  • Det
  • DynamicQuantizeLinear
  • EyeLike
  • If
  • GRU
  • Loop
  • LpPool
  • MatMulInteger
  • MaxUnpool
  • MeanVarianceNormalization
  • NegativeLogLikelihoodLoss
  • Optional
  • OptionalGetElement
  • OptionalHasElement
  • QLinearConv
  • QLinearMatMul
  • QuantizeLinear
  • ReverseSequence
  • RNN
  • Scan
  • SequenceAt
  • SequenceConstruct
  • SequenceEmpty
  • SequenceErase
  • SequenceInsert
  • SequenceLength
  • SoftmaxCrossEntropyLoss
  • SplitToSequence
  • StringNormalizer
  • TfIdfVectorizer
  • Unique

Additional resources