Class GPUCommandBufferOps

Reference GPU compute IOps implementation

Inheritance
Object
CPUOps
GPUCommandBufferOps
Inherited Members
Namespace: Unity.Sentis
Syntax
public class GPUCommandBufferOps : CPUOps, IOps, IDisposable

Constructors

GPUCommandBufferOps(ITensorAllocator)

Create GPUCommandBufferOps

Declaration
public GPUCommandBufferOps(ITensorAllocator allocator = null)
Parameters
Type Name Description
ITensorAllocator allocator

allocator

GPUCommandBufferOps(CommandBuffer, ITensorAllocator)

Create GPUCommandBufferOps

Declaration
public GPUCommandBufferOps(CommandBuffer cb, ITensorAllocator allocator = null)
Parameters
Type Name Description
CommandBuffer cb

commandbuffer
ITensorAllocator allocator

allocator

Fields

cb

Declaration
public CommandBuffer cb
Field Value
Type Description
CommandBuffer

Properties

deviceType

Returns the DeviceType for the ops.

Declaration
public override DeviceType deviceType { get; }
Property Value
Type Description
DeviceType
Overrides

Methods

Abs(TensorFloat)

Computes an output tensor by applying the element-wise Abs math function: f(x) = f(x) = |x|.

Declaration
public override TensorFloat Abs(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Abs(TensorInt)

Computes an output tensor by applying the element-wise Abs math function: f(x) = f(x) = |x|.

Declaration
public override TensorInt Abs(TensorInt X)
Parameters
Type Name Description
TensorInt X
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Acos(TensorFloat)

Computes an output tensor by applying the element-wise Acos trigonometric function: f(x) = acos(x).

Declaration
public override TensorFloat Acos(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Acosh(TensorFloat)

Computes an output tensor by applying the element-wise Acosh trigonometric function: f(x) = acosh(x).

Declaration
public override TensorFloat Acosh(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Add(TensorFloat, TensorFloat)

Performs an element-wise Add math operation: f(a, b) = a + b.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorFloat Add(TensorFloat A, TensorFloat B)
Parameters
Type Name Description
TensorFloat A

The first input tensor.
TensorFloat B

The second input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Add(TensorInt, TensorInt)

Performs an element-wise Add math operation: f(a, b) = a + b.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Add(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

And(TensorInt, TensorInt)

Declaration
public override TensorInt And(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A
TensorInt B
Returns
Type Description
TensorInt
Overrides

ArgMax(TensorFloat, Int32, Boolean, Boolean)

Computes the indices of the maximum elements of the input tensor along a given axis.

Declaration
public override TensorInt ArgMax(TensorFloat X, int axis, bool keepdim, bool selectLastIndex = false)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32 axis

The axis along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Boolean selectLastIndex

Whether to perform the operation from the back of the axis.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

ArgMax(TensorInt, Int32, Boolean, Boolean)

Computes the indices of the maximum elements of the input tensor along a given axis.

Declaration
public override TensorInt ArgMax(TensorInt X, int axis, bool keepdim, bool selectLastIndex = false)
Parameters
Type Name Description
TensorInt X

The input tensor.
Int32 axis

The axis along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Boolean selectLastIndex

Whether to perform the operation from the back of the axis.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

ArgMin(TensorFloat, Int32, Boolean, Boolean)

Computes the indices of the minimum elements of the input tensor along a given axis.

Declaration
public override TensorInt ArgMin(TensorFloat X, int axis, bool keepdim, bool selectLastIndex = false)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32 axis

The axis along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Boolean selectLastIndex

Whether to perform the operation from the back of the axis.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

ArgMin(TensorInt, Int32, Boolean, Boolean)

Computes the indices of the minimum elements of the input tensor along a given axis.

Declaration
public override TensorInt ArgMin(TensorInt X, int axis, bool keepdim, bool selectLastIndex = false)
Parameters
Type Name Description
TensorInt X

The input tensor.
Int32 axis

The axis along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Boolean selectLastIndex

Whether to perform the operation from the back of the axis.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Asin(TensorFloat)

Computes an output tensor by applying the element-wise Asin trigonometric function: f(x) = asin(x).

Declaration
public override TensorFloat Asin(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Asinh(TensorFloat)

Computes an output tensor by applying the element-wise Asinh trigonometric function: f(x) = asinh(x).

Declaration
public override TensorFloat Asinh(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Atan(TensorFloat)

Computes an output tensor by applying the element-wise Atan trigonometric function: f(x) = atan(x).

Declaration
public override TensorFloat Atan(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Atanh(TensorFloat)

Computes an output tensor by applying the element-wise Atanh trigonometric function: f(x) = atanh(x).

Declaration
public override TensorFloat Atanh(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

AveragePool(TensorFloat, Int32[], Int32[], Int32[])

Calculates an output tensor by pooling the mean values of the input tensor across its spatial dimensions according to the given pool and stride values.

Declaration
public override TensorFloat AveragePool(TensorFloat X, int[] pool, int[] stride, int[] pad)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] pool

The size of the kernel along each spatial axis.
Int32[] stride

The stride along each spatial axis.
Int32[] pad

The lower and upper padding values for each spatial dimension. For example, [pad_left, pad_right] for 1D, or [pad_top, pad_bottom, pad_left, pad_right] for 2D.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

AxisNormalization(TensorFloat, TensorFloat, TensorFloat, Single)

Declaration
public override TensorFloat AxisNormalization(TensorFloat X, TensorFloat S, TensorFloat B, float epsilon)
Parameters
Type Name Description
TensorFloat X
TensorFloat S
TensorFloat B
Single epsilon
Returns
Type Description
TensorFloat
Overrides

Bernoulli(TensorFloat, DataType, Nullable<Single>)

Generates an output tensor with values 0 or 1 from a Bernoulli distribution. The input tensor contains the probabilities to use for generating the output values.

Declaration
public override Tensor Bernoulli(TensorFloat X, DataType dataType, float? seed)
Parameters
Type Name Description
TensorFloat X
DataType dataType

The data type of the output tensor.
Nullable<Single> seed

The optional seed to use for the random number generation. If this is null the operation generates a seed using System.Random().
Returns
Type Description
Tensor

The computed output tensor.
Overrides

Cast(Tensor, DataType)

Computes the output tensor using an element-wise Cast function: f(x) = (float)x or f(x) = (int)x depending on the value of toType.

Declaration
public override Tensor Cast(Tensor X, DataType toType)
Parameters
Type Name Description
Tensor X
DataType toType

The data type to cast to as a DataType.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

Ceil(TensorFloat)

Computes an output tensor by applying the element-wise Ceil math function: f(x) = ceil(x).

Declaration
public override TensorFloat Ceil(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Celu(TensorFloat, Single)

Computes an output tensor by applying the element-wise Celu activation function: f(x) = max(0, x) + min(0, alpha * (exp(x / alpha) - 1)).

Declaration
public override TensorFloat Celu(TensorFloat X, float alpha)
Parameters
Type Name Description
TensorFloat X
Single alpha

The alpha value to use for the Celu activation function.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Clip(TensorFloat, Single, Single)

Computes an output tensor by applying the element-wise Clip math function: f(x) = clamp(x, min, max).

Declaration
public override TensorFloat Clip(TensorFloat X, float min, float max)
Parameters
Type Name Description
TensorFloat X
Single min

The lower clip value.
Single max

The upper clip value.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

CompressWithIndices(Tensor, TensorInt, Int32, Int32)

Declaration
protected override Tensor CompressWithIndices(Tensor X, TensorInt indices, int numIndices, int axis)
Parameters
Type Name Description
Tensor X
TensorInt indices
Int32 numIndices
Int32 axis
Returns
Type Description
Tensor
Overrides

Concat(Tensor[], Int32)

Calculates an output tensor by concatenating the input tensors along a given axis.

Declaration
public override Tensor Concat(Tensor[] tensors, int axis)
Parameters
Type Name Description
Tensor[] tensors

The input tensors.
Int32 axis

The axis along which to concatenate the input tensors.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

ConstantOfShape(TensorShape, Int32)

Generates a tensor with a given shape filled with a given value.

Declaration
public override TensorInt ConstantOfShape(TensorShape X, int value)
Parameters
Type Name Description
TensorShape X

The input tensor shape.
Int32 value

The fill value.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

ConstantOfShape(TensorShape, Single)

Generates a tensor with a given shape filled with a given value.

Declaration
public override TensorFloat ConstantOfShape(TensorShape X, float value)
Parameters
Type Name Description
TensorShape X

The input tensor shape.
Single value

The fill value.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Conv(TensorFloat, TensorFloat, TensorFloat, Int32, Int32[], Int32[], Int32[], FusableActivation)

Applies a convolution filter to an input tensor.

Declaration
public override TensorFloat Conv(TensorFloat X, TensorFloat K, TensorFloat B, int groups, int[] stride, int[] pad, int[] dilation, FusableActivation fusedActivation)
Parameters
Type Name Description
TensorFloat X

The input tensor.
TensorFloat K

The filter tensor.
TensorFloat B

The optional bias tensor.
Int32 groups

The number of groups that input channels and output channels are divided into.
Int32[] stride

The optional stride value for each spatial dimension of the filter.
Int32[] pad

The optional lower and upper padding values for each spatial dimension of the filter.
Int32[] dilation

The optional dilation value of each spatial dimension of the filter.
FusableActivation fusedActivation

The fused activation type to apply after the convolution.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Conv2DTrans(TensorFloat, TensorFloat, TensorFloat, Int32[], Int32[], Int32[], FusableActivation)

Applies a transpose convolution filter to an input tensor.

Declaration
public override TensorFloat Conv2DTrans(TensorFloat X, TensorFloat K, TensorFloat B, int[] stride, int[] pad, int[] outputAdjustment, FusableActivation fusedActivation)
Parameters
Type Name Description
TensorFloat X

The input tensor.
TensorFloat K

The filter tensor.
TensorFloat B

The optional bias tensor.
Int32[] stride

The optional stride value for each spatial dimension of the filter.
Int32[] pad

The optional lower and upper padding values for each spatial dimension of the filter.
Int32[] outputAdjustment

The output padding value for each spatial dimension in the filter.
FusableActivation fusedActivation

The fused activation type to apply after the convolution.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Cos(TensorFloat)

Computes an output tensor by applying the element-wise Cos trigonometric function: f(x) = cos(x).

Declaration
public override TensorFloat Cos(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Cosh(TensorFloat)

Computes an output tensor by applying the element-wise Cosh trigonometric function: f(x) = cosh(x).

Declaration
public override TensorFloat Cosh(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

CumSum(TensorFloat, Int32, Boolean, Boolean)

Performs the cumulative sum along a given axis.

Declaration
public override TensorFloat CumSum(TensorFloat X, int axis, bool reverse, bool exclusive)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32 axis

The axis along which to apply the cumulative sum.
Boolean reverse

Whether to perform the cumulative sum from the end of the axis.
Boolean exclusive

Whether to include the respective input element in the cumulative sum.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

CumSum(TensorInt, Int32, Boolean, Boolean)

Performs the cumulative sum along a given axis.

Declaration
public override TensorInt CumSum(TensorInt X, int axis, bool reverse, bool exclusive)
Parameters
Type Name Description
TensorInt X

The input tensor.
Int32 axis

The axis along which to apply the cumulative sum.
Boolean reverse

Whether to perform the cumulative sum from the end of the axis.
Boolean exclusive

Whether to include the respective input element in the cumulative sum.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Dense(TensorFloat, TensorFloat, TensorFloat, FusableActivation)

Performs a matrix multiplication operation: f(x, w, b) = X x W + B.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorFloat Dense(TensorFloat X, TensorFloat W, TensorFloat B, FusableActivation fusedActivation)
Parameters
Type Name Description
TensorFloat X

The input tensor.
TensorFloat W

The weights tensor.
TensorFloat B

The bias tensor.
FusableActivation fusedActivation

The fused activation to apply to the output tensor after the dense operation.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

DepthToSpace(TensorFloat, Int32, DepthToSpaceMode)

Computes the output tensor by permuting data from depth into blocks of spatial data.

Declaration
public override TensorFloat DepthToSpace(TensorFloat X, int blocksize, DepthToSpaceMode mode)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32 blocksize

The size of the blocks to move the depth data into.
DepthToSpaceMode mode

The ordering of the data in the output tensor as a DepthToSpaceMode.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Div(TensorFloat, TensorFloat)

Performs an element-wise Div math operation: f(a, b) = a / b.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorFloat Div(TensorFloat A, TensorFloat B)
Parameters
Type Name Description
TensorFloat A

The first input tensor.
TensorFloat B

The second input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Div(TensorInt, TensorInt)

Performs an element-wise Div math operation: f(a, b) = a / b.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Div(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Einsum(String, TensorFloat[])

Performs an Einsum math operation.

Declaration
public override TensorFloat Einsum(string equation, params TensorFloat[] operands)
Parameters
Type Name Description
String equation

The equation of the Einstein summation as a comma-separated list of subscript labels.
TensorFloat[] operands

The input tensors of the Einsum.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Elu(TensorFloat, Single)

Computes an output tensor by applying the element-wise Elu activation function: f(x) = x if x >= 0, otherwise f(x) = alpha * (e^x - 1).

Declaration
public override TensorFloat Elu(TensorFloat X, float alpha)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Single alpha

The alpha value to use for the Elu activation function.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Equal(TensorFloat, TensorFloat)

Performs an element-wise Equal logical comparison operation: f(a, b) = 1 if a == b, otherwise f(x) = 0.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Equal(TensorFloat A, TensorFloat B)
Parameters
Type Name Description
TensorFloat A

The first input tensor.
TensorFloat B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Equal(TensorInt, TensorInt)

Performs an element-wise Equal logical comparison operation: f(a, b) = 1 if a == b, otherwise f(x) = 0.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Equal(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Erf(TensorFloat)

Computes an output tensor by applying the element-wise Erf activation function: f(x) = erf(x).

Declaration
public override TensorFloat Erf(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Exp(TensorFloat)

Computes an output tensor by applying the element-wise Exp math function: f(x) = exp(x).

Declaration
public override TensorFloat Exp(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Expand(Tensor, TensorShape)

Calculates an output tensor by broadcasting the input tensor into a given shape.

Declaration
public override Tensor Expand(Tensor X, TensorShape newShape)
Parameters
Type Name Description
Tensor X

The input tensor.
TensorShape newShape
Returns
Type Description
Tensor

The computed output tensor.
Overrides

Floor(TensorFloat)

Computes an output tensor by applying the element-wise Floor math function: f(x) = floor(x).

Declaration
public override TensorFloat Floor(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

FMod(TensorFloat, TensorFloat)

Performs an element-wise Mod math operation: f(a, b) = a % b.

The sign of the remainder is the same as the sign of the dividend, as in C#.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorFloat FMod(TensorFloat A, TensorFloat B)
Parameters
Type Name Description
TensorFloat A

The first input tensor.
TensorFloat B

The second input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

FMod(TensorInt, TensorInt)

Performs an element-wise Mod math operation: f(a, b) = a % b.

The sign of the remainder is the same as the sign of the dividend, as in C#.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt FMod(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Gather(Tensor, TensorInt, Int32)

Takes values from the input tensor indexed by the indices tensor along a given axis and concatenates them.

Declaration
public override Tensor Gather(Tensor X, TensorInt indices, int axis)
Parameters
Type Name Description
Tensor X

The input tensor.
TensorInt indices

The indices tensor.
Int32 axis

The axis along which to gather.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

GatherElements(Tensor, TensorInt, Int32)

Takes values from the input tensor indexed by the indices tensor along a given axis and concatenates them.

Declaration
public override Tensor GatherElements(Tensor X, TensorInt indices, int axis)
Parameters
Type Name Description
Tensor X

The input tensor.
TensorInt indices

The indices tensor.
Int32 axis

The axis along which to gather.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

GatherND(Tensor, TensorInt, Int32)

Takes slices of values from the batched input tensor indexed by the indices tensor.

Declaration
public override Tensor GatherND(Tensor X, TensorInt indices, int batchDims)
Parameters
Type Name Description
Tensor X

The input tensor.
TensorInt indices

The indices tensor.
Int32 batchDims

The number of batch dimensions of the input tensor, the gather begins at the next dimension.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

Gelu(TensorFloat)

Computes an output tensor by applying the element-wise Gelu activation function: f(x) = x / 2 * (1 + erf(x / sqrt(2))).

Declaration
public override TensorFloat Gelu(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

GlobalAveragePool(TensorFloat)

Calculates an output tensor by pooling the mean values of the input tensor across all of its spatial dimensions. The spatial dimensions of the output are size 1.

Declaration
public override TensorFloat GlobalAveragePool(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

GlobalAverageVariancePool(TensorFloat, TensorFloat, Int32)

Declaration
public override void GlobalAverageVariancePool(TensorFloat O, TensorFloat X, int axis)
Parameters
Type Name Description
TensorFloat O
TensorFloat X
Int32 axis
Overrides

GlobalMaxPool(TensorFloat)

Calculates an output tensor by pooling the maximum values of the input tensor across all of its spatial dimensions. The spatial dimensions of the output are size 1.

Declaration
public override TensorFloat GlobalMaxPool(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Greater(TensorFloat, TensorFloat)

Performs an element-wise Greater logical comparison operation: f(a, b) = 1 if a > b, otherwise f(x) = 0.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Greater(TensorFloat A, TensorFloat B)
Parameters
Type Name Description
TensorFloat A

The first input tensor.
TensorFloat B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Greater(TensorInt, TensorInt)

Performs an element-wise Greater logical comparison operation: f(a, b) = 1 if a > b, otherwise f(x) = 0.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Greater(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

GreaterOrEqual(TensorFloat, TensorFloat)

Performs an element-wise GreaterOrEqual logical comparison operation: f(a, b) = 1 if a >= b, otherwise f(x) = 0.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt GreaterOrEqual(TensorFloat A, TensorFloat B)
Parameters
Type Name Description
TensorFloat A

The first input tensor.
TensorFloat B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

GreaterOrEqual(TensorInt, TensorInt)

Performs an element-wise GreaterOrEqual logical comparison operation: f(a, b) = 1 if a >= b, otherwise f(x) = 0.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt GreaterOrEqual(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Hardmax(TensorFloat, Int32)

Computes an output tensor by applying the Hardmax activation function along an axis: f(x, axis) = 1 if x is the first maximum value along the specified axis, otherwise f(x) = 0.

Declaration
public override TensorFloat Hardmax(TensorFloat X, int axis)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32 axis

The axis along which to apply the Hardmax activation function. The default value is -1.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

HardSigmoid(TensorFloat, Single, Single)

Computes an output tensor by applying the element-wise HardSigmoid activation function: f(x) = clamp(alpha * x + beta, 0, 1).

Declaration
public override TensorFloat HardSigmoid(TensorFloat X, float alpha, float beta)
Parameters
Type Name Description
TensorFloat X
Single alpha

The alpha value to use for the HardSigmoid activation function.
Single beta

The beta value to use for the HardSigmoid activation function.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

HardSwish(TensorFloat)

Computes an output tensor by applying the element-wise HardSwish activation function: f(x) = x * max(0, min(1, 1/6 * x + 0.5)).

Declaration
public override TensorFloat HardSwish(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

InstanceNormalization(TensorFloat, TensorFloat, TensorFloat, Single)

Computes the mean variance on the spatial dimensions of the input tensor and normalizes them according to scale and bias tensors.

Declaration
public override TensorFloat InstanceNormalization(TensorFloat X, TensorFloat S, TensorFloat B, float epsilon)
Parameters
Type Name Description
TensorFloat X

The input tensor.
TensorFloat S

The scale tensor.
TensorFloat B

The bias tensor.
Single epsilon

The epsilon value the layer uses to avoid division by zero.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

IsInf(TensorFloat, Boolean, Boolean)

Performs an element-wise IsInf logical operation: f(x) = 1 elementwise if x is +Inf and detectPositive is true, or x is -Inf and detectNegative is true. Otherwise f(x) = 0.

Declaration
public override TensorInt IsInf(TensorFloat X, bool detectNegative, bool detectPositive)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Boolean detectNegative

Whether to detect negative infinities in the IsInf function.
Boolean detectPositive

Whether to detect positive infinities in the IsInf function.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

IsNaN(TensorFloat)

Performs an element-wise IsNaN logical operation: f(x) = 1 if x is NaN, otherwise f(x) = 0.

Declaration
public override TensorInt IsNaN(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

LeakyRelu(TensorFloat, Single)

Computes an output tensor by applying the element-wise LeakyRelu activation function: f(x) = x if x >= 0, otherwise f(x) = alpha * x.

Declaration
public override TensorFloat LeakyRelu(TensorFloat X, float alpha)
Parameters
Type Name Description
TensorFloat X
Single alpha

The alpha value to use for the LeakyRelu activation function.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Less(TensorFloat, TensorFloat)

Performs an element-wise Less logical comparison operation: f(a, b) = 1 if a < b, otherwise f(x) = 0.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Less(TensorFloat A, TensorFloat B)
Parameters
Type Name Description
TensorFloat A

The first input tensor.
TensorFloat B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Less(TensorInt, TensorInt)

Performs an element-wise Less logical comparison operation: f(a, b) = 1 if a < b, otherwise f(x) = 0.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Less(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

LessOrEqual(TensorFloat, TensorFloat)

Performs an element-wise LessOrEqual logical comparison operation: f(a, b) = 1 if a <= b, otherwise f(x) = 0.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt LessOrEqual(TensorFloat A, TensorFloat B)
Parameters
Type Name Description
TensorFloat A

The first input tensor.
TensorFloat B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

LessOrEqual(TensorInt, TensorInt)

Performs an element-wise LessOrEqual logical comparison operation: f(a, b) = 1 if a <= b, otherwise f(x) = 0.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt LessOrEqual(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Log(TensorFloat)

Computes an output tensor by applying the element-wise Log math function: f(x) = log(x).

Declaration
public override TensorFloat Log(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

LogSoftmax(TensorFloat, Int32)

Computes an output tensor by applying the LogSoftmax activation function along an axis: f(x, axis) = log(Softmax(x, axis)).

Declaration
public override TensorFloat LogSoftmax(TensorFloat X, int axis)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32 axis

The axis along which to apply the LogSoftmax activation function. The default value is -1.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

MatMul(TensorFloat, TensorFloat)

Performs a multi-dimensional matrix multiplication operation: f(a, b) = a x b.

Declaration
public override TensorFloat MatMul(TensorFloat X, TensorFloat Y)
Parameters
Type Name Description
TensorFloat X

The first input tensor.
TensorFloat Y

The second input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

MatMul2D(TensorFloat, Boolean, TensorFloat, Boolean)

Performs a matrix multiplication operation with optional transposes: f(a, b) = a' x b'.

Declaration
public override TensorFloat MatMul2D(TensorFloat X, bool xTranspose, TensorFloat Y, bool yTranspose)
Parameters
Type Name Description
TensorFloat X

The first input tensor.
Boolean xTranspose

Whether to transpose the first input tensor before performing the matrix multiplication.
TensorFloat Y
Boolean yTranspose

Whether to transpose the second input tensor before performing the matrix multiplication.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Max(TensorFloat[])

Performs an element-wise Max math operation: f(x1, x2 ... xn) = max(x1, x2 ... xn).

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorFloat Max(TensorFloat[] tensors)
Parameters
Type Name Description
TensorFloat[] tensors

The input tensors.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Max(TensorInt[])

Performs an element-wise Max math operation: f(x1, x2 ... xn) = max(x1, x2 ... xn).

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Max(TensorInt[] tensors)
Parameters
Type Name Description
TensorInt[] tensors

The input tensors.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

MaxPool(TensorFloat, Int32[], Int32[], Int32[])

Calculates an output tensor by pooling the maximum values of the input tensor across its spatial dimensions according to the given pool and stride values.

Declaration
public override TensorFloat MaxPool(TensorFloat X, int[] pool, int[] stride, int[] pad)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] pool

The size of the kernel along each spatial axis.
Int32[] stride

The stride along each spatial axis.
Int32[] pad

The lower and upper padding values for each spatial dimension. For example, [pad_left, pad_right] for 1D, or [pad_top, pad_bottom, pad_left, pad_right] for 2D.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Mean(TensorFloat[])

Performs an element-wise Mean math operation: f(x1, x2 ... xn) = (x1 + x2 ... xn) / n.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorFloat Mean(TensorFloat[] tensors)
Parameters
Type Name Description
TensorFloat[] tensors

The input tensors.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

MemCopy(Tensor, Tensor, Int32, Int32, Int32)

Copy blocks of values from X to O, we copy 'count' blocks each of length 'length' values with initial offsets given by 'offsetX', 'offsetO' and with strides given by 'strideX', 'strideO'

Declaration
protected override void MemCopy(Tensor X, Tensor O, int length = -1, int offsetX = 0, int offsetO = 0)
Parameters
Type Name Description
Tensor X
Tensor O
Int32 length
Int32 offsetX
Int32 offsetO
Overrides

MemCopyStride(Tensor, Tensor, Int32, Int32, Int32, Int32, Int32, Int32)

Copy blocks of values from X to O, we copy 'count' blocks each of length 'length' values with initial offsets given by 'offsetX', 'offsetO' and with strides given by 'strideX', 'strideO'

Declaration
protected override void MemCopyStride(Tensor X, Tensor O, int strideX, int strideO, int length, int count, int offsetX = 0, int offsetO = 0)
Parameters
Type Name Description
Tensor X
Tensor O
Int32 strideX
Int32 strideO
Int32 length
Int32 count
Int32 offsetX
Int32 offsetO
Overrides

MemSet(Tensor, Int32, Int32, Int32)

Set values of O in indices [offset0, offsetO + length] to value

Declaration
protected override void MemSet(Tensor O, int value, int length = -1, int offsetO = 0)
Parameters
Type Name Description
Tensor O
Int32 value
Int32 length
Int32 offsetO
Overrides

Min(TensorFloat[])

Performs an element-wise Min math operation: f(x1, x2 ... xn) = min(x1, x2 ... xn).

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorFloat Min(TensorFloat[] tensors)
Parameters
Type Name Description
TensorFloat[] tensors

The input tensors.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Min(TensorInt[])

Performs an element-wise Min math operation: f(x1, x2 ... xn) = min(x1, x2 ... xn).

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Min(TensorInt[] tensors)
Parameters
Type Name Description
TensorInt[] tensors

The input tensors.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Mod(TensorInt, TensorInt)

Performs an element-wise Mod math operation: f(a, b) = a % b.

The sign of the remainder is the same as the sign of the divisor, as in Python.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Mod(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Mul(TensorFloat, TensorFloat)

Performs an element-wise Mul math operation: f(a, b) = a * b.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorFloat Mul(TensorFloat A, TensorFloat B)
Parameters
Type Name Description
TensorFloat A

The first input tensor.
TensorFloat B

The second input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Mul(TensorInt, TensorInt)

Performs an element-wise Mul math operation: f(a, b) = a * b.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Mul(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Neg(TensorFloat)

Computes an output tensor by applying the element-wise Neg math function: f(x) = -x.

Declaration
public override TensorFloat Neg(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Neg(TensorInt)

Computes an output tensor by applying the element-wise Neg math function: f(x) = -x.

Declaration
public override TensorInt Neg(TensorInt X)
Parameters
Type Name Description
TensorInt X

The input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

NewTensor(TensorShape, DataType, AllocScope)

Allocates a new tensor with the internal allocator.

Declaration
public override Tensor NewTensor(TensorShape shape, DataType dataType, AllocScope scope)
Parameters
Type Name Description
TensorShape shape
DataType dataType
AllocScope scope
Returns
Type Description
Tensor
Overrides

Not(TensorInt)

Performs an element-wise Not logical operation: f(x) = ~x.

Declaration
public override TensorInt Not(TensorInt X)
Parameters
Type Name Description
TensorInt X

The input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

OneHot(TensorInt, Int32, Int32, Int32, Int32)

Generates a one-hot tensor with a given depth, indices and on and off values.

Declaration
public override TensorInt OneHot(TensorInt X, int axis, int depth, int offValue, int onValue)
Parameters
Type Name Description
TensorInt X
Int32 axis

The axis along which the operation adds the one-hot representation.
Int32 depth

The depth of the one-hot tensor.
Int32 offValue

The value to use for an off element.
Int32 onValue

The value to use for an on element.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Or(TensorInt, TensorInt)

Performs an element-wise Or logical operation: f(a, b) = a | b.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Or(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Pad(TensorFloat, Int32[], PadMode, Single)

Calculates the output tensor by adding padding to the input tensor according to the given padding values and mode.

Declaration
public override TensorFloat Pad(TensorFloat X, int[] pad, PadMode padMode, float constant)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] pad

The lower and upper padding values for each dimension.
PadMode padMode

The PadMode to use when padding. The default value is PadMode.Constant.
Single constant

The constant value to fill with when using PadMode.Constant. The default value is 0.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

PinToDevice(Tensor, Boolean)

Prepares Tensor for use with CPU backend.

Declaration
public override Tensor PinToDevice(Tensor X, bool uploadCache = true)
Parameters
Type Name Description
Tensor X

Tensor to prepare for CPU backend.
Boolean uploadCache

Whether to copy tensor data to CPU backend.
Returns
Type Description
Tensor

Tensor once prepared for CPU backend.
Overrides

Pow(TensorFloat, TensorFloat)

Performs an element-wise Pow math operation: f(a, b) = pow(a, b).

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorFloat Pow(TensorFloat A, TensorFloat B)
Parameters
Type Name Description
TensorFloat A

The first input tensor.
TensorFloat B

The second input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Pow(TensorFloat, TensorInt)

Performs an element-wise Pow math operation: f(a, b) = pow(a, b).

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorFloat Pow(TensorFloat A, TensorInt B)
Parameters
Type Name Description
TensorFloat A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

PRelu(TensorFloat, TensorFloat)

Computes an output tensor by applying the element-wise PRelu activation function: f(x) = x if x >= 0, otherwise f(x) = slope * x.

Declaration
public override TensorFloat PRelu(TensorFloat X, TensorFloat S)
Parameters
Type Name Description
TensorFloat X
TensorFloat S
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

RandomNormal(TensorShape, Single, Single, Nullable<Single>)

Generates an output tensor of a given shape with random values in a normal distribution with given mean and scale, and an optional seed value.

Declaration
public override TensorFloat RandomNormal(TensorShape s, float mean, float scale, float? seed)
Parameters
Type Name Description
TensorShape s
Single mean

The mean of the normal distribution to use to generate the output.
Single scale

The standard deviation of the normal distribution to use to generate the output.
Nullable<Single> seed

The optional seed to use for the random number generation. If this is null the operation generates a seed using System.Random().
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

RandomUniform(TensorShape, Single, Single, Nullable<Single>)

Generates an output tensor of a given shape with random values in a uniform distribution between a given low and high, and an optional seed value.

Declaration
public override TensorFloat RandomUniform(TensorShape s, float low, float high, float? seed)
Parameters
Type Name Description
TensorShape s
Single low

The lower end of the interval of the uniform distribution to use to generate the output.
Single high

The upper end of the interval of the uniform distribution to use to generate the output.
Nullable<Single> seed

The optional seed to use for the random number generation. If this is null the operation generates a seed using System.Random().
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Range(Int32, Int32, Int32)

Generates a 1D output tensor where the values form an arithmetic progression defined by the start, limit, and delta values.

Declaration
public override TensorInt Range(int start, int limit, int delta)
Parameters
Type Name Description
Int32 start

The first value in the range.
Int32 limit

The limit of the range.
Int32 delta

The delta between subsequent values in the range.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Range(Single, Single, Single)

Generates a 1D output tensor where the values form an arithmetic progression defined by the start, limit, and delta values.

Declaration
public override TensorFloat Range(float start, float limit, float delta)
Parameters
Type Name Description
Single start

The first value in the range.
Single limit

The limit of the range.
Single delta

The delta between subsequent values in the range.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Reciprocal(TensorFloat)

Computes an output tensor by applying the element-wise Reciprocal math function: f(x) = 1 / x.

Declaration
public override TensorFloat Reciprocal(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ReduceL1(TensorFloat, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceL1 operation: f(x1, x2 ... xn) = |x1| + |x2| + ... + |xn|.

Declaration
public override TensorFloat ReduceL1(TensorFloat X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ReduceL1(TensorInt, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceL1 operation: f(x1, x2 ... xn) = |x1| + |x2| + ... + |xn|.

Declaration
public override TensorInt ReduceL1(TensorInt X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorInt X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

ReduceL2(TensorFloat, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceL2 operation: f(x1, x2 ... xn) = sqrt(x1² + x2² + ... + xn²).

Declaration
public override TensorFloat ReduceL2(TensorFloat X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ReduceLogSum(TensorFloat, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceLogSum operation: f(x1, x2 ... xn) = log(x1 + x2 + ... + xn).

Declaration
public override TensorFloat ReduceLogSum(TensorFloat X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ReduceLogSumExp(TensorFloat, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceLogSumExp operation: f(x1, x2 ... xn) = log(e^x1 + e^x2 + ... + e^xn).

Declaration
public override TensorFloat ReduceLogSumExp(TensorFloat X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ReduceMax(TensorFloat, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceMax operation: f(x1, x2 ... xn) = max(x1, x2, ... , xn).

Declaration
public override TensorFloat ReduceMax(TensorFloat X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ReduceMax(TensorInt, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceMean operation: f(x1, x2 ... xn) = max(x1, x2, ... , xn).

Declaration
public override TensorInt ReduceMax(TensorInt X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorInt X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

ReduceMean(TensorFloat, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceMean operation: f(x1, x2 ... xn) = (x1 + x2 + ... + xn) / n.

Declaration
public override TensorFloat ReduceMean(TensorFloat X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ReduceMin(TensorFloat, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceMin operation: f(x1, x2 ... xn) = min(x1, x2, ... , xn).

Declaration
public override TensorFloat ReduceMin(TensorFloat X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ReduceMin(TensorInt, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceMin operation: f(x1, x2 ... xn) = min(x1, x2, ... , xn).

Declaration
public override TensorInt ReduceMin(TensorInt X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorInt X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

ReduceProd(TensorFloat, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceProd operation: f(x1, x2 ... xn) = x1 * x2 * ... * xn.

Declaration
public override TensorFloat ReduceProd(TensorFloat X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ReduceProd(TensorInt, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceProd operation: f(x1, x2 ... xn) = x1 * x2 * ... * xn.

Declaration
public override TensorInt ReduceProd(TensorInt X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorInt X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

ReduceSum(TensorFloat, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceSum operation: f(x1, x2 ... xn) = x1 + x2 + ... + xn.

Declaration
public override TensorFloat ReduceSum(TensorFloat X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ReduceSum(TensorInt, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceSum operation: f(x1, x2 ... xn) = x1 + x2 + ... + xn.

Declaration
public override TensorInt ReduceSum(TensorInt X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorInt X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

ReduceSumSquare(TensorFloat, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceSumSquare operation: f(x1, x2 ... xn) = x1² + x2² + ... + xn².

Declaration
public override TensorFloat ReduceSumSquare(TensorFloat X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ReduceSumSquare(TensorInt, Int32[], Boolean)

Reduces an input tensor along the given axes using the ReduceSumSquare operation: f(x1, x2 ... xn) = x1² + x2² + ... + xn².

Declaration
public override TensorInt ReduceSumSquare(TensorInt X, int[] axes, bool keepdim)
Parameters
Type Name Description
TensorInt X

The input tensor.
Int32[] axes

The axes along which to reduce.
Boolean keepdim

Whether to keep the reduced axes in the output tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Relu(TensorFloat)

Computes an output tensor by applying the element-wise Relu activation function: f(x) = max(0, x).

Declaration
public override TensorFloat Relu(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Relu6(TensorFloat)

Computes an output tensor by applying the element-wise Relu6 activation function: f(x) = clamp(x, 0, 6).

Declaration
public override TensorFloat Relu6(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Resize(TensorFloat, Single[], InterpolationMode, NearestMode, CoordTransformMode)

Calculates an output tensor by resampling the input tensor along the spatial dimensions with given scales.

Declaration
public override TensorFloat Resize(TensorFloat X, float[] scale, InterpolationMode interpolationMode, NearestMode nearestMode = NearestMode.RoundPreferFloor, CoordTransformMode coordTransformMode = CoordTransformMode.HalfPixel)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Single[] scale

The factor to scale each dimension by.
InterpolationMode interpolationMode

The InterpolationMode to use for the operation.
NearestMode nearestMode

The NearestMode to use for the operation when using InterpolationMode.NearestMode. The default is NearestMode.RoundPreferFloor.
CoordTransformMode coordTransformMode

The CoordTransformMode to use for the operation. The default is CoordTransformMode.HalfPixel.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

RoiAlign(TensorFloat, TensorFloat, TensorInt, RoiPoolingMode, Int32, Int32, Int32, Single)

Calculates an output tensor by pooling the input tensor across each region of interest given by the rois tensor.

Declaration
public override TensorFloat RoiAlign(TensorFloat X, TensorFloat Rois, TensorInt Indices, RoiPoolingMode mode, int outputHeight, int outputWidth, int samplingRatio, float spatialScale)
Parameters
Type Name Description
TensorFloat X

The input tensor.
TensorFloat Rois

The region of interest input tensor.
TensorInt Indices

The indices input tensor.
RoiPoolingMode mode

The pooling mode of the operation as an RoiPoolingMode.
Int32 outputHeight

The height of the output tensor.
Int32 outputWidth

The width of the output tensor.
Int32 samplingRatio

The number of sampling points in the interpolation grid used to compute the output value of each pooled output bin.
Single spatialScale

The multiplicative spatial scale factor used to translate coordinates from their input spatial scale to the scale used when pooling.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Round(TensorFloat)

Computes an output tensor by applying the element-wise Round math function: f(x) = round(x).

If the fractional part is equal to 0.5, rounds to the nearest even integer.

Declaration
public override TensorFloat Round(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ScaleBias(TensorFloat, TensorFloat, TensorFloat)

Computes the output tensor with an element-wise ScaleBias function: f(x, s, b) = x * s + b.

Declaration
public override TensorFloat ScaleBias(TensorFloat X, TensorFloat S, TensorFloat B)
Parameters
Type Name Description
TensorFloat X

The input tensor.
TensorFloat S

The scale tensor.
TensorFloat B

The bias tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ScatterElements(Tensor, TensorInt, Tensor, Int32, ScatterReductionMode)

Copies the input tensor and updates values at indexes specified by the indices tensor with values specified by the updates tensor along a given axis.

ScatterElements updates the values depending on the reduction mode used.

Declaration
public override Tensor ScatterElements(Tensor X, TensorInt indices, Tensor updates, int axis, ScatterReductionMode reduction)
Parameters
Type Name Description
Tensor X

The input tensor.
TensorInt indices

The indices tensor.
Tensor updates

The updates tensor.
Int32 axis

The axis on which to perform the scatter.
ScatterReductionMode reduction

The reduction mode used to update the values as a ScatterReductionMode.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

ScatterND(TensorFloat, TensorInt, TensorFloat, ScatterReductionMode)

Copies the input tensor and updates values at indexes specified by the indices tensor with values specified by the updates tensor.

ScatterND updates the values depending on the reduction mode used.

Declaration
public override TensorFloat ScatterND(TensorFloat X, TensorInt indices, TensorFloat updates, ScatterReductionMode reduction)
Parameters
Type Name Description
TensorFloat X

The input tensor.
TensorInt indices

The indices tensor.
TensorFloat updates

The updates tensor.
ScatterReductionMode reduction

The reduction mode used to update the values as a ScatterReductionMode.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ScatterND(TensorInt, TensorInt, TensorInt, ScatterReductionMode)

Copies the input tensor and updates values at indexes specified by the indices tensor with values specified by the updates tensor.

ScatterND updates the values depending on the reduction mode used.

Declaration
public override TensorInt ScatterND(TensorInt X, TensorInt indices, TensorInt updates, ScatterReductionMode reduction)
Parameters
Type Name Description
TensorInt X

The input tensor.
TensorInt indices

The indices tensor.
TensorInt updates

The updates tensor.
ScatterReductionMode reduction

The reduction mode used to update the values as a ScatterReductionMode.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Selu(TensorFloat, Single, Single)

Computes an output tensor by applying the element-wise Selu activation function: f(x) = gamma * x if x >= 0, otherwise f(x) = (alpha * e^x - alpha).

Declaration
public override TensorFloat Selu(TensorFloat X, float alpha, float gamma)
Parameters
Type Name Description
TensorFloat X
Single alpha

The alpha value to use for the Selu activation function.
Single gamma

The alpha value to use for the Selu activation function.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Shrink(TensorFloat, Single, Single)

Computes an output tensor by applying the element-wise Shrink activation function: f(x) = x + bias if x < lambd. f(x) = x - bias if x > lambd. Otherwise f(x) = 0.

Declaration
public override TensorFloat Shrink(TensorFloat X, float bias, float lambd)
Parameters
Type Name Description
TensorFloat X
Single bias

The bias value to use for the Shrink activation function.
Single lambd

The lambda value to use for the Shrink activation function.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Sigmoid(TensorFloat)

Computes an output tensor by applying the element-wise Sigmoid activation function: f(x) = 1/(1 + e^(-x)).

Declaration
public override TensorFloat Sigmoid(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Sign(TensorFloat)

Performs an element-wise Sign math operation: f(x) = 1 if x > 0. f(x) = -1 if x < 0. Otherwise f(x) = 0.

Declaration
public override TensorFloat Sign(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Sign(TensorInt)

Performs an element-wise Sign math operation: f(x) = 1 if x > 0. f(x) = -1 if x < 0. Otherwise f(x) = 0.

Declaration
public override TensorInt Sign(TensorInt X)
Parameters
Type Name Description
TensorInt X

The input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Sin(TensorFloat)

Computes an output tensor by applying the element-wise Sin trigonometric function: f(x) = sin(x).

Declaration
public override TensorFloat Sin(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

SinglePassLSTM(TensorFloat, TensorFloat, TensorFloat, TensorFloat, TensorInt, TensorFloat, TensorFloat, TensorFloat, TensorFloat, RnnActivation[], Single[], Single[], Boolean, Single, Boolean, Int32, RnnLayout)

Computes a single pass LSTM either forward or reverse dirIndex and layout are used to calculate where to index the various tensors in bidirectional and batch first layout passes X has given layout W, R are cropped to single direction P, B are full number of directions Y has given layout and full number of directions (matches output of Layer) Y_h, Y_c are SequenceFirst layout and cropped to single direction HtxRT and XsixWT are temp vectors of the correct dimension for the intermediate results of the matmuls activations, activationAlpha and activationBeta have full number of dimensions

Declaration
protected override void SinglePassLSTM(TensorFloat X, TensorFloat W, TensorFloat R, TensorFloat B, TensorInt sequenceLens, TensorFloat P, TensorFloat Y, TensorFloat Y_h, TensorFloat Y_c, RnnActivation[] activations, float[] activationAlpha, float[] activationBeta, bool inputForget, float clip, bool isReverse, int dirIndex, RnnLayout layout)
Parameters
Type Name Description
TensorFloat X
TensorFloat W
TensorFloat R
TensorFloat B
TensorInt sequenceLens
TensorFloat P
TensorFloat Y
TensorFloat Y_h
TensorFloat Y_c
RnnActivation[] activations
Single[] activationAlpha
Single[] activationBeta
Boolean inputForget
Single clip
Boolean isReverse
Int32 dirIndex
RnnLayout layout
Overrides

Sinh(TensorFloat)

Computes an output tensor by applying the element-wise Sinh trigonometric function: f(x) = sinh(x).

Declaration
public override TensorFloat Sinh(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Slice(Tensor, Int32[], Int32[], Int32[], Int32[])

Calculates an output tensor by slicing the input tensor along given axes with given starts, ends, and steps.

Declaration
public override Tensor Slice(Tensor X, int[] starts, int[] ends, int[] axes, int[] steps)
Parameters
Type Name Description
Tensor X

The input tensor.
Int32[] starts

The start index along each axis.
Int32[] ends

The end index along each axis.
Int32[] axes

The axes along which to slice. If this is null, the layer slices all axes.
Int32[] steps

The step values for slicing. If this is null, the layer uses step size 1 throughout.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

Softmax(TensorFloat, Int32)

Computes an output tensor by applying the Softmax activation function along an axis: f(x, axis) = exp(X) / ReduceSum(exp(X), axis).

Declaration
public override TensorFloat Softmax(TensorFloat X, int axis)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32 axis

The axis along which to apply the Softmax activation function. The default value is -1.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Softplus(TensorFloat)

Computes an output tensor by applying the element-wise Softplus activation function: f(x) = ln(e^x + 1).

Declaration
public override TensorFloat Softplus(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Softsign(TensorFloat)

Computes an output tensor by applying the element-wise Softsign activation function: f(x) = x/(|x| + 1).

Declaration
public override TensorFloat Softsign(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

SpaceToDepth(TensorFloat, Int32)

Computes the output tensor by permuting data from blocks of spatial data into depth.

Declaration
public override TensorFloat SpaceToDepth(TensorFloat X, int blocksize)
Parameters
Type Name Description
TensorFloat X
Int32 blocksize

The size of the blocks to move the depth data into.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Split(Tensor, Int32, Int32, Int32)

Calculates an output tensor by splitting the input tensor along a given axis between start and end.

Declaration
public override Tensor Split(Tensor X, int axis, int start, int end)
Parameters
Type Name Description
Tensor X

The input tensor.
Int32 axis

The axis along which to split the input tensor.
Int32 start

The inclusive start value for the split.
Int32 end

The exclusive end value for the split.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

Sqrt(TensorFloat)

Computes an output tensor by applying the element-wise Sqrt math function: f(x) = sqrt(x).

Declaration
public override TensorFloat Sqrt(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Square(TensorFloat)

Computes an output tensor by applying the element-wise Square math function: f(x) = x * x.

Declaration
public override TensorFloat Square(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Sub(TensorFloat, TensorFloat)

Performs an element-wise Sub math operation: f(a, b) = a - b.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorFloat Sub(TensorFloat A, TensorFloat B)
Parameters
Type Name Description
TensorFloat A

The first input tensor.
TensorFloat B

The second input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Sub(TensorInt, TensorInt)

Performs an element-wise Sub math operation: f(a, b) = a - b.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorInt Sub(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides

Sum(TensorFloat[])

Performs an element-wise Sum math operation: f(x1, x2 ... xn) = x1 + x2 ... xn.

This supports numpy-style broadcasting of input tensors.

Declaration
public override TensorFloat Sum(TensorFloat[] tensors)
Parameters
Type Name Description
TensorFloat[] tensors

The input tensors.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Swish(TensorFloat)

Computes an output tensor by applying the element-wise Swish activation function: f(x) = sigmoid(x) * x = x / (1 + e^{-x}).

Declaration
public override TensorFloat Swish(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Tan(TensorFloat)

Computes an output tensor by applying the element-wise Tan trigonometric function: f(x) = tan(x).

Declaration
public override TensorFloat Tan(TensorFloat X)
Parameters
Type Name Description
TensorFloat X
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Tanh(TensorFloat)

Computes an output tensor by applying the element-wise Tanh activation function: f(x) = tanh(x).

Declaration
public override TensorFloat Tanh(TensorFloat X)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

ThresholdedRelu(TensorFloat, Single)

Computes an output tensor by applying the element-wise ThresholdedRelu activation function: f(x) = x if x > alpha, otherwise f(x) = 0.

Declaration
public override TensorFloat ThresholdedRelu(TensorFloat X, float alpha)
Parameters
Type Name Description
TensorFloat X
Single alpha

The alpha value to use for the ThresholdedRelu activation function.
Returns
Type Description
TensorFloat

The computed output tensor.
Overrides

Tile(Tensor, Int32[])

Calculates an output tensor by repeating the input layer a given number of times along each axis.

Declaration
public override Tensor Tile(Tensor X, int[] repeats)
Parameters
Type Name Description
Tensor X

The input tensor.
Int32[] repeats

The number of times to tile the input tensor along each axis.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

TopK(TensorFloat, Int32, Int32, Boolean, Boolean)

Calculates the top-K largest or smallest elements of an input tensor along a given axis.

Declaration
public override Tensor[] TopK(TensorFloat X, int k, int axis, bool largest, bool sorted)
Parameters
Type Name Description
TensorFloat X

The input tensor.
Int32 k

The number of elements to calculate.
Int32 axis

The axis along which to perform the top-K operation.
Boolean largest

Whether to calculate the top-K largest elements. If this is false, the layer calculates the top-K smallest elements.
Boolean sorted

Whether to return the elements in sorted order.
Returns
Type Description
Tensor[]

The computed output tensor.
Overrides

Transpose(Tensor)

Calculates an output tensor by reversing the dimensions of the input tensor.

Declaration
public override Tensor Transpose(Tensor X)
Parameters
Type Name Description
Tensor X
Returns
Type Description
Tensor

The computed output tensor.
Overrides

Transpose(Tensor, Int32[])

Calculates an output tensor by permuting the axes and data of the input tensor according to the given permutations.

Declaration
public override Tensor Transpose(Tensor X, int[] permutations)
Parameters
Type Name Description
Tensor X
Int32[] permutations

The axes to sample the output tensor from in the input tensor.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

Tril(Tensor, Int32)

Computes the output tensor by retaining the lower triangular values from an input matrix or matrix batch and setting the other values to zero.

Declaration
public override Tensor Tril(Tensor X, int k)
Parameters
Type Name Description
Tensor X

The input tensor.
Int32 k

The offset from the diagonal to keep.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

Triu(Tensor, Int32)

Computes the output tensor by retaining the upper triangular values from an input matrix or matrix batch and setting the other values to zero.

Declaration
public override Tensor Triu(Tensor X, int k)
Parameters
Type Name Description
Tensor X

The input tensor.
Int32 k

The offset from the diagonal to exclude.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

Where(TensorInt, Tensor, Tensor)

Performs an element-wise Where logical operation: f(condition, a, b) = a if condition is true, otherwise f(condition, a, b) = b.

Declaration
public override Tensor Where(TensorInt C, Tensor A, Tensor B)
Parameters
Type Name Description
TensorInt C

The condition tensor.
Tensor A

The first input tensor.
Tensor B

The second input tensor.
Returns
Type Description
Tensor

The computed output tensor.
Overrides

Xor(TensorInt, TensorInt)

Performs an element-wise Xor logical operation: f(a) = a ^ b.

Declaration
public override TensorInt Xor(TensorInt A, TensorInt B)
Parameters
Type Name Description
TensorInt A

The first input tensor.
TensorInt B

The second input tensor.
Returns
Type Description
TensorInt

The computed output tensor.
Overrides