DML_RESAMPLE_GRAD1_OPERATOR_DESC structure (directml.h)

Computes backpropagation gradients for DML_RESAMPLE2_OPERATOR_DESC.

DML_RESAMPLE2_OPERATOR_DESC rescales arbitrary dimensions of the input tensor using either nearest-neighbor sampling or bilinear interpolation. Given an InputGradientTensor with the same sizes as the output of an equivalent DML_RESAMPLE2_OPERATOR_DESC, this operator produces an OutputGradientTensor with the same sizes as the input of the DML_RESAMPLE2_OPERATOR_DESC.

As an example, consider a DML_RESAMPLE2_OPERATOR_DESC that performs a nearest-neighbor scaling of 1.5x in the width, and 0.5x in the height:

InputTensor           OutputTensor
[[1, 2],   Resample    [1, 1, 2]
 [3, 4]]      -->      

Notice how the 0th element of the input tensor (with value 1) contributes to two elements in the output; the 1st element (with value 2) contributes to one element in the output; and the 2nd and 3rd elements (with values 3 and 4) don't contribute to any elements of the output.

The corresponding DML_RESAMPLE_GRAD1_OPERATOR_DESC would perform the following:

InputGradientTensor           OutputGradientTensor
    [4, 5, 6]      ResampleGrad    [[9, 6],
                       -->          [0, 0]]

Notice that the values in the OutputGradientTensor represent the weighted contributions of that element to the OutputTensor during the original DML_RESAMPLE2_OPERATOR_DESC operator.

Syntax

struct DML_RESAMPLE_GRAD1_OPERATOR_DESC
{
    const DML_TENSOR_DESC* InputGradientTensor;
    const DML_TENSOR_DESC* OutputGradientTensor;
    DML_INTERPOLATION_MODE InterpolationMode;
    DML_AXIS_DIRECTION RoundingDirection;
    UINT DimensionCount;
    _Field_size_(DimensionCount) const FLOAT* Scales;
    _Field_size_(DimensionCount) const FLOAT* InputPixelOffsets;
    _Field_size_(DimensionCount) const FLOAT* OutputPixelOffsets;
};

Members

InputGradientTensor

Type: const DML_TENSOR_DESC*

The incoming gradient tensor. This is typically obtained from the output of backpropagation of a preceding layer. Typically this tensor would have the same sizes as the output of the corresponding DML_RESAMPLE2_OPERATOR_DESC in the forward pass.

OutputGradientTensor

Type: const DML_TENSOR_DESC*

An output tensor containing the backpropagated gradients. Typically this tensor would have the same sizes as the input of the corresponding DML_RESAMPLE2_OPERATOR_DESC in the forward pass.

InterpolationMode

Type: DML_INTERPOLATION_MODE

See DML_RESAMPLE2_OPERATOR_DESC::InterpolationMode.

RoundingDirection

Type: DML_AXIS_DIRECTION

See DML_RESAMPLE2_OPERATOR_DESC::RoundingDirection.

DimensionCount

Type: UINT

The number of elements in the Scales, InputPixelOffsets, and OutputPixelOffsets arrays. This value must equal the DimensionCount provided in the InputGradientTensor and OutputGradientTensor.

Scales

Type: _Field_size_(DimensionCount) const FLOAT*

See DML_RESAMPLE2_OPERATOR_DESC::Scales.

InputPixelOffsets

Type: _Field_size_(DimensionCount) const FLOAT*

See DML_RESAMPLE2_OPERATOR_DESC::InputPixelOffsets.

OutputPixelOffsets

Type: _Field_size_(DimensionCount) const FLOAT*

See DML_RESAMPLE2_OPERATOR_DESC::OutputPixelOffsets.

Remarks

This operator is equivalent to DML_RESAMPLE_GRAD_OPERATOR_DESC when InterpolationMode is set to DML_INTERPOLATION_MODE_LINEAR; or when InterpolationMode is set to DML_INTERPOLATION_MODE_NEAREST_NEIGHBOR, and RoundingDirection to DML_AXIS_DIRECTION_DECREASING, and OutputPixelOffsets are adjusted an additional -0.5.

Availability

This operator was introduced in DML_FEATURE_LEVEL_5_1.

Tensor constraints

InputGradientTensor and OutputGradientTensor must have the same DataType and DimensionCount.

Tensor support

Requirements