DML_RESAMPLE_GRAD_OPERATOR_DESC structure (directml.h)

Computes backpropagation gradients for Resample (see DML_RESAMPLE1_OPERATOR_DESC).

DML_RESAMPLE1_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_RESAMPLE1_OPERATOR_DESC, this operator produces an OutputGradientTensor with the same sizes as the input of the DML_RESAMPLE1_OPERATOR_DESC.

As an example, consider a DML_RESAMPLE1_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) contribute to no elements of the output.

The corresponding DML_RESAMPLE_GRAD_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_RESAMPLE1_OPERATOR_DESC operator.

Syntax

struct DML_RESAMPLE_GRAD_OPERATOR_DESC {
  const DML_TENSOR_DESC  *InputGradientTensor;
  const DML_TENSOR_DESC  *OutputGradientTensor;
  DML_INTERPOLATION_MODE InterpolationMode;
  UINT                   DimensionCount;
  const FLOAT            *Scales;
  const FLOAT            *InputPixelOffsets;
  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_RESAMPLE1_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_RESAMPLE1_OPERATOR_DESC in the forward pass.

InterpolationMode

Type: DML_INTERPOLATION_MODE

See InterpolationMode in DML_RESAMPLE1_OPERATOR_DESC.

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 Scales in DML_RESAMPLE1_OPERATOR_DESC.

InputPixelOffsets

Type: _Field_size_(DimensionCount) const FLOAT*

See InputPixelOffsets in DML_RESAMPLE1_OPERATOR_DESC.

OutputPixelOffsets

Type: _Field_size_(DimensionCount) const FLOAT*

See OutputPixelOffsets in DML_RESAMPLE1_OPERATOR_DESC.

Availability

This operator was introduced in DML_FEATURE_LEVEL_3_0.

Tensor constraints

InputGradientTensor and OutputGradientTensor must have the same DataType and DimensionCount.

Tensor support

DML_FEATURE_LEVEL_5_1 and above

DML_FEATURE_LEVEL_3_0 and above

Requirements