DML_MAX_POOLING_GRAD_OPERATOR_DESC structure (directml.h)

Article
10/20/2021

Computes backpropagation gradients for max pooling (see DML_MAX_POOLING2_OPERATOR_DESC).

Consider a 2x2 DML_MAX_POOLING2_OPERATOR_DESC without padding nor dilations and a stride of 1, which performs the following.

InputTensor             OutputTensor    IndicesTensor
[[1, 2, 3],   MaxPool    [[4, 4],        [[4, 4],
 [2, 4, 2],     -->       [6, 7]]         [7, 8]]
 [5, 6, 7]]

The largest element of each 2x2 window in the input tensor produces one element of the output. Below is an example of the output of DML_MAX_POOLING_GRAD_OPERATOR_DESC, given similar parameters.

InputTensor   InputGradientTensor            OutputGradientTensor
[[1, 2, 3],       [[1, 2],     MaxPoolGrad   [[0, 0, 0],
 [2, 4, 2],        [4, 5]]         -->        [0, 3, 0],
 [5, 6, 7]]                                   [0, 4, 5]]

In effect, this operator uses the InputTensor to determine the index of the largest element from each window, and distributes the values of InputGradientTensor into the OutputGradientTensor based on these indices. Where indices overlap, the values are summed. Any unreferenced output elements are zeroed.

In the case of a tie (where more than one element in a window has the same maximum value), the element with the lowest logical element index is chosen.

Syntax

struct DML_MAX_POOLING_GRAD_OPERATOR_DESC {
  const DML_TENSOR_DESC *InputTensor;
  const DML_TENSOR_DESC *InputGradientTensor;
  const DML_TENSOR_DESC *OutputGradientTensor;
  UINT                  DimensionCount;
  const UINT            *Strides;
  const UINT            *WindowSize;
  const UINT            *StartPadding;
  const UINT            *EndPadding;
  const UINT            *Dilations;
};

Members

InputTensor

Type: const DML_TENSOR_DESC*

The input feature tensor. This is typically the same tensor that was provided as the InputTensor to DML_MAX_POOLING2_OPERATOR_DESC in the forward pass.

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_MAX_POOLING2_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_MAX_POOLING2_OPERATOR_DESC in the forward pass.

DimensionCount

Type: UINT

The number of elements in the Strides, WindowSize, StartPadding, EndPadding, and Dilations arrays. This value must equal the spatial dimension count (InputTensor's DimensionCount - 2). As this operator only supports 4D tensors, the only valid value for this parameter is 2.

Strides