Edit

Share via

DML_BATCH_NORMALIZATION_TRAINING_OPERATOR_DESC structure (directml.h)

  • Article

Performs a batch normalization on the input. This operator performs the following computation: Output = FusedActivation(Scale * ((Input - Mean) / sqrt(Variance + Epsilon)) + Bias + FusedAdd).

Any dimension in ScaleTensor and BiasTensor can be set to 1, and be automatically broadcast to match InputTensor, but otherwise must equal the corresponding dimension's size from InputTensor. MeanTensor and VarianceTensor are computed on the input across the set of dimensions for which ScaleTensor and BiasTensor sizes equal one.

Syntax

struct DML_BATCH_NORMALIZATION_TRAINING_OPERATOR_DESC {
  const DML_TENSOR_DESC   *InputTensor;
  const DML_TENSOR_DESC   *ScaleTensor;
  const DML_TENSOR_DESC   *BiasTensor;
  const DML_TENSOR_DESC   *FusedAddTensor;
  const DML_TENSOR_DESC   *OutputTensor;
  const DML_TENSOR_DESC   *OutputMeanTensor;
  const DML_TENSOR_DESC   *OutputVarianceTensor;
  FLOAT                   Epsilon;
  const DML_OPERATOR_DESC *FusedActivation;
};

Members

InputTensor

Type: const DML_TENSOR_DESC*

A tensor containing the Input data.

ScaleTensor

Type: const DML_TENSOR_DESC*

A tensor containing the Scale data.

BiasTensor

Type: const DML_TENSOR_DESC*

A tensor containing the Bias data.

FusedAddTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor containing data that is added to the result prior to FusedActivation, if any.

OutputTensor

Type: const DML_TENSOR_DESC*

A tensor to write the results to.

OutputMeanTensor

Type: const DML_TENSOR_DESC*

A tensor to write the mean of the input to.

OutputVarianceTensor

Type: const DML_TENSOR_DESC*

A tensor to write the variance of the input to.

Epsilon

Type: FLOAT

The epsilon value to use to avoid division by zero.

FusedActivation

Type: _Maybenull_ const DML_OPERATOR_DESC*

An optional fused activation layer to apply after the normalization. For more info, see Using fused operators for improved performance.

Remarks

Availability

This operator was introduced in DML_FEATURE_LEVEL_4_1.

Tensor constraints

  • BiasTensor, FusedAddTensor, InputTensor, OutputMeanTensor, OutputTensor, OutputVarianceTensor, and ScaleTensor must have the same DataType and DimensionCount.
  • BiasTensor, OutputMeanTensor, OutputVarianceTensor, and ScaleTensor must have the same Sizes.
  • FusedAddTensor, InputTensor, and OutputTensor must have the same Sizes.

Tensor support

DML_FEATURE_LEVEL_4_1 and above

Tensor Kind Dimensions Supported dimension counts Supported data types
InputTensor Input { InputDimensions[] } 1 to 8 FLOAT32, FLOAT16
ScaleTensor Input { ScaleDimensions[] } 1 to 8 FLOAT32, FLOAT16
BiasTensor Input { ScaleDimensions[] } 1 to 8 FLOAT32, FLOAT16
FusedAddTensor Optional input { InputDimensions[] } 1 to 8 FLOAT32, FLOAT16
OutputTensor Output { InputDimensions[] } 1 to 8 FLOAT32, FLOAT16
OutputMeanTensor Output { ScaleDimensions[] } 1 to 8 FLOAT32, FLOAT16
OutputVarianceTensor Output { ScaleDimensions[] } 1 to 8 FLOAT32, FLOAT16

Requirements

Requirement Value
Header directml.h

See also