DML_MEAN_VARIANCE_NORMALIZATION_OPERATOR_DESC structure (directml.h)

Article
05/09/2023

Performs a mean variance normalization function on the input tensor. This operator will calculate the mean and variance of the input tensor to perform normalization. This operator performs the following computation.

Output = FusedActivation(Scale * ((Input - Mean) / sqrt(Variance + Epsilon)) + Bias).

Syntax

struct DML_MEAN_VARIANCE_NORMALIZATION_OPERATOR_DESC {
  const DML_TENSOR_DESC   *InputTensor;
  const DML_TENSOR_DESC   *ScaleTensor;
  const DML_TENSOR_DESC   *BiasTensor;
  const DML_TENSOR_DESC   *OutputTensor;
  BOOL                    CrossChannel;
  BOOL                    NormalizeVariance;
  FLOAT                   Epsilon;
  const DML_OPERATOR_DESC *FusedActivation;
};

Members

InputTensor

Type: const DML_TENSOR_DESC*

A tensor containing the Input data. This tensor's dimensions should be { BatchCount, ChannelCount, Height, Width }.

ScaleTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor containing the Scale data. This tensor's dimensions should be { BatchCount, ChannelCount, Height, Width }. Any dimension can be replaced with 1 to broadcast in that dimension. If DML_FEATURE_LEVEL is less than DML_FEATURE_LEVEL_5_2, then this tensor is required if BiasTensor is present. If DML_FEATURE_LEVEL is greater than or equal to DML_FEATURE_LEVEL_5_2, then this tensor can be null regardless of the value of BiasTensor.

BiasTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor containing the bias data. This tensor's dimensions should be { BatchCount, ChannelCount, Height, Width }. Any dimension can be replaced with 1 to broadcast in that dimension. If DML_FEATURE_LEVEL is less than DML_FEATURE_LEVEL_5_2, then this tensor is required if ScaleTensor is present. If DML_FEATURE_LEVEL is greater than or equal to DML_FEATURE_LEVEL_5_2, then this tensor can be null regardless of the value of ScaleTensor.

OutputTensor

Type: const DML_TENSOR_DESC*

A tensor to write the results to. This tensor's dimensions are { BatchCount, ChannelCount, Height, Width }.

CrossChannel

Type: BOOL

When TRUE, the MeanVariance layer includes channels in the Mean and Variance calculations, meaning they are normalized across axes {ChannelCount, Height, Width}. When FALSE, Mean and Variance calculations are normalized across axes {Height, Width} with each channel being independent.

NormalizeVariance

Type: BOOL

TRUE if the Normalization layer includes Variance in the normalization calculation. Otherwise, FALSE. If FALSE, then normalization equation is Output = FusedActivation(Scale * (Input - Mean) + Bias).

Epsilon

Type: FLOAT

The epsilon value to use to avoid division by zero. A value of 0.00001 is recommended as default.

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

A newer version of this operator, DML_MEAN_VARIANCE_NORMALIZATION1_OPERATOR_DESC, was introduced in DML_FEATURE_LEVEL_2_1.

Availability

This operator was introduced in DML_FEATURE_LEVEL_1_0.

Tensor constraints

InputTensor and OutputTensor must have the same Sizes.
BiasTensor, InputTensor, OutputTensor, and ScaleTensor must have the same DataType.

Tensor support

Tensor	Kind	Dimensions	Supported dimension counts	Supported data types
InputTensor	Input	{ BatchCount, ChannelCount, Height, Width }	4	FLOAT32, FLOAT16
ScaleTensor	Optional input	{ ScaleBatchCount, ScaleChannelCount, ScaleHeight, ScaleWidth }	4	FLOAT32, FLOAT16
BiasTensor	Optional input	{ BiasBatchCount, BiasChannelCount, BiasHeight, BiasWidth }	4	FLOAT32, FLOAT16
OutputTensor	Output	{ BatchCount, ChannelCount, Height, Width }	4	FLOAT32, FLOAT16

Requirements

Requirement	Value
Header	directml.h

Share via