Layer Normalization module. More...

Inheritance diagram for Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >:

Collaboration diagram for Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >:

[legend]

Public Types
using	ModuleBase = Module< TDeviceType, TInput, TOutput >
	Alias for base module type.

using	MR = std::conditional_t< TDeviceType==DeviceType::Cuda, CudaMemoryResource, CpuMemoryResource >
	Memory resource type used for tensors, selected based on device type.

Public Types inherited from Mila::Dnn::Module< TDeviceType, TInput, TOutput >
using	MR = std::conditional_t< TDeviceType==DeviceType::Cuda, CudaMemoryResource, CpuMemoryResource >

Public Member Functions
	LayerNorm (const std::string &device_name, const LayerNormConfig &config)
	Constructs a new LayerNorm module with a device name.

	LayerNorm (std::shared_ptr< DeviceContext > device_context, const LayerNormConfig &config)
	Constructs a new LayerNorm module with a provided device context.

void	backward (const Tensor< TInput, MR > &input, const Tensor< TOutput, MR > &output_grad, Tensor< TInput, MR > &input_grad)
	Performs the backward pass of the Layer Normalization operation.

void	forward (const Tensor< TInput, MR > &input, Tensor< TOutput, MR > &output)
	Performs the forward pass of the Layer Normalization operation.

std::shared_ptr< Tensor< TInput, MR > >	getBias ()
	Gets the bias tensor used after normalization and scaling.

std::shared_ptr< Tensor< TInput, MR > >	getWeight ()
	Gets the weight tensor used for scaling after normalization.

bool	hasBias () const
	Gets whether the module has a bias tensor.

void	load (ModelArchive &archive) override
	Deserializes the module state from a ZIP archive.

size_t	parameterCount () const override
	Gets the number of trainable parameters in this module.

void	save (ModelArchive &archive) const override
	Serializes the module state to a ZIP archive.

std::string	toString () const override
	Generates a string representation of this module's configuration.

Public Member Functions inherited from Mila::Dnn::Module< TDeviceType, TInput, TOutput >
	Module (const std::string &device_name, const ComponentConfig &config)
	Constructor with device name.

	Module (std::shared_ptr< DeviceContext > context, const ComponentConfig &config)
	Constructor with a specific device context.

virtual	~Module ()=default
	Virtual destructor for proper cleanup in derived classes.

std::shared_ptr< Compute::DeviceContext >	getDeviceContext () const
	Get the device context for this module.

Compute::DeviceType	getDeviceType () const
	Get the device type of the current device context.

std::string	getName () const
	Get the name of the module.

const auto &	getParameterTensors () const
	Get the parameter tensors of this module.

const ComputePrecision::Policy &	getPrecision () const

const auto &	getStateTensors () const
	Get the state tensors of this module.

bool	isTraining () const
	Check if the module is in training mode.

virtual void	setTraining (bool is_training)
	Set the training mode of this module.

Private Member Functions
void	createOperation ()
	Creates the appropriate Layer Normalization operation based on the current device context.

void	initializeTensors ()
	Initializes the tensors needed for the Layer Normalization operation.

Private Attributes
std::shared_ptr< Tensor< TOutput, MR > >	bias_ { nullptr }
	The bias tensor added after normalization and scaling.

LayerNormConfig	config_
	Configuration for the LayerNorm module.

std::shared_ptr< Tensor< TOutput, MR > >	mean_ { nullptr }
	The mean tensor used for normalization.

std::shared_ptr< UnaryOperation< TDeviceType, TInput, TOutput > >	operation_ { nullptr }
	The underlying operation that implements Layer Normalization.

std::vector< std::shared_ptr< Tensor< TOutput, MR > > >	output_state_
	Collection of output state tensors for caching.

std::vector< std::shared_ptr< Tensor< TOutput, MR > > >	parameters_
	Collection of trainable parameters for this module.

OperationAttributes	properties_
	Operation attributes and configuration.

std::shared_ptr< Tensor< TOutput, MR > >	rstd_ { nullptr }
	The reciprocal standard deviation tensor.

std::shared_ptr< Tensor< TOutput, MR > >	weight_ { nullptr }
	The weight tensor for scaling after normalization.

Additional Inherited Members
Protected Member Functions inherited from Mila::Dnn::Module< TDeviceType, TInput, TOutput >
const std::string	parametersToString () const
	Helper method to convert parameters to string representation.

const std::string	stateToString () const
	Helper method to convert state tensors to string representation.

Protected Attributes inherited from Mila::Dnn::Module< TDeviceType, TInput, TOutput >
std::unordered_map< std::string, std::shared_ptr< Tensor< TOutput, MR > > >	parameter_map_ = {}
	Map of parameter names to parameter tensors.

std::unordered_map< std::string, std::shared_ptr< Tensor< TOutput, MR > > >	state_map_ = {}
	Map of state names to state tensors.

Detailed Description

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>
requires ValidTensorTypes<TInput, TOutput>
class Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >

Layer Normalization module.

Layer Normalization is a technique used to normalize the inputs across features for each data sample in a batch. It helps stabilize and accelerate deep neural network training by reducing internal covariate shift.

The operation can be expressed as: y = ((x - mean) / sqrt(variance + epsilon)) * weight + bias

Unlike Batch Normalization, Layer Normalization computes statistics independently for each sample in a batch, making it well-suited for variable-length sequences and recurrent neural networks.

Template Parameters

TDeviceType	The device type (CPU or CUDA) on which to perform computations.
TInput	Data type of the input tensor elements.
TOutput	Data type of the output tensor elements, defaults to TInput.

Member Typedef Documentation

◆ ModuleBase

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

using Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::ModuleBase = Module<TDeviceType, TInput, TOutput>

export

Alias for base module type.

◆ MR

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

using Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::MR = std::conditional_t<TDeviceType == DeviceType::Cuda, CudaMemoryResource, CpuMemoryResource>

export

Memory resource type used for tensors, selected based on device type.

Constructor & Destructor Documentation

◆ LayerNorm() [1/2]

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::LayerNorm	(	const std::string &	device_name,
		const LayerNormConfig &	config
	)

inlineexplicitexport

Constructs a new LayerNorm module with a device name.

Creates a new DeviceContext internally using the provided device name. This constructor is useful for creating standalone modules without pre-existing device contexts.

Parameters

device_name	The name of the device to use (e.g., "CPU", "CUDA:0").
config	Configuration parameters for the LayerNorm module.

Exceptions

std::invalid_argument	If the device name is invalid or the configuration is invalid
std::runtime_error	If device type doesn't match template parameter TDeviceType

Here is the call graph for this function:

◆ LayerNorm() [2/2]

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::LayerNorm	(	std::shared_ptr< DeviceContext >	device_context,
		const LayerNormConfig &	config
	)

inlineexplicitexport

Constructs a new LayerNorm module with a provided device context.

Uses a pre-existing DeviceContext instance. This constructor is useful when integrating the module into a larger network that shares device contexts across modules.

Parameters

device_context	The device context to use for this module.
config	Configuration parameters for the LayerNorm module.

Exceptions

std::invalid_argument	If device_context is null or configuration is invalid
std::runtime_error	If device context type doesn't match template parameter TDeviceType

Here is the call graph for this function:

Member Function Documentation

◆ backward()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

void Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::backward	(	const Tensor< TInput, MR > &	input,
		const Tensor< TOutput, MR > &	output_grad,
		Tensor< TInput, MR > &	input_grad
	)

inlineexport

Performs the backward pass of the Layer Normalization operation.

Computes gradients with respect to input and parameters. Layer Normalization's backward pass requires computing:

Gradient with respect to weight and bias (if present)
Gradient with respect to input, which is more complex due to the normalization operation's chain rule derivatives

Parameters

input	The input tensor from the forward pass.
output_grad	The gradient of the loss with respect to the output.
input_grad	The tensor to store the gradient with respect to input.

◆ createOperation()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

void Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::createOperation ( )

inlineexportprivate

Creates the appropriate Layer Normalization operation based on the current device context.

This method initializes the operation_ member with the appropriate implementation of Layer Normalization for either CPU or CUDA, based on the current device context. It also passes the config object to the operation.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ forward()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

void Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::forward	(	const Tensor< TInput, MR > &	input,
		Tensor< TOutput, MR > &	output
	)

inlineexport

Performs the forward pass of the Layer Normalization operation.

Normalizes the input tensor across the specified axis, then scales and shifts the result using the weight and bias tensors.

Parameters

input	The input tensor to be normalized.
output	The output tensor where the results will be stored.

◆ getBias()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

std::shared_ptr< Tensor< TInput, MR > > Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::getBias ( )

inlineexport

Gets the bias tensor used after normalization and scaling.

The bias tensor is added after normalization and scaling.

Returns: std::shared_ptr<Tensor<TInput, MR>> Shared pointer to the bias tensor.

◆ getWeight()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

std::shared_ptr< Tensor< TInput, MR > > Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::getWeight ( )

inlineexport

Gets the weight tensor used for scaling after normalization.

The weight tensor is applied as a scale factor to the normalized values.

Returns: std::shared_ptr<Tensor<TInput, MR>> Shared pointer to the weight tensor.

◆ hasBias()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

bool Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::hasBias ( ) const

inlineexport

Gets whether the module has a bias tensor.

Returns: bool True if the module has a bias tensor, false otherwise.

Here is the call graph for this function:

◆ initializeTensors()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

void Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::initializeTensors ( )

inlineexportprivate

Initializes the tensors needed for the Layer Normalization operation.

Creates and initializes:

weight tensor (initialized to ones)
bias tensor (initialized to zeros)
mean tensor (for storing means during forward pass)
reciprocal standard deviation tensor (for storing 1/std during forward pass)

Here is the call graph for this function:

Here is the caller graph for this function:

◆ load()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

void Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::load ( ModelArchive & archive )

inlineoverrideexportvirtual

Deserializes the module state from a ZIP archive.

Loads the trainable parameters (weight, bias) from the provided archive.

Parameters

zip	The ZIP archive to load the module state from.

Implements Mila::Dnn::Module< TDeviceType, TInput, TOutput >.

Here is the call graph for this function:

◆ parameterCount()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

size_t Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::parameterCount ( ) const

inlineoverrideexportvirtual

Gets the number of trainable parameters in this module.

Counts the total number of trainable parameters, which includes the weight tensor and, if present, the bias tensor.

Returns: size_t The total number of parameters.

Implements Mila::Dnn::Module< TDeviceType, TInput, TOutput >.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ save()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

void Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::save ( ModelArchive & archive ) const

inlineoverrideexportvirtual

Serializes the module state to a ZIP archive.

Saves the trainable parameters (weight, bias) to the provided archive.

Parameters

zip	The ZIP archive to save the module state to.

Implements Mila::Dnn::Module< TDeviceType, TInput, TOutput >.

Here is the call graph for this function:

◆ toString()

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

std::string Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::toString ( ) const

inlineoverrideexportvirtual

Generates a string representation of this module's configuration.

Returns: std::string A formatted string with module information

Implements Mila::Dnn::Module< TDeviceType, TInput, TOutput >.

Here is the call graph for this function:

Member Data Documentation

◆ bias_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

std::shared_ptr<Tensor<TOutput, MR> > Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::bias_ { nullptr }

exportprivate

The bias tensor added after normalization and scaling.

◆ config_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

LayerNormConfig Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::config_

exportprivate

Configuration for the LayerNorm module.

◆ mean_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

std::shared_ptr<Tensor<TOutput, MR> > Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::mean_ { nullptr }

exportprivate

The mean tensor used for normalization.

Stores the mean values computed during the forward pass.

◆ operation_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

std::shared_ptr<UnaryOperation<TDeviceType, TInput, TOutput> > Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::operation_ { nullptr }

exportprivate

The underlying operation that implements Layer Normalization.

◆ output_state_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

std::vector<std::shared_ptr<Tensor<TOutput, MR> > > Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::output_state_

exportprivate

Collection of output state tensors for caching.

◆ parameters_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

std::vector<std::shared_ptr<Tensor<TOutput, MR> > > Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::parameters_

exportprivate

Collection of trainable parameters for this module.

◆ properties_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

OperationAttributes Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::properties_

exportprivate

Operation attributes and configuration.

◆ rstd_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

std::shared_ptr<Tensor<TOutput, MR> > Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::rstd_ { nullptr }

exportprivate

The reciprocal standard deviation tensor.

Stores the reciprocal of the standard deviation values (1/sqrt(variance + epsilon)) computed during the forward pass.

◆ weight_

template<DeviceType TDeviceType = DeviceType::Cuda, typename TInput = float, typename TOutput = TInput>

std::shared_ptr<Tensor<TOutput, MR> > Mila::Dnn::LayerNorm< TDeviceType, TInput, TOutput >::weight_ { nullptr }

exportprivate

The weight tensor for scaling after normalization.

The documentation for this class was generated from the following file:

/home/runner/work/Mila/Mila/Mila/Src/Dnn/Modules/Normalization/LayerNorm.ixx

Public Types

Public Member Functions

Private Member Functions

Private Attributes

Additional Inherited Members

Detailed Description

Member Typedef Documentation

◆ ModuleBase

◆ MR

Constructor & Destructor Documentation

◆ LayerNorm() [1/2]

◆ LayerNorm() [2/2]

Member Function Documentation

◆ backward()

◆ createOperation()

◆ forward()

◆ getBias()

◆ getWeight()

◆ hasBias()

◆ initializeTensors()

◆ load()

◆ parameterCount()

◆ save()

◆ toString()

Member Data Documentation

◆ bias_

◆ config_

◆ mean_

◆ operation_

◆ output_state_

◆ parameters_

◆ properties_

◆ rstd_

◆ weight_