mlpack::ann::MultiheadAttention< InputDataType, OutputDataType, RegularizerType > Class Template Reference

Multihead Attention allows the model to jointly attend to information from different representation subspaces at different positions. More...

#include <multihead_attention.hpp>

Public Member Functions
	MultiheadAttention ()
	Default constructor.
	MultiheadAttention (const size_t tgtSeqLen, const size_t srcSeqLen, const size_t embedDim, const size_t numHeads)
	Create the MultiheadAttention object using the specified modules. More...
void	Reset ()
	Reset the layer parameters.
template<typename eT >
void	Forward (const arma::Mat< eT > &input, arma::Mat< eT > &output)
	Ordinary feed forward pass of a neural network, evaluating the function f(x) by propagating the activity forward through f. More...
template<typename eT >
void	Backward (const arma::Mat< eT > &, const arma::Mat< eT > &gy, arma::Mat< eT > &g)
	Ordinary feed backward pass of a neural network, calculating the function f(x) by propagating x backwards trough f. More...
template<typename eT >
void	Gradient (const arma::Mat< eT > &input, const arma::Mat< eT > &error, arma::Mat< eT > &gradient)
	Calculate the gradient using the output delta and the input activation. More...
size_t	WeightSize () const
	Get the size of the weights.
template<typename Archive >
void	serialize (Archive &ar, const uint32_t)
	Serialize the layer.
size_t	TgtSeqLen () const
	Get the target sequence length.
size_t &	TgtSeqLen ()
	Modify the target sequence length.
size_t	SrcSeqLen () const
	Get the source sequence length.
size_t &	SrcSeqLen ()
	Modify the source sequence length.
size_t	EmbedDim () const
	Get the embedding dimension.
size_t &	EmbedDim ()
	Modify the embedding dimension.
size_t	NumHeads () const
	Get the number of attention heads.
size_t &	NumHeads ()
	Modify the number of attention heads.
OutputDataType const &	AttentionMask () const
	Get the two dimensional Attention Mask.
OutputDataType &	AttentionMask ()
	Modify the two dimensional Attention Mask.
OutputDataType const &	KeyPaddingMask () const
	Get Key Padding Mask.
OutputDataType &	KeyPaddingMask ()
	Modify the Key Padding Mask.
OutputDataType const &	OutputParameter () const
	Get the output parameter.
OutputDataType &	OutputParameter ()
	Modify the output parameter.
OutputDataType const &	Delta () const
	Get the delta.
OutputDataType &	Delta ()
	Modify the delta.
OutputDataType const &	Gradient () const
	Get the gradient.
OutputDataType &	Gradient ()
	Modify the gradient.
OutputDataType const &	Parameters () const
	Get the parameters.
OutputDataType &	Parameters ()
	Modify the parameters.
size_t	InputShape () const

Detailed Description

template<typename InputDataType = arma::mat, typename OutputDataType = arma::mat, typename RegularizerType = NoRegularizer>
class mlpack::ann::MultiheadAttention< InputDataType, OutputDataType, RegularizerType >

Multihead Attention allows the model to jointly attend to information from different representation subspaces at different positions.

With a single attention head, averaging inhibits this. [arxiv.org:1706.03762v5]

The MultiheadAttention class takes concatenated form of query, key and value. The query, key and value are concatenated into single matrix and fed to the Forward function as input.

The query, key and value are matrices of shapes (embedDim * tgtSeqLen, batchSize), (embedDim * srcSeqLen, batchSize) and (embedDim * srcSeqLen, batchSize) respectively. The output is a matrix of shape (embedDim * tgtSeqLen, batchSize). The embeddings are stored consequently.

Template Parameters

InputDataType	Type of the input data (arma::colvec, arma::mat, arma::sp_mat or arma::cube).
OutputDataType	Type of the output data (arma::colvec, arma::mat, arma::sp_mat or arma::cube).
RegularizerType	Type of the regularizer to be used.

Constructor & Destructor Documentation

MultiheadAttention()

template<typename InputDataType , typename OutputDataType , typename RegularizerType >

mlpack::ann::MultiheadAttention< InputDataType, OutputDataType, RegularizerType >::MultiheadAttention	(	const size_t	tgtSeqLen,
		const size_t	srcSeqLen,
		const size_t	embedDim,
		const size_t	numHeads
	)

Create the MultiheadAttention object using the specified modules.

Parameters

tgtSeqLen	Target sequence length.
srcSeqLen	Source sequence length.
embedDim	Total dimension of the model.
numHeads	Number of parallel attention heads.

Member Function Documentation

Backward()

template<typename InputDataType , typename OutputDataType , typename RegularizerType >

template<typename eT >

void mlpack::ann::MultiheadAttention< InputDataType, OutputDataType, RegularizerType >::Backward	(	const arma::Mat< eT > &	,
		const arma::Mat< eT > &	gy,
		arma::Mat< eT > &	g
	)

Ordinary feed backward pass of a neural network, calculating the function f(x) by propagating x backwards trough f.

Using the results from the feed forward pass.

Parameters

gy	The backpropagated error.
g	The calculated gradient.

Forward()

template<typename InputDataType , typename OutputDataType , typename RegularizerType >

template<typename eT >

void mlpack::ann::MultiheadAttention< InputDataType, OutputDataType, RegularizerType >::Forward	(	const arma::Mat< eT > &	input,
		arma::Mat< eT > &	output
	)

Ordinary feed forward pass of a neural network, evaluating the function f(x) by propagating the activity forward through f.

Parameters

input	The query matrix.
output	Resulting output activation.

Gradient()

template<typename InputDataType , typename OutputDataType , typename RegularizerType >

template<typename eT >

void mlpack::ann::MultiheadAttention< InputDataType, OutputDataType, RegularizerType >::Gradient	(	const arma::Mat< eT > &	input,
		const arma::Mat< eT > &	error,
		arma::Mat< eT > &	gradient
	)

Calculate the gradient using the output delta and the input activation.

Parameters

input	The input data used for evaluating specified function.
error	The calculated error.
gradient	The calculated gradient.

---

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

• src/mlpack/methods/ann/layer/layer_types.hpp
• src/mlpack/methods/ann/layer/multihead_attention.hpp
• src/mlpack/methods/ann/layer/multihead_attention_impl.hpp