Fusion of several torch modules implemented for inference in Eigen. More...

#include <deeplima/eigen_wrp/bilstm_and_dense.h>

Inheritance diagram for deeplima::eigen_impl::Op_BiLSTM_Dense_ArgMax< M, V, T, AuxScalar, IFracBits, WFracBits >:

Classes
struct	params_bilstm_dense_argmax_t

struct	workbench_t

Public Types
typedef params_bilstm_dense_argmax_t	params_t

Public Member Functions
virtual std::shared_ptr< Op_Base::workbench_t >	create_workbench (uint32_t input_size, const std::shared_ptr< param_base_t > params, bool precomputed_input=false) const override

virtual bool	supports_precomputing () const

virtual void	precompute_inputs (const std::shared_ptr< param_base_t > params, const M &inputs, M &outputs, int64_t first_column)

virtual size_t	execute (std::shared_ptr< Op_Base::workbench_t > pwb, const M &input_matrix, const std::shared_ptr< param_base_t > pparams, std::vector< std::vector< uint8_t > > &final_output, size_t input_begin, size_t, size_t output_begin, size_t output_end)

Public Member Functions inherited from deeplima::eigen_impl::Op_Base
virtual	~Op_Base ()

Protected Types
typedef AuxScalar	fixed_point_t

using	MatrixWeight = typename std::conditional< std::is_integral_v< AuxScalar > &&std::is_signed_v< AuxScalar >, Eigen::Matrix< fixed_point_t, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor >, M >::type

using	MatrixInput = typename std::conditional< std::is_integral_v< AuxScalar > &&std::is_signed_v< AuxScalar >, Eigen::Matrix< fixed_point_t, Eigen::Dynamic, Eigen::Dynamic, Eigen::ColMajor >, M >::type

Protected Member Functions
void	forward_pass (const size_t hidden_size, Eigen::Ref< const M > input, const MatrixWeight &weight_hh, V &s, V &c, M &output)

void	backward_pass (const size_t hidden_size, Eigen::Ref< const M > input, const MatrixWeight &weight_hh, V &s, V &c, M &output)

void	update_c (const size_t hidden_size, const V &s, V &c)

void	step_fw (const size_t hidden_size, const size_t t, const V &s, V &c, M &output)

void	step_bw (const size_t hidden_size, const size_t t, const V &s, V &c, M &output)

Static Protected Attributes
static constexpr fixed_point_t	WEIGHT_FRACTION_MULT = fixed_point_t(1) << WFracBits

static constexpr fixed_point_t	DATA_FRACTION_MULT = fixed_point_t(1) << IFracBits

static constexpr fixed_point_t	WEIGHT_DATA_FRACTION_MULT = fixed_point_t(1) << (WFracBits + IFracBits)

Detailed Description

template<typename M, typename V, typename T, typename AuxScalar = float, uint8_t IFracBits = 4, uint8_t WFracBits = 4>
class deeplima::eigen_impl::Op_BiLSTM_Dense_ArgMax< M, V, T, AuxScalar, IFracBits, WFracBits >

Fusion of several torch modules implemented for inference in Eigen.

Precomputed inputs. The x of each gate is fasttext embeddings. They can be computed during model loading. This is done in precompute_inputs. This takes a very long time. This can be done in the first layer only

forward_pass is different for fixed point and non-fixed point

The implementation with factorization is less efficient while it should be more.

In linear layer on top of RNN outputs.

Could bequicker in fixed point than in floating point but not sure. Must be tested
For pos tagging, could avoid to compute everything because it is very sparse but the code must stay generic and cannot use information specific to morphologic model.

In fixed, only last step is in fixed because some necessary functions (tanh…) are not implemented in fixed but quicker nevertheless

Definition at line 55 of file bilstm_and_dense.h.