opennn::LSTMOperator Struct Reference

#include <opennn/network/layers/lstm_layer.h>

Inheritance diagram for opennn::LSTMOperator:

Public Types
enum	ForwardSlot {   InputSlot = 0 , ForgetGateSlot , InputGateSlot , CandidateGateSlot ,   OutputGateSlot , CellStateSlot , HiddenStateSlot , CellActivationSlot ,   CudnnInputSequenceSlot , CudnnOutputSequenceSlot , OutputSlot }
enum	BackwardSlot {   OutputDeltaSlot = 0 , InputDeltaSlot , HiddenDeltaScratchSlot , CellDeltaScratchSlot ,   ForgetDeltaScratchSlot , InputDeltaScratchSlot , CandidateDeltaScratchSlot , OutputDeltaScratchSlot ,   CudnnOutputDeltaScratchSlot , CudnnInputDeltaScratchSlot }
enum	Gate {   Forget , Input , Candidate , Output ,   GateCount }
using	GateViews = array<const TensorView*, GateCount>

Public Member Functions
void	set (Index, Index, Index, ActivationFunction new_activation_function=ActivationFunction::Tanh, ActivationFunction new_recurrent_activation_function=ActivationFunction::Sigmoid, Type new_compute_dtype=Type::FP32)
GateViews	gate_biases () const
GateViews	gate_weights () const
GateViews	gate_recurrent_weights () const
GateViews	gate_bias_gradients () const
GateViews	gate_weight_gradients () const
GateViews	gate_recurrent_weight_gradients () const
vector< TensorSpec >	parameter_specs () const override
vector< ParameterSlot >	parameter_slots () override
void	set_parameters_random () override
void	set_parameters_glorot () override
void	set_parameters_pytorch () override
void	forward_propagate (ForwardPropagation &, size_t, ForwardPropagationMode) override
void	back_propagate (ForwardPropagation &, BackPropagation &, size_t) const override
Public Member Functions inherited from opennn::Operator
virtual	~Operator ()=default
virtual vector< TensorSpec >	state_specs () const
virtual vector< SlotQuantization >	parameter_quantization () const
virtual void	link_parameters (span< const TensorView > views)
virtual void	link_gradients (span< const TensorView > views)
virtual void	link_states (span< const TensorView >)
virtual void	link_parameter_scales (span< const TensorView >)
bool	link_slots (span< const TensorView > views, TensorView *ParameterSlot::*member)
virtual void	initialize_states ()
virtual void	set_weights_dtype (Type new_weights_dtype)
virtual void	to_JSON (JsonWriter &) const
virtual void	from_JSON (const Json *)
virtual void	load_state_from_JSON (const Json *)
TensorView &	get_input (ForwardPropagation &forward_propagation, size_t layer, size_t slot_index=0) const noexcept
vector< TensorView > &	get_inputs (ForwardPropagation &forward_propagation, size_t layer) const noexcept
TensorView &	get_output (ForwardPropagation &forward_propagation, size_t layer, size_t slot_index=0) const noexcept
TensorView &	get_output_delta (BackPropagation &back_propagation, size_t layer, size_t slot_index=0) const noexcept
TensorView &	get_input_delta (BackPropagation &back_propagation, size_t layer, size_t slot_index=0) const noexcept

Public Attributes
Index	input_features = 0
Index	output_features = 0
Index	time_steps = 0
bool	return_sequences = false
ActivationFunction	activation_function = ActivationFunction::Tanh
ActivationFunction	recurrent_activation_function = ActivationFunction::Sigmoid
TensorView	forget_bias
TensorView	input_bias
TensorView	candidate_bias
TensorView	output_bias
TensorView	forget_weights
TensorView	input_weights
TensorView	candidate_weights
TensorView	output_weights
TensorView	forget_recurrent_weights
TensorView	input_recurrent_weights
TensorView	candidate_recurrent_weights
TensorView	output_recurrent_weights
TensorView	forget_bias_gradient
TensorView	input_bias_gradient
TensorView	candidate_bias_gradient
TensorView	output_bias_gradient
TensorView	forget_weight_gradient
TensorView	input_weight_gradient
TensorView	candidate_weight_gradient
TensorView	output_weight_gradient
TensorView	forget_recurrent_weight_gradient
TensorView	input_recurrent_weight_gradient
TensorView	candidate_recurrent_weight_gradient
TensorView	output_recurrent_weight_gradient
Public Attributes inherited from opennn::Operator
Type	compute_dtype = Type::FP32
Type	weights_dtype = Type::FP32
vector< size_t >	input_slots = {0}
vector< size_t >	output_slots = {1}
vector< size_t >	input_delta_slots = {1}
vector< size_t >	output_delta_slots = {0}

Additional Inherited Members
Protected Member Functions inherited from opennn::CudnnRnnState
unique_lock< mutex >	lock_backend_state () const
void	cudnn_rnn_forward_ (const CudnnRnnShapeSlot &, bool is_training, bool has_cell_state, const void *x, void *y, Buffer &forward_state, const function< CudnnRnnShapeSlot &()> &reconfigure) const
void	cudnn_rnn_backward_ (const CudnnRnnShapeSlot &, bool has_cell_state, const void *x, const void *y, const void *dy, void *dx, const Buffer &forward_state, Buffer &backward_scratch) const
CudnnRnnShapeSlot &	cudnn_setup_ (const CudnnRnnConfig &, Index input_features, Index output_features, Index time_steps, Index batch_size, bool for_training) const
CudnnRnnShapeSlot &	cudnn_setup_attempt_ (const CudnnRnnConfig &, Index input_features, Index output_features, Index time_steps, Index batch_size, bool for_training) const
void	cudnn_copy_weight_regions_ (int num_linear_layers, Index input_features, Index output_features, const TensorView *const *matrices, const TensorView *const *vectors, Buffer &packed_weights, bool to_cudnn) const
void	cudnn_pack_weights_ (int num_linear_layers, Index input_features, Index output_features, const TensorView *const *weights, const TensorView *const *biases, Buffer &forward_state, uint64_t parameters_version) const
void	cudnn_unpack_gradients_ (int num_linear_layers, Index input_features, Index output_features, const TensorView *const *weight_gradients, const TensorView *const *bias_gradients, Buffer &backward_scratch) const
void	prepare_cudnn_forward_state_ (Buffer &, bool is_training, const CudnnRnnShapeSlot &) const
string	cudnn_rnn_configuration_ (const CudnnRnnShapeSlot &, const char *mode) const
void	drive_cudnn_forward_ (const CudnnRnnDims &, const TensorView &input, TensorView &sequence_output, TensorView &output, TensorView &cudnn_input_sequence, TensorView &cudnn_output_sequence, Buffer &forward_state, bool is_training, uint64_t parameters_version) const
void	drive_cudnn_backward_ (const CudnnRnnDims &, const TensorView &input, const TensorView &sequence_output, const TensorView &output_delta, const TensorView &cudnn_input_sequence, const TensorView &cudnn_output_sequence, TensorView &input_delta, TensorView &sequence_delta_scratch, TensorView &input_delta_scratch, const Buffer &forward_state, Buffer &backward_scratch) const
Protected Attributes inherited from opennn::CudnnRnnState
BackendState	backend_state

Member Typedef Documentation

GateViews

using opennn::LSTMOperator::GateViews = array<const TensorView*, GateCount>

Member Enumeration Documentation

BackwardSlot

enum opennn::LSTMOperator::BackwardSlot

Enumerator
OutputDeltaSlot
InputDeltaSlot
HiddenDeltaScratchSlot
CellDeltaScratchSlot
ForgetDeltaScratchSlot
InputDeltaScratchSlot
CandidateDeltaScratchSlot
OutputDeltaScratchSlot
CudnnOutputDeltaScratchSlot
CudnnInputDeltaScratchSlot

ForwardSlot

enum opennn::LSTMOperator::ForwardSlot

Enumerator
InputSlot
ForgetGateSlot
InputGateSlot
CandidateGateSlot
OutputGateSlot
CellStateSlot
HiddenStateSlot
CellActivationSlot
CudnnInputSequenceSlot
CudnnOutputSequenceSlot
OutputSlot

Gate

enum opennn::LSTMOperator::Gate

Enumerator
Forget
Input
Candidate
Output
GateCount

Member Function Documentation

back_propagate()

void opennn::LSTMOperator::back_propagate ( ForwardPropagation & , BackPropagation & , size_t  ) const

overridevirtual

Reimplemented from opennn::Operator.

forward_propagate()

void opennn::LSTMOperator::forward_propagate ( ForwardPropagation & , size_t , ForwardPropagationMode  )

overridevirtual

Reimplemented from opennn::Operator.

gate_bias_gradients()

GateViews opennn::LSTMOperator::gate_bias_gradients ( ) const

inline

gate_biases()

GateViews opennn::LSTMOperator::gate_biases ( ) const

inline

gate_recurrent_weight_gradients()

GateViews opennn::LSTMOperator::gate_recurrent_weight_gradients ( ) const

inline

gate_recurrent_weights()

GateViews opennn::LSTMOperator::gate_recurrent_weights ( ) const

inline

gate_weight_gradients()

GateViews opennn::LSTMOperator::gate_weight_gradients ( ) const

inline

gate_weights()

GateViews opennn::LSTMOperator::gate_weights ( ) const

inline

parameter_slots()

vector< ParameterSlot > opennn::LSTMOperator::parameter_slots ( )

overridevirtual

Reimplemented from opennn::Operator.

parameter_specs()

vector< TensorSpec > opennn::LSTMOperator::parameter_specs ( ) const

overridevirtual

Reimplemented from opennn::Operator.

set()

void opennn::LSTMOperator::set	(	Index	,
		Index	,
		Index	,
		ActivationFunction	new_activation_function = ActivationFunction::Tanh,
		ActivationFunction	new_recurrent_activation_function = ActivationFunction::Sigmoid,
		Type	new_compute_dtype = Type::FP32 )

set_parameters_glorot()

void opennn::LSTMOperator::set_parameters_glorot ( )

overridevirtual

Reimplemented from opennn::Operator.

set_parameters_pytorch()

void opennn::LSTMOperator::set_parameters_pytorch ( )

overridevirtual

Reimplemented from opennn::Operator.

set_parameters_random()

void opennn::LSTMOperator::set_parameters_random ( )

overridevirtual

Reimplemented from opennn::Operator.

Member Data Documentation

activation_function

ActivationFunction opennn::LSTMOperator::activation_function = ActivationFunction::Tanh

candidate_bias

TensorView opennn::LSTMOperator::candidate_bias

candidate_bias_gradient

TensorView opennn::LSTMOperator::candidate_bias_gradient

candidate_recurrent_weight_gradient

TensorView opennn::LSTMOperator::candidate_recurrent_weight_gradient

candidate_recurrent_weights

TensorView opennn::LSTMOperator::candidate_recurrent_weights

candidate_weight_gradient

TensorView opennn::LSTMOperator::candidate_weight_gradient

candidate_weights

TensorView opennn::LSTMOperator::candidate_weights

forget_bias

TensorView opennn::LSTMOperator::forget_bias

forget_bias_gradient

TensorView opennn::LSTMOperator::forget_bias_gradient

forget_recurrent_weight_gradient

TensorView opennn::LSTMOperator::forget_recurrent_weight_gradient

forget_recurrent_weights

TensorView opennn::LSTMOperator::forget_recurrent_weights

forget_weight_gradient

TensorView opennn::LSTMOperator::forget_weight_gradient

forget_weights

TensorView opennn::LSTMOperator::forget_weights

input_bias

TensorView opennn::LSTMOperator::input_bias

input_bias_gradient

TensorView opennn::LSTMOperator::input_bias_gradient

input_features

Index opennn::LSTMOperator::input_features = 0

input_recurrent_weight_gradient

TensorView opennn::LSTMOperator::input_recurrent_weight_gradient

input_recurrent_weights

TensorView opennn::LSTMOperator::input_recurrent_weights

input_weight_gradient

TensorView opennn::LSTMOperator::input_weight_gradient

input_weights

TensorView opennn::LSTMOperator::input_weights

output_bias

TensorView opennn::LSTMOperator::output_bias

output_bias_gradient

TensorView opennn::LSTMOperator::output_bias_gradient

output_features

Index opennn::LSTMOperator::output_features = 0

output_recurrent_weight_gradient

TensorView opennn::LSTMOperator::output_recurrent_weight_gradient

output_recurrent_weights

TensorView opennn::LSTMOperator::output_recurrent_weights

output_weight_gradient

TensorView opennn::LSTMOperator::output_weight_gradient

output_weights

TensorView opennn::LSTMOperator::output_weights

recurrent_activation_function

ActivationFunction opennn::LSTMOperator::recurrent_activation_function = ActivationFunction::Sigmoid

return_sequences

bool opennn::LSTMOperator::return_sequences = false

time_steps

Index opennn::LSTMOperator::time_steps = 0