#include <opennn/training/adam.h>

Inheritance diagram for opennn::Adam:

Public Types
enum	DataSlot { GradientMoment , SquareGradientMoment , GraphScalars }

Public Types inherited from opennn::Optimizer
using	StoppingCondition = opennn::StoppingCondition

Public Member Functions
	Adam (Loss *=nullptr)

void	configure_for_task (NetworkTask) override

void	set_learning_rate (const float new_learning_rate)

float	get_learning_rate () const

void	set_beta_1 (const float)

void	set_beta_2 (const float)

void	set_epsilon (float)

float	get_epsilon () const noexcept

void	set_bf16_first_moment (const bool enabled)

bool	get_bf16_first_moment () const noexcept

void	update_parameters (BackPropagation &, OptimizerData &, UpdateMode=UpdateMode::Standard) override

void	on_epoch_begin (Index, OptimizerData &) override

void	from_JSON (const JsonDocument &) override

void	to_JSON (JsonWriter &) const override

Public Member Functions inherited from opennn::Optimizer
	Optimizer (Loss *=nullptr)

virtual	~Optimizer ()

const Loss *	get_loss () const noexcept

bool	get_display () const noexcept

void	set (Loss *new_loss)

virtual void	set_display (bool new_display)

void	set_display_period (const Index new_display_period)

void	set_workers_number (int new_workers_number)

void	set_cuda_graph (bool enabled)

bool	get_cuda_graph_capture_failed () const noexcept

void	set_shuffle (bool enabled)

void	set_batch_pool_size (int size)

void	set_maximum_epochs (const Index new_maximum_epochs)

Index	get_maximum_epochs () const noexcept

void	set_maximum_time (const float new_maximum_time)

void	set_loss_goal (const float new_loss_goal)

void	set_maximum_validation_failures (const Index new_maximum_validation_failures)

void	set_validation_period (const Index n)

void	set_restore_best (bool enabled)

void	set_gradient_clip_norm (const float new_clip)

void	set_joint_gradient_arena (bool enabled) noexcept

Index	get_batch_size () const noexcept

void	set_batch_size (const Index new_batch_size)

virtual TrainingResult	train ()

Index	get_maximum_batch_size () const

const string &	get_name () const noexcept

void	save (const filesystem::path &) const

void	load (const filesystem::path &)

Additional Inherited Members
Public Attributes inherited from opennn::Optimizer
function< void(Index, float, float, Network *)>	post_epoch_callback

function< void(Network *)>	post_batch_callback

function< void(Index, float)>	post_best_callback

Protected Types inherited from opennn::Optimizer
enum class	UpdateMode { Standard , Capturable }

Protected Member Functions inherited from opennn::Optimizer
void	set_names ()

void	prepare_training_artifacts ()

Index	get_epoch_budget () const noexcept

optional< StoppingCondition >	find_stopping_condition (const EpochReport &, float minimum_loss_decrease=-MAX) const

void	close_training_result (TrainingResult &, const EpochReport &, bool has_validation) const

void	update_best_parameters (Network *, float, Index, Index &, BestModelSnapshot &)

void	restore_best_parameters (Network *, TrainingResult &, const BestModelSnapshot &)

void	write_common_json (JsonWriter &) const

void	read_common_json (const Json *)

void	setup_device_training ()

void	teardown_device_training ()

void	warmup_device_training (TrainingContext &, ThreadSafeQueue< Batch * > &, const vector< vector< Index > > &, const FeatureSelection &, TrainingSession &, OptimizerData &, ForwardPropagation validation_forward_propagation=nullptr, ThreadSafeQueue< Batch > validation_empty_queue=nullptr, const vector< vector< Index > > validation_batches=nullptr)

void	prefetch_batch (Batch &)

void	sync_device (bool on_gpu, bool has_recurrent_layers, TrainingSession &)

bool	should_display (Index epoch) const noexcept

void	display_epoch_results (Index epoch, float training_error, float training_accuracy, float validation_error, float validation_accuracy, bool has_validation, bool validation_fresh, bool is_token_cross_entropy, float elapsed_time) const

void	setup_batch_pools (BatchPools &pools, Dataset &dataset, Network &network, Index training_batch_size, Index validation_batch_size, bool has_validation, TrainingSession &training_session)

unique_ptr< BatchPrefetchSession >	start_batch_prefetch (ThreadSafeQueue< Batch * > &, const vector< vector< Index > > &, const FeatureSelection &, FillMode, WorkerProfileCounters *profile_counters=nullptr)

int	get_batch_pool_size (const Network &) const

Loss::EvaluationResult	run_epoch_loop (EpochLoopContext &)

void	prepare_full_batch_training (FullBatchContext &, const char *banner)

TrainingResult	train_full_batch (FullBatchContext &, const FullBatchHooks &)

bool	network_has_active_dropout () const

bool	can_use_cuda_graph () const

bool	uses_joint_gradient_arena () const noexcept

Loss::EvaluationResult	run_graph_epoch (TrainingSession &, OptimizerData &, ForwardPropagation &, BackPropagation &, ThreadSafeQueue< Batch * > &, const vector< vector< Index > > &, const FeatureSelection &)

Loss::EvaluationResult	train_epoch (TrainingContext &, ThreadSafeQueue< Batch * > &, const vector< vector< Index > > &, const FeatureSelection &, TrainingSession &, OptimizerData &)

Loss::EvaluationResult	evaluate_epoch (ForwardPropagation &, ThreadSafeQueue< Batch * > &, const vector< vector< Index > > &, const FeatureSelection &, TrainingSession &)

Static Protected Member Functions inherited from opennn::Optimizer
static void	clip_gradient_norm (BackPropagation &, float)

Protected Attributes inherited from opennn::Optimizer
Loss *	loss = nullptr

float	training_loss_goal = 0.0f

Index	maximum_validation_failures = numeric_limits<Index>::max()

Index	validation_period = 1

float	gradient_clip_norm = 0.0f

bool	joint_gradient_arena = false

bool	restore_best = true

Index	maximum_epochs = 10000

Index	batch_size = 0

float	maximum_time = 360000.0f

Index	display_period = 10

bool	display = true

bool	shuffle_samples = true

int	batch_pool_size_override = 0

int	workers_number = default_workers_number

bool	use_cuda_graph = false

bool	cuda_graph_capture_failed = false

string	name

Static Protected Attributes inherited from opennn::Optimizer
static const int	default_workers_number

Member Enumeration Documentation

◆ DataSlot

enum opennn::Adam::DataSlot

Enumerator
GradientMoment
SquareGradientMoment
GraphScalars

Constructor & Destructor Documentation

◆ Adam()

opennn::Adam::Adam ( Loss * = nullptr )

explicit

Member Function Documentation

◆ configure_for_task()

void opennn::Adam::configure_for_task ( NetworkTask )

overridevirtual

Reimplemented from opennn::Optimizer.

◆ from_JSON()

void opennn::Adam::from_JSON ( const JsonDocument & )

overridevirtual

Reimplemented from opennn::Optimizer.

◆ get_bf16_first_moment()

bool opennn::Adam::get_bf16_first_moment ( ) const

inlinenoexcept

◆ get_epsilon()

float opennn::Adam::get_epsilon ( ) const

inlinenoexcept

◆ get_learning_rate()

float opennn::Adam::get_learning_rate ( ) const

inline

◆ on_epoch_begin()

void opennn::Adam::on_epoch_begin	(	Index	,
		OptimizerData &	)

overridevirtual

Reimplemented from opennn::Optimizer.

◆ set_beta_1()

void opennn::Adam::set_beta_1 ( const float )

◆ set_beta_2()

void opennn::Adam::set_beta_2 ( const float )

◆ set_bf16_first_moment()

void opennn::Adam::set_bf16_first_moment ( const bool enabled )

inline

◆ set_epsilon()

void opennn::Adam::set_epsilon ( float )

◆ set_learning_rate()

void opennn::Adam::set_learning_rate ( const float new_learning_rate )

inline

◆ to_JSON()

void opennn::Adam::to_JSON ( JsonWriter & ) const

overridevirtual

Reimplemented from opennn::Optimizer.

◆ update_parameters()

void opennn::Adam::update_parameters	(	BackPropagation &	,
		OptimizerData &	,
		UpdateMode	= UpdateMode::Standard )

overridevirtual

Reimplemented from opennn::Optimizer.

Public Types

Public Member Functions

Additional Inherited Members

Member Enumeration Documentation

◆ DataSlot

Constructor & Destructor Documentation

◆ Adam()

Member Function Documentation

◆ configure_for_task()

◆ from_JSON()

◆ get_bf16_first_moment()

◆ get_epsilon()

◆ get_learning_rate()

◆ on_epoch_begin()

◆ set_beta_1()

◆ set_beta_2()

◆ set_bf16_first_moment()

◆ set_epsilon()

◆ set_learning_rate()

◆ to_JSON()

◆ update_parameters()