dismec::init Namespace Reference

Classes
class	ConstantInitializer
class	ConstantInitializationStrategy
	An initialization strategy that sets the weight vector to a given constant. More...
class	MeanOfFeaturesInitializer
class	MeanOfFeaturesStrategy
struct	TypeLookup
struct	TypeLookup< false >
struct	TypeLookup< true >
class	MultiPosMeanInitializer
class	MultiPosMeanStrategy
class	NumpyInitializer
class	NumpyInitializationStrategy
class	PreTrainedInitializer
class	PreTrainedInitializationStrategy
class	SubsetFeatureMeanInitializer
class	SubsetFeatureMeanStrategy
class	ZeroInitializer
class	ZeroInitializationStrategy
class	WeightsInitializer
	Base class for all weight initializers. More...
class	WeightInitializationStrategy
	Base class for all weight init strategies. More...

Functions
std::shared_ptr< WeightInitializationStrategy >	create_zero_initializer ()
	Creates an initialization strategy that initializes all weight vectors to zero. More...
std::shared_ptr< WeightInitializationStrategy >	create_constant_initializer (DenseRealVector vec)
std::shared_ptr< WeightInitializationStrategy >	create_pretrained_initializer (std::shared_ptr< model::Model > model)
	Creates an initialization strategy that uses an already trained model to set the initial weights. More...
std::shared_ptr< WeightInitializationStrategy >	create_numpy_initializer (const std::filesystem::path &weights, std::optional< std::filesystem::path > biases)
	Creates an initialization strategy that uses weights loaded from a npy file. More...
std::shared_ptr< WeightInitializationStrategy >	create_feature_mean_initializer (std::shared_ptr< DatasetBase > data, real_t pos=1, real_t neg=-2)
	Creates an initialization strategy based on the mean of positive and negative features. More...
std::shared_ptr< WeightInitializationStrategy >	create_multi_pos_mean_strategy (std::shared_ptr< DatasetBase > data, int max_pos, real_t pos=1, real_t neg=-2)
	Creates an initialization strategy based on the mean of positive and negative features. More...
std::shared_ptr< WeightInitializationStrategy >	create_ova_primal_initializer (const std::shared_ptr< DatasetBase > &data, RegularizerSpec regularizer, LossType loss)

Function Documentation

create_constant_initializer()

std::shared_ptr< WeightInitializationStrategy > dismec::init::create_constant_initializer

(

DenseRealVector

vec

)

Creates an initialization strategy that initializes all weight vectors to the given vector. TODO allow both dense and sparse vectors.

Definition at line 56 of file constant.cpp.

Referenced by create_ova_primal_initializer(), TrainingProgram::make_config(), and register_init().

create_feature_mean_initializer()

std::shared_ptr< WeightInitializationStrategy > dismec::init::create_feature_mean_initializer	(	std::shared_ptr< DatasetBase >	data,
		real_t	pos = 1,
		real_t	neg = -2
	)

Creates an initialization strategy based on the mean of positive and negative features.

Definition at line 90 of file msi.cpp.

Referenced by TrainingProgram::make_config(), and register_init().

create_multi_pos_mean_strategy()

std::shared_ptr< WeightInitializationStrategy > dismec::init::create_multi_pos_mean_strategy	(	std::shared_ptr< DatasetBase >	data,
		int	max_pos,
		real_t	pos = 1,
		real_t	neg = -2
	)

Creates an initialization strategy based on the mean of positive and negative features.

Definition at line 212 of file multi_pos.cpp.

Referenced by TrainingProgram::make_config(), and register_init().

create_numpy_initializer()

std::shared_ptr< WeightInitializationStrategy > dismec::init::create_numpy_initializer	(	const std::filesystem::path &	weights,
		std::optional< std::filesystem::path >	biases
	)

Creates an initialization strategy that uses weights loaded from a npy file.

Definition at line 58 of file numpy.cpp.

References dismec::io::load_matrix_from_npy().

Referenced by TrainingProgram::make_config().

create_ova_primal_initializer()

std::shared_ptr< WeightInitializationStrategy > dismec::init::create_ova_primal_initializer	(	const std::shared_ptr< DatasetBase > &	data,
		RegularizerSpec	regularizer,
		LossType	loss
	)

Creates an initialization strategy based on

Huang Fang et al. “Fast training for large-scale one-versus-all linear classi- fiers using tree-structured initialization”. In: Proceedings of the 2019 SIAM International Conference on Data Mining.

Definition at line 15 of file ova-primal.cpp.

References create_constant_initializer(), dismec::objective::LinearClassifierBase::get_label_ref(), dismec::make_loss(), dismec::objective::make_regularizer(), and dismec::types::visit().

Referenced by TrainingProgram::make_config(), and register_init().

create_pretrained_initializer()

std::shared_ptr< WeightInitializationStrategy > dismec::init::create_pretrained_initializer

(

std::shared_ptr< model::Model >

model

)

Creates an initialization strategy that uses an already trained model to set the initial weights.

Definition at line 48 of file pretrained.cpp.

create_zero_initializer()

std::shared_ptr< WeightInitializationStrategy > dismec::init::create_zero_initializer

(

)

Creates an initialization strategy that initializes all weight vectors to zero.

Definition at line 33 of file zero.cpp.

Referenced by dismec::create_cascade_training(), dismec::create_dismec_training(), and register_init().