deeppavlov.core.models¶

Abstract model classes and interfaces.

class deeppavlov.core.models.component.Component[source]¶: Abstract class for all callables that could be used in Chainer’s pipe.

class deeppavlov.core.models.serializable.Serializable(save_path: Optional[Union[pathlib.Path, str]], load_path: Optional[Union[pathlib.Path, str]] = None, mode: str = 'infer', *args, **kwargs)[source]¶: Abstract base class that expresses the interface for all models that can serialize data to a path.

class deeppavlov.core.models.estimator.Estimator(save_path: Optional[Union[pathlib.Path, str]], load_path: Optional[Union[pathlib.Path, str]] = None, mode: str = 'infer', *args, **kwargs)[source]¶: Abstract class for components that could be fitted on the data as a whole.

class deeppavlov.core.models.nn_model.NNModel(save_path: Optional[Union[pathlib.Path, str]], load_path: Optional[Union[pathlib.Path, str]] = None, mode: str = 'infer', *args, **kwargs)[source]¶: Abstract class for deep learning components.

class deeppavlov.core.models.lr_scheduled_model.LRScheduledModel(learning_rate: Union[None, float, Tuple[float, float]] = None, learning_rate_decay: Union[str, deeppavlov.core.models.lr_scheduled_model.DecayType, Tuple[Union[str, deeppavlov.core.models.lr_scheduled_model.DecayType], float]] = <DecayType.NO: 1>, learning_rate_decay_epochs: int = 0, learning_rate_decay_batches: int = 0, learning_rate_drop_div: float = 2.0, learning_rate_drop_patience: Optional[int] = None, momentum: Union[None, float, Tuple[float, float]] = None, momentum_decay: Union[str, deeppavlov.core.models.lr_scheduled_model.DecayType, Tuple[Union[str, deeppavlov.core.models.lr_scheduled_model.DecayType], float]] = <DecayType.NO: 1>, momentum_decay_epochs: int = 0, momentum_decay_batches: int = 0, fit_batch_size: Union[None, int, str] = None, fit_learning_rate: Tuple[float, float] = (1e-07, 100), fit_learning_rate_div: float = 10.0, fit_beta: float = 0.98, fit_min_batches: int = 10, fit_max_batches: Optional[int] = None, load_before_drop: bool = False, *args, **kwargs)[source]¶

Abstract model enhanced with optimizer, learning rate and momentum management and search.

Parameters

learning_rate – learning rate value or ranges
learning_rate_decay – learning rate decay type. Set of values: “linear”, “onecycle”, “trapezoid”, “exponential”, “cosine”, [“polynomial”, K], where K is a polynomial power
learning_rate_decay_epochs – number of epochs for learning rate decay process
learning_rate_decay_batches – number of batches for learning rate decay process
learning_rate_drop_div – division coefficient for learning rate in case of exceeding patience learning_rate_drop_patience
learning_rate_drop_patience – patience limit of loss increase
momentum – range of momentum values
momentum_decay – momentum decay type. Set of values: “linear”, “onecycle”, “trapezoid”, “exponential”, “cosine”, [“polynomial”, K], where K is a polynomial power
momentum_decay_epochs – number of epochs for momentum decay process
momentum_decay_batches – number of batches for momentum decay process
fit_batch_size – batch size when fitting learning rate
fit_learning_rate – range of learning rate values to explore
fit_learning_rate_div – division coefficient for best learning rate obtained from fitting, divided learning rate value will be used when training model
fit_beta – smoothing coefficient for loss calculation when fitting learning rate
fit_min_batches – number of batches to train model on before fitting learning rate
fit_max_batches – number of batches to train model on when fitting learning rate
load_before_drop – set True to load saved model from disk when learning rate is dropped, set False to continue training current model
*args – other parameters
**kwargs – other parameters