Trainers Module

The trainers module provides training utilities and callback management.

class xflow.trainers.BaseTrainer(model: Any, data_pipeline: Any, output_dir: str, *, callbacks: List[Callback] | None = None, model_io: ModelIO | None = None, config: Dict[str, Any] | None = None)

Bases: ABC

Thin orchestrator: runs loops, dispatches callbacks, collects history. Creation/compilation of model/optimizer is outside; inject everything in.

__init__(model: Any, data_pipeline: Any, output_dir: str, *, callbacks: List[Callback] | None = None, model_io: ModelIO | None = None, config: Dict[str, Any] | None = None)

save_history(path: str | None = None)

save_model(path: str | None = None, **extra)

abstractmethod fit(*, epochs: int, train_loader: Iterable | None = None, val_loader: Iterable | None = None) → Dict[str, List[Any]]

abstractmethod predict(loader: Iterable, **kwargs) → Any

class xflow.trainers.CallbackRegistry

Bases: object

Registry for callback handlers or factories.

classmethod register(name: str)

classmethod get_handler(name: str) → Callable

classmethod list_handlers() → List[str]

xflow.trainers.build_callbacks_from_config(config: List[Dict[str, Any]], framework: str, name_key: str = 'name', params_key: str = 'params') → List[Any]

Build a list of callbacks (native or unified) from a config list.

Parameters:

• config – List of callback config dicts. Each dict should have at least a ‘name’ key and optionally a ‘params’ dict.

• framework – Which framework to use (‘tf’, ‘pl’, or ‘torch’).

• name_key – Key in each config dict for the callback/factory name (default: ‘name’).

• params_key – Key in each config dict for the callback/factory parameters (default: ‘params’).

Returns:

List of instantiated callback objects.

Each config entry may either: 1) Define only ‘name’ + ‘params’ → handler must return a Callback instance (native/factory style) 2) Define ‘events’ (list of {event, name, params}) → use unified wrapper for event-based callbacks