State

class torchtnt.framework.state.State(*, entry_point: EntryPoint, timer: Optional[TimerProtocol] = None, train_state: Optional[PhaseState] = None, eval_state: Optional[PhaseState] = None, predict_state: Optional[PhaseState] = None)

Parent State class which can contain up to 3 instances of PhaseState, for the 3 phases. Modified by the framework, read-only for the user.

property active_phase: ActivePhase

Current active phase of the loop. (One of TRAIN, EVALUATE, PREDICT).

property entry_point: EntryPoint

Entry point used to start loop execution. (One of FIT, TRAIN, EVALUATE, PREDICT).

property eval_state: Optional[PhaseState]

A PhaseState object which contains meta information about the eval phase.

property predict_state: Optional[PhaseState]

A PhaseState object which contains meta information about the predict phase.

property should_stop: bool

Read-only property for whether to terminate the loop after the current step completes.

stop() → None

Signal to the loop to end after the current step completes.

property timer: Optional[TimerProtocol]

A TimerProtocol object which can be used for debugging to record latencies of key events during loop execution.

property train_state: Optional[PhaseState]

A PhaseState object which contains meta information about the train phase.

PhaseState

class torchtnt.framework.state.PhaseState(*, dataloader: Iterable[Any], max_epochs: Optional[int] = None, max_steps: Optional[int] = None, max_steps_per_epoch: Optional[int] = None, evaluate_every_n_steps: Optional[int] = None, evaluate_every_n_epochs: Optional[int] = None)

State for each phase (train, eval, predict). Modified by the framework, read-only for the user.

property dataloader: Iterable[Any]

Dataloader defined by the user.

property evaluate_every_n_epochs: Optional[int]

Frequency with which to evaluate in terms of training epochs, when running fit(). Defined by the user.

property evaluate_every_n_steps: Optional[int]

Frequency with which to evaluate in terms of training steps, when running fit(). Defined by the user.

property iteration_timer: TimerProtocol

An always-on TimerProtocol object which contains CPU timings (without synchronisation) of the iterations.

property max_epochs: Optional[int]

Maximum number of epochs to train, defined by the user.

property max_steps: Optional[int]

Maximum number of steps to train, defined by the user.

property max_steps_per_epoch: Optional[int]

Maximum number of steps to run per epoch, defined by the user.

property step_output: Any

Output of the last step.