Engine

Events

class argus.engine.Events(value, names=None, *, module=None, qualname=None, type=None, start=1, boundary=None)

Events that are fired by the argus.engine.Engine during running.

Built-in events:

• START: triggered when the engine’s run is started.

• COMPLETE: triggered when the engine’s run is completed.

• EPOCH_START: triggered when the epoch is started.

• EPOCH_COMPLETE: triggered when the epoch is ended.

• ITERATION_START: triggered when an iteration is started.

• ITERATION_COMPLETE: triggered when the iteration is ended.

• CATCH_EXCEPTION: triggered on catching of an exception.

class argus.engine.EventEnum(value, names=None, *, module=None, qualname=None, type=None, start=1, boundary=None)

Base class for engine events. User-defined custom events should also inherit this class. An example of creating custom events is available here.

State

class argus.engine.State(step_method: Callable[[Any, State], Any], engine: Engine | None = None, phase_states: Dict[str, State] | None = None, **kwargs)

A state used to store internal and user-defined variables during a run of argus.engine.Engine. The class is highly inspired by the State from pytorch-ignite.

Parameters:

• step_method (Callable) – Method of argus.model.Model that takes batch, state and returns step output.

• engine (Engine, optional) – argus.engine.Engine that uses this object as a state.

• phase_states (dict, optional) – Dictionary with states for each training phase.

• **kwargs – Initial attributes of the state.

By default, the state contains the following attributes.

iteration

Iteration, the first iteration is 0.

Type:

int

epoch

Epoch, the first iteration is 0.

Type:

int

model

argus.Model that uses argus.engine.State.engine and this object as a state.

Type:

argus.Model

data_loader

A data passed to the argus.engine.Engine.

Type:

Iterable, optional

logger

Logger.

Type:

logging.Logger

exception

Catched exception.

Type:

BaseException, optional

engine

argus.engine.Engine that uses this object as a state.

Type:

Engine, optional

phase

A phase of training this state was created for. The value takes from the name of the method step_method. If the step_method name ends with _step, the postfix will be removed. For default steps of argus model values are ‘train’ and ‘val’.

Type:

str

phase_states

Dictionary with states for each training phase.

Type:

dict, optional

batch

Batch sample from a data loader on the current iteration.

Type:

Any

step_output

Current output from step_method on current iteration.

Type:

Any

metrics

Dictionary with metrics values.

Type:

dict

stopped

Boolean indicates argus.engine.Engine is stopped or not.

Type:

bool

update(**kwargs)

Update state attributes.

Parameters:

**kwargs – Update attributes using kwargs

Engine

class argus.engine.Engine(step_method: Callable[[Any, State], Any], phase_states: Dict[str, State] | None = None, **kwargs)

Runs step_method over each batch of a data loader with triggering event handlers. The class is highly inspired by the Engine from pytorch-ignite.

Parameters:

• step_method (Callable) – Method of argus.model.Model that takes batch, state and returns step output.

• phase_states (dict, optional) – Dictionary with states for each training phase.

• **kwargs – Initial attributes of the state.

state

Stores internal and user-defined variables during a run of the engine.

Type:

State

step_method

Method of argus.model.Model that takes batch, state and returns step output.

Type:

Callable

event_handlers (dict of EventEnum

list): Dictionary that stores event handlers.

add_event_handler(event: EventEnum, handler: Callable, *args, **kwargs)

Add an event handler to be executed when the event is triggered.

Parameters:

• event (EventEnum) – An event that will be associated with the handler.

• handler (Callable) – A callable handler that will be executed on the event. The handler should take argus.engine.State as the first argument.

• *args – optional args arguments to be passed to the handler.

• **kwargs – optional kwargs arguments to be passed to the handler.

raise_event(event: EventEnum)

Execute all the handlers associated with the given event.

Parameters:

event (EventEnum) – An event that will be triggered.

run(data_loader: Iterable, start_epoch: int = 0, end_epoch: int = 1) → State

Run step_method on each batch from data loader end_epoch - start_epoch times.

Parameters:

• data_loader (Iterable) – An iterable collection that returns batches.

• start_epoch (int) – The first epoch number.

• end_epoch (int) – One above the largest epoch number.

Returns:

An engine state.

Return type:

State