Engine is a coordinator for running algorithms and resolving ordering conflicts among them for composition.

The order in which algorithms are run matters significantly during composition. For example, the SelectiveBackprop algorithm runs on the Event.AFTER_DATALOADER event and must run before any data augmentations. Engine runs re-ordering passes to resolve such ordering issues or conflicts.


  • An instance of Engine is automatically constructed by the Trainer constructor. A user need not instantiate the Engine class.

  • The design of Engine is subject to change in future releases to accommodate more complexity as we investigate composition of algorithms.

Currently, the following passes are registered:

  • LIFO order for events

    For the events that follow the before_* (e.g., Event.BEFORE_LOSS) and after_* (e.g., Event.AFTER_LOSS) pattern, the ordering of algorithms is reversed for the after_* events. For example, four given algorithms A, B, C, and D will run in ABCD ordering on the before_* event while DCBA ordering on the after_* event.

    This allows algorithms to โ€œclean upโ€ their changes. For example, LabelSmoothing will smooth the labels upon the Event.BEFORE_LOSS event and then restore the original unsmoothed labels on the Event.AFTER_LOSS event.

  • Run Selective Backprop first

    SelectiveBackprop runs after the dataloader returns the batch and executes an extra forward pass to rank and prune the examples in the batch by loss. To ensure a clean estimate of loss, SelectiveBackprop should run before any other data augmentations (e.g., MixUp) on the Event.AFTER_DATALOADER event.


Traces record whether an algorithm ran at a particular step and event combination and also the order of such executions. These are logged with the key <algorithm_name>/<event>.

For example, the algorithm LayerFreezing, which runs at the end of every epoch on Event.EPOCH_END, will emit a series of traces:

[STEP=3][layer_freezing/EPOCH_END=1]  # <-- layer freezing ran on step 3 here!



Coordinator for running algorithms and resolving ordering conflicts among them for composition.


Record of an algorithm's execution.


class composer.core.engine.Engine(state, logger)[source]#

Coordinator for running algorithms and resolving ordering conflicts among them for composition.

  • state (State) โ€“ The initial State of the trainer. state will be modified in-place.

  • logger (Logger) โ€“ A Logger instance to be used for logging algorithm and callback specific metrics.


Shutdown the engine.

As part of the shutdown procedure, Callback.close() and Callback.post_close() are invoked for each callback. Note that Callback.post_close() is invoked only for callbacks that did not raise an exception during Callback.close().

This method does not re-raise any exceptions from Callback.close() and Callback.post_close(). Instead, these exceptions are logged as errors.


Runs the sequence of algorithms and callbacks (see Callback).

Filters algorithms by calling each oneโ€™s Algorithm.match() method, internally checks for conflicting algorithms, then runs each algorithmโ€™s Algorithm.apply() method to make in-place changes to the state.

The default order of execution for algorithms is determined by the provided list. However, Engine makes changes to this order internally to resolve ordering conflicts.

Returns Traces of the execution, a dictionary with keys formatted as <algorithm_name>/<event> (e.g., Blurpool/INIT), and values are an instance of Trace.

Callbacks are always run after algorithms and do not return a trace.

This method can be called with either the Event enum member values or a string of the event name.


>>> engine = Engine(state, logger)
>>> engine.run_event(Event.BEFORE_LOSS)
>>> # calling with a string of the event name also works
>>> engine.run_event('before_loss')

event (Event | str) โ€“ The current Event. It can be the enum member values or a string with the event value.


traces (Traces) โ€“ Ordered dictionary of trace for each algorithm.

class composer.core.engine.Trace(name='', event=None, exit_code=None, order=None, run=False)[source]#

Record of an algorithmโ€™s execution.


The name of the algorithm.




The current event.




Optional return value from an algorithm. Default: None.


int | None


Order in which the algorithm was executed in the list of algorithms. None means algorithm was not run.


int | None


Whether the algorithm was run. Default: False




The default traces of an entire run is an OrderedDict. The keys are of format <algorithm_name>/<event> (e.g., Blurpool/INIT) and values are an instance of Trace.

alias of Dict[str, Trace]