Observers¶

Olympus use a set of Observer objects to monitor a training in progress. This allow Olympus to keep the training code short and focused on its purpose.

For example, all metrics, validation, online accuracy, speed are computed through observer, making that logic easily reusable for different tasks.

Available Observers¶

More Observers are defined as metrics.

Progress

Observer Interface¶

olympus.observers.observer.Metric¶: alias of olympus.observers.observer.Observer

class olympus.observers.observer.Observer(frequency_new_epoch: int = 0, frequency_new_batch: int = 0, frequency_new_trial: int = 0, priority: int = 0) → None[source]¶

Bases: object

Metrics are observers that receives events periodically

Attributes:	frequency_epoch: int Controls how often `on_new_epoch` is called, 0 disables it frequency_batch: int Controls how often `on_new_batch` is called, 0 disables it frequency_trial: int Controls how often `on_new_trial` is called, 0 disables it priority: int Controls which metric is called first

Methods

`every`(*args[, epoch, batch])	Define how often this metric should be called
`load_state_dict`(state_dict)	Load a state dictionary to resume a previous training
`on_end_train`(task[, step])	Called at the end of training after the last epoch
`on_new_batch`(task, step[, input, context])	Called after a batch has been processed
`on_new_epoch`(task, epoch, context)	Called at the end of an epoch, before a new epoch starts
`on_new_trial`(task, step, parameters, uid)	Called after a trial has been processed
`on_start_train`(task[, step])	Called on ce the training starts
`state_dict`()	Return a state dictionary used to checkpointing and resuming
`value`()	Return the key values that metrics computes

every(*args, epoch=None, batch=None)[source]¶: Define how often this metric should be called

frequency_new_batch = 0¶

frequency_new_epoch = 0¶

frequency_new_trial = 0¶

load_state_dict(state_dict)[source]¶: Load a state dictionary to resume a previous training

on_end_train(task, step=None)[source]¶: Called at the end of training after the last epoch

on_new_batch(task, step, input=None, context=None)[source]¶: Called after a batch has been processed

on_new_epoch(task, epoch, context)[source]¶: Called at the end of an epoch, before a new epoch starts

on_new_trial(task, step, parameters, uid)[source]¶: Called after a trial has been processed

on_start_train(task, step=None)[source]¶

Called on ce the training starts

Notes

You should not rely on this function to initialize your metric as it will not be called if the training is resumed from a previous state

priority = 0¶

state_dict()[source]¶: Return a state dictionary used to checkpointing and resuming

value()[source]¶: Return the key values that metrics computes

Observer List¶

olympus.observers.MetricList¶: alias of olympus.observers.ObserverList

class olympus.observers.ObserverList(*args, task=None, name=None)[source]¶

Bases: object

MetricList relays the Event to the Metrics/Observers

Methods

`append`(m[, key])	Insert a new metric to compute
`end_batch`(step[, input, context])	Broadcast a `new_batch` event to all metrics
`end_epoch`(epoch[, context])	Broadcast a `new_epoch` event to all metrics
`end_train`()	Broadcast a finish event to all metrics
`get`(key[, default])	Retrieve a metric from its key
`load_state_dict`(state_dict[, strict])	Resume all children metrics using a state_dict
`new_batch`(step[, input, context])	Broadcast a `new_batch` event to all metrics
`new_epoch`(epoch[, context])	Broadcast a `new_epoch` event to all metrics
`new_trial`(parameters, uid)	Broadcast a `new_trial` event
`report`([pprint, print_fun])	Pretty prints all the metrics
`resume_train`(start_epoch)	Broadcast a resume event to all metrics
`start_train`()	Broadcast a `start` event to all metrics
`state_dict`([destination, prefix, keep_vars])	Save all the children states
`value`()	Returns a dictionary of all computed metrics

broadcast_event
should_run

append(m: olympus.observers.observer.Observer, key=None)[source]¶

Insert a new metric to compute

Parameters:	m: Metric new metric to insert key: Optional[str] optional key used to retrieve the metric by default the type name will be used as key

broadcast_event(event_name, task, step, *args, **kwargs)[source]¶

end_batch(step, input=None, context=None)[source]¶: Broadcast a new_batch event to all metrics

end_epoch(epoch, context=None)[source]¶: Broadcast a new_epoch event to all metrics

end_train()[source]¶: Broadcast a finish event to all metrics

get(key, default=None)[source]¶

Retrieve a metric from its key

Parameters:	key: Union[str, int] default: any default object returned if not found

load_state_dict(state_dict, strict=True)[source]¶: Resume all children metrics using a state_dict

new_batch(step, input=None, context=None)[source]¶: Broadcast a new_batch event to all metrics

new_epoch(epoch, context=None)[source]¶: Broadcast a new_epoch event to all metrics

new_trial(parameters, uid)[source]¶: Broadcast a new_trial event

report(pprint=True, print_fun=<built-in function print>)[source]¶: Pretty prints all the metrics

resume_train(start_epoch)[source]¶: Broadcast a resume event to all metrics

static should_run(name, step)[source]¶

start_train()[source]¶: Broadcast a start event to all metrics

state_dict(destination=None, prefix='', keep_vars=False)[source]¶: Save all the children states

value()[source]¶: Returns a dictionary of all computed metrics