Improve Frequency

🚀 Feature

If we would like to log datapoints/second every 100 iterations, we most probably do like this

wps_metric = Frequency(output_transformer=lambda x: x['ntokens'])
wps_metric.attach(trainer, name='wps', event_name=Events.ITERATION_COMPLETED(every=100))

however, seems like this wont take into account all other iterations while computing the total number of tokens.

class Frequency(Metric):
    ....
    def attach(self, engine, name, event_name=Events.ITERATION_COMPLETED):
        engine.add_event_handler(Events.EPOCH_STARTED, self.started)
        engine.add_event_handler(event_name, self.iteration_completed)
        engine.add_event_handler(event_name, self.completed, name)

IMO, should be

class Frequency(Metric):
    ....
    def attach(self, engine, name, event_name=Events.ITERATION_COMPLETED):
        engine.add_event_handler(Events.EPOCH_STARTED, self.started)
        engine.add_event_handler(Events.ITERATION_COMPLETED, self.iteration_completed)
        engine.add_event_handler(event_name, self.completed, name)

cc @erip

[erip]

Not sure how I missed the cc here - sorry about that! Let me reacquaint myself quickly and I’ll try to submit a PR.

[vfdev-5]

@erip I’m on Mac too, the following test passes on master on my computer.

Anyway, please send the PR and we can see what happens. I think we also need to improve the docs and explain better the purpose of event_name argument.