Monitoring consumption on single call inside loop?

[Kominaru]

Hi to the community,

CodeCarbon's high-level functions seem to become a computational bottleneck when trying to isolatedly monitor a single inexpensive function inside a loop with many iterations

For instance, this is common in streaming/online ML tasks:

tracker = codecarbon.OfflineEmissionsTracker()

for instance in tqdm.tqdm(stream):
    prediction = ob_learner.predict(instance)

    tracker.start_task() 
    ob_learner.train(instance)
    tracker.stop_task()

    ob_evaluator.update(instance.y_index, prediction)

Adding the start_task and stop_task calls in the snippet above increases the execution runtime by a factor of 1000 (from a few seconds, to several days). Using lower-level or private methods like flush() and _do_measurements() also adds this overhead.

Is there any way to do this efficiently, even if it's hacky or more manual?

Cheers,