Task naming confusion about autogluon

[jwmueller]

FYI Matthias, task in autogluon is current a very special frontend term reserved to refer to high-level applications like: image_classification, text_classification, object_detection, predict_table_column. Once the task is specified by user, autogluon makes many choices on behalf of the user to make things as easy as possible for ML novices.

So I'm not sure task should be the name for any user-accessible internal autogluon variables, since it's reserved for one of the most important user concepts.

If there still ends up being a naming conflict with trial, here's some random name options:
job, training_job, training_run, config_evaluation, config_trial, score_config, search_iter

[mseeger]

Hello,

we will contribute a scheduler (Hyperband in the ASHA sense) for which tasks do not equate to trials. There, a task is moving a trial (or config evaluation) forward, but a trial may be driven by a sequence of tasks, being paused in between.

If you now call a task a trial, then the meaning of num_trials is not correct anymore.

[zhanghang1989]

Hello,

we will contribute a scheduler (Hyperband in the ASHA sense) for which tasks do not equate to trials. There, a task is moving a trial (or config evaluation) forward, but a trial may be driven by a sequence of tasks, being paused in between.

If you now call a task a trial, then the meaning of num_trials is not correct anymore.

Thanks Matthias. This is a good point.

Currently, we have naming issue with backend 'Task' and frontend 'task' (applications, such as image classification). We may need to think about a better name.

[cgraywang]

As discussed with @jwmueller and @zhanghang1989 , we could decide between trial and job. Thanks.

[mseeger]

I'd vote for job then, because trial has a different meaning in schedulers and searchers.

[jwmueller]

addressed in the new core API PR: #31

In all tutorial/documentations / text written for the user, please reserve the term "trial" to refer to this concept (ie. evaluating performance of trained model under a single setting of the hyperparameter values). It is fine if "trial" is not used to name this variable in the code (or as the name of a different variable as in hyperband); but we need to ensure all user-facing writing is consistent and uses "trial" if and only if referring to this concept.