feature request: allow setting the current step without calling `logger.report_*`

Currently, automatic reporting (e.g. CPU / GPU / memory utilization) relies on the last reported step when logger.report_* was called. If that doesn’t happen for some time, it falls back to using milliseconds since start for the x-axis: https://github.com/allegroai/clearml/blob/5a9155b2039413280f13dfded1121470c4c4323d/clearml/utilities/resource_monitor.py#L110-L111

This is problematic in a few scenarios:

1. the user doesn’t want to artifically create dummy reports and the first desired report (if any) would take longer than the time-out
2. the worker has died and gets restarted. This is affected by #439.

In both cases, it would be nice to be able to do something like

clearml_task.set_iteration(current_iteration)

If the worker resumes and recovers (externally, maybe from a checkpoint on disk) from iteration 433, we call

clearml_task.set_iteration(433)

Or in the default use-case:

for iteration in range(max_steps):
    clearml_task.set_iteration(iteration)
    # do some work
    loss = train_step()
    if iteration % 1000:  # very infrequent reporting
        clearml_task.report_scalar(loss, "loss", iteration)

Links:

• automatic logging is configured through Task.init

[bmartinn]

Regarding contribution: could you point me to where I should start looking and contribution instructions.

https://github.com/allegroai/clearml/blob/master/docs/contributing.md

I can only work on client-side code.

Client SDK will be great! (server-side is definitely more complicated to PR to 😃

[patzm]

It is slightly different. If we follow through with your proposal in https://github.com/allegroai/clearml/issues/439#issuecomment-907882223 (either (1) or (2)), then I think we would implement two methods:

1. set_last_checkpoint
2. set_iteration: this is like reporting without a reporting payload.

And maybe also getters for the two.

Regarding contribution: could you point me to where I should start looking and contribution instructions. Ideally I can only work on client-side code. I think setting up the server backend as well would be too much. I don’t have much experience with it and also not a lot of time 😉.