aim icon indicating copy to clipboard operation
aim copied to clipboard

Filtering runs by the number of steps/epochs

Open avkudr opened this issue 3 years ago • 2 comments

🚀 Feature

As a user, I would like to be able to filter the runs in metrics explorer by the number of steps/epochs.

Motivation

I have a lot of runs and some of them have only a few steps because I was only debugging the scripts. So ignoring the runs with a number of steps fewer than X would be helpful.

Pitch

Enter something like (run.steps > 100) in the filter of metrics explorer and only see the runs that lasted more than that

Alternatives

In my particular case, the alternative would be to select all runs that performed less than 100 steps and delete(archive) them all at once

Thanks for considering!

avkudr avatar Jan 18 '22 20:01 avkudr

@avkudr thanks for opening this issue, it seems super useful 🙌

I think if steps count will be stored inside runs metadata during the training(along with other properties of runs), it will enable searching runs by steps count both from explorers and programmatically via SDK.

But the thing is that usually metrics have different lengths, e.g. train loss has more steps than validation loss. Hence, I think the proposed query syntax could be modified a bit. For example: run["loss", {"subset": "train"}].steps > 100. Searching by metric last step could be enabled as well e.g. run["loss", {"subset": "train"}].value.last <= 0.001.

However, I believe the above mentioned syntax is too complicated(and ugly :D). @avkudr @roubkar @alberttorosyan @mahnerak Any better alternatives that come to your mind?

gorarakelyan avatar Jan 18 '22 22:01 gorarakelyan

@gorarakelyan I have the following suggestion. since metric in the context of the query refers to a unique metric sequence, steps count can be added as a property of the metric. So the syntax would be like:

metric.steps > 100

There's yet one thing to clarify. Since metrics are sparse, should we take the last step or the number of tracked values?

alberttorosyan avatar Jan 19 '22 06:01 alberttorosyan