treeverse / treeverse/dvc

dvc experiments checkout: option to select the best experiment regarding one metric

Open
#4,283 2 comments 3 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

A: experiments p3-nice-to-have ui
Dominant language
Python
Stars
15.9k
Forks
1.3k
PR merge metrics
No merged PRs in 30d

Description

I had an idea while trying the new experiment feature. For the use case of grid-search or hyper-parameter optimization, could we imagine an option for dvc experiments checkout to checkout the best experiment regarding one of the metrics? Something like dvc experiments checkout --best metrics.json:accuracy. With an option like this, you can run (manually or automatically) a lot of experiments and ask dvc to return the best experiment, which can be then commited to git. Or instead of --best, it could be --lowest or --highest to take into account that some metrics should be minimized and some others maximized.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with the dvc experiments checkout entry point and trace how experiment metrics are selected and compared. Clarify whether the interface should support --best, --lowest, or --highest, then define behavior for metrics.json:accuracy and metric direction before implementing tests and the checkout result.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
cli, machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.