Skip to content

Latest commit

 

History

History
72 lines (39 loc) · 2.16 KB

3-online-reproducibility.md

File metadata and controls

72 lines (39 loc) · 2.16 KB
Raw

Online Reproducibility

Add new secret PERSONAL_GITHUB_TOKEN.

  • Create a personal access token

https://docs.github.com/en/authentication/keeping-your-account-and-data-secure/creating-a-personal-access-token

  • Create a new secret and name it PERSONAL_GITHUB_TOKEN

Grant GitHub access to DVC Remote

You need to grant GitHub access to the DVC Remote.

Get the credentials.

cat ".dvc/tmp/gdrive-user-credentials.json"

And create a new GitHub secret called GDRIVE_CREDENTIALS_DATA to store them.

With this, GitHub runners will be able to pull and push all the changes generated by the pipeline.

Pull Request workflow

You can create a new GitHub actions workflow that runs when a new Pull Request is created.

This workflow will use DVC to reproduce the pipeline and update the large artifacts tracked by DVC.

In addition it will use CML to post a report with the DVC metrics, params, and plots (cml send-comment). It will also update the artifacts tracked by Git (cml pr)

Report Metrics

Report Plots

Create and fill `.github/workflows/on_pr.yml`

https://github.com/iterative/workshop-uncool-mlops-solution/blob/main/.github/workflows/on_pr.yaml

Reproduce Online

And now you can reproduce the pipeline from the web:

From GitHub UI

  • Edit params.yaml from the GitHub Interface.

  • Change train.epochs.

  • Select Create a new branch for this commit and start a pull request

From Studio

More info: https://dvc.org/doc/studio

  • Click on Run new experiment button.

More compute

In the above workflow we are using the default GitHub runners to train our model.

While this is enough for our use case (small dataset, small model), your project would often require more compute resources.

CML Self-Hosted Runners allows you to allocate cloud instances (or on-premise machines) and use them in your GitHub actions workflow.