open-mmlab/mmengine

[Feature] A Checkpoint hook for saving model checkpoints as Weights & Biases Artifacts

Opened this issue · 0 comments

What is the feature?

A checkpoint hook similar to CheckpointHook that enables us to save checkpoints periodically as Weights & Biases Artifacts.

Any other context?

Storing model checkpoints as W&B artifacts offers several advantages:

  • Version Control: Artifacts allow you to version your model checkpoints, making it easy to track changes and revert to previous versions if necessary.
  • Reproducibility: By storing model checkpoints alongside your runs, you can easily reproduce experiments at any point in time.
  • Centralization: The Model Registry centralizes your best models, facilitating easy access and management across your team.
  • Lifecycle Management: Manage the lifecycle of your models within the W&B ecosystem, from development to deployment.
  • Automation: Automate downstream actions such as model deployment or further analysis through webhooks or jobs linked to your artifacts.