Model Management
Use the W&B Model Registry as a central system of record for models.
- Create Registered Models to organize your best model versions for a given task
- Track a model moving into staging and production
- See a history of all changes, including who moved a model to production
Model Registry Quickstart
Open your Model Registry and create a registered model.
In your script, log a model as an artifact version.
art = wandb.Artifact("my-object-detector", type="model")
art.add_file("saved_model_weights.pt")
wandb.log_artifact(art)From the Artifact page, link the artifact version to the registry.
Model Registry Features
Model Versioning
Iterate to get the best model version for a task, and catalog all the changes along the way.
- Track every model version in a central repository
- Browse and compare model versions
- Capture training metrics and hyperparameters
Model Lineage
Document and reproduce the complete pipeline of model training and evaluation.
- Identify the exact dataset version the model trained on
- Restore the training code, including git commit and diff patch
- Get back to the model’s hyperparameters and other metadata for reproducibility
- Dig in to upstream jobs that can affect model performance
Model Lifecycle
Manage the process as a model moves from training through staging to production.
- Highlight the best model versions that are being evaluated for production
- Communicate where a model version is in the process — staging, production etc
- Review the history of model versions that moved through each stage
Model Registry Pilot Limits
This new feature is now turned on for all users to try for free, up to:
- 5 Registered Models, with unlimited versions linked to each model
- 10 most recent steps of Action History shown in the UI for each registered model
Explore more
- Read Model Management Concepts for more information on basic Model Management concepts.
- Follow the steps in the Walkthrough to learn how to use Model Management.