Model Versioning And Registry

Created: 2023-02-13 15:18
#note

ML models evolve over time as new training data and different hyperparameters are used.

Benefits:

• linking with training data versions;
• linking with model hyperparameters;
• performance improvement analytics;
• integration and testing;
• model life cycle management

A model registry is a database containing a model and metadata associated to it. Contents:

• serialized model;
• model ID, version, description;
• hyperparameters;
• model status;
• model history;
• link to input data versions, code versions and performance results.

Thanks to the registry we obtain traceability, reproducibility, life cycle management, automation, governance, and we facilitate communication. Tools like MLflow provide comprehensive model registry capabilities.

Tags

#mlops #ml