Model Versioning: Why, When, and How
summary
Argues that production ML needs true model versioning and describes ModelDB and Verta workflows that track the model artifact plus code, data, config and environment so models are reproducible and rollbacks are possible. Shows a demo versioning S3 datasets by checksum, logging git SHAs and requirements, locking hyperparameters, comparing model diffs and metrics, and exposing a model API spec for monitoring and data drift detection. Notes top Kaggle teams build 100+ model iterations and recommends a git plus database approach.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 2:07Why productionizing AI is hard
- 6:20Why model versioning matters now
- 9:51Code vs model versioning
- 13:44When to start versioning models
- 15:29Demo of model versioning
- 22:31Benefits of model versioning
- 23:12Summary and Q&A
related talks
32mReproducibility and versioning of ML systems
32mMachine Learning Models Versioning Using Open Source Tools
24mData Versioning Explained the Open Source Way
36mData Versioning - does it mean what we think it means?
01h 25mBest practices and solution design using data versioning for machine learning
59mDevOps for Machine Learning Projects
33mDealing with the versioning of production-ready models
31mGetting ML Right in a Complex Data World