Better Care and Feeding of Machine Learning Models
summary
Defines reproducibility as data plus code plus documentation and shows a working project template that uses cookiecutter-data-science, scripted downloads and processing, and an isolated environment (virtualenv/conda) so running a few scripts delete artifacts and rerun to recreate analysis actually rebuilds figures and trained models. Demonstrates containerizing models, tracking data provenance (examples like DVC/Pachyderm mentioned), and using SHAP per-data-point importance to audit predictions; claims this enables CI/CD, monitoring, and accountable model lifecycle management.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:03Why reproducibility matters
- 4:41Reproducibility definition
- 7:10Levels of reproducibility
- 8:59Organized folders setup
- 13:10Full reproducibility test
- 14:36Data provenance with containers
- 16:26Interpretability tools
- 18:34MLOps and observability
related talks
59mDevOps for Machine Learning Projects
30mReproducible Machine Learning at Scale
39mApplying deployment oriented mindset for building Machine Learning models
32mReproducibility and versioning of ML systems
01h 02mDocker for Data Science: Reproducibility and Deployment
39mEnd to End Machine learning pipelines for Python driven organizations
33mCompliant Data Management and Machine Learning on Kubernetes
29mReproducibility in ML Systems: A Netflix Original