Notebook To Production
summary
Seven concrete practices shorten research to production: host code in Python modules so functions are testable and reusable; version everything — code, notebooks, large data and models with Git plus data-versioning tools; track experiments with tools like Weights & Biases or MLflow; use task-oriented notebooks and peer reviews (nbdime/diffing) ; record environment and OS/GPU/CUDA dependencies and Dockerize; convert pipelines into scripts for training and deployment.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:25Jupyter notebooks challenges
- 3:10Dark side of Jupyter
- 4:57Best practices workflows
- 7:26Hosting code in modules
- 9:12Versioning ML projects
- 12:00Peer review bias reduction
- 13:24Tracking experiments
- 15:13Task-oriented notebooks
- 17:02Reproducibility tracking
- 18:48No notebooks in production
- 20:36Scripts for training deployment
- 20:57Recap best practices
- 22:42Q&A
related talks
28mUnlock the Full Potential of Jupyter Notebooks
41mBetter Code for Data Science
01h 00mFrom notebook to production using
01h 23mImproving Your Data Modeling Work Through Open-Source Software
34mNotebooker: Production and Scheduling for your Jupyter Notebooks
38mBuilding Data Pipelines in Python
39mClean Code in Jupyter notebooks, using Python
28mGoing From Notebooks to Scalable Systems