Leveraging Git for ML Experiment Management
summary
Claims a reproducible, Git-first way to manage the dozens or tens or hundreds of experiments data teams run: use Git for code + DVC for data, models and pipelines, push artifacts to remotes (AWS GCP Azure or even Google Drive), and view results in a VS Code extension. Demo classifies ~1,000 Pokémon (legendary vs normal) with LightGBM, DVC Live logging metrics and plots; shows smart caching skips unchanged stages and DVC tracks data and models, and the workflow works end to end.
AI-written summary · about
chapters (21)
- 0:00Experiment tracking with Git
- 1:47Reproducible ML experiments matter
- 3:33Challenges with tracking methods
- 4:16DVC for experiment tracking
- 5:22Reproducible workflows with DVC
- 6:06Data and artifact versioning
- 8:10Remote storage and team collaboration
- 8:53DVC pipelines for ML
- 9:54Demo: Pokémon classification
- 10:56Project structure and params
- 11:37Creating DVC pipelines
- 12:40Running experiments with DVC
- 14:30Exploring results and outputs
- 16:20Alternative DVC experiment runs
- 17:48VS Code extension for experiments
- 18:30Experiment visualization and comparison
- 21:00Promoting experiments to Git
- 21:22Visualizing metrics and plots
- 22:05Queueing and parallelizing experiments
- 23:06Git with DVC advantages
- 24:11Q&A on DVC tracking
related talks
37mOpen Source Tools for ML Experiments Management
46mFlexible ML Experiment Tracking System for Python Coders with DVC and Streamlit
33mFlexible ML Experiment Tracking System for Python Coders with DVC and St...
42mApplications in ML Drug Discovery pipelines
35mData versioning in machine learning projects
02h 10mBuild End-To-End MLOps Platform with Open Source DVC Ecosystem
59mDevOps for Machine Learning Projects
31mBecoming a Pokémon Master with DVC: reproducible machine learning experiments