Becoming a Pokémon Master with DVC: reproducible machine learning experiments
summary
A working Pokémon sprites classifier gets built to predict water types and returns a sample probability of 0.87 with a small confusion matrix of false positives and negatives, showing it mostly works; DVC is used with Git for reproducibility, replacing image files with metadata and storing blobs in a cache named by md5 so commits tie data code and params together; pipelines skip unchanged stages and the demo shows an Pokemon sprites classifier, an experiment table and VS Code extension, and a DVC cache with metadata and md5 hashes.
AI-written summary · about
chapters (22)
- 0:00Pokémon and MLops intro
- 1:52Speaker and agenda
- 2:13Pokémon type classifier
- 4:21Evolving datasets in ML
- 5:46Notebook-based experimentation problems
- 7:52DVC for reproducible ML
- 8:56DVC data versioning
- 11:04DVC pipelines for ML
- 13:55ML beyond Pokémon
- 15:43Moving to DVC from notebooks
- 17:30Tracking experiments
- 18:57Recap and reproducibility
- 20:27Closing and resources
- 21:31Q&A: Kubernetes and AWS
- 22:57Q&A: DVC vs Git LFS
- 24:22Q&A: Git commits for DVC
- 25:24Q&A: Parameter changes
- 26:29Q&A: DVC cache
- 27:12Q&A: Fun technical questions
- 28:14Q&A: DVC blame and commits
- 28:59Q&A: Feature store integration
- 30:26Final Q&A and conclusion
related talks
27mLeveraging Git for ML Experiment Management
33mHow to Track your Development Process with DVC
33mFlexible ML Experiment Tracking System for Python Coders with DVC and St...
48mHow I found my Pokémon cards thanks to Postgres: an AI journey
16mMake your data science workflows reproducible
38mUsing Reproducible Experiments To Create Better Models
46mFlexible ML Experiment Tracking System for Python Coders with DVC and Streamlit
26mTransforming a Jupyter Notebook into a reproducible pipeline for ML experiments