The Model Review: improving transparency, reproducibility, & knowledge sharing...
summary
A concrete process is proposed: require a model review before deploy, share a link to an MLflow run, and hold a 30-minute review meeting where reviewers inspect code, metrics, artifacts and notes. The team builds common training infrastructure that embeds MLflow Tracking so runs record parameters, datasets, conda environment, saved model and plots. autologging one-liner captures metrics and curves automatically; demos show comparing six runs and training a neural net for 40 epochs. shared run links make reviews reproducible and searchable.
AI-written summary · about
chapters (6)
- 0:00The Model Review
- 1:38Motivation
- 3:40Why do we Review Code?
- 7:12First Example
- 15:31Autologging Model Training
- 24:15Step 4: Model Review
related talks
36mHow to review a model
38mHidden Gems in MLFlow that Improve Experiment Reproducibility
57mManaging Machine Learning Lifecycle with MLflow
33mOperationalizing Model Lifecycle with MLFlow
32mReproducibility and versioning of ML systems
29mML Model Traceability and Reproducibility by Design
26mTrust Fall: Three Hidden Gems in MLFlow
43mMachine Learning Engineering principles with Python and MLFlow