Making MLOps uncool again
summary
Workflow claims reproduceable ML using Git tooling and actually shows it: a HuggingFace transformer trains on 3,000 issues and predicts GitHub issue labels, pipeline wired with DVC to version data and models, CML to post PR reports and GitHub Actions to run CI. Demo builds a Docker image with the latest model and deploys automated labeling; some runs hit an actions checkout bug but solution repo proves the flow. Uses a DVC remote on Google Drive and can spin cloud runners via GitHub Actions + CML for heavier training.
AI-written summary · about
chapters (21)
- 0:00Intro
- 1:05What is uncool again
- 3:12What is envelopes
- 4:07Reproducibility
- 5:27Cloud Providers
- 8:09DevOps
- 10:16Automatic USU Labeler
- 11:05Workshop
- 12:53DVC Pipeline
- 16:11Running VC Pipeline
- 17:06Running VC Repo
- 18:16Metrics and plots
- 19:26Studio
- 20:21Running the pipeline
- 21:06dbc remote
- 24:31dbc config
- 30:11share reproducibility
- 38:16grant dbc remote access
- 45:41create a workflow
- 50:41create a PR
- 57:41selfhosted runners
related talks
23mAccelerating Model Deployment with MLOps
24mFrom Model to Production: Automating the ML Lifecycle with CI/CD
31m🤑 Paying Off Technical Debt in the Development of Machine Learning Apps
53mExperience with an end-to-end MLOps pipeline: Practical insights and 3D architectural perspectives
30mHow to make continuous integration work with machine learning
13mMLOps at Scale
14mWhy Your DevOps Skills Are Your MLOps
25mBuilding Scalable MLOps Pipelines That Actually Work