Pythonsevilla2019 - Introduction to MLFlow
Download as PPTX, PDF0 likes326 views
The document provides an overview of MLflow, a platform designed to streamline the machine learning lifecycle by addressing key challenges such as experiment reproducibility, model comparison, data sharing, and deployment. It highlights the components of MLflow, including tracking, projects, and models, emphasizing its compatibility with various ML libraries and environments. Recent updates and future directions for MLflow are also discussed, showcasing its growing capabilities and community support.
Pythonsevilla2019 - Introduction to MLFlow
- 1. Introduction to MLFLOW Accelerating the Machine Learning Lifecycle
- 2. Who am I? Fernando Ortega Gallego Data Enginer @ Plain Concepts UK • Interests & hobbies: – Machine Learning and NLP – Python, Azure and my daughter – Honda biker @FernanOrtega [email protected]
- 3. Roadmap Introduction MLFlow Tracking MLFlow Projects MLFlow Models Conclusions
- 4. Roadmap Introduction MLFlow Tracking MLFlow Projects MLFlow Models Conclusions
- 6. Problems • Reproduce experiments • Compare experiments • Fine tune previous experiments across teams • Share data, parameters or metrics • Deploy trained models 7
- 7. It is difficult to productionize and share 8
- 8. Custom ML Platforms • Facebook FBLearner, Google TFX, Uber Michelangelo • Advantages: – Standarise the ML loop • Disadvantages: – Limited to a few algorithms or frameworks – Tied to the company’s infrastructure 9
- 9. Are there any similar solutions in an open manner? 10
- 10. 11
- 11. Introducing MLFlow • It works with any ML library & language • It runs the same way anywhere • It is designed to be useful both for 1 person, small teams or big teams
- 14. MLFlow components 15 Tracking Projects Models
- 15. Roadmap Introduction MLFlow Tracking MLFlow Projects MLFlow Models Conclusions
- 16. MLFlow Tracking • Record and query experiments: – Data – Code – Model parameters – Results (performance metrics) – Model 17 Tracking
- 17. Motivation 18 Centralised repository of useful information to analyse several runs of training.
- 18. How it is working 19
- 19. Key concepts • Parameters • Metrics • Tags and notes • Artifacts • Source • Version 20
- 20. Code example 21
- 21. 22 Demo
- 22. Roadmap Introduction MLFlow Tracking MLFlow Projects MLFlow Models Conclusions
- 23. MLFlow Projects • Packaging format for reproducible ML runs • Defines dependencies for reproducibility • Execution API for running projects locally or remote 24 Projects
- 24. Motivation 25 Diverse set of tools and environments involves difficulty to productionalize and share ML work
- 25. How it is working 26
- 26. Key concepts • MLproject file • Entry points • Environments – Conda – Docker – System • Run 27
- 27. Code example 28 mlflow run [email protected]:mlflow/mlflow-example.git -P alpha=0.5 mlflow run <uri> -m databricks --cluster-spec <json-cluster-spec>
- 28. 29 Demo
- 29. Roadmap Introduction MLFlow Tracking MLFlow Projects MLFlow Models Conclusions
- 30. MLFlow Models • Packaging format for ML Models • Defines dependencies for reproducibility • Deployment APIs 31 Models
- 31. Motivation 32 ML Frameworks Batch & Stream Scoring Serving tools Inference code
- 32. How it is working 33
- 33. Key concepts • MLmodel file • Storage format • Entry points • Flavours • Custom model 34
- 34. Code example 35
- 35. 36 Demo
- 36. Roadmap Introduction MLFlow Tracking MLFlow Projects MLFlow Models Conclusions
- 37. MLFlow rocks! • Log important parameters, metrics, and other data that is important to the machine learning model • Track the environment a model is run on • Run any machine learning codes on that environment • Deploy and export models to various platforms with multiple packaging formats 38
- 38. MLFlow 1.0 (4-jun) • Support for step tracking • Improved Search features • Batched logging of metrics • Support for HDFS • Windows support for the client • Build Docker images to deploy • ONNX model flavour 39
- 39. MLFlow last updates (1.4: 31-oct) • Windows support • Tags and descriptions • Google Cloud run models • Log directories as artifacts • CLI command to export to CSV • Keras compatibility with TF 2.0 • Model Registry in preview 40
Editor's Notes
- #20: Backend stores are entity store (filestore and database) and artifact repository (s3, azure blob, Google cloud, dbfs)
- #21: Parameters: key-value inputs to your code Metrics: numeric values Tags and notes: additional information about a run Artifacts: arbitrary files, including data and models Source: code Version: version of the code
- #27: Backend stores are entity store (filestore and database) and artifact repository (s3, azure blob, Google cloud, dbfs)
- #34: Backend stores are entity store (filestore and database) and artifact repository (s3, azure blob, Google cloud, dbfs)
- #39: Conclusion Workflow tools can greatly simplify the ML lifecycle • Improve usability for both data scientists and engineers • Same way software dev lifecycle tools simplify development Learn more about MLflow at mlflow.org