Machine Learning Engineering principles with Python and MLFlow by Natu Lauchande

Machine Learning is a very hyped topic of the moment. While a lot of the talks and presentations cover the data science component, very few cover the nity gritty details of a machine learning pipeline. This talk will focus on the engineering part of Machine Learning by covering different Machine Learning systems architecture best practices, strategies including design. We will delve into the essence of Uber's Michelangelo, Airbnbs s Bighead and Facebooks FB Learner. During the talk, I will use MLFlow and Python as platforms to create an open-source based solution similar to the ones from the big tech companies for the everyday tech startup. The entirety of the cycle of training, deployment, monitoring, champion/challenger testing, and serving layer will be addressed. Technical debt prevention is another topic that will be addressed in the end of the talk.

Transcript

1. Machine Learning Engineering Principles with Python and MLFLow (PyconZA-2019) Natu Lauchande (Principal Data Engineer @ Jumo.World) 1
2. AGENDA In Scope: 1. ML Engineering pain points 2. Sample of relevant ML Pipeline architectures used across industries 3. ML Engineering principles to address some of the pain points in the ML development process 4. Examples of usage of MLFlow and other tools to implement principles Out of Scope: 1. Detailed review of ML systems 2. Any specific data science or algorithmic solution 4. Tutorial on using any specific tool ( eg: MLFlow, etc.) 2
3. Context : Machine Learning easily accessible
4. Pain points - Issues getting models from dev to production
5. Pain points - Data Pipeline Entanglement
6. Pain Points - Experimentation Management
7. Machine Learning Engineering Pain - Points - Metrics
8. How to crack this issues?
9. ML Engineering
10. Idealized Workflow of a Machine Learning Project
11. ML Engineering
12. ML code in the Context of ML backed System
13. Relevant Machine Learning Engineering Principles • Software engineering approach, platform view of your ML functioning - ( CD4ML, Testing, MLOps) • Integrated reliable data lifecycle • Reproducible and reliable model management • Experimentation and rapid prototyping facilities
14. Uber Michelangelo
15. Uber Michelangelo PyML
16. Uber Michelangelo Feature Store
17. Facebook FBLearner
18. Airbnb Bighead
19. Machine Learning Platforms Comparative View
20. DIY ML Engineering Pipeline
21. Ideal Pipeline - Real Time Monitoring Perspective ( D. Sculley - https://www.youtube.com/watch?v=V18AsBIHlWs) ML Score - https://storage.googleapis.com/pub-tools-public-publication-data/pdf/45742.pdf) 21
22. Data Lifecycle 22 Examples of data versioning tools: 1. Apache Delta Lake 2. DVC
23. Data Lifecycle Tooling 23 Examples of tools : 1. Deequ 2. Great Expectations
24. Data Lifecycle Tooling 24 Feature Management: By gojek/feast and google cloud (still WIP): • Support ingesting feature data via batch or streaming • Low latency API • Enable discovery and documentation of features • Provide an overview of the general health of features in the system
25. 25 Open source machine learning platform that addresses the following issues : 1. Experimentation Management : UI to manage and log different models 2. Reproducibility : Runs he same way anywhere 3. Standardization of ML Projects: Designed to scale for 1 or 100 000 orgs 4. Standardization of deployment operations into multiple runtimes ( AWS Sagemaker, Azure ML, Google Cloud, Kubernetes etc.)
26. 26
27. 27
28. Machine Learning Principes : Databricks Blue Print for an Idealized Pipeline and the role of MLFlow 28
29. 29
30. 30 Demo Example I - Bitpred - Data https://github.com/nlauchande/bitpred
31. 31 ML Problem - Bitpred - Problem Toy Problem Statement : Will the bitcoin move up or down in the next 24 hours ? - https://github.com/nlauchande/bitpred
32. 32 ML Problem - Bitpred - Project Structure - MLProject file
33. 33 ML Problem - Bitpred - Project Structure - Conda File
34. 34 ML Problem - Bitpred - train.py/ w/ experimentation log
35. 35 ML Problem - Bitpred - Running your training pipeline in a reproducible manner
36. 36 ML Problem - Demo ML Flow UI
37. 37 ML Problem - Demo ML Flow UI
38. 38 ML Problem - MLFlow Immediate Access to a Serving Layer ( Batch, Rest API ) Running the prediction service : Calling the rest API
39. 39 Predictions Monitoring
40. 40 Predictions Monitoring Dashboard Example
41. 41 Demo Example II - ML Pipeline for a PTSD Chatbot risk classifier
42. 42 Demo Example - Sample sklearn pipeline
43. 43 Demo Example - Log your model
44. 44 Demo Example - Track Experiments
45. 45 Demo Example - Model Tracking and Versioning
46. 46 Demo Example - Model Tracking and Versioning
47. 47 Demo Example - Serving Layer
48. 48 Demo Example - Serving Layer
49. 49 Demo Example - ML Package manager
50. Summary / Lessons learned • It’s painful to move from development to production • ML ecosystem is very recent and tools and frameworks are still maturing • Devops and standard software engineering practices are extremely useful • Cross functional working on ML projects at the same time 50
51. 51 Wrap UP & QA