TensorFlow Extended (TFX): Machine Learning Pipelines and Model Understanding

on YouTube ↗38m

" summary

End-to-end production pipelines here use TFX components (ExampleGen, Transform, Trainer, ModelValidator) wired through ML Metadata so artifacts and lineage are queryable; pipelines run locally or scale to Dataflow, Airflow or Kubeflow with caching and warm-starting to save compute. Data issues surface with TF Data Validation (distribution shifts, missing values); model slices and time-series performance use TF Model Analysis; a what-if tool lets analysts tweak inputs. Example builds include exporting predictions to SQL via a custom component to compute real business cost; systems are demonstrably deployed at Google scale.

AI-written summary · about

" chapters (40)

1. 0:00Introduction
2. 0:51Overview
3. 1:15TFX Introduction
4. 5:27Accessibility
5. 8:49Contract
6. 9:32Systems View
7. 10:14Monitoring Systems
8. 11:04Building Better Software
9. 12:54User Journey
10. 13:44Data Quality
11. 14:35Massage
12. 15:51Validation
13. 16:25Separation
14. 17:19Components
15. 18:56How many know TensorFlow
16. 19:22Model Validator
17. 20:01MLMetadata
18. 21:04Component Configuration
19. 22:28Dependency Graph
20. 22:49DataDependency Graph
21. 23:19Metadata Store
22. 23:34Lineage
23. 24:29Warm Starting
24. 24:50Model Analysis
25. 25:09Reuse Components
26. 25:35Developing Models
27. 26:52Open Source Orchestrators
28. 27:50Putting it all together
29. 28:51Model Understanding
30. 29:55What Went Wrong
31. 30:25Data Validation
32. 30:40Data Validation Example
33. 31:30Questions
34. 31:41Slices
35. 32:22Whatif
36. 33:38CTR
37. 34:16Custom Executor
38. 34:51Assumptions
39. 35:51Baseline Trainer
40. 36:43Conclusion

" related talks

• 37mEnd-to-End ML Pipelines with KubeFlow and TensorFlow Extended (TFX)Chris Fregly · Scale By The Bay 2019
• 13mAn introduction to MLOps with TensorFlow Extended (TFX)Robert Crowe · Google I/O 2022
• 38mFrom Experimentation to Products: The Production Machine Learning JourneyRobert Crowe · GOTOpia February 2021
• 08mManage MLOps and deploy Machine Learning to production with the new and improved TFXRobert Crowe, Tris Warkentin · Google's Machine Learning Virtual Community Day 2021
• 31mMake your app a product with the TFX teamRobert Crowe, Kaz Sato, Anusha Ramesh, Zhitao Li · Google I/O 2021
• 42mDeep learning beyond the learningJörg Schad · Codemotion Amsterdam 2018
• 38mDistributed Processing for Machine Learning Production PipelinesRobert Crowe, Ahmet Altay, REZA ROKNI · Flink Forward Virtual 2020
• 30mThe art of machine learning operations in TensorFlow ecosystemTaha Bouhsin · BlaBlaConf 2022