Bringing Decades of DevOps Experience To Deep Learning

on YouTube ↗29m

" summary

Argues that DevOps practices must become DeepOps around three concrete pillars: code, data, and machines. Demonstrates an experiment pipeline that auto-commits to a shadow repo and records the exact training command so every run is reproducible. Shows data versioning needs with the Chest X-Ray 14 45GB dataset and automates train/val/test splits. Describes spinning up many instances to run parallel jobs (eg. 20 concurrent instances) and quantifies waste when GPUs stay idle costing hundreds–thousands monthly.

AI-written summary · about

" chapters (48)

1. 0:00Introduction
2. 1:03Jesse Freemans background
3. 2:25Automating deep learning
4. 2:58The struggles of data scientists
5. 4:17Training a model
6. 4:55Debugging a model
7. 6:10Computer coordination
8. 6:43Reproducing experiments
9. 7:32AI first companies
10. 9:20Common problem
11. 9:37First computers
12. 10:28Lost at sea
13. 10:53No way to reproduce
14. 11:08Computers were tied to software
15. 11:28Larger software teams
16. 11:56Two different branches
17. 12:38DevOps
18. 13:30TestFirst Culture
19. 14:01Automation
20. 14:21Version Control
21. 14:51Shipping Smaller Changes
22. 15:37Netflix DevOps
23. 16:16How did we get here
24. 16:37DevOps culture
25. 17:13Empower the data scientist
26. 17:45What is deep ops
27. 18:02Tracking changes
28. 19:01No knowledge transfer
29. 19:49The risks
30. 20:08Automating it
31. 21:24Committing code
32. 21:51Tracking code
33. 22:14Organizing data
34. 22:35Folders
35. 23:09Database
36. 23:31Query
37. 24:14Version
38. 24:32Managing Data
39. 24:47Managing Computers
40. 25:04Running GPUs
41. 25:38Waste of money
42. 25:54Configuring machines
43. 26:27Running projects locally
44. 27:26Rerunning experiments
45. 27:51Data scientists vs engineers
46. 28:20Creating a culture
47. 28:39Applying concepts to deep learning
48. 28:55Outro

" related talks

• 56mDevOps for Machine LearningDamian Brady · NDC Minnesota 2019
• 59mDevOps for Data ScienceDamian Brady · NDC Sydney 2018
• 59mDevOps for Machine Learning ProjectsDmitry Pretrov, Ivan Shcheklein · The Fifth Elephant 2020
• 29mDevOps for Data Science? - automate the boring stuff and leverage the OSS ecosystemTania Allard · PyConline AU 2020
• 30mDevOps For Data TeamsTobias Macey · DevOpsDays Boston 2018
• 26mApplying DevOps practices in Data and ML EngineeringAntoni Ivanov · Conf42 Kube Native 2023
• 44mDevOps for Machine Learning - why is it different?Ryan Dawson · DataXDay 2020
• 57mDevOps for Machine LearningDamian Brady · NDC Oslo 2019