ML Models and Dataset Versioning

Transcript

1. ML MODELS AND
   DATASET
   VERSIONING
   Kurian Benoy
   

   View Slide

2. $ WHOAMI
   Open source contributor
   FOSSASIA OpenTechNights Winner
   Kaggle Expert in Kernels
   

   View Slide

3. $ WHOAMI
   Open source contributor
   FOSSASIA OpenTechNights Winner
   Kaggle Expert
   Final Year BTech student @MEC
   

   View Slide

4. OUTLINE
   Start up Adventures
   Challenges
   Model and Dataset versioning
   How I discovered DVC?
   Use case: Versioning dogs and Cats
   Conclusion
   

   View Slide

5. Startup Adventures
   

   View Slide

6. CHALLENGE 1:
   ML IS SLOW
   

   View Slide

7. CHALLENGE 2:
   WORKING WITH ML PROJECTS
   Most software products take a few
   seconds to execute.
   $ git clone project-repo
   $ pip install -r requirements.txt
   

   View Slide

8. View Slide

9. CHALLENGE 3:
   METRIC DRIVEN
   

   View Slide

10. CHALLENGE 4:
    NOT ABLE TO USE GIT
    git not suitable for projects > 1GB
    git clone becomes slow
    

    View Slide

11. MODEL
    VERSIONING
    

    View Slide

12. TRACKING EXPERIMENTS
    TRACKING
    METRICS
    

    View Slide

13. Why Model Versioning?
    > To keep track of experiments
    > Choose the best ideas
    >> EXPERIMENTS = CODE + OUTPUTS
    Models are outputs
    

    View Slide

14. DATASET
    VERSIONING
    

    View Slide

15. View Slide

16. 4 TB/day
    

    View Slide

17. View Slide

18. Why Dataset management?
    > Moving Datasets around
    > Datasets evolve, so versioning required
    >> EXPERIMENTS = CODE + DATA + OUTPUTS
    Source code, Datasets
    

    View Slide

19. HOW I DISCOVERED
    DVC
    

    View Slide

20. DATA VERSION
    CONTROL(DVC)
    

    View Slide

21. > Experiment and Dataset tracking
    > Open-source(3500+ stars)
    > Build to adopt the best practises of ML
    > Works well with git
    > Language and framework agnostic
    

    View Slide

22. VERSIONING CATS &
    DOGS
    

    View Slide

23. DEMO TIME
    

    View Slide

24. DVC WORKFLOW
    

    View Slide

25. Tracking data
    1 Tracking 1000 cats and dogs
    2 Add 1000 more labelled images of cats & dogs
    

    View Slide

26. SWITCHING VERSIONS
    

    View Slide

27. CONCLUSION
    

    View Slide

28. "Data science as different from software
    as software was different from hardware."
    Nick Elprin,
    CEO, DominoLabs.
    

    View Slide

29. Think about your processes(ML projects)
    

    View Slide

30. Think about your processes
    Try to version control for your projects
    

    View Slide

31. Try it out in your ML project!
    

    View Slide

32. THANK YOU
    Twitter: kurianbenoy2
    Email : kurian.bkk@gmail.com
    Speaker Deck: bit.ly/mlversion19
    

    View Slide

33. APPENDIX
    

    View Slide

34. Other Tools for versioning
    ML Flow - Tracking Models, Metrics
    Git-LFS - Tracking Large files
    Jovian - JupyterNB based tracking
    Neptune.Ml
    Hangar Py - Versioning Tensor Data
    

    View Slide