Using Reproducible Experiments To Create Better Models

Transcript

1. Using Reproducible Experiments To Create Better Models Milecia McGregor @FlippedCoding Developer Advocate @DVCorg 1
3. Common issues 3
4. Find the best combo of hyperparams, algorithms, and datasets Common issues 4
5. Keep track of all of the changes Common issues 5
6. Hard to follow changes over time Common issues 6
7. How to fix them 7
8. 8 How to fix them
9. Background on tuning hyperparameters 9
10. Parameters that define the model Background on tuning hyperparameters 10
11. Grid search and random search Background on tuning hyperparameters 11
12. Using DVC 12
13. Using DVC ✓ Open-Source ✓ Works on top of Git ✓ No libraries, no API calls 13
14. Using DVC dvc exp run 14
15. Using DVC 15
16. Using DVC 16
17. Hyperparameter tuning 17
18. dvc exp run --queue 18
19. ... 19
20. dvc exp run --run-all dvc exp show 20
21. ... 21
22. dvc plots diff exp-6182a exp-cb998 22
24. dvc exp run --queue 24
25. ... 25
26. dvc exp run --run-all dvc exp show 26
27. ... 27
28. dvc exp run --queue 28
29. ... 29
30. dvc exp run --run-all dvc exp show 30
31. ... 31
32. dvc plots diff exp-9d023 exp-cb998 32
35. Key takeaways • Adding reproducibility to experiments is important • Using DVC helps you track every part of your experiments • Don’t be afraid to try new tools Milecia McGregor @FlippedCoding https://discord.gg/zpCsrscfMW 35
36. Resources • dvc.org/docs • https://discord.gg/zpCsrscfMW • https://studio.iterative.ai Milecia McGregor @FlippedCoding https://discord.gg/zpCsrscfMW 36