[meta] Track experiments with the tracking platform W&B

[marco-c]

Bugs

Preview Give feedback

1. Only 3 evaluation datasets are visible in custom charts firefox-translations-training#688
   bug weights and biases +2
2. Some metric tables are empty firefox-translations-training#716
   bug weights and biases +2
3. Stalled metrics are not reported for backward model firefox-translations-training#846
   bug weights and biases +2
4. COMET is still logged in [0,1] interval firefox-translations-training#862
   bug weights and biases +2
5. Check that COMET is published from the recently archived GCP experiments
6. Models are missing in group logs firefox-translations-training#874
   bug weights and biases +2

Reuploading

Preview Give feedback

1. Perform a comprehensive testing before the final reuploading #793
   weights and biases +1
2. Upload GCP experiments #863
   weights and biases +1
3. Actualize moz-translations when online uploading is feature complete #574
   weights and biases +1

Experiment tracking

Preview Give feedback

1. Actualize documentation on experiment tracking firefox-translations-training#845
   documentation weights and biases +2
2. Publish training charts from Taskcluster firefox-translations-training#333
   weights and biases +1
3. Publish original Marian configs firefox-translations-training#313
   weights and biases +1
4. Display corpus size in W&B firefox-translations-training#529
   weights and biases +1
5. W&B runs published from GCP experiments should be suffixed with the Task Group ID when possible firefox-translations-training#734
   weights and biases +1
6. Add publishing to CI firefox-translations-training#792
   weights and biases +1
7. Support uploading GCP Taskcluster artifacts to W&B firefox-translations-training#809
   weights and biases +1
8. Publish task logs from Taskcluster firefox-translations-training#522
   weights and biases +1
9. Duplicate runs in W&B firefox-translations-training#729
   bug weights and biases +2
10. COMET results are not visible on custom charts firefox-translations-training#731
    weights and biases +1
11. Tune workspace dashboard to enable comparison across models and experiments firefox-translations-training#408
    weights and biases +1
12. Publish group logs evals in online mode firefox-translations-training#697
    weights and biases +1
13. Ensure naming for runs is consistent firefox-translations-training#573
    weights and biases +1
14. [tracking ERROR] Publication failed: Invalid config section: while scanning a simple key firefox-translations-training#554
    bug weights and biases +2
15. Publish evals from Taskcluster firefox-translations-training#519
    weights and biases +1
16. Publish experiment config from Taskcluster firefox-translations-training#521
    weights and biases +1
17. Support COMET metric in the parser firefox-translations-training#615
    weights and biases +1
18. Create a new W&B project if it doesn't exist firefox-translations-training#639
    bug weights and biases +2
19. Support extra metrics from Tensorboard firefox-translations-training#534
    weights and biases +1
20. tests fail with numpy 2.0 firefox-translations-training#689
    bug weights and biases +2
    
21. group_logs is incomplete firefox-translations-training#575
    bug weights and biases +2
22. Add an author field firefox-translations-training#578
    weights and biases +1
23. Backward steps are duplicated firefox-translations-training#474
    bug weights and biases +2
24. Publish experiment config to W&B firefox-translations-training#332
    weights and biases +1
25. Add missing steps for experiment tracking firefox-translations-training#334
    weights and biases +1
26. Teacher ensemble step is missing in Taskcluster publication firefox-translations-training#452
    weights and biases +1
27. Inconsistent dataset naming publishing Taskcluster group firefox-translations-training#449
    weights and biases +1
28. Traversal tasks from Taskcluster task group publication firefox-translations-training#448
    weights and biases +1
29. Support overriding experiment on Weight&Biases firefox-translations-training#366
    weights and biases +1
30. Update parser to work with Marian 1.12 firefox-translations-training#509
    weights and biases +1
31. Publish past experiment directly from Taskcluster firefox-translations-training#312
    weights and biases +1
32. Support taskcluster secrets for W&B secret token firefox-translations-training#481
    taskcluster weights and biases +2
33. training runs don't log properly to stdout/stderr anymore firefox-translations-training#552
    bug weights and biases +2
34. CI pollutes W&B firefox-translations-training#553
    bug weights and biases +2
    
35. Publish Marian config from Taskcluster firefox-translations-training#520
    weights and biases +1
36. Issues with uploaded experiments firefox-translations-training#518
    bug weights and biases +2
37. Publish all evaluation results on experiment dashboard firefox-translations-training#336
    weights and biases +1
38. Publish experiment metrics separetely from Marian training logs firefox-translations-training#346
    weights and biases +1
39. Show parser publication failures firefox-translations-training#345
    weights and biases +1
40. Use real time publication for parser generic publishers firefox-translations-training#347
    weights and biases +1
41. Refactor translations_parser.cli.experiments.main firefox-translations-training#344
    weights and biases +1
42. weights and biases reporting doesn't work correctly when resuming training after a spot termination firefox-translations-training#594
    weights and biases +1

Docs:

1. Initial spec
2. Spec by Bastien