Seed NumPy using np.random.SeedSequence() in pl_worker_init_function() to robustly seed NumPy-dependent dataloader workers

[amorehead]

What does this PR do?

Fixes an issue where small changes to a user's random seed (specified via seed_everything(), e.g., incrementing seed by 1) would not result in a new (unique) random seed for NumPy (set within pl_worker_init_function()). This could cause multi-process dataloader workers to not change their random state based on a user's desired seed set via seed_everything() before training. In other words, if users were relying on seed_everything() to change e.g., the order of the indices sampled by each dataloader worker (if their dataloader(s) use NumPy's random module for index sampling), this would not work until now.

Below is an example of a seed_sequence that previously would not have generated a distinct random seed for NumPy. For example, 7768447584330995212 & 0xFFFFFFFF and 13249712275147347468 & 0xFFFFFFFF yield the same seed even though these former numbers were generated using different base_seeds. To reproduce more of these issues, set base_seed=42,worker_id=0,global_rank=0 within seed_sequence = _generate_seed_sequence(base_seed, worker_id, global_rank, count=4) and then trying incrementing or decrementing base_seed. Most of these small changes to base_seed yield the same random seed for NumPy.

0 =
7900221878587883521
1 =
15926613439302434606
2 =
2812539921844481303
3 =
7768447584330995212

and

0 =
14998723003968126977
1 =
17951608614340099886
2 =
3289043221663264023
3 =
13249712275147347468

-> 

seed_sequence[3] & 0xFFFFFFFF
1078613516

???

Before submitting

• Was this discussed/agreed via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or minor internal changes/refactors)

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing, make sure you have read the review guidelines. In short, see the following bullet-list:

Reviewer checklist

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

---

📚 Documentation preview 📚: https://pytorch-lightning--20369.org.readthedocs.build/en/20369/

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 88%. Comparing base (28d3afd) to head (b9ba94c).
Report is 55 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master   #20369   +/-   ##
=======================================
  Coverage      88%      88%           
=======================================
  Files         267      267           
  Lines       23203    23203           
=======================================
  Hits        20313    20313           
  Misses       2890     2890           

[lantiga]

Nice catch! Can you please add a test that verifies the desired behavior so we know it won't regress in the future?

[adosar]

@amorehead Does this mean that previously seed_everything didn't work as expected, e.g. when using random augmentations etc. based on NumPy?

[amorehead]

@adosar, if your multiprocessing dataloaders (plural) use NumPy's random module for dataset index sampling, then yes, until this PR was merged in, Lightning was not handling random seed definitions correctly.

[adosar]

@amorehead If I understand correctly, this affect any dataset with a np.random.* call under __getitem__?

[amorehead]

@adosar, yes, any np.random.* calls under __getitem__ (if you used multiple dataloader workers) would not be seeded properly (i.e., uniquely for each worker) until now.