Link to https://pytorch.org/docs/stable/notes/randomness.html

[cowwoc]

Suggestion: https://pytorch-lightning.readthedocs.io/en/latest/api/pytorch_lightning.utilities.seed.html should reference https://pytorch.org/docs/stable/notes/randomness.html

When I initially starting using PyTorch Lightning I (wrongfully) assumed that invoking seed_everything() was sufficient. Now I see I need to invoke much more. It would be useful if the PyTorch Lightning documentation linked to the PyTorch documentation on this topic.

[justusschock]

Hi @cowwoc ,

Thanks for the suggestion. Do you want to send a PR that applies that change?

[carmocca]

Mind explaining why your assumption was wrong? Perhaps there is something we missed and we can improve.

[cowwoc]

@justusschock It's not clear to me which file needs to be edited to update the docs.

@carmocca As far as I can tell, seed_everything() does not invoke torch.use_deterministic_algorithms() and the cuda-related deterministic flags. There might be other methods to be invoked but those are the big two that were missing for me.

[ananthsub]

Hi @cowwoc - thanks for raising this. The Trainer has a deterministic flag which is used to set cudnn deterministic flags. I believe it makes sense to use this flag to determine when to set torch.use_deterministic_algorithms. I sent out #9121 for this.

@justusschock @edenafek - given the seed & determinism could go together, maybe we could combine these in the docs somewhere on a page more discoverable than the Trainer API: https://pytorch-lightning.readthedocs.io/en/latest/common/trainer.html#reproducibility

[tchaton]

Dear @cowwoc,

I believe there is a little confusion :) seed_everything does exactly what the function says. It seed python, numpy and torch + the workers for the dataloaders, so your transform are reproducible up to the determinism of cuda.

Furthermore, if you want deterministic benchmarking, you can add Trainer(deterministic=True) as stated before.

Finally, I agree a reference to the PyTorch page would be good to avoid further confusion.

Mind doing a PR ?

Best,
T.C

[cowwoc]

@tchaton I agree with everyone you said but I am not comfortable doing a PR. I am very new to Python not to mention this project specifically. I'm not sure how/where to best insert the aforementioned reference. Is anyone else willing to handle this PR?

[rishawsingh]

Hi @tchaton can I work on this issue if this is not assigned to anybody? If not, cann you give me some good first issues

[justusschock]

Hi @rishawsingh , The PR linked to this issue (#9141) already implements a fix for this by @tchaton .

In general we do have the issue labels good_first_issue which you can use to filter those kind of issues :)

[cowwoc]

The PR doesn't look correct. If you open https://pytorch-lightning.readthedocs.io/en/latest/benchmarking/benchmarks.html you will see the link at the bottom renders as plain text (not a hyperlink). Also, I was hoping/expecting to see this link near the top of https://pytorch-lightning.readthedocs.io/en/latest/api/pytorch_lightning.utilities.seed.html since users looking there are likely interested in reproducibility.