seed_everything() runs _select_seed_randomly() even if PL_GLOBAL_SEED is set

[bprus]

🐛 Bug

There is a bug in pl.utilities.seed.seed_everything().

The problem is here:
https://github.com/PyTorchLightning/pytorch-lightning/blob/1.0.5/pytorch_lightning/utilities/seed.py#L44

If None is passed as a seed:
seed = os.environ.get("PL_GLOBAL_SEED", _select_seed_randomly(min_seed_value, max_seed_value))
is run. And _select_seed_randomly is run even if PL_GLOBAL_SEED is present in os.environ.

The seed is set properly, but there is a warning displayed that suggests a different seed had been set. This introduces some confusion, especially during training with ddp.

To Reproduce

import pytorch_lightning as pl
import os

pl.utilities.seed.seed_everything()
seed = os.environ.get('PL_GLOBAL_SEED')
print(seed)

pl.utilities.seed.seed_everything()
seed = os.environ.get('PL_GLOBAL_SEED')
print(seed)

Expected behavior

The code above produces the following output:

No correct seed found, seed set to 3315818503
3315818503
No correct seed found, seed set to 1668269197
3315818503

_select_seed_randomly(min_seed_value, max_seed_value) should only be run when there is no PL_GLOBAL_SEED set.

Environment

• PyTorch Lightning Version (e.g., 1.0): 1.0.5

[github-actions]

Hi! thanks for your contribution!, great first issue!

[SeanNaren]

Thanks for the issue! Could you verify #4743 fixes this?

[bprus]

It seems that this fixes my problem. But it introduces another one as any seed is now incorrect because of:
if not (seed < min_seed_value <= seed <= max_seed_value):
(someone already pointed it in PR comments)

[stale]

This issue has been automatically marked as stale because it hasn't had any recent activity. This issue will be closed in 7 days if no further activity occurs. Thank you for your contributions, Pytorch Lightning Team!

[bprus]

Still waiting for a fix.