Increase flexibility of setting seed

I was running some code that has randomness during evaluation. However I found that my, supposedly, random preprocessing was always done in exactly the same way.

This is related to setting the random seed by default to 12: https://github.com/pytorch/ignite/blob/master/ignite/engine/engine.py#L583

An alternative mechanism, that respects the user’s choice of setting a random seed is to change seed =12 to:

seed = torch.randint(100)

which reads from the manual seed if it was set.

On Slack it was suggested to use int(time.time() * 100 % 1000), however this still surprises a user that wants to run the same script with two different random seeds and have both seeds reproducible.

I think it’s okay if torch.randint changes the random state, because many things in the dataloader already do as well.

[y0ast]

Aside from using torch.randint there are a bunch of other options: https://pytorch.org/docs/stable/torch.html#random-sampling

Spelling out the two cases:

1. users sets no random seed: need seed from ignite for proper restarting
2. users sets random seed: ignite shouldn’t overwrite using default seed, it is unexpected behaviour for a training framework. Instead it should support restarting and respect the user seed (so it can be reproduced).

if we do torch.randint then 1) is safeguarded, but not reproducible between runs. Also 2) is safeguarded, but we do advance the random state by sampling from torch.randint.

I think ignite shouldn’t care about reproducibility between runs if user doesn’t set a seed (model parameters are different).

[y0ast]

Actually, you don’t need to. As long as you set a manual seed in your script then it’ll work as expected.

Because the new mechanism reads from global random state: https://github.com/pytorch/ignite/pull/799/files