Deterministic Data Pre-processing

[moonforsun]

I am wondering how to make the MONAI's random transforms deterministic?
I am using RandCropByLabelClassesd in my transforms. But everytime the random cropped data is different with the same global seed. The seed_everything is called from pytorch_lightning

Besdies the following setting, what else should I add to ensure the deterministic training?

seed_everything(cfg.seed, workers=True)
torch.backends.cudnn.determinstic = True
torch.backends.cudnn.benchmark = False

Thank you in advance!

[KumoLiu]

Hi @moonforsunm,
You could use set_determinism in MONAI. You could also refer to any tutorials, most of them have this part.

from monai.utils import set_determinism
set_determinism(seed=0)

Hope it can help you, thanks!

[moonforsun]

This line indeed ensures the deterministic training although the documentation itself indicates this function will not affect the randomizable objects in monai.transforms.Randomizable (https://docs.monai.io/en/stable/transforms.html#monai.transforms.Randomizable).

Thank you very much!

[KumoLiu]

Hi @moonforsun, I think the note in set_determinism means that if you are using monai.data.DataLoader, it will automatically set the seeds according to the global random state. Therefore just as Wenqi mentioned, if you are not using monai.data.DataLoader, you need to set the transform' random state by yourself.
https://docs.monai.io/en/stable/utils.html#monai.utils.misc.set_determinism

Hope I have explained it clearly. Thanks!

[wyli]

if you are not using monai.data.DataLoader, you should set the transform's random state as well:

crop = RandCropByLabelClassesd(keys=..., ...)
crop.set_random_state(cfg.seed)

[moonforsun]

Thank you for your reply. Yes, I did indeed use MONAI's Dataloader rather than the torch's Dataloader when I reported the issue.