Reproducibility

[javierlorenzod]

Hi! thanks for sharing your code with the community. I am using your fine tune example. I set all random seeds properly to ensure reproducibility as it is explained on pytorch docs but I obtain slightly different results between experiments with the same hyperparameters, do you know the reason?

I only changed the following lines in example_finetune.py:

from pytorch_lightning import seed_everything
...
if __name__ == "__main__":
    seed_everything(47)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False
...