Can't reproduce training process even if i have fix random seed on different platforms

1

I want to reproduce on two different platforms. I fixed the random seed with the following code:

def FixedSeed(seed: int = 1122) -> None:
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed(seed)
        torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.benchmark = False
    torch.backends.cudnn.deterministic = True

The training process is reproducible on each platform, but the training process is not exactly the same on different platforms. Does anyone know what is causing the problem?

2

Hi Jim!

This is to be expected. Pytorch does not offer any assurance that
computations will be exactly reproducible across platforms.

As long as your results agree within reasonable round-off error (say,
after a single forward and backward pass – deviations may accumulate
if you train for many iterations), you are doing as well as can be expected.

Best.

K. Frank

3

Thanks for your answer!