Fixing the seed for torch random_split()

Question

Is it possible to fix the seed for torch.utils.data.random_split() when splitting a dataset so that it is possible to reproduce the test results?

Answer 1

You can use torch.manual_seed function to seed the script globally:

import torch
torch.manual_seed(0)

See reproducibility documentation for more information.

If you want to specifically seed torch.utils.data.random_split you could "reset" the seed to it's initial value afterwards. Simply use torch.initial_seed() like this:

torch.manual_seed(torch.initial_seed())

AFAIK pytorch does not provide arguments like seed or random_state (which could be seen in sklearn for example).

Answer 2

As you can see from the documentation is possible to pass a generator to random_split

random_split(range(10), [3, 7], generator=torch.Generator().manual_seed(42))