Why is it required to set seeds separately for each library?

PyTorch Live
1

I have set the seed for torch in my code as:

torch.manual_seed(3)

Now, since I am importing numpy and tensorflow as well, is it required set the seeds for each library separately? If yes, why?

2

Yes, this is required since each library uses their own pseudorandom number generator, which needs to be seeded.

3

Thank you for the clarification.
So does the number I use matter? For example, since I used 3 for torch, should I use 3 for the other libraries also or it can be any random number?

4

No, I don’t think it matters since different PRNG implementations could be used as already seen in this small example:

torch.manual_seed(2809)
torch.randn(1)
# tensor([-2.0748])

np.random.seed(2809)
np.random.randn(1)
# array([0.23747478])

Also note that not all libraries support the same backends (e.g. numpy doesn’t support the GPU) so a parity isn’t even possible for all use cases.

1 Like