Even if the seed is fixed, it is difficult to reproduce the experiment

1

Hi,

I have set seed everything, but the results were very different from experiment to experiment.

How do explain this strange phenomenon?

2

In general determinism can be difficult to guarantee as many operations (e.g., if you are using a GPU) are not necessarily deterministic. Have you tried the advice in the reproducibility/randomness post? Reproducibility — PyTorch 1.11.0 documentation

1 Like
3

Thank you for your advice!

After reading the link you provided, I set the following parameters

torch.backends.cudnn.deterministic

Now, everything is ok

1 Like