I don’t get the exact same results, but I don’t know that the model is trained once on another GPU.
I am thinking of training the model once on one GPU and then on another GPU and then evaluate and compare.
I think that it is important to use the same initial seed (torch.initial_seed — PyTorch 1.9.0 documentation) to make these GPUs more comparable. Can you give me a suggestion, how to initial it?
It looks like that now:
Is there anything else that I can have a better comparison of the results?