search

DeepGraphLearning / KnowledgeGraphEmbedding

MIT License
1.24k stars 264 forks source link

About the seed setting. #21

Closed Zawo1995 closed 4 years ago

Zawo1995 commented 4 years ago

Thanks for sharing the wonderful code :) When I reproduce the RototE，I find the result is different each time(but just a little),such as loss.mr,mrr,etc.So I find that your code didn't contain ant code about seed setting code,like seed = 42 np.random.seed(seed) torch.manual_seed(seed) if is_cuda: torch.cuda.manual_seed_all(seed)

It is different from the tutorial ever I did. Could you give some explanation about your thinking about that?

Looking forward to your repeat,sorry for asking so shallow issues QAQ.

Edward-Sun commented 4 years ago

Yes. We didn't set random seeds because we find that the results are insensitive to different random seeds. As we report in the paper, the variance is usually only 0.000-0.001.

It is common that when the hidden size is large enough, the random initialization won't influence much on the final results.