Discuss on the reproducibility

[baileytrang]

Hi authors,

I am trying to reproduce the results of your method and the baselines.
It looks good since the results I got are very close to your reported ones, in multiple random seeds.

However, my observations are:

• With the same random seed and same GPU, different runs will give different losses, valid results, and test results (around 1-2%)
• With the same GPU and trained model, testing in different times will give different results sometimes.
• Easy to see that the problems above will happen when we use different GPUs too.
  (different GPUs but the same GPU model and configurations)

Personally, I think this issue did occur from the baseline's implementation...

Could you please guide me on how to fix this issue?

Thank you in advance!