Lack of reproducibility despite using the same seed

#1

Hello,
We’ve been having some discussion concerning a lack of reproducibility despite using a consistent seed in xgboost when employing Bayesian optimization for the hyperparameters. After some digging, the issue seems to be in the xgboost package itself, but I’m still unclear as to what’s driving this. Could you please take a look? a sample syntax to reproduce the problem is available here: https://github.com/AnotherSamWilson/ParBayesianOptimization/issues/43

Thank you!

#2

Any chance to review this? Thanks!

#3

There is now a fix available: https://github.com/dmlc/xgboost/pull/7303

#4

Thanks for reviewing this! So unfortunately we are somewhat limited in terms of server access. Any estimation as to when this fix will be incorporated in the R release?

Thanks again

#5

It will be part of 1.6.0 release.