issues with reproducing pLogP benchmark for GPCN

[chrisvdwerf]

Hello,

I am trying to reproduce the logP benchmark for GCPN: Coming close to a top-1 property score of 6.560 Penalized LogP.

For this, I ran the same code provided in the tutorial multiple times. However, the highest Penalized LogP values I seem to achieve is ~ 2.7 which is way lower than the benchmark of 6.560.

Did anyone manage to reproduce the benchmark?

Thank you!