Impact of first-stage diversity filter on scaffold appearance in second stage learning

[yukisoya]

In a staged learning process, consider a case where a diversity filter was implemented in the first stage. At the midpoint of the first stage, the number of compounds containing scaffold 'A' reached the bucket_size. Under these circumstances, would scaffold 'A' still be expected to appear in the second stage of learning?

[halx]

Hi,

you are basically asking if the diversity memory is carried over into a follow-up run. This is controlled by use_checkpoint (in [parameters]) which is by default false. If you want to read in the scaffolds from the previous stage, set this to true and you will continue with the diversity filter state as it was in the final epoch of the previous stage. This is the case when you run multiple stages manually through separate configuration files.

If you set up multiple stages in a single configuration file you would need to set purge_memories. By default it is true which means that the diversity filter state is not carried over to the next stage.

Cheers,
Hannes.