RL for libinvent

[vivianzheng-st]

Hello, I run a RL for libinvent transformer model for my molecules based on the config default provided in REINVENT4 example for stage 1 and stage 2, however, the csv file of stage 1 shows that I have about 3000 molecules and stage 2 which is 12000 molecules so I am interested in why this happens because aren't stage 2 supposed to generate less molecules then stage 1? Since my understanding is that RL works like a funnel that over each stage it generate much less molecules but more targeted and diverse than stage 1. I have attached the csv file for both stage 1 and stage 2 learning and I used exactly same code in the configs/staged_learning.toml both for stage 1 and stage 2.
staged1_learning_1.csv
staged2_learning_1.csv

[halx]

The stages run for as many steps as specifified unless a termination criterion is always given. But termination is based on the current total score. Funneling could be a potentail extension depending on how it is done but at the moment you would need to use a post-processing step to reduce the number and pick out the molecules you like.