[QUESTION]: training dataloader is currently defined with shuffle=False. Is this intentional? about molpal HOT 1 CLOSED

Hey @coparks2012 - thanks for poring through the code. You're right that it's weird that we don't shuffle the data and that it might cause training instability based on the ordering of the dataset. However, I can tell you that the MPNN code was adapted from the legacy chemprop code and there were quite a few odd design choices in there. When I wrote this, my goal was to simply interface the two codebases with as little changes to chemprop as possible. Of course, it resulted in instances like this, but it seems to work fairly well across a variety of tasks. If we were to change it, we'd have to redo a bunch of the original studies to assess the impact (positive, negative, or neutral) of shuffling on overall optimization performance, and that's just something I've never gotten around to.

from molpal.