Odd shape for 3D flex train graph

[forrest]

Hi,

I’ve been playing around with a dataset that has significant flexibility around a homodimerization domain and one of the things I wanted to try was 3D flexible refinement. I’ve generally followed the steps outlined in the recent end-to-end processing guide but when I run 3D flex training, things get weird around iteration 500 (see representative example below) and don’t recover for the remainder of the training.

My dataset is pretty well-pruned via a few rounds of iterative heterogeneous refinement and I’m pretty confident that I set up the mesh and data prep jobs correctly. Is this blip in the plot representative of something wonky, and if so is there a way to correct for this?

[hbridges1]

Hi @forrest,

Thanks for your question! The learning rate schedule for 3D Flex Training changes around iteration 500, so we typically expect to see a blip in the validation loss in this region. Your plots look similar in this regard to those that we got for the training in our CGRP receptor end-to-end processing guide (shown below). The downstream Flex Generate and/or Flex Reconstruct results should not be affected by the blip that you observe early in the training, and hopefully ran successfully with useful outputs!