OC20 S2EF 200K Training Errors

[pgeitner]

Hello,

I apologize that this question is about OC20. I am a student looking to explore with this problem. Given my lack of compute resources, I am working with the S2EF 200K split from OC20. After following the tutorials I am able to download the split and build the TrajectoryLMDBDataset to interact with the data (e.g., Data(atomic_numbers=[66], cell=[1, 3, 3], fid=237, fixed=[66], force=[66, 3], id=“0_0”, natoms=66, pos=[66, 3], sid=1716871, tags=[66], y=-263.89420902)).

However, when I go to train any of the provided models (GemNet, SchNet, etc.), I get errors due to the fact that the data does not contain edge indices:

Is this error related to how I’ve downloaded/pre-processed the data? If so, how can I fix this? Do I need to store edges when downloading? Again, my apologies for asking about OC20 and such a simple problem, but I am just trying to get started. Thanks in advance.

[mshuaibi]

Hi - No worries. This occurs when S2EF data is processed without edge information, specifically not including --get-edges in the download script. This is perfectly fine, however you must define otf_graph=True in the model configs in order to train properly and compute edges on-the-fly.

[Sejal2002]

Hello team,

I was training running-ML relaxations for the IS2RS task using a pre-trained S2EF checkpoint on my IS2RES dataset.
However, after the relaxations were complete, I got the same error as above, indicating that the “pos” field was not present in the resulting trajectories.
May I please know any possible reasons for this and how I can fix it?
Thank you!