Problem When Submitting Result Files

[yusowa0716]

Dear OCP Team,

I am encountering a problem when attempting to submit my oc20 s2ef result files. The specific issue is

Object arrays cannot be loaded when allow_pickle=False.

I mimicked the way the results file is generated in trainer and used make_submission_file.py to generate the final submission file. To help me better understand how to properly format and submit the result files, I would appreciate it if the team could provide a sample file as a reference.

Thank you very much for your help and support!

Best regards,

Yusong Wang

[mshuaibii]

Hi -

Looking at the submission you made, you can do the following to manually modify the npz file so you don't have to re-run predictions:

• For all force predictions in your file, concatenate them - i.e. np.vstack(data["id_forces"])
• For all energy predictions in your file, flatten them - i.e. data["id_energy"].flatten()

You can download a sample prediction file here to give you an example of what it should look like.

To get to the bottom of why this didn't save as expected, can you let me know what commit hash you are working out of for the repo? I know we had made some changes with prediction and something may have changed there.

Hope this helps!

[yusowa0716]

Hi,

After I did the above, I got the submission file in the same shape as the reference submission file and I didn't get the previous error, however I got a new error:

The size of tensor a (24750831) must match the size of tensor b (29447516) at non-singleton dimension 0

BTW, I'm not using ocp's repo, but my own trainer, but I'm following the latest code.

[mshuaibii]

It looks like your chunk_idx is stored incorrectly. It should be a a cumulative sum, try doing np.cumsum() on all your *_chunk_idx keys.

[yusowa0716]

Thanks for your help!