TypeError: BrainLMForPretraining.forward() got an unexpected keyword argument 'xyzt_vectors'

[zeyunZhao]

Hello authors,

When I tried to run "tutorial_predict_next_steps.ipynb", I couldn't find the model: "/home/sr2464/Desktop/BrainLM/training-runs/2023-04-21-10_29_26/checkpoint-181" and got the bug

TypeError: BrainLMForPretraining.forward() got an unexpected keyword argument 'xyzt_vectors'.

Can you share the weight of models like "/home/sr2464/Desktop/BrainLM/training-runs/2023-04-21-10_29_26/checkpoint-181", which contains xyzt_projector described in your paper?

Any reply will be appreciated!

[matteomattersberger]

I'm not an author, so take my advice with a grain of salt. The 'xyzt_vectors' seems to have been used in an older version of brainLM, for which the models are not publicly available (at least I haven't found them anywhere). If you check the model config, you'll see that the newer versions require an 'xyz' tensor object instead, with the time dimension being dropped. This works (for me) using the (also old by now) 13M parameter model on huggingface, and also the model version stored on GitHub. I adjusted the model preprocessing scripts to account for this, and now it seems to work.

However, the newest versions (111M and 650M) seem to require input of yet another shape, which I myself am currently stuck with. Great though as this project is, version control and documentation is often very confusing. Hope this was helpful.