Finetuning on datasets with continuous electrode positions

[alveare]

Hi, thank you for sharing this great work.

I'm trying to finetune your model on intracranial datasets other than BrainTreeBank, and I have a question regarding the channel position embedding.

In the current implementation, it seems that channel positions are handled through discrete indices (e.g., 0–4999 lookup embedding) with a 'pe' tensor of size [1, 5000, 128].
For datasets where electrode positions are available as continuous 3D coordinates (e.g., x,y,z in standard brain space), is it possible to fine-tune PopT on such datasets?