DIVER-1

EPHYS FOUNDATION MODEL

(Organized by @juliaaayun66)

Paper arxiv link

SETUP

• Before finetuning, set the following requirements.

conda create --name my-env-name --file requirements.txt 
conda activate my-env-name

PREPROCESSING

• We have implemented our own preprocessing methods for the downsteam task datasets, therefore to utilize our model, use our preprocessing files.
• Standard elc files for preprocessing are from https://github.com/fieldtrip/fieldtrip/blob/master/template/electrode/standard_1005.elc

MODEL

• DIVER model: models/diver.py
• Finetuning model: models/finetune_model.py

FINETUNE

• Example scripts are currently implemented on the following downstream tasks; Neuroprobe, FACED, Physionet-MI, MentalArithmetic.
• For Neuroprobe downstream tasks, cloning the Neuroprobe's github is needed; (https://github.com/insight-neuro/neuroprobe/tree/main). After cloning the directory, paste this code in to the ./datasets/datasets_loaders.py file.

#datasets/datasets_loaders.py 
import sys
module_path = "/your/path/to/neuroprobe"
if module_path not in sys.path:
    sys.path.append(module_path)
import neuroprobe

WEIGHTS

Option 1 : Via Google Drive

gdrive link

Option 2: Via lfs

• pretrained weights for iEEG downstream tasks: ./weights/ieeg_pretrained_weights.pt
• pretrained weights for EEG downstream tasks: ./weights/i_eeg_pretrained_weights.pt

NOTE

• If you would like 1s iEEG weights please raise an issue
• Also note that when using the models, please use the mup mode such that the attention to be scaled by d_model not its square root.