This work try to reproduce Défossez's work (Decoding speech from non-invasive brain recordings)

Dataset

Please first load data from https://osf.io/ag3kj/, you will get like sub01/ses0/meg/... save it to /data/meg/meg.gwilliams2022neural/ like ../data/meg.gwilliams2022neural/sub-01/ses-0/meg/sub-01_ses-0_task-0_meg.con

Preprocess

it is structured as python session.py python run.py python subject.py python dataset.py But you can directly run dataset.py

models

the defossez2022decoding is for brain_decoding model, you can find subjectlayer, spatialattention, etc. in layers file the wav2vec is for speech_decoding model

prarams

initialized parameters for model

utils

all kinds of metric, data load/save,... Before you run the train, first get the dataset. you should run gwilliams2022neural.py in utils/data/meg

train

Finally, run defossez2022.py please.

In brief, run dataset.py, gwilliams2022neural.py, defossez2022.py