In-progress -- this repo is under active development in the MedARC discord server. https://medarc.ai/fmri
-
Run setup.sh to create a new "foundation_env" virtual environment
-
Activate the virtual environment with "source foundation_env/bin/activate"
- main.ipynb (use accel.slurm to allocate multi-gpu Slurm job)
Save latents to hdf5 / parquet:
Then evaluate downstream performance using the saved latents:
- mindeye_downstream.ipynb
- HCP_downstream.ipynb
This requires having access to train_subj01.hdf5 which is saved in "/weka/proj-fmri/paulscotti/fMRI-foundation-model/src".
If you cannot access this file, the commented out code shows how to create this file yourself.
- mindeye_finetuning.ipynb