Code for "Contrastive learning to fine-tune feature extraction models for the visual cortex".

1. Download the images, fMRI data, and ROI masks from the Algonauts Project website: [http://algonauts.csail.mit.edu/challenge.html] (http://algonauts.csail.mit.edu/challenge.html). Put the fMRI data in a subfolder called training_fmri in the root project directory, and the images in a subfolder called training_images. Both of those folders should have one folder for each subject x named "Subjx". For the fMRI data, each subject folder should have two files: subjx_lh_training_fmri.npy and subjx_rh_training_fmri.npy. Each subject folder in the training_images directory should have all of the images for that subject (with names like train-0001_nsd-00013, as downloaded). The ROI masks should be in a roi_masks folder in the root project directory, also with one folder for each subject called Subjx that contains all of the mask files (lh.all-vertices_fsaverage_space, rh.floc-faces_challenge_space, etc.)
2. Before running any of the other scripts, run the generate_voxel_roi_mappings.py script, which creates a folder called voxel_roi_mappings in the root directory, with a folder for each subject containing files that map all the voxels in the left and right hemisphere to the ROIs they are in.
3. Train CL, regression, or pooled models using cl_model_training.py, cl_model_training_all_rois.py, reg_model_training.py, reg_model_training_all_rois.py, pooled_model_training.py, and pooled_model_training_all_rois.py. Options for each are at the top of the scripts.
4. Once the models are trained, they can be used to get encoding results for NSD with the fit_encoding_models.py script. Similarly, the trained models can be used to get encoding results for NOD using the fit_encoding_models_nod.py script.

For all runnable scripts, options are specified at the begining of the file.
generate_voxel_roi_mappings.py -- script that should be run first to create mappings from voxels to the ROIs they are in for all voxels in each subject
utils.py - contains helper functions for dataset loading, mapping voxels to correct ROIs, extracting features from AlexNet models, etc.
results_utils.py - contains helper functions for downstream tasks, mutual information estimates, more.
models.py - defines classes and functions for CL models and regression models
cl_model_training.py, cl_model_training_all_rois.py, reg_model_training.py, reg_model_training_all_rois.py, pooled_model_training.py - scripts that can be run from the command line to train CL or regression models for each a single ROI (model_training files) or all the ROIs in a hemisphere for the specified subject (all_rois files), or for the pooled approach (using all subjects' data).
fit_encoding_models.py - script that used untuned, cl-tuned, reg-tuned, or pooled-tuned AlexNet as feature extractor to fit encoding models for NSD dataset.
fit_encoding_models_nod.py - script that used untuned, cl-tuned, reg-tuned, or pooled-tuned AlexNet as feature extractor to fit encoding models for NOD dataset.
generate_results.py - script that can be used to generate results for downstream image classification tasks, mutual information lower bound estimates, more. feature_extraction_model_landscape_imagenet.ipynb - notebook that can be used to generate model landscape plots