Code for the PointTransformer was taken from https://github.com/qq456cvb/Point-Transformers and adjusted by Marga Don for the purposes of this project.

Process the data

To convert the cortical data to point clouds, first make sure to download the data. It should be structured as such:

├── cortical_data
│   ├── diffusion
│   ├── structural
│   │   ├── cortical_thickness
│   │   ├── curvature
│   │   ├── midthickness
│   │   ├── myelin_map
│   │   ├── sulcal_depth
│   ├── train_labels.csv

To preprocess the data, run

python process_cortical_data.py --src_dir {SRC_DIR} --save_dir {SAVE_DIR}

where SRC_DIR is the path to the directory where you saved the data and SAVE_DIR is the path to the directory where you want to save the processed data. This step should take about 2 minutes.

Training

Run

python train.py

to train the model.

The code includes Weights & Biases (wandb) logging code, either log in to your wandb account or uncomment the lines relating to wandb.