Deep Learning Models for Eye Tracking Trials

Installation

We recommend using a conda environment with Python >= 3.12 :

conda create -n precog python=3.12
conda activate precog

Clone the repository and install the dependencies:

git clone https://github.com/usc-sail/precog-eye-dl
cd precog-eye-dl
pip install -r requirements.txt

You should have access to the data folder named eyelink_processed. You should copy this folder in this directory.

Code Structure

• config.yaml contains sample values for the experiment variables.
• dataset.py contains the dataset class for the eye tracking data.
• model.py contains the model classes to be used for modeling (adapted from TimesNet).
• trainer.py ontains the class to perform train and evaluation.
• script.py is the main script to run for model traning.
• preprocess_*.py creates the input data for the model from ASCII files.

Data Structure

The data for this code is located under /PATH/TO/eyelink-processed/.

Pre-processing

You do not need to run pre-processing, since all the data have been processed and are available in the data folder input_30trials. To reproduce this sub-folder you need to run the preprocess.py and preprocess_1_7.py (for first 7 participants) scripts. These scripts read the raw asc_files.

Running the models

To run the deep learning pipeline for C v DS training (adjust for C v S inside the code), you should run

python script.py

Most parameters can be tuned withing the config.yaml file.