Kindly download the data for replication at https://drive.google.com/drive/folders/159L1pnh1kArIgZMNxlWp2Pv2hihgB0Mf?usp=share_link

From the google drive folder, extract humandata.zip and either place the contents at /nfs/projects/humanattn/data/eyesum or change this locaton in train.py inside the humantrain/ directory in this repo

Using the following command(s) inside the humantrain directory of this repo, a model can be trained over human eye gazing data

python3 train.py --model-type=astgnn --epochs=70 --batch-size=200 

Alternatively,

./runner.sh>output.txt

may be used to extract average correlations for training human attention

From the google drive folder, extract q90dats.zip and either place the contents at /nfs/projects/humanattn/data/javastmt/q90 or change this location inside biodatmaker.py inside the makebiodat/ directory in this repo.

Using the following command inside the makebiodat directory , create the biodats

python3 biodatmaker.py

You may need to edit the model location and name inside the biodatmaker.py file These biodats are also all provided in the dataset at biodats.zip

Using the following command inside the main repository, you may train the model that mimicks human attention

python3 train.py --model-type=attendgru-bio --with-graph --with-biodats=biodats_q90_astgnn.pkl --epochs=10

You may need to change the directories inside train.py to read in the q90data as well as the biodats.

python3 predict.py /path/to/model/ --with-graph --with-biodats=biodats_q90_astgnn.pkl

gpu=X can be used with any of these scripts to make use of gpu instead of cpu depending on availability, we used a titan rtx 24gb to train these models with tensorflow 2.4

Resulting predictions are available inside the predictions directory of this repository.

Attention plots and the script to generate these attentions are provided in the plotter subdirectory

python3 plotter.py --fid=X