Awesome

GraFITi

This is the source code for the paper GraFITi: Graphs for Forecasting of Irregularly sampled Time Series

Requirements

python 3.8.11

Pytorch 1.9.0

sklearn 0.0

numpy 1.19.3

pandas 1.5

Training and Evaluation

We provide an example for physionet for observing 36 hrs and predicting 12 hrs. All the datasets can be run in the similar manner.

train_grafiti.py --epochs 200 --learn-rate 0.001 --batch-size 128 --attn-head 1 --latent-dim 128 --nlayers 4 --dataset physionet2012 --fold 0 -ct 36 -ft 12

Remaining datasets can be run similarly. MIMIC-IV and MIMIC-III require permissions to download the data. Once, the datasets are downloaded, you can add them to the folder .tsdm/rawdata/ and use the TSDM package to extract the folds. We use TSDM package provided by Scholz .et .al from [https://openreview.net/forum?id=a-bD9-0ycs0]