WinIT

Code for paper "Temporal Dependencies in Feature Importance for Time Series Predictions" presented at the ICML 2021 Time Series Workshop

After cloning make sure to initialize the git submodules by using:

git submodule init
git submodule update

Environment

This requires:

Python 3.7
Anaconda
(GPU) CUDA 10.1

Tested on Ubuntu 18.04.

Anaconda

Install the individual edition:

wget https://repo.anaconda.com/archive/Anaconda3-2021.05-Linux-x86_64.sh
chmod +x Anaconda3-2021.05-Linux-x86_64.sh
./Anaconda3-2021.05-Linux-x86_64.sh

Restart your shell after installing so that you are using the base conda environment.

Setup Conda Environment

conda env create -f linux_environment.yml

Run paper experiments

Generate synthetic datasets

Generate spike datasets:

python -m FIT.data_generator.simulations_threshold_spikes

This will generate five datasets and store them in data/:

The original spike dataset (data/simulated_spike_data)
Four spike datasets with delays of 1 through 4 (data/simulated_spike_data_delay_X).

Run benchmarks

python experiments.py

And will store the results in the csv file specified at the beginning of the script.