Home

Awesome

Whittle Networks: A Deep Likelihood Model for Time Series

This repository is the official implementation of Whittle Networks introduced in Whittle Networks: A Deep Likelihood Model for Time Series by Zhongjie Yu, Fabrizio Ventola, and Kristian Kersting, published at ICML 2021.

Setup

This will clone the repo, install a Python virtual env (requires Python 3.6), and the required packages.

git clone https://github.com/ml-research/WhittleNetworks.git
./setup.sh

Datasets

Download datasets from TU Datalib, and unzip:

wget https://tudatalib.ulb.tu-darmstadt.de/bitstream/handle/tudatalib/2887/data.zip
unzip data.zip

Demos

Activate the virtual environment:

source ./venv_wnet/bin/activate

To train WSPN on some of the datasets:

./run_WSPN.sh Sine

"Sine" can be replaced with "MNIST", "SP", "Stock", or "Billiards".

This will train and evaluate WSPNs with 1d, pair, and 2d Gaussian leaf nodes. Details can be found in Table 1 in our paper.

To extract the conditional independence structure from a WSPN:

python script_graph.py --data_type=Sine --graph_type=bn

"Sine" can be replaced with "SP", "Stock", or "VAR".

--graph_type can be either "bn" -- directed graph, or "mn" -- undirected graph.

Bayesian information criterion will be enabled with --BIC

Pre-trained WSPN models are in results/

To train and evaluate conditional WSPN for forecasting on the Mackey-Glass dataset:

python script_wcspn.py

To train and evaluate Whittle AE:

python train_WhittleAE.py
python test_WhittleAE.py

Citation

If you find this code useful in your research, please consider citing:

@inproceedings{yu2021wspn,
  title = {Whittle Networks: A Deep Likelihood Model for Time Series}, 
  author = {Yu, Zhongjie and Ventola, Fabrizio and Kersting, Kristian}, 
  booktitle = { Proceedings of the International Conference on Machine Learning (ICML) },
  pages = {12177--12186},
  year = {2021}
}

Acknowledgments