Fig 1. (Top) Generative model of Neural Continuous-Discrete State Space Model. (Bottom) Amortized inference for auxiliary variables and continuous-discrete Bayesian inference for states.

This repository contains code for reproducing the experiments presented in the ICML 2023 paper Neural Continuous-Discrete State Space Models for Irregularly-Sampled Time Series.

Installation

NCDSSM requires Python 3.8 or higher.

• Create a conda environment (optional, but recommended). sh conda create --name ncdssm python=3.9 conda activate ncdssm
• Clone this repo and install the package. sh # Install package with all optional dependencies for running experiments pip install --editable ".[exp]"

Usage

Generate or download datasets

• For bouncing ball, damped pendulum, climate and Pymunk datasets, generate/download them by running the scripts in data/. ```sh # Bouncing Ball dataset python data/bouncing_ball.py

Damped Pendulum dataset

python data/damped_pendulum.py

USHCN Climate dataset

python data/climate.py

Pymunk datasets

python data/box.py python data/pong.py `` * For CMU MoCap dataset, download the preprocessed dataset (the file namedmocap35.mat) available at [cagatayyildiz/ODE2VAE](https://github.com/cagatayyildiz/ODE2VAE) and place it indata/mocap/mocap35.mat`.

Training

Train models using train_pymunk.py --config /path/to/config for the Pymunk datasets and train_ts.py --config /path/to/config for all other datasets. The yaml config files can be found in configs/. The keys in the config files can also be passed as command line options to the training script as --key value. ```sh

Train NCDSSM-NL on Damped Pendulum with 30% missing

python traints.py --config configs/pendulumncdssmnl.yaml --trainmissingp 0.3

Train NCDSSM-LL on Bouncing Ball with 80% missing

python traints.py --config configs/ballncdssmll.yaml --trainmissingp 0.8

Train NCDSSM-NL on CMU MoCap (setup 1)

python traints.py --config configs/mocapncdssmnl.yaml

Train NCDSSM-LL on CMU MoCap (setup 2)

python traints.py --config configs/mocap2ncdssmll.yaml

Train NCDSSM-NL on USHCN Climate (first fold)

--sporadic means that data is missing both in time and feature dimensions

python traints.py --sporadic --config configs/climatencdssmnl.yaml --data_fold 0

Train NCDSSM-LL on Box

python trainpymunk.py --config configs/boxncdssmll.yaml ```

Evaluation

Once trained, models can be evaluated on the test set using the eval_pymunk.py and eval_ts.py. Alternatively, you may download our pretrained checkpoints from this repository's Releases or from :hugs: hugging face and place them in ./checkpoints/. ```sh

Download pretrained checkpoint from Releases

wget https://github.com/clear-nus/NCDSSM/releases/download/v0.0.1/checkpoints.tar.gz \ && tar -xzf checkpoints.tar.gz \ && rm checkpoints.tar.gz ```

```sh

Evaluate NCDSSM-NL on Damped Pendulum with 30% missing

python evalts.py --ckpt checkpoints/dampedpendulum0.3ncdssmnl.pt --smooth

Evaluate NCDSSM-LL on Bouncing Ball with 80% missing

python evalts.py --ckpt checkpoints/bouncingball0.8ncdssmll.pt --smooth

Evaluate NCDSSM-NL on CMU MoCap (Setup 1)

python evalts.py --ckpt checkpoints/mocapncdssmnl.pt --nostatesampling --seed 0

Evaluate NCDSSM-LL on CMU MoCap (Setup 2)

python evalts.py --ckpt checkpoints/mocap2ncdssmll.pt --nostatesampling --seed 0

Evaluate NCDSSM-NL on USHCN Climate (first fold)

python evalts.py --sporadic --ckpt checkpoints/climatencdssmnl.pt

Evaluate NCDSSM-LL on Box

python evalpymunk.py --ckpt checkpoints/boxncdssmll.pt --wass --smooth --nostatesampling --seed 0 ```

Questions

For any questions regarding the code or the paper, please email Fatir.

BibTeX

If you find this repository or the ideas presented in our paper useful, please consider citing. @inproceedings{ ansari2023neural, title={Neural Continuous-Discrete State Space Models for Irregularly-Sampled Time Series}, author={Ansari, Abdul Fatir and Heng, Alvin and Lim, Andre and Soh, Harold}, booktitle={International Conference on Machine Learning}, year={2023} }

Acknowledgement

This repo contains parts of code based on the following repos.

| Repo | Copyright (c) | License | | ------------- | ---------- | ------------- | | edebrouwer/gruodebayes | Edward De Brouwer | MIT License | | cagatayyildiz/ODE2VAE | Çağatay Yıldız | NA | | mannyray/KalmanFilter | @mannyray | MIT License | | simonkamronn/kvae | Simon Kamronn | MIT License