Skip to content

Files

Latest commit

 

History

History

Physionet

Attentive ODEs for Irregularly-Sampled Time Series

Prerequisites

Install torchdiffeq from https://github.com/rtqichen/torchdiffeq.

conda env create -f attentive_ode.yml

Experiments on different datasets

By default, the dataset are downloadeded and processed when script is run for the first time.

Raw datasets:

[Physionet]

Running Attentive ODE

  • Attentive ODE
sh test.sh

or

python3 run_models.py  --niters 20 -n 8000 -l 20 --dataset physionet --attentive-ode --rec-dims 40 --rec-layers 3 --gen-layers 3 --units 50 --gru-units 50 --quantization 0.016 --classif

Latent_ODE

  • DATA : Physionet
Time Series Latent(Seed) AUC Score
1991 (91194) 0.8470
2022 (93669) 0.8430
2021 (53879) 0.8496
2020 (6152) 0.8454
2019 (51159) 0.8470

Test MSE on PhysioNet. Encoder-decoder models.

Model Interp(x 10^-3) Extrap(x 10^-3)
RNN-VAE 5.930±0.249 3.055±0.145
LatentODE (RNN enc.) 3.907±0.252 3.162±0.052
LatentODE (ODE enc.) 2.118±0.271 2.231±0.029
LatentODE + Poisson 2.789±0.771 2.208±0.050
ACE-NODE(ODE enc.) -- 2.045±0.039