Generative Modeling of Regular and Irregular Time Series Data via Koopman VAEs

Authors: Ilan Naiman, N. Benjamin Erichson, Pu Ren, Michael W. Mahoney, Omri Azencot

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we detail the results of our extensive experiments. Details related to datasets and baselines are provided in App. E. Our code is available at Git Hub.
Researcher Affiliation Academia Ilan Naiman Ben-Gurion University, UC Berkeley naimani@post.bgu.ac.il N. Benjamin Erichson LBNL and ICSI erichson@lbl.gov Pu Ren LBNL pren@lbl.gov Michael W. Mahoney ICSI, LBNL, and UC Berkeley mmahoney@stat.berkeley.edu Omri Azencot Ben-Gurion University azencot@cs.bgu.ac.il
Pseudocode No The paper describes the steps of its model and training objective in text, but it does not include a distinct figure, block, or section labeled "Pseudocode" or "Algorithm" with structured steps.
Open Source Code Yes Our code is available at Git Hub.
Open Datasets Yes Energy, is a multivariate UCI appliance energy prediction dataset (Candanedo, 2017)... Finally, Mu Jo Co (Multi-Joint dynamics with Contact) (Todorov et al., 2012)... The dataset in this task comprises temperature at 2-meter height above the surface from ERA5 reanalysis dataset (Hersbach et al., 2020).
Dataset Splits No We split the train and test sets with a ratio of 80% and 20%. (Appendix G). This explicitly states train/test splits for the weather data, but a validation split is not explicitly mentioned for all experiments or any general case.
Hardware Specification Yes The software environments we use are: Cent OS Linux 7 (Core) and PYTHON 3.9.16, and the hardware is: NVIDIA RTX 3090.
Software Dependencies Yes The software environments we use are: Cent OS Linux 7 (Core) and PYTHON 3.9.16, and the hardware is: NVIDIA RTX 3090.
Experiment Setup Yes Combining the penalties from Sec. 3 and the loss Eq. 7, we arrive at the following training objective which includes a reconstruction term, a prediction term, and a regularization term, L = Ez1:T q[log p(xt1:t N |z1:T )] + αLpred(z1:T , z1:T ) βKL[q(z1:T |xt1:t N ) p(z1:T )] , (8) where α, β R+ are user weights... We also explore how stable our model is to hyperparameter choice. To this end, we perform an extensive grid search over the following space for: α, β {1.0, 0.9, 0.7, 0.5, 0.3, 0.1, 0.09, 0.07, 0.05, 0.03, 0.01, 0.009, 0.007, 0.005, 0.003, 0.001, 0.0009, 0.0007, 0.0005}2 for the stocks dataset. (Appendix J).