HIPPOSVGP

Official PyTorch Codebase for the paper titled Recurrent Memory for Online Interdomain Gaussian Processes (NeurIPS 2025)

Installation

# if using anaconda
# conda create -n test_HIPPOSVGP python=3.10 
# conda activate test_HIPPOSVGP
# install in editable mode
python -m pip install -e /path/to/HIPPOSVGP
# https://cds.climate.copernicus.eu/how-to-api follow instructions to install cdsapi and get an api key

Running

1D Solar or Audio dataset

python -m scripts.onedim.solar.train_full_first2 # run this frist 
python -m scripts.onedim.solar.test_vff_online # OVFF
python -m scripts.onedim.solar.train_svgp_online # OSGPR
python -m scripts.onedim.solar.train_svgp_online_pivini # OVC
python -m scripts.onedim.solar.test_hsgp_online_fixedkernel #OHSGPR

1D COVID

python -m scripts.onedim.covid.test_osvgp_piv # OVC
python -m scripts.onedim.covid.test_osvgp # OSVGP
python -m scripts.onedim.covid.test_ohsgp # OHSVGP

UCI

python -m scripts.multidim.gp_ini # obtain kernel hyperparameters from a full GP trained on the first task
python -m scripts.multidim.main_hsgp_online_toy
python -m scripts.multidim.main_hsgp_powerplant
python -m scripts.multidim.main_hsgp_skill
python -m scripts.multidim.main_svgp_online_toy
python -m scripts.multidim.main_svgp_powerplant_piv
python -m scripts.multidim.main_svgp_powerplant
python -m scripts.multidim.main_svgp_skill_piv
python -m scripts.multidim.main_svgp_skill

GPVAE

python -m scripts.gpvae.test_osvgpvae_era5 
python -m scripts.gpvae.test_osvgpvae_ovc_era5
python -m scripts.gpvae.test_ohsvgpvae_era5 # run osvgpvae_era5 first before hippo because some hyperparameters for task 1 are initiated the same

Citing the paper (bib)

@inproceedings{chen2025recurrent,
  title = {Recurrent Memory for Online Interdomain Gaussian Processes},
  author = {Chen, Wenlong and Kiyohara, Naoki and Zhu, Harrison Bo Hua and Curran-Sebastian, Jacob and Bhatt, Samir and Li, Yingzhen},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2025}
}

Third-party code and licenses

This project includes code adapted from the HiPPO project by HazyResearch.

• Source: HazyResearch/hippo-code model/hippo.py and model/op.py
• Adapted into: hipposvgp/hippo.py
• License: Apache-2.0 (see upstream LICENSE)

For full third-party license text and details, see THIRD_PARTY_NOTICES.md in this repository.