Rényi Neural Processes

This repository accompanies the paper "Rényi Neural Processes"

Xuesong Wang, He Zhao, and Edwin V. Bonilla . ICML 2025 (Oral, ~1 % of total submissions)

---

1D Regression

---

Training

python regression/main_gp.py --data_name=RBF --model_name=NP --mode=train --divergence=Renyi_0.7    

The config of hyperparameters of each model is saved in regression/configs/gp. If training for the first time, evaluation data will be generated and saved in regression/evalsets/gp. Model weights and logs are saved in regression/results/{data_name}/{model_name}/.

Evaluation

python regression/main_gp.py --data_name=RBF --model_name=NP --mode=eval --divergence=Renyi_0.7    

Plotting

python regression/main_gp.py --data_name=RBF --model_name=NP --mode=plot --divergence=Renyi_0.7    

Some codes are borrowed from https://github.com/tung-nd/TNP-pytorch , please refer to their documentation for more details for the dataset.

Lotka Volterra & Hare Lynx datasets

---

Simulation Data Generation (Lotka Volterra)

python3 data/lotka_volterra.py --filename=train --num_batches=10000 --trajectory_all=0

python3 data/lotka_volterra.py --filename=eval --num_batches=1000 --trajectory_all=0

The code will generate datasets/lotka_volterra/train.tar and `'datasets/lotka_volterra/eval.tar'

Realworld Data Generation (Hare Lynx)

python3 data/hare_lynx.py

The code will download the dataset to datasets/lotka_volterra/LynxHare.txt

Training

python regression/main_lotka_volterra.py --data_name=lotka_volterra --model_name=NP --mode=train --divergence=Renyi_0.7    

---

Citation

Please cite us if you use this work:

Xuesong Wang, He Zhao, Edwin V. Bonilla. The Forty-Second International Conference on Machine Learning (ICML), 2025.

@article{wang2025rnp,
  title={R$\backslash$'enyi Neural Processes},
  author={Wang, Xuesong and Zhao, He and Bonilla, Edwin V},
  booktitle={International Conference on Machine Learning},
  year={2025}
}

Acknowledgements

• Transformer neural processes: https://github.com/tung-nd/TNP-pytorch
• Lotka volterra: https://github.com/juho-lee/bnp