chen-hao-chao / ebflow

[NeurIPS 2023] Training Energy-Based Normalizing Flow with Score-Matching Objectives

Training Energy-Based Normalizing Flow with Score-Matching Objectives

This repository contains the code implementation of the experiments presented in the paper Training Energy-Based Normalizing Flow with Score-Matching Objectives.

Directory Structure

• Use the code in ebflow/toy_examples to reproduce the results presented in Sections 5.1 and A5.
• Use the code in ebflow/real_world to reproduce the results presented in Sections 5.2, 5.3, and 5.4.

Dependencies

(Optional) Launch a Docker Container

# assume the current directory is the root of this repository
docker run --rm -it --gpus all --ipc=host -v$(pwd):/app nvcr.io/nvidia/pytorch:20.12-py3
# inside the docker container, run:
cd /app

Install Dependencies

Setup the conda environment with conda_environment.yml:

conda env create -f conda_environment.yml

Launch ebflow conda environment:

source activate
conda activate ebflow

References

This code implementation is developed based on the following repositories:

• taufikxu/FD-ScoreMatching (at commit 9df0789) is licensed under the MIT License.
• akandykeller/SelfNormalizingFlows (at commit 9feebb3) is licensed under the MIT License.
• yang-song/score_sde_pytorch (at commit cb1f359) is licensed under the Apache-2.0 License.
• ermongroup/sliced_score_matching (at commit 880c047) is licensed under the GPL-3.0 license.
• kamenbliznashki/normalizing_flows (at commit 97a73a0).
• fissoreg/relative-gradient-jacobian (at commit d2e03ca).

Citing EBFlow

If you find this code useful, please consider citing our paper.

@inproceedings{chao2023ebflow,
      title={{Training Energy-Based Normalizing Flow with Score-Matching Objectives}}, 
      author={Chen-Hao Chao and Wei-Fang Sun and Yen-Chang Hsu and Zsolt Kira and Chun-Yi Lee},
      year={2023},
      booktitle={Proceedings of International Conference on Neural Information Processing Systems (NeurIPS)}
}