NEO: Non Equilibrium Sampling on the orbit of a deterministic transform

Description of the code

This repo describes the NEO estimator described in the paper NEO: Non Equilibrium Sampling on the orbit of a deterministic transform published at NeurIPS 2021 and available https://papers.nips.cc/paper/2021/file/8dd291cbea8f231982db0fb1716dfc55-Paper.pdf.

Three notebooks describe typical experiments of the main paper.

• Mix_gaussian, the normalizing constant estimation on a mixture of Gaussian distributions
• Sampler the sampling of a mixture of Gaussian distributions or on Funnel distribution.
• Experiments_colab the training of VAE.

Requirements

Mainly uses pytorch, pyro-ppl. Later tensorboard

pip install -r requirements.txt

To cite this work

If you use this repository, please reference our article e.g. using bibtex

@inproceedings{thin2021neo, title={NEO: Non Equilibrium Sampling on the Orbits of a Deterministic Transform}, author={Thin, Achille and El Idrissi, Yazid Janati and Le Corff, Sylvain and Ollion, Charles and Moulines, Eric and Doucet, Arnaud and Durmus, Alain and Robert, Christian P}, booktitle={Thirty-Fifth Conference on Neural Information Processing Systems}, year={2021} }

or other formats available at https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&authuser=1&q=neo+non+equilibrium&btnG=&oq=neo+no#d=gs_cit&u=%2Fscholar%3Fq%3Dinfo%3AeV5WBKEHRfkJ%3Ascholar.google.com%2F%26output%3Dcite%26scirp%3D0%26hl%3Den%26authuser%3D1.