This is the repository contains the implementation for the Denoising Diffusion Samplers paper published at ICLR 2023.
This code uses python 3.10. To install we recommend running:
$ pip install -e .
in the root directory of the repository and within a conda environment. Once this is complete please install jaxline and the annealed flow montecarlo repos by following the installation instructions at:
For quick comparison we stored all of our raw unproccessed results in a dictionary named results_full that can be found at denoising_diffusion_samplers/dds_data/results/dds_results.py . Similarly our SMC baselines can be found at denoising_diffusion_samplers/dds_data/results/smc_results.py . For each method, task, steps combination there are 30 different estimates for ln Z using different random seeds.
A notebook that summarises/visualizes all the results form the results dictionary can be found at denoising_diffusion_samplers/notebooks/line_plots_results.ipynb. This is the plot we use in the main section of our paper.
In the notebook folders we have prepared notebooks that immediately reproduce the results for the Funnel and ION tasks at a given number of steps. In order to train from scratch and reproduce the results the correct tuned hpyperparameters must be used for each method. To facilitate this we have provided a full dictionary containing hyperparaeters for each method, tasks, steps combination. This dictionary can be found at denoising_diffusion_samplers/opt_hyperparams.py.
In order to see examples how to train / run, please read any of the following notebooks:
denoising_diffusion_samplers/notebooks/Simple Funnel Test Run.ipynbdenoising_diffusion_samplers/notebooks/Mixture Well.ipynbdenoising_diffusion_samplers/notebooks/Logistic Regression.ipynb
If you use this code in your work, please cite the corresponding paper.
@inproceedings{
vargas2023denoising,
title={Denoising Diffusion Samplers},
author={Francisco Vargas and Will Sussman Grathwohl and Arnaud Doucet},
booktitle={International Conference on Learning Representations},
year={2023},
url={https://openreview.net/forum?id=8pvnfTAbu1f}
}