Divide-and-Conquer Posterior Sampling for Denoising Diffusion priors

Authors: Yazid Janati, Badr MOUFAD, Alain Durmus, Eric Moulines, Jimmy Olsson

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the versatility and effectiveness of our approach for a wide range of Bayesian inverse problems. The code is available at https://github.com/Badr-MOUFAD/dcps
Researcher Affiliation Collaboration Yazid Janati ,1 Badr Moufad ,1 Alain Durmus1 Eric Moulines1,3 Jimmy Olsson2 1 CMAP, Ecole polytechnique 2 KTH Royal Institute of Technology 3 MBZUAI
Pseudocode Yes The pseudo-code of the DCPS algorithm is in Algorithm 1.
Open Source Code Yes The code is available at https://github.com/Badr-MOUFAD/dcps
Open Datasets Yes For these imaging experiments, we use the FFHQ256 [23] and Image Net256 [14] datasets and the publicly available pre-trained models of [8] and [15].
Dataset Splits Yes The UCY-student dataset was split int a train and a validation sets with 1450 and 140 trajectories respectively.
Hardware Specification Yes The memory consumption is measured by how many samples each algorithm can generate in parallel on a single 48GB L40S NVIDIA GPU for the Diffusion model trained on FFHQ [15].
Software Dependencies No The paper does not provide specific version numbers for ancillary software dependencies such as Python, PyTorch, or CUDA.
Experiment Setup Yes For all the experiments we implement Algorithm 1. We use the same parameters K = 2, L = 3 and ΞΆ = 1 for all the experiments. For the number of Langevin steps, we set it to M = 50 and M = 500 (respectively) for the Gaussian mixture experiment and M = 5 for the imaging and trajectory inpainting experiments.