Provably Convergent Plug-and-Play Quasi-Newton Methods

Code for the paper "Provably Convergent Plug-and-Play Quasi-Newton Methods", submitted to SIAM Imaging Sciences.

[Paper]

Prerequisites

The code was computed using Python 3.10, Pytorch 1.12.1. The code was originally computed with Python 3.8.10, PyTorch Lightning 1.2.6, PyTorch 1.7.1.

Prox-Denoiser (Prox-DRUNet)

The code relative to the Proximal (Gradient Step) Denoiser can be found in the GS_denoising directory. This repository contains, along with the proposed PnP-LBFGS1 and PnP-LBFGS2 methods, the following competing methods:

• PnP-PGD, PnP-DRS, PnP-DRSdiff from the Prox-PnP papers
• Relaxed PnP-PGD, or PnP-aPGD from this followup paper.
• PnP-FISTA
• (Optional) DPIR, which is PnP-HQS. For this, clone the DPIR repository into the main folder (so the top level looks like ``GS_denoising, PnP_restoration, DPIR, ...```.
  • For DPIR, please copy utils_model, utils_image from their utils folder to the utils folder in PnP_restoration/utils, and download the drunet_color.pth model from their page.

Pre-trained model

• Download pretrained checkpoint from https://plmbox.math.cnrs.fr/f/faf7d62213e449fa9c8a/?dl=1 and save it as GS_denoising/ckpts/Prox_DRUNet.ckpt

Implementation

The implementation for the BFGS method can be found in PnP_restoration/prox_PnP_restoration.

Usage

A batch file containing a script to test all methods, as well as produce a log for hyperparameter grid search, can be found in the repository.

Example:

cd PnP_restoration
python (deblur|SR).py --dataset_name (set3c|CBSD10|CBSD68) --PnP_algo (BFGS|BFGS2|PGD|aPGD|DRS|DRSdiff) --noise_level_img (2.25|7.65|12.75) (--extract_curves) (--extract_images) (--sf 2) (--params)

CUDA_VISIBLE_DEVICES=1 python SR.py --dataset_name CBSD68 --PnP_algo BFGS2 --noise_level_img 2.55 --extract_curves --extract_images --sigma_multi=2.0 --lamb=4.0 --gamma=1.0 --beta=0.01 --alpha=0.5 --maxitr=100 --sf=2


python deblur.py --dataset_name CBSD68 --PnP_algo BFGS --noise_level_img 2.25 --extract_curves --extract_images --sigma_multi=1.0 --lamb=1.0 --gamma=1.0 --beta=0.01 --alpha=0.5 --maxitr=100
python deblur.py --dataset_name CBSD68 --PnP_algo BFGS --noise_level_img 7.65 --extract_curves --extract_images --sigma_multi=0.75 --lamb=1.0 --gamma=0.85 --beta=0.01 --alpha=0.5 --maxitr=100
python deblur.py --dataset_name CBSD68 --PnP_algo BFGS --noise_level_img 12.75 --extract_curves --extract_images --sigma_multi=0.75 --lamb=1.0 --gamma=1.0 --beta=0.01 --alpha=0.7 --maxitr=100

Other things

We updated the fft transforms from Pytorch 1.7.1 to Pytorch 1.12.1. We also added the relaxed proximal denoiser as in this followup paper.

Acknowledgments

This repo is based on the following repo:

• Proximal denoiser for convergent plug-and-play optimization with nonconvex regularization : https://github.com/samuro95/Prox-PnP

This repo contains parts of code taken from :

• Deep Plug-and-Play Image Restoration (DPIR) : https://github.com/cszn/DPIR
• Gradient Step Denoiser for convergent Plug-and-Play (GS-PnP) : https://github.com/samuro95/GSPnP