Training firmly nonexpansive denoisers

This repository is an updated version of our training codes based on the training framework and dataset of Hurault et al.

Summary

We train a denoiser while imposing a firm-nonexpansiveness penalization through a Jacobian regularization term. Given the denoiser J, we impose a 1-Lipschitz regularization on Q = 2J-I, where I is the identity. The Lipschitz constant of Q is estimated via the spectral norm of its Jacobian.

New weights for a DnCNN trained with small noise level (2.0/255.) are currently available here.

This repository only focuses on training denoisers; for the PnP algorithm, we refer you to our older code. However, a small google colab notebook with a demo PnP-FB example is available online. Try it out here!

Link to paper: https://arxiv.org/pdf/2012.13247.pdf

Requirements

This code was tested in the following configuration:

- numpy = 1.23.4
- torch = 1.13.0
- torchvision = 0.14.0
- pytorch-lightning = 1.8.1
- torchmetrics = 0.11.0
- opencv-python = 4.6.0.66

Training

• Download training dataset from https://drive.google.com/file/d/1WVTgEBZgYyHNa2iVLUYwcrGWZ4LcN4--/view?usp=sharing and unzip DRUNET in a somewhere/datasets folder. Important: Update the path accordingly here.

• To train with the regularised loss:

python3 main_train.py --model_name DNCNN \
                      --nc_in 3 --nc_out 3 \
                      --min_sigma_train 15 --max_sigma_train 15 --sigma_list_test 15 \
                      --jacobian_loss_weight 1e-2 --jacobian_compute_type 'nonsymmetric' --power_method_nb_step 20 \
                      --act_mode 'R' --optimizer_lr 1e-4 --loss_name 'l1'

Monitoring training with tensorboard

Training logs are saved in hparams.log_folder (log_folder argument of the parser). cd to this folder and launch

tensorboard --logdir .

If launching on a server, forward this to the appropriate port with the --port argument in the above command.

Exporting the trained weights to a standard state_dict

Saved checkpoints (.ckpt files) follow the structure of a DenoisingModel class. In order to export them to a standard state_dict torch nomenclature (more versatile), you can use the template save_dict.py code.

Acknowledgments

This repo contains parts of code taken from :

• Prox-PnP : https://github.com/samuro95/Prox-PnP
• 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
• Learning Maximally Monotone Operators for image recovery https://github.com/basp-group/PnP-MMO-imaging

Citation

This code is based on the following work:

@article{pesquet2021learning,
  title={Learning maximally monotone operators for image recovery},
  author={Pesquet, Jean-Christophe and Repetti, Audrey and Terris, Matthieu and Wiaux, Yves},
  journal={SIAM Journal on Imaging Sciences},
  volume={14},
  number={3},
  pages={1206--1237},
  year={2021},
  publisher={SIAM}
}

Don't hesitate to contact me if you have any question!

License: GNU General Public License v3.0.