A Neural-Network-Based Convex Regularizer for Inverse Problems

New Check the extension of CRR-NNs to weakly-convex regularizers preprint/repo. Remark: the latter repository also includes a multi-noise-level deep-equilibrium training of CRR-NNs with a slightly simpler procedure (shared activations + no spline regularization), with same denoising performance.

Implementation of the CRR-NNs as presented in this paper (or open access version). For any bug report/question/help needed please contact us.

Nb: detailed information is provided in the README.me of each folder.

The repository is organized as follows:

• trained_models: contains a few instances of trained CRR-NNs.
• tutorial: to understand and deploy trained CRR-NNs for solving inverse problems.
• inverse_problems: contains the MRI and CT experiments + generic utils for solving inverse problems and for the automated validation and testing.
• denoising: to reproduce the denoising results on BSD68 with the provided trained models.
• training: to reproduce training.
• under_the_hood: to visualize the NNs (filters and activations).
• models: contains the spline modules, the CRR-NNs class, some optimization schemes (t-step denoiser and accelerated gradient descent)...
• hyperparameter_tuning: helpers and discussion on the tuning of $\lambda$ and $\mu$ for solving inverse problems. See inverse_problems folder for usage.

Some requirements (depending on what you use)

• python >= 3.8
• pytorch >= 1.12
• (optional) CUDA
• numpy
• pandas
• tensorboard
• matplotlib
• Pillow

For the CT experiments:

• astra-toolbox

For the MRI expeirements:

• fastmri
• bart