Carpet: Neural Net based solver for the 1d-TV problem

An adaptive optimization package for the 1d-TV problem for research purpose, proposed in the following paper:

Hamza Cherkaoui, Jeremias Sulam and Thomas Moreau "Learning to solve TV regularised problems with unrolled algorithms", accepted at NeurIPS 2020.

This package implements:

• Classical solver for TV-regularized optimization problems: (primal-dual) ISTA, (primal-dual) FISTA, Condat-Vu.
• Learnable algorithms: All the iterative versions cited above.

Important links

• Official source code repo: https://github.com/hcherkaoui/carpet

Dependencies

The required dependencies to use the software are:

• Numpy >= 1.14.0
• Scipy >= 1.0.0
• Joblib >= 0.11
• Torch >= 1.4.0
• Matplotlib >= 2.1.2
• Prox_tv

License

All material is Free Software: BSD license (3 clause).

Installation

In order to perform the installation, run the following command from the carpet directory:

python3 setup.py install --user

To run all the tests, run the following command from the carpet directory:

pytest

Development

Code

GIT

You can check the latest sources with the command:

git clone git://github.com/hcherkaoui/carpet

or if you have write privileges:

git clone [email protected]:hcherkaoui/carpet