Disclaimer

This code is a proof of concept and part of an ongoing work. It is provided for reproducing the results of a published paper.

Caution

This code is NOT intended for production use

ANR DynaSTI: Paper Gretsi 2025

This repository contains the code to reproduce the results of the paper "Correction matricielle de l'indétermination d'échelle pour l'optimisation alternée" Submitted to the conference Gretsi 2025

Acknowlegement

This work has been supported by the ANR (French research agency "Agence Nationale de la Recherche") project DynaSTI: ANR-22-CE45-0008

Code structure

The code is structured as follows:

├──data
├──external
|   └──prox_TV_Condat
├──notebooks
|   ├──figs
|   ├──figs_Gretsi2025_soumission
|   ├──Convergence_tests.ipynb
|   ├──Convergence_tests.py
|   ├──Plot_simulation_data.ipynb
|   ├──Plot_simulation_data.py
|   ├──evaluate_mutlirealization.ipynb
|   ├──evaluate_mutlirealization.py
|   ├──optimization_hyperparameters.ipynb
|   └──optimization_hyperparameters.py
├──scripts
|   ├──optimization_results
|   ├──optimization_results_Gretsi_soumission
|   ├──evaluate_model_mutliple_data.py
|   └──optimize_hyperparameters_simulation.py
├──src
├──__init__.py
├──LICENCE_FRENCH
├──ReadMe.md
├──gretsi_soumission_ID1416.pdf
├──licence
├──requirements.txt
└──run_simulations.sh

The data folder contains the example dataset used in the paper.

The external folder contains the C code of the function to compute the proximal operator of the total variation proposed and implemented by L. Condat.

The notebooks folder contains the Jupyter notebooks to analyze the results and plot the figures

The folder notebooks/figs_GRETSI_soumission contains the exact figures that are used in the manuscript

The figures generated when running the code are stored in folder notebooks/figs

The codes to run the methods for multiple datasets and on a grid of hyperparameters are in the scripts folder.

The source code of the methods are in the src folder

The PDF of the manuscript submitted to Gretsi 2025 is gretsi_soumission_ID1416.pdf

Important

All the notebooks (.ipnb) in notebooks where converted into scripts (.py) of the same names to allow automated execution. The scripts should NOT be modified.

How to reproduce the results:

After cloning the repository, open a terminal in the root folder of the project.

• First install the requirements in the file requiremnts.txt by running the following commands in the terminal: This requires a valid Miniconda (or Anaconda) installation

  # Create a new environment
  conda create --name Grest2025_ID1416 python==3.11.11
  # Activate environment
  conda activate Grest2025_ID1416
  # Install dependencies
  conda install --yes --file requirements.txt
  # Compile the c library of L. Condat used to compute the proximal operator of total variation
  gcc -fpic -O3 -march=native -Wall -shared -o src/Estimation_A/lib_prox_TV_1D_Condat.so external/prox_TV_Condat/Condat_TV_1D_v2.c
  

• You can run all the scripts at once with the bash script run_simulations.sh

  # Activate environment if it is not activated
  conda activate Grest2025_ID1416
  # Run the simulations
  bash run_simulations.sh

Running the script will:

• Plot the example of simulation data used in the paper (Figure1)

• Run the optimization for the three methods on a reduced grid of hyperparameters to produce Figure 2. (To run on the full grid of parameters as in the paper use the command bash run_simulations.sh full)

• Evaluate the three methods for the best set of hyperparameters for 25 realizations of the simulation data (and with 100 realizations with the option bash run_simulations.sh full).

• Plot the box plot of the reconstruction error on these realizations (Figure 3)

• Evaluate the effect of initialization scaling on the convergence of the methods and plot the corresponding Figure 4.

• Save all the figures presented in the paper in the folder notebooks/figs

Important

It should be noted that for the study of the effect of hyperparameters, we evaluated 1008 combinations and the box plot Figure 3 is generated with 100 realizations. Running these evaluations took 5 hours while using 52 CPU cores. With fewer cores, it will take much longer. The default behavior of the provided code is to run on a reduced grip of 108 hyperparameters and generate the box plot with 25 realizations. You can run the full evaluation by adding the parameter full to the command: bash run_simulations.sh full