Pyoneer is a Python 3 package for the continuous recovery of non-bandlimited periodic signals with finite rates of innovation (e.g. Dirac streams) from generalised measurements. This package contains notably the reference implementation of the Cadzow Plug-and-play Gradient Descent (CPGD) algorithm, proposed by Matthieu Simeoni, Adrien Besson, Paul Hurley and Martin Vetterli in the paper Cadzow Plug-and-Play Gradient Descent for Generalised FRI [1].
The results of the paper are provided in the folder ./results and can be fully reproduced using the routines in the folder ./benchmarking.
Finite-rate-of-innovation (FRI) is a powerful reconstruction framework enabling the recovery of sparse Dirac streams from uniform low-pass filtered samples. An extension of this framework, called generalised FRI (genFRI), has been recently proposed for handling cases with arbitrary linear measurement models. In this context, signal reconstruction amounts to solving a joint constrained optimisation problem, yielding estimates of both the Fourier series coefficients of the Dirac stream and its so-called annihilating filter, intervening in the regularisation term. This optimisation problem is however highly non convex and non linear in the data. Moreover, the proposed numerical solver is computationally intensive and without any convergence guarantee.
In this work, we revisit the genFRI problem and propose an implicit formulation of the latter. To this end, we leverage a novel regularisation term which does not depend explicitly on the unknown annihilating filter but yet enforces sufficient structure in the solution for stable recovery. The resulting optimisation problem is still non convex, but simpler since linear in the data and with less unknowns. We solve it by means of a provably convergent proximal gradient descent (PGD) method. Since the proximal step does not admit a simple closed-form expression, we propose an inexact PGD method, coined as Cadzow plug-and-play gradient descent (CPGD). The latter approximates the proximal steps by means of Cadzow denoising, a well-known denoising algorithm in FRI. We provide local fixed-point convergence guarantees for CPGD. Through an extensive number of numerical simulations, we demonstrate the superiority of CPGD against the state-of-the-art in the case of non uniform time samples.
./benchmarkingcontains the main routines for (re)-generating the results of the paper [1]:- The script
./benchmarking/reproduce_simulation_results.pyassesses the performances (positioning errors, execution times, number of iterations) of LS-Cadzow, CPGD and GenFRI for various oversampling parameters, peak signal-to-noise (PSNR) ratios and noise levels. - The script
./benchmarking/reproduce_execution_times.pycompares the execution times of CPGD and GenFRI for various oversampling parameters.
- The script
./resultscontains the results of the simulations (plots and pickle files)../pyoneer/algorithmscontains the class definitions of the three reconstruction/denoising algorithms Cadzow, CPGD and GenFRI (see [1] for more details)../pyoneer/modelcontains routines for generating and sampling Dirac streams../pyoneer/operatorscontains classes and routines for linear operators used in generalised FRI problems../pyoneer/plotscontains plotting routines as well as the custom matplotlib style file used to generate the plots of the paper../pyoneer/utilscontains routines for FRI reconstruction as well as miscellaneous routines defining useful mathematical functions.
Note: The following instructions have been tested on a MacBook Pro (16-inch, 2019) with macOS Catalina version 10.15.2 (19C57).
Pyoneer is a Python 3 package relying on multiple open source packages written by third-partites (you can see all the import calls of pyoneer using the command grep -h -R -e "^import" --include '*.py' ./ | sort -h | uniq). To facilitate the management of cross-dependencies, we make use of Miniconda, a light-weight version of the package and environment management open source software Anaconda.
If you are on macOS Catalina and had Miniconda or Anaconda installed prior to upgrading to the latest macOS, then your installation is probably broken due to new restrictions from Apple. To properly install Miniconda on macOS Catalina, download the binary file at this link. From the terminal, navigate to the location where the file was downloaded and execute
bash Miniconda3-latest-MacOSX-x86_64.sh
Follow the prompts. Then, run source ~/miniconda3/bin/activate followed by conda init zsh. Finally, reopen the terminal and conclude the installation with conda update.
We are now ready to create a conda environment named pyoneer with Python 3.7 as the default python version:
conda create -n pyoneer python=3.7
We will set the channel priority to strict, and add conda-forge as channel with higher priority.
conda config --set channel_priority strict
conda config --add channels conda-forge
These settings are stored in the hidden configuration file ~/.condarc. Execute cat ~/.condarc to check that this is indeed the case.
We can now install the relevant packages in the created environment:
conda install -n pyoneer -c conda-forge matplotlib numpy astropy h5py scipy joblib=0.13.0 numba pandas dask
We are done with the installation of dependencies.
Note: On versions of joblib
>0.13.0, we experienced worse performances with the backendmultiprocessing. We hence recommend sticking with version0.13.0for now.
To use the custom pyoneer package, we need to tell to Python where to look for it. To this end, navigate with your terminal to the root of the repository and type:
echo "export PYTHONPATH=$PYTHONPATH:${PWD}" >> ~/.zshrc
Close and reopen you terminal for the changes to take place.
M. Simeoni and A. Besson have contributed equally to this work.
- Matthieu Simeoni, Laboratoire de Communications Audiovisuelles (LCAV), École Polytechnique Fédérale de Lausanne (EPFL), CH-1015, Lausanne, Switzerland.
- Adrien Besson, E-Scopics, Saint-Cannat, France.
- Paul Hurley, Western Sydney University (WSU), Australia.
- Martin Vetterli, Laboratoire de Communications Audiovisuelles (LCAV), École Polytechnique Fédérale de Lausanne (EPFL), CH-1015, Lausanne, Switzerland.
For questions related to this package please contact: [email protected]
MIT License
Copyright (c) 2020 Matthieu SIMEONI, Adrien BESSON, Paul HURLEY and Martin VETTERLI
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
We would like to thank Léa Freydier for designing the logo of the package.