Inverse Problems Benchmark

Benchopt is a package to simplify and make more transparent and reproducible comparisons of optimization methods. This benchmark is dedicated to solvers of inverse problems. It contains

Datasets:

• CBSD68/Set3c (train/test)
• BSD500/CBSD68 (train/test)
• BSD500

Solvers:

• DPIR
• Wavelet
• DRUNet

Tasks:

• Denoising
• Deblurring
• Super-Resolution
• Inpainting

Install

Set up Benchopt and Clone repository

This benchmark can be retrieved using the following commands:

$ pip install -U benchopt
$ git clone https://github.com/benchopt/benchmark_inverse_problems

Customize config file

Edit the config.yml file in the root of the benchmark

data_home: ____________          # Path to the main folder that will contain the data
data_paths:
    Set3c_CBSD68: ____________   # Path where the DeepInv built dataset and physic will be saved
    CBSD68_BSD500: ____________  # Path where the DeepInv built dataset and physic will be saved
    BSD500: ____________         # Path to the folder `images` from `data_home` of the BSD500 dataset (see : Download data section)

For more information on how config file works see :

• https://benchopt.github.io/benchmark_workflow/config_benchopt.html
• https://benchopt.github.io/user_guide/tweak_datasets.html

Download data

Set3c

This dataset is automatically downloaded from HuggingFace deepinv/set3c. You have nothing to do.

CBSD68

This dataset is automatically downloaded from HuggingFace deepinv/CBSD68. You have nothing to do.

BSD500

Please download the dataset by following this url : https://web.archive.org/web/20120508113820/http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/BSR/BSR_bsds500.tgz

Extract the archive somewhere in your data_home path and update the BSD500 key of the config file to point towards the images directory in BSR/BSDS500/data/images

Run the benchmark

You can now run the benchmark by using the following command

benchopt run path/to/the/benchmark

Apart from the problem, options can be passed to benchopt run, to restrict the benchmarks to some solvers or datasets, e.g.:

$ benchopt run benchmark_inverse_problems -s solver1 -d dataset2 --max-runs 10 --n-repetitions 10

Use benchopt run -h for more details about these options, or visit https://benchopt.github.io/api.html.