PyLops Benchmarks

A zoo of inverse problems built with PyLops, each solved with different combinations of solvers and regularizers/preconditioners, and compared on the same metrics.

Every benchmark is defined by:

• a true model \(\mathbf{x}\);

• a PyLops operator \(\mathbf{A}\);

• the data \(\mathbf{y} = \mathbf{A}\,\mathbf{x}\,(+\,\mathbf{n})\), possibly with noise \(\mathbf{n}\) drawn with a fixed seed.

Each solver estimates \(\hat{\mathbf{x}}\) from \(\mathbf{y}\) and \(\mathbf{A}\), and scores it against the true model with PyLops’ Mean Absolute Error (pylops.utils.metrics.mae) and Peak Signal-to-Noise Ratio (pylops.utils.metrics.psnr, computed over the full dynamic range of the true model). Times are the wall-clock time of the inversion only (operator construction is excluded); when a benchmark is run with several repetitions, the median time is reported.

Benchmarks are run with BenchOpt on GitHub Actions and their results are stored as CSV files in the results/ folder of the repository.

• Benchmarks
• Contributing