This repository contains benchmarking and plotting scripts used to generate performance comparisons for dask-image, SciPy and CuPy for a lightning talk at Dask Summit 2021. The functions tested were restricted to those present in all three libraries.
There is an example notebook for this located at benchmarks/skimage_apply_parallel.ipynb
Recent versions of Dask and scikit-image will be required. It has only been tested with
scikit-image>=0.18 dask>=2020.12.0
A recent PR for scikit-image fixed multi-threaded use of denoise_tv_bregman
and denoise_bilateral, but has not yet appeared in a released version of the
library:
scikit-image/scikit-image#5400
The benchmarking scripts named dask_cupyx_scipy_* were adapted from ones
previously created for the cuCIM library.
These scripts generate Markdown tables and CSV format outputs containing benchmark results.
Selected results were manually copied into plotting scripts in benchmarks/viz.
CuPy >= 9 SciPy >= 1.6 pandas dask-image (pre-release)
Running the benchmarks currently requires a branch of dask-image incorporating a few recent PRs that were opened after the release of v0.6.0:
These may be included in the next release of dask-image.