fastLBP

Highly parallel LBP implementation

Installation

Note: it is not recommended to proceed on a head node; consider starting an ijob or a jupyter instance.

• Activate a Python 3.11 environment (e.g. using conda create -y -n p11 python=3.11 && conda activate p11)
• Verify you are using the right env
  • python --version та pip --version
• Install a stable version from PyPI
  • pip install fastlbp-imbg
• Or build the latest version from sources

git clone git@github.com:imbg-ua/fastLBP.git
cd fastLBP
# git checkout <branchname> # if you need a specific branch
pip install . # this will install the fastlbp_imbg package in the current env

• You can use import fastlbp_imbg now

Cython2 branch details

TODO:

• Data Contiguity (e3293cd)
• Job sorting (from slow to quick) in fastlbp.py (8a21a04)
• Change data types in lbp.pyx from float64 to something smaller (f177823)
• Add mask feature to lbp.pyx (d70bc93)
• Add mask feature to fastlbp.py (d812831)
• Fill all commit numbers

Implemented modules

run_skimage

Computes multiradial LBP of a single multichannel image in a parallel fashion.
This is a quick sample implementation that could be a baseline for further benchmarking.

Features:

• Powered by skimage.feature.local_binary_pattern
• Concurrency is managed by Python's multiprocessing module
• Parallel computation via multiprocessing.Pool of size ncpus
• Efficient memory usage via multiprocessing.shared_memory to make sure processes do not create redundant copies of data
• It computes everything in RAM, no filesystem usage

TODO:

• Use max_ram parameter to estimate optimal number of sub-processes and collect memory stats. Now max_ram is ignored.

Planned modules

run_chunked_skimage

Similar to 1. run_skimage, but each subprocess should compute LBP for its image chunk, not the whole image.

run_dask and run_chunked_dask

Similar to 1. run_skimage, but use Dask and dask.array.map_overlap for parallelisation instead of multiprocessing and manual data wrangling

Other notable things to try

• Perform benchmarking of an in-house optimised cython version of skimage.feature.local_binary_pattern (see skimage_lbp.pyx at imbg-ua/nf-img-benchmarking)
• Do some research on Numba - is it applicable here?
• Add annotations to run_skimage results using anndata