cuda-slic: A CUDA implementation of the SLIC Superpixel algorithm

SLIC

SLIC stands for simple linear iterative clustering. SLIC uses tiled k-means clustering to segment an input image into a set of superpixels/supervoxels.

This library was designed to segment large 2D/3D images coming from different detectors at the Diamond Light Source. These images can be very large so using a serial CPU code is out of the question.

To speed up processing we use GPU acceleration to achieve great speed improvements over alternative implementations. cuda-slic borrows its API from skimage.segmentation.slic.

Benchmark

Machine: 8 Core Intel Xeon(R) W-2123 CPU @ 3.60GHz with NVIDIA Quadro P2000

from skimage import data
from cuda_slic.slic import slic as cuda_slic
from skimage.segmentation import slic as skimage_slic

blob = data.binary_blobs(length=500, n_dim=3, seed=2)
n_segments = 500**3/5**3 # super pixel shape = (5,5,5)

%timeit -r1 -n1 skimage_slic(blob, n_segments=n_segments, multichannel=False, max_iter=5)
# 2min 28s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)

%timeit -r1 -n1 cuda_slic(blob, n_segments=n_segments, multichannel=False, max_iter=5)
# 13.1 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)

CUDA JIT Compilation

cuda-slic uses JIT compilation to covert CUDA kernels into GPU machine-code (PTX). Two options are available for JIT compiliing CUDA code with python: Cupy or PyCUDA. If PyCUDA is installed in the virtutalenv it is used by default. Otherwise Cupy is used.

To ease distribution cuda-slic is packaged into two independent packages that share the same source-code but depend on a different libraries for JIT compilation:

1. cuda-slic uses pycuda for JIT compilation.
2. gpu-slic uses cupy for JIT compilation.

Installing cuda-slic (with PyCUDA)

pip install cuda-slic

cuda-slic uses pycuda which has the following non-python build dependencies:

1. gcc and g++/gcc-c++ on Linux. MSVC++ compiler and C++ build-tools on Windows.
2. the cudatoolkit for linking with cuda.h.

and the following runtime dependencies:

1. gcc and g++/gcc-c++ on Linux. MSVC++ compiler and C++ build-tools on Windows.
2. the cudatoolkit for linking with cuda libraries.
3. the nvcc compiler. Ships with newer cudatoolkit versions.

See the pycuda docs for installation instructions.

Installing gpu-slic (with Cupy)

pip install gpu-slic

gpu-slic uses Cupy which has the following non-python build dependencies:

1. gcc and g++/gcc-c++ on Linux.
2. the cudatoolkit for linking with cuda libraries.
3. the nvcc compiler. Ships with newer cudatoolkit versions.

Note that when pip installing gpu-slic, cupy is installed as an sdist meaning that your host must meet the compiling and linking requirements of cupy.

Check if gpu-slic is available on conda-forge to get precompiled binaries of Cupy.

See also cupy docs for installation instructions.

Usage

from cuda_slic import slic
from skimage import data

# 2D RGB image
img = data.astronaut() 
labels = slic(img, n_segments=100, compactness=10)

# 3D gray scale
vol = data.binary_blobs(length=50, n_dim=3, seed=2)
labels = slic(vol, n_segments=100, multichannel=False, compactness=0.1)

# 3D multi-channel
# volume with dimentions (z, y, x, c)
# or video with dimentions (t, y, x, c)
vol = data.binary_blobs(length=33, n_dim=4, seed=2)
labels = slic(vol, n_segments=100, multichannel=True, compactness=1)

Development

Prerequisites:

1. gcc and g++/gcc-c++ installed and available on PATH.
2. cudatoolkit installed and CUDA_HOME pointing to its location.
3. nvcc compiler. Ships with recent versions of cudatoolkit.

Dependency Management

We use conda as a dependency installer and virtual env manager. A development environment can be created with

conda env create -f environment.yml

now you can activate the virtual env with conda activate cuda-slic, and deactivate with conda deactivate. To add a dependency, add it to the environment.yml file, then you can run

conda env update -f environment.yml

to keep environment.yml file as lean as possible, development dependencies are kept in requirements_dev.txt and can be installed with

conda install --file requirements_dev.txt -c conda-forge

or

pip install -r requirements_dev.txt

Tests

In the notebooks folder there are Jupyter notebooks where the segmentation algorithm can be visually inspected.

Our unit-testing framework of choice is Py.test. The unit-tests can be run with

conda activate cuda-slic
pip install pytest
pytest

or

conda activate cuda-slic
pip install tox
tox