sambitdash / Somoclu.jl

Julia package for Somoclu -- Massively parallel self-organizing maps: accelerate training on multicore CPUs and GPUs

Somoclu.jl - Julia Interface for Somoclu

Somoclu is a massively parallel implementation of self-organizing maps. It relies on OpenMP for multicore execution and it can be accelerated by CUDA. The topology of map is either planar or toroid, the grid is rectangular or hexagonal.

Key features of the Julia interface:

• Fast execution by parallelization: OpenMP and CUDA are supported.
• Planar and toroid maps.
• Rectangular and hexagonal grids.
• Gaussian or bubble neighborhood functions.
• PCA initialization of the codebook.

Usage

A simple example is as follows.

using Somoclu

ncolumns, nrows = 40, 30;
ndimensions, nvectors = 2, 50;
c1 = Array{Float32}(rand(ndimensions, nvectors) ./ 5);
c2 = Array{Float32}(rand(ndimensions, nvectors) ./ 5 .+ [0.2; 0.5]);
c3 = Array{Float32}(rand(ndimensions, nvectors) ./ 5 .+ [0.4; 0.1]);
data = hcat(c1, c2, c3);
som = Som(ncolumns, nrows, maptype="toroid")
train!(som, data);

Then you can plot the result with a plotting package of your choice. For example:

using PyPlot
class_colors = vcat(["red" for _=1:50], ["blue" for _=1:50], ["green" for _=1:50]);
imshow(som.umatrix, cmap="Spectral_r")
scatter(som.bmus[1, :], som.bmus[2, :], c=class_colors)

Citation

1. Peter Wittek, Shi Chao Gao, Ik Soo Lim, Li Zhao (2015). Somoclu: An Efficient Parallel Library for Self-Organizing Maps. arXiv:1305.1422.