Massively parallel self-organizing maps: accelerate training on multicore CPUs, GPUs, and clusters
Somoclu is a massively parallel implementation of self-organizing maps. It exploits multicore CPUs, it is able to rely on MPI for distributing the workload in a cluster, and it can be accelerated by CUDA. A sparse kernel is also included, which is useful for training maps on vector spaces generated in text mining processes.
For more information, refer to the manuscript about the library .
Somoclu takes a plain text input file -- either dense or sparse data. Example files are included.
$ [mpirun -np NPROC] somoclu [OPTIONs] INPUT_FILE OUTPUT_PREFIX
-c FILENAME Specify an initial codebook for the map. -d NUMBER Coefficient in the Gaussian neighborhood function exp(-||x-y||^2/(2*(coeff*radius)^2)) (default: 0.5) -e NUMBER Maximum number of epochs -g TYPE Grid type: square or hexagonal (default: square) -h, --help This help text -k NUMBER Kernel type 0: Dense CPU 1: Dense GPU 2: Sparse CPU -l NUMBER Starting learning rate (default: 0.1) -L NUMBER Finishing learning rate (default: 0.01) -m TYPE Map type: planar or toroid (default: planar) -n FUNCTION Neighborhood function (bubble or gaussian, default: gaussian) -p NUMBER Compact support for Gaussian neighborhood (0: false, 1: true, default: 0) -r NUMBER Start radius (default: half of the map in direction min(x,y)) -R NUMBER End radius (default: 1) -s NUMBER Save interim files (default: 0): 0: Do not save interim files 1: Save U-matrix only 2: Also save codebook and best matching -t STRATEGY Radius cooling strategy: linear or exponential (default: linear) -T STRATEGY Learning rate cooling strategy: linear or exponential (default: linear) -v NUMBER Verbosity level, 0-2 (default: 0) -x, --columns NUMBER Number of columns in map (size of SOM in direction x) -y, --rows NUMBER Number of rows in map (size of SOM in direction y)
$ somoclu data/rgbs.txt data/rgbs $ mpirun -np 4 somoclu -k 0 --rows 20 --columns 20 data/rgbs.txt data/rgbs
With random initialization, the initial codebook will be filled with random numbers ranging from 0 to 1. Either supply your own initial codebook or normalize your data to fall in this range.
If the range of the values of the features includes negative numbers, the codebook will eventually adjust. It is, however, not advised to have negative values, especially if the codebook is initialized from 0 to 1. This comes from the batch training nature of the parallel implementation. The batch update rule will change the codebook values with weighted averages of the data points, and with negative values, the updates can cancel out.
The maps generated by the GPU and the CPU kernels are likely to be different. For computational efficiency, Somoclu uses single-precision floats. This occasionally results in identical distances between a data instance and the neurons. The CPU version will pick the best matching unit with the lowest coordinate values. Such sequentiality cannot be guaranteed in the reduction kernel of the GPU variant. This is not a bug, but it is better to be aware of it.
The CPU kernels use OpenMP to load multicore processors. On a single node, this is more efficient than launching tasks with MPI to match the number of cores. The MPI tasks replicated the codebook, which is especially inefficient for large maps.
For instance, given a single node with eight cores, the following execution will use 1/8th of the memory, and will run 10-20% faster:
$ somoclu -x 200 -y 200 data/rgbs.txt data/rgbs
$ OMP_NUM_THREADS=8 somoclu -x 200 -y 200 data/rgbs.txt data/rgbs
Avoid the following on a single node:
$ OMP_NUM_THREADS=1 mpirun -np 8 somoclu -x 200 -y 200 data/rgbs.txt data/rgbs
The same caveats apply for the sparse CPU kernel.
The primary purpose of generating a map is visualisation. Apart from the Python interface, Somoclu does not come with its own functions for visualisation, since there are numerous generic tools that are capable of plotting high-quality figures. The R version integrates with kohonen and the MATLAB version with somtoolbox.
The output formats U-matrix and the codebook of the command-line version are compatible with Databionic ESOM Tools for more advanced visualisation.
One sparse and two dense data formats are supported. All of them are plain text files. The entries can be separated by any white-space character. One row represents one data instance across all formats. Comment lines starting with a hash mark are ignored.
The sparse format follows the libsvm guidelines. The first feature is zero-indexed. For instance, the vector [ 1.2 0 0 3.4] is represented as the following line in the file: 0:1.2 3:3.4. The file is parsed twice: once to get the number of instances and features, and the second time to read the data in the individual threads.
The basic dense format includes the coordinates of the data vectors, separated by a white-space. Just like the sparse format, this file is parsed twice to get the basic dimensions right.
The .lrn file of Databionic ESOM Tools is also accepted and it is parsed only once. The format is described as follows:
% s1 s2 .. sm
% varname1 varname2 .. var_namem
x11 x12 .. x1m
x21 x22 .. x2m
. . . .
. . . .
xn1 xn2 .. xnm
Here n is the number of rows in the file, that is, the number of data instances. Parameter m defines the number of columns in the file. The next row defines the column mask: the value 1 for a column means the column should be used in the training. Note that the first column in this format is always a unique key, so this should have the value 9 in the column mask. The row with the variable names is ignore by Somoclu. The elements of the matrix follow -- from here, the file is identical to the basic dense format, with the addition of the first column as the unique key.
If the input file is sparse, but a dense kernel is invoked, Somoclu will execute and results will be incorrect. Invoking a sparse kernel on a dense input file is likely to lead to a segmentation fault.
Python, Julia, R, and MATLAB interfaces are available for the dense CPU and GPU kernels. MPI and the sparse kernel are not support through the interfaces. For respective examples, see the folders in src.
The Python version is also available in PyPI. You can install it with
$ pip install somoclu
Alternatively, it is also available on conda-forge:
$ conda install somoclu
ImportError: DLL load failed: The specified module could not be foundwhen
import somoclu, you may need to use Dependency Walker as shown here on
_somoclu_wrap.pydto find out missing DLLs and place them at the write place. Usually right version (32/64bit) of
vcomp90.dll, msvcp90.dll, msvcr90.dllshould be put to
The wheel binaries for macOS are compiled with the system
clang++, which means by default it is not parallelized. To use the parallel version on Mac, you can either use the version in conda-forge or compile it from source with your favourite OpenMP-friendly compiler. To get it working with the GPU kernel, you might have to follow the instructions at the Somoclu - Python Interface.
The R version is available on CRAN. You can install it with
To get it working with the GPU kernel, download the source zip file and specify your CUDA directory the following way:
R CMD INSTALL src/Rsomoclu_version.tar.gz --configure-args=/path/to/cuda
The Julia version is available on GitHub. The standard
For using the MATLAB toolbox, install SOM-Toolbox following the instructions at ilarinieminen/SOM-Toolbox and define the location of your MATLAB install to the configure script:
./configure --without-mpi --with-matlab=/usr/local/MATLAB/R2014a
For the GPU kernel, specify the location of your CUDA library for the configure script. More detailed instructions are in the MATLAB source folder.
These are the instructions for compiling the core library and the command line interface. The only dependency is a C++ compiler chain -- GCC, ICC, clang, and VC were tested.
Multicore execution is supported through OpenMP -- the compiler must support this. Distributed systems are supported through MPI. The package was tested with OpenMPI. It should also work with other MPI flavours. CUDA support is optional.
If you have just cloned the git repository first run
Then follow the standard POSIX procedure:
$ ./configure [options] $ make $ make install
Options for configure
--prefix=PATH Set directory prefix for installation
By default Somoclu is installed into /usr/local. If you prefer a different location, use this option to select an installation directory.
--without-mpi Disregard any MPI installation found. --with-mpi=MPIROOT Use MPI root directory. --with-mpi-compilers=DIR or --with-mpi-compilers=yes use MPI compiler (mpicxx) found in directory DIR, or in your PATH if =yes --with-mpi-libs="LIBS" MPI libraries [default "-lmpi"] --with-mpi-incdir=DIR MPI include directory [default MPIROOT/include] --with-mpi-libdir=DIR MPI library directory [default MPIROOT/lib]
The above flags allow the identification of the correct MPI library the user wishes to use. The flags are especially useful if MPI is installed in a non-standard location, or when multiple MPI libraries are available.
--with-cuda=/path/to/cuda Set path for CUDA
Somoclu looks for CUDA in /usr/local/cuda. If your installation is not there, then specify the path with this parameter. If you do not want CUDA enabled, set the parameter to
src/Windows/somocluas an example Visual Studio 2015 solution. Modify the CUDA version or VC compiler version according to your needs.
The default solution enables all of OpenMP, MPI, and CUDA. The default MPI installation path is
C:\Program Files (x86)\Microsoft SDKs\MPI\, modify the settings if yours is in a different path. The configuration default CUDA version is 9.1. Disable MPI by removing
HAVE_MPImacro in the project properties (
Properties -> Configuration Properties -> C/C++ -> Preprocessor). Disable CUDA by removing
CUDAmacro in the solution properties and uncheck CUDA in
Project -> Custom Build Rules. If you open the solution without CUDA installed, please remove the following sections in
or change the version number according to which you installed.
The usage is identical to the Linux version through command line (see the relevant section).
This work was supported by the European Commission Seventh Framework Programme under Grant Agreement Number FP7-601138 PERICLES and by the AWS in Education Machine Learning Grant award.