Efficient Implementation of Self-Organizing Map for Sparse Input Data.
This program uses an algorithm especially intended for sparse data, which much faster than the classical one on very sparse datasets (time-complexity depend to non-zero values only).
- Highly optimized for sparse data (LIBSVM format).
- Support both online and batch SOM algorithms.
- Parallel batch implementation (OpenMP).
- OS independent.
- Python support.
The simplest way to build the cli tools from the main directory : cd src && make all.
After the compilation terminates, the resulting executables may be found in the build directory.
GCC is reccomended, but you can use another compiler if you want. C++11 support is required. OpenMP support is required to take advantage of parallelism (sparse-bsom).
No install required.
To install the python version, simply run pip install sparse-som.
To use the online version :
Usage: sparse-som
-i infile - input file at libsvm sparse format
-y nrows - number of rows in the codebook
-x ncols - number of columns in the codebook
[ -u ] - one based column indices (default is zero based)
[ -N ] - normalize the input vectors
[ -l codebook ] - load codebook from binary file
[ -o|O codebook ] - output codebook to filename (o:binary, O:text)
[ -c|C classes ] - output classification (c:without counts, C:with counts)
[ -n neighborhood ] - neighborhood topology: 4=circ, 6=hexa, 8=rect (default 8)
[ -t tmax | -T epochs ] - number of training iterations (epoch=nb. of samples)
[ -r radius0 -R radiusN ] - radius at start and end (default r=(x+y)/2, R=0.5)
[ -a alpha0 -A alphaN ] - learning rate at start and end (default a=0.5, A=1.e-37)
[ -H radiusCool ] - radius cooling: 0=linear, 1=exponential (default 0)
[ -h alphaCool ] - alpha cooling: 0=linear, 1=exponential (default 0)
[ -s stdCoeff ] - sigma = radius * coeff (default 0.3)
[ -q ] - quiet
To use the batch version :
Usage: sparse-bsom
-i infile - input file at libsvm sparse format
-y nrows - number of rows in the codebook
-x ncols - number of columns in the codebook
[ -u ] - one based column indices (default is zero based)
[ -N ] - normalize the input vectors
[ -l codebook ] - load codebook from binary file
[ -o|O codebook ] - output codebook to filename (o:binary, O:text)
[ -c|C classes ] - output classification (c:without counts, C:with counts)
[ -n neighborhood ] - neighborhood topology: 4=circ, 6=hexa, 8=rect (default 8)
[ -T epochs ] - number of epochs (default 10)
[ -r radius0 -R radiusN ] - radius at start and end (default r=(x+y)/2, R=0.5)
[ -H radiusCool ] - radius cooling: 0=linear, 1=exponential (default 0)
[ -s stdCoeff ] - sigma = radius * coeff (default 0.3)
[ -q ] - quiet
To control the number of threads used by OpenMP, set to OMP_NUM_THREADS variable to the desired value, for example :
OMP_NUM_THREADS=4 sparse-bsom ...
If undefined one thread per CPU is used.
import numpy as np
from scipy.sparse import csr_matrix
from sklearn.datasets import load_digits
from sklearn.metrics import classification_report
from sparse_som import *
# load some dataset
dataset = load_digits()
# convert to sparse CSR format
X = csr_matrix(dataset.data)
# setup SOM dimensions
H, W = 12, 15 # Network height and width
N = X.shape[1] # Nb. features (vectors dimension)
################ Simple usage ################
# setup SOM network
som = Som(H, W, N, topology.HEXA) # , verbose=True
print(som.nrows, som.ncols, som.dim)
# reinit the codebook (not needed)
som.codebook = np.random.rand(H, W, N).\
astype(som.codebook.dtype, copy=False)
# train the SOM
som.train(X)
# get bmus for the data
bmus = som.bmus(X)
################ Use classifier ################
# setup SOM classifier (using batch SOM)
cls = SomClassifier(BSom, H, W, N)
# use SOM calibration
cls.fit(X, labels=dataset.target)
# make predictions
y = cls.predict(X)
print(classification_report(dataset.target, y))Other examples are available in the python/examples directory.
Input files must be at LIBSVM format.
<label> <index1>:<value1> <index2>:<value2> ...
.
.
.
Each line contains an instance and is ended by a '\n' character. The pair <index>:<value> gives a feature (attribute) value: <index> is an integer starting from 0 and <value> is a real number. Indices must be in ASCENDING order. Labels in the file are only used for network calibration. If they are unknown, just fill the first column with any numbers.
The python documentation can be found at: http://sparse-som.readthedocs.io/en/latest/
The C++ API is not public yet, because things still may change.
@InProceedings{melka-mariage:ijcci17,
author={Melka, Josu{\'e} and Mariage, Jean-Jacques},
title={Efficient Implementation of Self-Organizing Map for Sparse Input Data},
booktitle={Proceedings of the 9th International Joint Conference on Computational Intelligence: IJCCI},
volume={1},
month={November},
year={2017},
address={Funchal, Madeira, Portugal},
pages={54-63},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0006499500540063},
isbn={978-989-758-274-5},
url={http://www.ai.univ-paris8.fr/~jmelka/IJCCI_2017_20.pdf}
}