som_seqtrain
Purpose
Trains a Self-Organizing Map using the sequential algorithm.
Syntax
sM = som_seqtrain(sM,D);
sM = som_seqtrain(sM,sD);
sM = som_seqtrain(...,'argID',value,...);
sM = som_seqtrain(...,value,...);
[sM,sT] = som_seqtrain(M,D,...);
Description
Trains the given SOM (sM or M above) with the given training data
(sD or D) using sequential SOM training algorithm. If no optional
arguments (argID, value) are given, a default training is done, the
parameters are obtained from SOM_TRAIN_STRUCT function. Using
optional arguments the training parameters can be specified. Returns
the trained and updated SOM and a train struct which contains
information on the training.
References
Kohonen, T., "Self-Organizing Map", 2nd ed., Springer-Verlag,
Berlin, 1995, pp. 78-82.
Kohonen, T., "Clustering, Taxonomy, and Topological Maps of
Patterns", International Conference on Pattern Recognition
(ICPR), 1982, pp. 114-128.
Kohonen, T., "Self-Organized formation of topologically correct
feature maps", Biological Cybernetics 43, 1982, pp. 59-69.
Required input arguments
sM The map to be trained.
(struct) map struct
(matrix) codebook matrix (field .data of map struct)
Size is either [munits dim], in which case the map grid
dimensions (msize) should be specified with optional arguments,
or [msize(1) ... msize(k) dim] in which case the map
grid dimensions are taken from the size of the matrix.
Lattice, by default, is 'rect' and shape 'sheet'.
D Training data.
(struct) data struct
(matrix) data matrix, size [dlen dim]
Optional input arguments
argID (string) Argument identifier string (see below).
value (varies) Value for the argument (see below).
The optional arguments can be given as 'argID',value -pairs. If an
argument is given value multiple times, the last one is
used. The valid IDs and corresponding values are listed below. The values
which are unambiguous (marked with '*') can be given without the
preceeding argID.
'mask' (vector) BMU search mask, size dim x 1. Default is
the one in sM (field '.mask') or a vector of
ones if only a codebook matrix was given.
'msize' (vector) map grid dimensions. Default is the one
in sM (field sM.topol.msize) or
'si = size(sM); msize = si(1:end-1);'
if only a codebook matrix was given.
'radius' (vector) neighborhood radius
length = 1: radius_ini = radius
length = 2: [radius_ini radius_fin] = radius
length > 2: the vector given neighborhood
radius for each step separately
trainlen = length(radius)
'radius_ini' (scalar) initial training radius
'radius_fin' (scalar) final training radius
'alpha' (vector) learning rate
length = 1: alpha_ini = alpha
length > 1: the vector gives learning rate
for each step separately
trainlen is set to length(alpha)
alpha_type is set to 'user defined'
'alpha_ini' (scalar) initial learning rate
'tracking' (scalar) tracking level: 0, 1 (default), 2 or 3
0 - estimate time
1 - track time and quantization error
2 - plot quantization error
3 - plot quantization error and two first
components
'trainlen' (scalar) training length (see also 'tlen_type')
'trainlen_type' *(string) is the trainlen argument given in 'epochs'
or in 'samples'. Default is 'epochs'.
'sample_order'*(string) is the sample order 'random' (which is the
the default) or 'ordered' in which case
samples are taken in the order in which they
appear in the data set
'train' *(struct) train struct, parameters for training.
Default parameters, unless specified,
are acquired using SOM_TRAIN_STRUCT (this
also applies for 'trainlen', 'alpha_type',
'alpha_ini', 'radius_ini' and 'radius_fin').
'sTrain', 'som_train' (struct) = 'train'
'neigh' *(string) The used neighborhood function. Default is
the one in sM (field '.neigh') or 'gaussian'
if only a codebook matrix was given. Other
possible values is 'cutgauss', 'ep' and 'bubble'.
'topol' *(struct) topology of the map. Default is the one
in sM (field '.topol').
'sTopol', 'som_topol' (struct) = 'topol'
'alpha_type'*(string) learning rate function, 'inv', 'linear' or 'power'
'lattice' *(string) map lattice. Default is the one in sM
(field sM.topol.lattice) or 'rect'
if only a codebook matrix was given.
'shape' *(string) map shape. Default is the one in sM
(field sM.topol.shape) or 'sheet'
if only a codebook matrix was given.
Output arguments
sM the trained map
(struct) if a map struct was given as input argument, a
map struct is also returned. The current training
is added to the training history (sM.trainhist).
The 'neigh' and 'mask' fields of the map struct
are updated to match those of the training.
(matrix) if a matrix was given as input argument, a matrix
is also returned with the same size as the input
argument.
sT (struct) train struct; information of the accomplished training
Examples
Simplest case:
sM = som_seqtrain(sM,D);
sM = som_seqtrain(sM,sD);
To change the tracking level, 'tracking' argument is specified:
sM = som_seqtrain(sM,D,'tracking',3);
The change training parameters, the optional arguments 'train',
'neigh','mask','trainlen','radius','radius_ini', 'radius_fin',
'alpha', 'alpha_type' and 'alpha_ini' are used.
sM = som_seqtrain(sM,D,'neigh','cutgauss','trainlen',10,'radius_fin',0);
Another way to specify training parameters is to create a train struct:
sTrain = som_train_struct(sM,'dlen',size(D,1),'algorithm','seq');
sTrain = som_set(sTrain,'neigh','cutgauss');
sM = som_seqtrain(sM,D,sTrain);
By default the neighborhood radius goes linearly from radius_ini to
radius_fin. If you want to change this, you can use the 'radius' argument
to specify the neighborhood radius for each step separately:
sM = som_seqtrain(sM,D,'radius',[5 3 1 1 1 1 0.5 0.5 0.5]);
By default the learning rate (alpha) goes from the alpha_ini to 0
along the function defined by alpha_type. If you want to change this,
you can use the 'alpha' argument to specify the learning rate
for each step separately:
alpha = 0.2*(1 - log([1:100]));
sM = som_seqtrain(sM,D,'alpha',alpha);
You don't necessarily have to use the map struct, but you can operate
directly with codebook matrices. However, in this case you have to
specify the topology of the map in the optional arguments. The
following commads are identical (M is originally a 200 x dim sized matrix):
M = som_seqtrain(M,D,'msize',[20 10],'lattice','hexa','shape','cyl');
M = som_seqtrain(M,D,'msize',[20 10],'hexa','cyl');
sT= som_set('som_topol','msize',[20 10],'lattice','hexa','shape','cyl');
M = som_seqtrain(M,D,sT);
M = reshape(M,[20 10 dim]);
M = som_seqtrain(M,D,'hexa','cyl');
The som_seqtrain also returns a train struct with information on the
accomplished training. This is the same one as is added to the end of the
trainhist field of map struct, in case a map struct is given.
[M,sTrain] = som_seqtrain(M,D,'msize',[20 10]);
[sM,sTrain] = som_seqtrain(sM,D); % sM.trainhist{end}==sTrain
See also