sparse-som documentation

Python package for the efficient Self-Organizing Map (SOM, Kohonen map) implementation of sparse-som, intended for sparse, high-dimensional data (e.g. text).

Module contents

enums

class sparse_som.cooling(*values)
LINEAR = 0
EXPONENTIAL = 1
class sparse_som.topology(*values)
CIRC = 4
HEXA = 6
RECT = 8

classes

Self-Organizing Maps wrappers for python, intended for sparse input data.

class sparse_som.BSom

Uses the batch algorithm and can take advantage from multi-core processors to learn efficiently.

Parameters:

  • h (int) – the network height

  • w (int) – the network width

  • d (int) – the dimension of input vectors

  • topol (topology.RECT or topology.HEXA) – the network topology

  • verbose (int (0..2)) – verbosity parameter

train(data, epochs=10, r0=0, rN=0.5, std=0.3, cool=cooling.LINEAR)

Train the network with data.

Parameters:

  • data (scipy.sparse.spmatrix) – sparse input matrix (ideally csr_matrix of numpy.single)

  • epochs (int) – number of epochs

  • r0 (float) – radius at the first iteration

  • rN (float) – radius at the last iteration

  • cool (cooling.LINEAR or cooling.EXPONENTIAL) – cooling strategy

bmus(data, tg_error=False, qt_error=False)

Return the best match units for data.

Parameters:

  • data (scipy.sparse.spmatrix) – sparse input matrix (ideally csr_matrix of numpy.single)

  • tg_error (bool) – indicates whether to calculate the topographic error (default: False)

  • qt_error (bool) – indicates whether to calculate the quantization error (default: False)

Returns:

an array of the bmus coordinates (y,x)

Return type:

2D numpy.ndarray

activation_map(data)

Return the distance between each data sample and each codebook unit.

Parameters:

data (scipy.sparse.spmatrix) – sparse input matrix (ideally csr_matrix of numpy.single)

Returns:

an array with shape (nsample, nunits)

Return type:

2D numpy.ndarray

codebook
Returns:

a view of the internal codebook.

Return type:

3D numpy.ndarray

dim
Returns:

the dimension of the input vectors.

Return type:

int

ncols
Returns:

the number of columns in the network.

Return type:

int

nrows
Returns:

the number of rows in the network.

Return type:

int

umatrix
Returns:

the network U-matrix.

Return type:

2D numpy.ndarray

class sparse_som.Som

Uses the SD-SOM algorithm (online learning).

Parameters:

  • h (int) – the network height

  • w (int) – the network width

  • d (int) – the dimension of input vectors

  • topol (topology.RECT or topology.HEXA) – the network topology

  • verbose (int (0..2)) – verbosity parameter

train(data, tmax, r0=0, a0=0.5, rN=0.5, aN=0., std=0.3, rcool=cooling.LINEAR, acool=cooling.LINEAR)

Train the network with data.

Parameters:

  • data (scipy.sparse.spmatrix) – sparse input matrix (ideally csr_matrix of numpy.single)

  • tmax (int) – number of iterations

  • r0 (float) – radius at the first iteration

  • a0 (float) – learning-rate at the first iteration

  • rN (float) – radius at the last iteration

  • aN (float) – learning-rate at the last iteration

  • rcool (cooling.LINEAR or cooling.EXPONENTIAL) – radius cooling strategy

  • acool (cooling.LINEAR or cooling.EXPONENTIAL) – alpha cooling strategy

bmus(data, tg_error=False, qt_error=False)

Return the best match units for data.

Parameters:

  • data (scipy.sparse.spmatrix) – sparse input matrix (ideally csr_matrix of numpy.single)

  • tg_error (bool) – indicates whether to calculate the topographic error (default: False)

  • qt_error (bool) – indicates whether to calculate the quantization error (default: False)

Returns:

an array of the bmus coordinates (y,x)

Return type:

2D numpy.ndarray

activation_map(data)

Return the distance between each data sample and each codebook unit.

Parameters:

data (scipy.sparse.spmatrix) – sparse input matrix (ideally csr_matrix of numpy.single)

Returns:

an array with shape (nsample, nunits)

Return type:

2D numpy.ndarray

codebook
Returns:

a view of the internal codebook.

Return type:

3D numpy.ndarray

dim
Returns:

the dimension of the input vectors.

Return type:

int

ncols
Returns:

the number of columns in the network.

Return type:

int

nrows
Returns:

the number of rows in the network.

Return type:

int

umatrix
Returns:

the network U-matrix.

Return type:

2D numpy.ndarray

Submodules

sparse_som.classifier module

class sparse_som.SomClassifier(cls=<class 'sparse_som.som.BSom'>, height=10, width=10, dim=None, **kwargs)[source]
__init__(cls=<class 'sparse_som.som.BSom'>, height=10, width=10, dim=None, **kwargs)[source]
Parameters:

  • cls (Som or BSom) – SOM constructor

  • height (int) – SOM height (default 10)

  • width (int) – SOM width (default 10)

  • dim (int) – SOM nodes nb of dimensions (if None, detected when fit called)

  • **kwargs – named parameters for the constructor

setup(**kwargs)[source]

Add / Change SOM constructor parameters.

params()[source]

Get the SOM constructor parameters.

fit(data, labels, **kwargs)[source]

Training the SOM on the the data and calibrate itself.

After the training, self.quant_error and self.topog_error are respectively set.

Parameters:

  • data (scipy.sparse.csr_matrix) – sparse input matrix (ideal dtype is numpy.float32)

  • labels (iterable) – the labels associated with data

  • **kwargs – optional parameters for train()

predict(data, unkown=None)[source]

Classify data according to previous calibration.

Parameters:

  • data (scipy.sparse.csr_matrix) – sparse input matrix (ideal dtype is numpy.float32)

  • unkown – the label to attribute if no label is known

Returns:

the labels guessed for data

Return type:

numpy.array

fit_predict(data, labels, unkown=None)[source]

Fit and classify data efficiently.

Parameters:

  • data (scipy.sparse.csr_matrix) – sparse input matrix (ideal dtype is numpy.float32)

  • labels (iterable) – the labels associated with data

  • unkown – the label to attribute if no label is known

Returns:

the labels guessed for data

Return type:

numpy.array

get_precision()[source]
Returns:

the ratio part of the dominant label for each unit.

Return type:

2D numpy.ndarray

histogram(bmus=None)[source]

Return a 2D histogram of bmus.

Parameters:

bmus (numpy.ndarray) – the best-match units indexes for underlying data.

Returns:

the computed 2D histogram of bmus.

Return type:

numpy.ndarray

Indices and tables