som_example.py

import numpy as np
import som
from grid import hexagonal_grid # comes with som repository
from distance import euclidean # also comes with the som repository

##### Create a funky, curved data set in 3 dimensions. #####
data = np.transpose(np.matrix([
	np.random.normal(0,1,20000),
	np.random.normal(0,1,20000),
	np.random.normal(0,1,20000)
]))
indices = []
n, p = data.shape
for i in range(n):
    if (euclidean(np.array(data[i,:])[0], np.array([0,0,0])) > .5) and \
       (data[i,0] > 0 and data[i,1] > 1):
        indices.append(i)
new_data = np.zeros([len(indices), 3])
for i, j in enumerate(indices):
    new_data[i,:] = data[j,:]
data = np.matrix(new_data)

######################## Let's make a SOM! #####################

# First, create a SOM with 3 dimensions, and width and height 20 cells 
# (that's 400 cells / prototypes / models). 
hex_som = som.SelfOrganizingMap(3, 20, 20, 
                                grid_type = hexagonal_grid)

# if you've got a numpy matrix, training is as simple as this.
hex_som.train(data)

# this special function plots 3 dimensional data, so you can see what your
# map looks like hanged on the original dataset.
hex_som.preview_3d_data_and_map(data)

# get the prototype vectors, in the original R^n space:
pvs = hex_som.prototype_vector

# get the grid coordinates, in the new R^2 space:
som_coords = hex_som.som

# the indices of the two are identical. pv[i,:] corresponds to som_coords[i,:].