pythonでSOM的なもの。一応sklearn互換っぽい。

som.py

from sklearn.base import BaseEstimator, TransformerMixin
import numpy


class SOM(BaseEstimator, TransformerMixin):
	def __init__(self, mapsize, input_dim, ini=None, fin=1, coefficient=0.005, verbose=False):
		self.input_dim = input_dim
		self.mapsize = numpy.array(list(mapsize))
		self.mapdim = len(self.mapsize)
		self.idx = numpy.mgrid[[slice(x) for x in self.mapsize]].swapaxes(0, self.mapdim).swapaxes(0, 1)

		if ini is None:
			ini = max(1, self.mapsize.min() // 2)
		self.ini = ini
		self.fin = fin
		self.coefficient = coefficient

		self.verbose = verbose

		self.initialize()

	def initialize(self):
		self.weight = numpy.random.rand(self.mapsize.prod(), self.input_dim).reshape(list(self.mapsize) + [self.input_dim])

	def transform(self, X):
		return numpy.array([
			numpy.unravel_index(
				numpy.sqrt((self.weight - x)**2).sum(axis=self.mapdim).argmin(),
				self.mapsize
			)[::-1]
			for x in X
		])

	def fit_iteration(self, X, n_loop=2000):
		for t in range(n_loop):
			if self.verbose:
				print('fitting SOM: {0}/{1}({2:.2%})'.format(t, n_loop, t/n_loop))

			delta = (self.ini * (self.fin / self.ini) ** (t / n_loop)) ** 2

			self.weight += sum((self.coefficient * numpy.exp(
					-((self.idx - win)**2).sum(axis=self.mapdim) / delta
				) * (x - self.weight).swapaxes(0, self.mapdim)).swapaxes(0, self.mapdim)
				for x, win in zip(X, self.transform(X))
			)
			yield self
		return self

	def fit(self, X, n_loop=2000):
		self.initialize()
		for _ in self.fit_iteration(X, n_loop=n_loop):
			pass
		return self