def get_nearest_cluster (clusters, point):
""" Retrieves the closest cluster of a point. """
# Initially the result is set to the first neuron.
nearest_cluster = clusters[0]
best_distance = distance.euclidian(clusters[0], point)
# We look for a closer neuron in the neuron list.
for i in xrange(1, len(clusters)):
tmp_distance = distance.euclidian(clusters[i], point)
if tmp_distance < best_distance:
nearest_cluster = clusters[i]
best_distance = tmp_distance
return nearest_cluster
def get_nearest_cluster(self, point):
""" Retrieves the closest cluster of a point. """
# Initially the result is set to the first neuron.
nearest_cluster = self._clusters[0]
best_distance = distance.euclidian(self._clusters[0], point)
# We look for a closer neuron in the neuron list.
for i in xrange(1, len(self._clusters)):
tmp_distance = distance.euclidian(self._clusters[i], point)
if tmp_distance < best_distance:
nearest_cluster = self._clusters[i]
best_distance = tmp_distance
return nearest_cluster
def class_probability (self, point, c):
distances = []
# Compute distances with all dataset
for i in xrange (len (self.__data)):
dist = distance.euclidian (self.__data[i], point)
distances.append ((dist, self.__labels[i]))
# Keep only K nearest points
distances.sort (dist_cmp)
distances = distances[:self.k]
# Compute the most frequent label in neighbours
labels_freq = {}
freq_tot = 0.0
ref_freq = 0.0
for i in xrange (len (distances)):
freq_tot += 1
if distances[i][1] == c:
ref_freq += 1
if not labels_freq.has_key (distances[i][1]):
labels_freq[distances[i][1]] = 0
else:
labels_freq[distances[i][1]] += 1
i = 0
for label in labels_freq:
if i == 0:
label_max = label
freq_max = labels_freq[label]
i += 1
else:
if labels_freq[label] > freq_max:
label_max = label
freq_max = labels_freq[label]
return ref_freq / freq_tot
def class_probability(self, point, c):
distances = []
# Compute distances with all dataset
for i in xrange(len(self.__data)):
dist = distance.euclidian(self.__data[i], point)
distances.append((dist, self.__labels[i]))
# Keep only K nearest points
distances.sort(dist_cmp)
distances = distances[: self.k]
# Compute the most frequent label in neighbours
labels_freq = {}
freq_tot = 0.0
ref_freq = 0.0
for i in xrange(len(distances)):
freq_tot += 1
if distances[i][1] == c:
ref_freq += 1
if not labels_freq.has_key(distances[i][1]):
labels_freq[distances[i][1]] = 0
else:
labels_freq[distances[i][1]] += 1
i = 0
for label in labels_freq:
if i == 0:
label_max = label
freq_max = labels_freq[label]
i += 1
else:
if labels_freq[label] > freq_max:
label_max = label
freq_max = labels_freq[label]
return ref_freq / freq_tot
def clusterize(neurons, data_dimension, clusters, dataset, nb_clusters=0):
# Each neuron is put in its own cluster.
for neuron in neurons:
c = cluster.Cluster(data_dimension)
c.add_neuron(neuron)
c.compute_average()
clusters.append(c)
# Every element from the dataset is classified and the corresponding
# cluster stores its label.
for z in dataset:
nearest_cluster = clusters[0]
best_distance = distance.euclidian(clusters[0], z.data())
for i in xrange(1, len(clusters)):
tmp_distance = distance.euclidian(clusters[i], z.data())
if tmp_distance < best_distance:
nearest_cluster = clusters[i]
best_distance = tmp_distance
nearest_cluster.add_label(z.labels())
nearest_cluster._neurons[0]._labels.append(z.labels())
# Here we reduce the number of cluster to the one specified. To do
# that we look for the two closest clusters and we merge them until
# we have the good number of clusters.
if nb_clusters > 0:
while len(clusters) > nb_clusters:
best_distance = clusters[0].distance_with(clusters[1].get_center())
cluster1 = 0
cluster2 = 1
for i in xrange(len(clusters)):
for j in xrange(len(clusters)):
if i != j:
tmp_distance = clusters[i].distance_with(
clusters[j].get_center())
if tmp_distance < best_distance:
cluster1 = i
cluster2 = j
best_distance = tmp_distance
clusters[i].merge_with(clusters[j])
clusters.pop(j)
def clusterize (neurons, data_dimension, clusters, dataset, nb_clusters=0):
# Each neuron is put in its own cluster.
for neuron in neurons:
c = cluster.Cluster(data_dimension)
c.add_neuron(neuron)
c.compute_average()
clusters.append(c)
# Every element from the dataset is classified and the corresponding
# cluster stores its label.
for z in dataset:
nearest_cluster = clusters[0]
best_distance = distance.euclidian(clusters[0], z.data())
for i in xrange(1, len(clusters)):
tmp_distance = distance.euclidian(clusters[i], z.data())
if tmp_distance < best_distance:
nearest_cluster = clusters[i]
best_distance = tmp_distance
nearest_cluster.add_label(z.labels())
nearest_cluster._neurons[0]._labels.append(z.labels())
# Here we reduce the number of cluster to the one specified. To do
# that we look for the two closest clusters and we merge them until
# we have the good number of clusters.
if nb_clusters > 0:
while len(clusters) > nb_clusters:
best_distance = clusters[0].distance_with(clusters[1].get_center())
cluster1 = 0
cluster2 = 1
for i in xrange(len(clusters)):
for j in xrange(len(clusters)):
if i != j:
tmp_distance = clusters[i].distance_with(clusters[j].get_center())
if tmp_distance < best_distance:
cluster1 = i
cluster2 = j
best_distance = tmp_distance
clusters[i].merge_with(clusters[j])
clusters.pop(j)
def learn (self):
if self._verbose:
self.draw_all()
# Parameters initialization
neurons = self._neurons
sigma_init = distance.euclidian(neurons[len(neurons) / 2].get_position(), \
neurons[len(neurons) - 1].get_position(), \
dimension=len(neurons[0].get_position()))
t1 = 1000 / math.log(sigma_init)
nu0 = 0.1
t2 = 1000
nu = lambda n: nu0 * math.exp(-n / t2)
iter = 1
data = self._dataset.data()
while (iter <= ITERATION_MAX):
z = data[int(random.random() * len(data))]
sigma = sigma_init * math.exp(-iter / t1)
# Get nearest neuron for observation z
nearest_neuron = neurons[0]
best_distance = distance.euclidian(neurons[0], z)
for i in xrange(1, len(neurons)):
tmp_distance = distance.euclidian(neurons[i], z)
if tmp_distance < best_distance:
nearest_neuron = neurons[i]
best_distance = tmp_distance
if (iter % len(neurons) == 0):
changed_neurons = 0
else:
changed_neurons = None
for neuron in neurons:
d = distance.euclidian(neuron.get_position(), \
nearest_neuron.get_position(), \
dimension=len(neuron.get_position()))
h = gaussian_kernel(d, sigma)
dmoved = 0
for i in xrange(len(neuron)):
step = nu(iter) * h * (z[i] - neuron[i])
neuron[i] += step
dmoved += abs(step)
if not changed_neurons is None:
if dmoved > 0:
changed_neurons += 1
if not changed_neurons is None:
perc = float(changed_neurons) / len(neurons) * 100
if perc <= PERC_NEURONS_CHANGED:
break
if iter % 1000 == 0 and self._verbose:
print "%f%% of neurons moved at last check" % perc
print "iteration %i" % iter
self.draw_all()
iter += 1
if self._verbose:
print "%i%% of neurons moved at last check" % int(perc)
self.draw_all()
clusterize(self._neurons, self._data_dimension, self._clusters, self._dataset, self._nb_clusters)
return iter - 1
def learn(self):
if self._verbose:
self.draw_all()
# Parameters initialization
neurons = self._neurons
sigma_init = distance.euclidian(neurons[len(neurons) / 2].get_position(), \
neurons[len(neurons) - 1].get_position(), \
dimension=len(neurons[0].get_position()))
t1 = 1000 / math.log(sigma_init)
nu0 = 0.1
t2 = 1000
nu = lambda n: nu0 * math.exp(-n / t2)
iter = 1
data = self._dataset.data()
while (iter <= ITERATION_MAX):
z = data[int(random.random() * len(data))]
sigma = sigma_init * math.exp(-iter / t1)
# Get nearest neuron for observation z
nearest_neuron = neurons[0]
best_distance = distance.euclidian(neurons[0], z)
for i in xrange(1, len(neurons)):
tmp_distance = distance.euclidian(neurons[i], z)
if tmp_distance < best_distance:
nearest_neuron = neurons[i]
best_distance = tmp_distance
if (iter % len(neurons) == 0):
changed_neurons = 0
else:
changed_neurons = None
for neuron in neurons:
d = distance.euclidian(neuron.get_position(), \
nearest_neuron.get_position(), \
dimension=len(neuron.get_position()))
h = gaussian_kernel(d, sigma)
dmoved = 0
for i in xrange(len(neuron)):
step = nu(iter) * h * (z[i] - neuron[i])
neuron[i] += step
dmoved += abs(step)
if not changed_neurons is None:
if dmoved > 0:
changed_neurons += 1
if not changed_neurons is None:
perc = float(changed_neurons) / len(neurons) * 100
if perc <= PERC_NEURONS_CHANGED:
break
if iter % 1000 == 0 and self._verbose:
print "%f%% of neurons moved at last check" % perc
print "iteration %i" % iter
self.draw_all()
iter += 1
if self._verbose:
print "%i%% of neurons moved at last check" % int(perc)
self.draw_all()
clusterize(self._neurons, self._data_dimension, self._clusters,
self._dataset, self._nb_clusters)
return iter - 1
