class GAAClusterer(VectorSpaceClusterer):
"""
The Group Average Agglomerative starts with each of the N vectors as singleton
clusters. It then iteratively merges pairs of clusters which have the
closest centroids. This continues until there is only one cluster. The
order of merges gives rise to a dendogram: a tree with the earlier merges
lower than later merges. The membership of a given number of clusters c, 1
<= c <= N, can be found by cutting the dendogram at depth c.
This clusterer uses the cosine similarity metric only, which allows for
efficient speed-up in the clustering process.
"""
def __init__(self, num_clusters=1, normalise=True, svd_dimensions=None):
VectorSpaceClusterer.__init__(self, normalise, svd_dimensions)
self._num_clusters = num_clusters
self._dendogram = None
self._groups_values = None
def cluster(self, vectors, assign_clusters=False, trace=False):
# stores the merge order
self._dendogram = Dendogram(
[numpy.array(vector, numpy.float64) for vector in vectors])
return VectorSpaceClusterer.cluster(self, vectors, assign_clusters, trace)
def cluster_vectorspace(self, vectors, trace=False):
# create a cluster for each vector
clusters = [[vector] for vector in vectors]
# the sum vectors
vector_sum = copy.copy(vectors)
while len(clusters) > max(self._num_clusters, 1):
# find the two best candidate clusters to merge, based on their
# S(union c_i, c_j)
best = None
for i in range(len(clusters)):
for j in range(i + 1, len(clusters)):
sim = self._average_similarity(
vector_sum[i], len(clusters[i]),
vector_sum[j], len(clusters[j]))
if not best or sim > best[0]:
best = (sim, i, j)
# merge them and replace in cluster list
i, j = best[1:]
vsum = clusters[i] + clusters[j]
if trace: print 'merging %d and %d' % (i, j)
clusters[i] = vsum
del clusters[j]
vector_sum[i] = vector_sum[i] + vector_sum[j]
del vector_sum[j]
self._dendogram.merge(i, j)
self.update_clusters(self._num_clusters)
def update_clusters(self, num_clusters):
clusters = self._dendogram.groups(num_clusters)
self._centroids = []
for cluster in clusters:
assert len(cluster) > 0
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
centroid /= float(len(cluster))
self._centroids.append(centroid)
self._num_clusters = len(self._centroids)
def compute_rss(self, num_clusters):
clusters = self._dendogram.groups(num_clusters)
rss = 0
for cluster in clusters:
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
centroid = centroid / float(len(cluster))
for vector in cluster:
diff = vector - centroid
rss = rss + numpy.sqrt(numpy.vdot(diff, diff))
return rss
def classify_vectorspace(self, vector):
best = None
for i in range(self._num_clusters):
centroid = self._centroids[i]
sim = self._average_similarity(vector, 1, centroid, 1)
if not best or sim > best[0]:
best = (sim, i)
return best[1]
def dendogram(self):
"""
@return: The dendogram representing the current clustering
@rtype: Dendogram
"""
return self._dendogram
def num_clusters(self):
return self._num_clusters
def _average_similarity(self, v1, l1, v2, l2):
asum = v1 + v2
length = l1 + l2
return (numpy.dot(asum, asum) - length) / (length * (length - 1))
def __repr__(self):
return '<GroupAverageAgglomerative Clusterer n=%d>' % self._num_clusters
class OptCentroidClusterer(VectorSpaceClusterer):
def __init__(self, vector_names = None, num_clusters=1, normalise=True, svd_dimensions=None, msg_handle=None):
VectorSpaceClusterer.__init__(self, normalise, svd_dimensions)
self._num_clusters = num_clusters
self._dendogram = None
self._groups_values = None
self._names = vector_names
self._name_dendogram = None
self._reassigned_clusters = {}
self.msg_handle = msg_handle
def array_max(self, ar):
for i in range(ar.shape[0]):
ar[i, i] = -9e9
location = ar.argmax()
r = int(round (location / ar.shape[1]))
c = int(numpy.mod(location, ar.shape[1]))
return [r, c]
def cluster(self, vectors, assign_clusters=False, trace=False):
# stores the merge order
if self.msg_handle is not None:
self.msg_handle.dm(str(len(vectors)))
self.msg_handle.tile_yield()
self._dendogram = Dendogram(
[numpy.array(vector, numpy.float64) for vector in vectors])
if self._names:
self._name_dendogram = Dendogram(self._names)
self._vectors_to_cluster = vectors
return VectorSpaceClusterer.cluster(self, vectors, assign_clusters, trace)
def cluster_vectorspace(self, vectors, trace=False):
# create a cluster for each vector
clusters = [[vector] for vector in vectors]
# This copy module and function is from the python standard library
vector_sum = copy.copy(vectors)
norm_sum = [normalize(vsum) for vsum in vector_sum]
cluster_matrix = numpy.zeros([len(vectors[0]), len(vectors)])
i = 0
for v in norm_sum:
cluster_matrix[:, i] = v
i = i + 1
if self.msg_handle is not None:
self.msg_handle.dm("initializing dot_store_matrix")
dot_store_matrix = numpy.dot(cluster_matrix.transpose(), cluster_matrix)
start = time.time()
while len(clusters) > max(self._num_clusters, 1):
# find the two best candidate clusters to merge
max_sim = self.array_max(dot_store_matrix)
i = max_sim[0]
j = max_sim[1]
if i == j:
print "got stuck when at " + str(len(clusters)) + " clusters"
break;
vsum = clusters[i] + clusters[j]
if trace:
print 'merging %d and %d' % (i, j)
clusters[i] = vsum
del clusters[j]
vector_sum[i] = vector_sum[i] + vector_sum[j]
norm_sum[i] = normalize(vector_sum[i])
cluster_matrix[:, i] = norm_sum[i]
del vector_sum[j]
del norm_sum[j]
cluster_matrix = numpy.delete(cluster_matrix, j, 1)
dot_store_matrix = numpy.dot(cluster_matrix.transpose(), cluster_matrix)
self._dendogram.merge(i, j)
if self._names:
self._name_dendogram.merge(i, j)
end = time.time()
if end - start > 5:
if self.msg_handle is not None:
self.msg_handle.dm(str(len(clusters)))
start = end
else:
print len(clusters)
self.update_clusters(len(clusters))
def update_clusters(self, num_clusters):
print "entering update clusters with num_clusters = " + str(num_clusters)
clusters = self._dendogram.groups(num_clusters)
self._centroids = []
for cluster in clusters:
assert len(cluster) > 0
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
# centroid /= float(len(cluster)) # was this supposed to be some sort of normalizing?
norm_centroid = normalize(centroid)
self._centroids.append(norm_centroid)
self._num_clusters = len(self._centroids)
def compute_rss(self, num_clusters):
clusters = self._dendogram.groups(num_clusters)
rss = 0
for cluster in clusters:
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
if self._should_normalise:
centroid = self._normalise(centroid)
for vector in cluster:
diff = vector - centroid
rss = rss + numpy.sqrt(numpy.vdot(diff, diff))
return rss
def get_iteratively_reassigned_clusters(self, num_clusters, max_iterations=100, saveit=True):
if num_clusters in self._reassigned_clusters:
return self._reassigned_clusters[num_clusters]
self.update_clusters(num_clusters)
new_centroids = self._centroids
clusters = self._dendogram.groups(num_clusters)
for iter in range(max_iterations):
number_reassigned = 0
new_clusters = [[] for i in range(len(self._centroids))]
for cluster_number, cluster in enumerate(clusters):
for vec in cluster:
dps = [numpy.dot(numpy.transpose(centroid), vec) for centroid in new_centroids]
new_cluster_index = dps.index(max(dps))
new_clusters[new_cluster_index].append(vec)
if new_cluster_index != cluster_number:
number_reassigned += 1
new_centroids = []
for cluster in new_clusters:
assert len(cluster) > 0
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
# centroid /= float(len(cluster)) # was this supposed to be some sort of normalizing?
norm_centroid = normalize(centroid)
new_centroids.append(norm_centroid)
print [len(cluster) for cluster in new_clusters]
print "Number reassigned = %i" % number_reassigned
if number_reassigned == 0:
print "Stable after %i iterations" % iter
break;
clusters = new_clusters
if saveit:
self._reassigned_clusters[num_clusters] = (new_centroids,[len(cluster) for cluster in new_clusters], clusters)
class CentroidClusterer(VectorSpaceClusterer):
def __init__(self, vector_names=None, num_clusters=1, normalise=True, svd_dimensions=None):
VectorSpaceClusterer.__init__(self, normalise, svd_dimensions)
self._num_clusters = num_clusters
self._dendogram = None
self._groups_values = None
self._names = vector_names
self._name_dendogram = None
def cluster(self, vectors, assign_clusters=False, trace=False):
# stores the merge order
self._dendogram = Dendogram([numpy.array(vector, numpy.float64) for vector in vectors])
if self._names:
self._name_dendogram = Dendogram(self._names)
self._vectors_to_cluster = vectors
return VectorSpaceClusterer.cluster(self, vectors, assign_clusters, trace)
def cluster_vectorspace(self, vectors, trace=False):
# create a cluster for each vector
clusters = [[vector] for vector in vectors]
# This copy module and function is from the python standard library
vector_sum = copy.copy(vectors)
norm_sum = [normalize(vsum) for vsum in vector_sum]
while len(clusters) > max(self._num_clusters, 1):
# find the two best candidate clusters to merge
best = None
for i in range(len(clusters)):
for j in range(i + 1, len(clusters)):
sim = numpy.dot(norm_sum[i], norm_sum[j])
if not best or sim > best[0]:
best = (sim, i, j)
# merge them and replace in cluster list
i, j = best[1:]
csum = clusters[i] + clusters[j]
if trace:
print 'merging %d and %d' % (i, j)
# print len(clusters)
clusters[i] = csum
del clusters[j]
vector_sum[i] = vector_sum[i] + vector_sum[j]
norm_sum[i] = normalize(vector_sum[i])
del vector_sum[j]
del norm_sum[j]
self._dendogram.merge(i, j)
if self._names:
self._name_dendogram.merge(i, j)
if len(clusters) % 50 == 0:
print len(clusters)
self.update_clusters(self._num_clusters)
def update_clusters(self, num_clusters):
clusters = self._dendogram.groups(num_clusters)
# print clusters
self._centroids = []
for cluster in clusters:
assert len(cluster) > 0
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
centroid /= float(len(cluster)) # was this supposed to be some sort of normalizing?
if self._should_normalise:
self._centroids.append(normalize(centroid))
else:
self._centroids.append(centroid)
self._num_clusters = len(self._centroids)
def compute_rss(self, num_clusters):
clusters = self._dendogram.groups(num_clusters)
rss = 0
for cluster in clusters:
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
centroid /= float(len(cluster))
for vector in cluster:
diff = vector - centroid
rss = rss + numpy.sqrt(numpy.vdot(diff, diff))
return rss
def classify_vectorspace(self, vector):
best = None
for i in range(self._num_clusters):
centroid = self._centroids[i]
sim = self._similarity(vector, centroid)
if not best or sim > best[0]:
best = (sim, i)
return best[1]
def dendogram(self):
"""
@return: The dendogram representing the current clustering
@rtype: Dendogram
"""
return self._dendogram
def name_dendogram(self):
return self._name_dendogram
def num_clusters(self):
return self._num_clusters
def _similarity(self, v1, v2):
return (numpy.dot(v1, v2))
def __repr__(self):
return '<Centroid Clusterer n=%d>' % self._num_clusters
class GAAClusterer(VectorSpaceClusterer):
"""
The Group Average Agglomerative starts with each of the N vectors as singleton
clusters. It then iteratively merges pairs of clusters which have the
closest centroids. This continues until there is only one cluster. The
order of merges gives rise to a dendogram: a tree with the earlier merges
lower than later merges. The membership of a given number of clusters c, 1
<= c <= N, can be found by cutting the dendogram at depth c.
This clusterer uses the cosine similarity metric only, which allows for
efficient speed-up in the clustering process.
"""
def __init__(self, num_clusters=1, normalise=True, svd_dimensions=None):
VectorSpaceClusterer.__init__(self, normalise, svd_dimensions)
self._num_clusters = num_clusters
self._dendogram = None
self._groups_values = None
def cluster(self, vectors, assign_clusters=False, trace=False):
# stores the merge order
self._dendogram = Dendogram(
[numpy.array(vector, numpy.float64) for vector in vectors])
return VectorSpaceClusterer.cluster(self, vectors, assign_clusters,
trace)
def cluster_vectorspace(self, vectors, trace=False):
# create a cluster for each vector
clusters = [[vector] for vector in vectors]
# the sum vectors
vector_sum = copy.copy(vectors)
while len(clusters) > max(self._num_clusters, 1):
# find the two best candidate clusters to merge, based on their
# S(union c_i, c_j)
best = None
for i in range(len(clusters)):
for j in range(i + 1, len(clusters)):
sim = self._average_similarity(vector_sum[i],
len(clusters[i]),
vector_sum[j],
len(clusters[j]))
if not best or sim > best[0]:
best = (sim, i, j)
# merge them and replace in cluster list
i, j = best[1:]
vsum = clusters[i] + clusters[j]
if trace: print 'merging %d and %d' % (i, j)
clusters[i] = vsum
del clusters[j]
vector_sum[i] = vector_sum[i] + vector_sum[j]
del vector_sum[j]
self._dendogram.merge(i, j)
self.update_clusters(self._num_clusters)
def update_clusters(self, num_clusters):
clusters = self._dendogram.groups(num_clusters)
self._centroids = []
for cluster in clusters:
assert len(cluster) > 0
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
centroid /= float(len(cluster))
self._centroids.append(centroid)
self._num_clusters = len(self._centroids)
def compute_rss(self, num_clusters):
clusters = self._dendogram.groups(num_clusters)
rss = 0
for cluster in clusters:
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
for vector in cluster:
diff = vector - centroid
rss = rss + numpy.sqrt(numpy.vdot(diff, diff))
return rss
def classify_vectorspace(self, vector):
best = None
for i in range(self._num_clusters):
centroid = self._centroids[i]
sim = self._average_similarity(vector, 1, centroid, 1)
if not best or sim > best[0]:
best = (sim, i)
return best[1]
def dendogram(self):
"""
@return: The dendogram representing the current clustering
@rtype: Dendogram
"""
return self._dendogram
def num_clusters(self):
return self._num_clusters
def _average_similarity(self, v1, l1, v2, l2):
asum = v1 + v2
length = l1 + l2
return (numpy.dot(asum, asum) - length) / (length * (length - 1))
def __repr__(self):
return '<GroupAverageAgglomerative Clusterer n=%d>' % self._num_clusters
class OptCentroidClusterer(VectorSpaceClusterer):
def __init__(self,
vector_names=None,
num_clusters=1,
normalise=True,
svd_dimensions=None):
VectorSpaceClusterer.__init__(self, normalise, svd_dimensions)
self._num_clusters = num_clusters
self._dendogram = None
self._groups_values = None
self._names = vector_names
self._name_dendogram = None
def array_max(self, ar):
for i in range(ar.shape[0]):
ar[i, i] = -9e9
location = ar.argmax()
r = int(round(location / ar.shape[1]))
c = int(numpy.mod(location, ar.shape[1]))
return [r, c]
def cluster(self, vectors, assign_clusters=False, trace=False):
# stores the merge order
self._dendogram = Dendogram(
[numpy.array(vector, numpy.float64) for vector in vectors])
if self._names:
self._name_dendogram = Dendogram(self._names)
self._vectors_to_cluster = vectors
return VectorSpaceClusterer.cluster(self, vectors, assign_clusters,
trace)
def cluster_vectorspace(self, vectors, trace=False):
# create a cluster for each vector
clusters = [[vector] for vector in vectors]
# This copy module and function is from the python standard library
vector_sum = copy.copy(vectors)
norm_sum = [normalize(vsum) for vsum in vector_sum]
cluster_matrix = numpy.zeros([len(vectors[0]), len(vectors)])
i = 0
for v in norm_sum:
cluster_matrix[:, i] = v
i = i + 1
dot_store_matrix = numpy.dot(cluster_matrix.transpose(),
cluster_matrix)
while len(clusters) > max(self._num_clusters, 1):
# find the two best candidate clusters to merge
max_sim = self.array_max(dot_store_matrix)
i = max_sim[0]
j = max_sim[1]
if i == j:
print "got stuck when at " + str(len(clusters)) + " clusters"
break
vsum = clusters[i] + clusters[j]
if trace:
print 'merging %d and %d' % (i, j)
clusters[i] = vsum
del clusters[j]
vector_sum[i] = vector_sum[i] + vector_sum[j]
norm_sum[i] = normalize(vector_sum[i])
cluster_matrix[:, i] = norm_sum[i]
del vector_sum[j]
del norm_sum[j]
cluster_matrix = numpy.delete(cluster_matrix, j, 1)
dot_store_matrix = numpy.dot(cluster_matrix.transpose(),
cluster_matrix)
self._dendogram.merge(i, j)
if self._names:
self._name_dendogram.merge(i, j)
if len(clusters) % 50 == 0:
print len(clusters)
self.update_clusters(len(clusters))
def update_clusters(self, num_clusters):
print "entering update clusters with num_clusters = " + str(
num_clusters)
clusters = self._dendogram.groups(num_clusters)
self._centroids = []
for cluster in clusters:
assert len(cluster) > 0
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
# centroid /= float(len(cluster)) # was this supposed to be some sort of normalizing?
norm_centroid = normalize(centroid)
self._centroids.append(norm_centroid)
self._num_clusters = len(self._centroids)
def compute_rss(self, num_clusters):
clusters = self._dendogram.groups(num_clusters)
rss = 0
for cluster in clusters:
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
if self._should_normalise:
centroid = self._normalise(centroid)
for vector in cluster:
diff = vector - centroid
rss = rss + numpy.sqrt(numpy.vdot(diff, diff))
return rss
def classify_vectorspace(self, vector):
best = None
for i in range(self._num_clusters):
centroid = self._centroids[i]
sim = self._similarity(vector, centroid)
if not best or sim > best[0]:
best = (sim, i)
return best[1]
def dendogram(self):
"""
@return: The dendogram representing the current clustering
@rtype: Dendogram
"""
return self._dendogram
def name_dendogram(self):
return self._name_dendogram
def num_clusters(self):
return self._num_clusters
def _similarity(self, v1, v2):
return (numpy.dot(v1, v2))
def __repr__(self):
return '<Opt Centroid Clusterer n=%d>' % self._num_clusters
class CentroidClusterer(VectorSpaceClusterer):
def __init__(self,
vector_names=None,
num_clusters=1,
normalise=True,
svd_dimensions=None):
VectorSpaceClusterer.__init__(self, normalise, svd_dimensions)
self._num_clusters = num_clusters
self._dendogram = None
self._groups_values = None
self._names = vector_names
self._name_dendogram = None
def cluster(self, vectors, assign_clusters=False, trace=False):
# stores the merge order
self._dendogram = Dendogram(
[numpy.array(vector, numpy.float64) for vector in vectors])
if self._names:
self._name_dendogram = Dendogram(self._names)
self._vectors_to_cluster = vectors
return VectorSpaceClusterer.cluster(self, vectors, assign_clusters,
trace)
def cluster_vectorspace(self, vectors, trace=False):
# create a cluster for each vector
clusters = [[vector] for vector in vectors]
# This copy module and function is from the python standard library
vector_sum = copy.copy(vectors)
norm_sum = [normalize(vsum) for vsum in vector_sum]
while len(clusters) > max(self._num_clusters, 1):
# find the two best candidate clusters to merge
best = None
for i in range(len(clusters)):
for j in range(i + 1, len(clusters)):
sim = numpy.dot(norm_sum[i], norm_sum[j])
if not best or sim > best[0]:
best = (sim, i, j)
# merge them and replace in cluster list
i, j = best[1:]
csum = clusters[i] + clusters[j]
if trace:
print 'merging %d and %d' % (i, j)
# print len(clusters)
clusters[i] = csum
del clusters[j]
vector_sum[i] = vector_sum[i] + vector_sum[j]
norm_sum[i] = normalize(vector_sum[i])
del vector_sum[j]
del norm_sum[j]
self._dendogram.merge(i, j)
if self._names:
self._name_dendogram.merge(i, j)
if len(clusters) % 50 == 0:
print len(clusters)
self.update_clusters(self._num_clusters)
def update_clusters(self, num_clusters):
clusters = self._dendogram.groups(num_clusters)
# print clusters
self._centroids = []
for cluster in clusters:
assert len(cluster) > 0
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
# centroid /= float(len(cluster)) # was this supposed to be some sort of normalizing?
if self._should_normalise:
self._centroids.append(normalize(centroid))
else:
self._centroids.append(centroid)
self._num_clusters = len(self._centroids)
def compute_rss(self, num_clusters):
clusters = self._dendogram.groups(num_clusters)
rss = 0
for cluster in clusters:
if self._should_normalise:
centroid = self._normalise(cluster[0])
else:
centroid = numpy.array(cluster[0])
for vector in cluster[1:]:
if self._should_normalise:
centroid += self._normalise(vector)
else:
centroid += vector
for vector in cluster:
diff = vector - centroid
rss = rss + numpy.sqrt(numpy.vdot(diff, diff))
return rss
def classify_vectorspace(self, vector):
best = None
for i in range(self._num_clusters):
centroid = self._centroids[i]
sim = self._similarity(vector, centroid)
if not best or sim > best[0]:
best = (sim, i)
return best[1]
def dendogram(self):
"""
@return: The dendogram representing the current clustering
@rtype: Dendogram
"""
return self._dendogram
def name_dendogram(self):
return self._name_dendogram
def num_clusters(self):
return self._num_clusters
def _similarity(self, v1, v2):
print "othertest"
return (numpy.dot(v1, v2))
def __repr__(self):
return '<Centroid Clusterer n=%d>' % self._num_clusters