def __kmedoids_compression(self, clustering, matrix_handler):
"""
"""
representatives = []
for cluster in clustering.clusters:
# Guess 'correct' number of elements for this cluster
cluster_size = cluster.get_size()
expected_cluster_elements = cluster_size * (float(self.parameters["final_number_of_frames"]) / clustering.total_number_of_elements)
expected_cluster_elements = int(math.ceil(expected_cluster_elements))
remapped_matrix = get_submatrix(matrix_handler.distance_matrix, cluster.all_elements)
# Prepare and run kmedoids algorithm
kmedoids = KMedoidsAlgorithm(remapped_matrix)
# print "KMEDOIDS:: EXPECTED", expected_cluster_elements, cluster_size, clustering.total_number_of_elements, self.parameters["final_number_of_frames"]
new_clustering = kmedoids.perform_clustering({
"k": expected_cluster_elements,
"seeding_type": "EQUIDISTANT"
})
# print "NEW CLUSTERING SIZE clusters: %d elements: %d"%(len(new_clustering.clusters), new_clustering.total_number_of_elements)
# reverse the remapping and add it to representatives
remapped_representatives = new_clustering.get_medoids(remapped_matrix)
fake_cluster = Cluster(None, remapped_representatives)
representatives.extend(Refiner.redefine_cluster_with_map(cluster, fake_cluster).all_elements)
return representatives
def test_gromos_seeding(self):
points = [(0,0),(0,1),(0,-1),(1,0),
(3,0),(3,1),
(6,0),(7,0),(7,1),(7,-1)]
# Maximum distance of connected components is 1, for 1.5 it may discover 3 clusters.
# For 3.2 it may discover only 2
#
# 1 5 8
# | | |
# 0 - 3 4 6 - 7
# | |
# 2 9
#
matrix = CondensedMatrix(pdist(points))
kmed_alg = KMedoidsAlgorithm(matrix, rand_seed = 10)
# With a small cutoff we get all 3 connected components
numpy.testing.assert_array_equal( kmed_alg.gromos_seeding(3, 1.4),[0, 7, 4]) # if it's 1.5 we'll have 6 instead of 7 as medoid (as it is the first one to have 3 neighbours)
# With a bigger cutoff it has to try to find the minimum cutoff for 3 clusters, then 6 is returned instead of 7
numpy.testing.assert_array_equal( kmed_alg.gromos_seeding(3, 3.2),[3, 6, 5])
# This one is regression
numpy.testing.assert_array_equal( kmed_alg.gromos_seeding(2, 3.2),[4, 7])
# This one should return a random sequence, so is only testable because of the rand_seed
numpy.testing.assert_array_equal(kmed_alg.gromos_seeding(2, 0), [5, 3])
def test_update_medoids(self):
points = [(0, 0), (0, 1), (0, -1), (1, 0), (6, 0), (7, 0), (7, 1),
(7, -1)]
matrix = CondensedMatrix(pdist(points))
kmed_alg = KMedoidsAlgorithm(matrix)
kmed_alg.class_list = [0, 0, 0, 0, 1, 1, 1, 1]
numpy.testing.assert_array_equal(kmed_alg.update_medoids(), [0, 5])
def test_update_medoids(self):
points = [(0,0),(0,1),(0,-1),(1,0),
(6,0),(7,0),(7,1),(7,-1)]
matrix = CondensedMatrix(pdist(points))
kmed_alg = KMedoidsAlgorithm(matrix)
kmed_alg.class_list = [0, 0, 0, 0,
1, 1, 1, 1]
numpy.testing.assert_array_equal( kmed_alg.update_medoids(), [0,5])
def test_naive_case(self):
# 1 5 8
# | | |
# 0 - 3 4 6 - 7
# | |
# 2 9
points = [(0,0),(0,1),(0,-1),(1,0),
(3,0),(3,1),
(6,0),(7,0),(7,1),(7,-1)]
matrix = CondensedMatrix(pdist(points))
s_algo = KMedoidsAlgorithm(matrix, 10)
clusters = s_algo.perform_clustering({'k':3, 'seeding_type':'RANDOM'}).clusters
for c in clusters:
self.assertIn(c.prototype, [0, 4, 6])
self.assertIn(c.all_elements, [[0, 1, 2, 3],[6, 7, 8, 9],[4, 5]])
def test_naive_case(self):
# 1 5 8
# | | |
# 0 - 3 4 6 - 7
# | |
# 2 9
points = [(0, 0), (0, 1), (0, -1), (1, 0), (3, 0), (3, 1), (6, 0),
(7, 0), (7, 1), (7, -1)]
matrix = CondensedMatrix(pdist(points))
s_algo = KMedoidsAlgorithm(matrix, 10)
clusters = s_algo.perform_clustering({
'k': 3,
'seeding_type': 'RANDOM'
}).clusters
for c in clusters:
self.assertIn(c.prototype, [0, 4, 6])
self.assertIn(c.all_elements, [[0, 1, 2, 3], [6, 7, 8, 9], [4, 5]])
def test_gromos_seeding(self):
points = [(0, 0), (0, 1), (0, -1), (1, 0), (3, 0), (3, 1), (6, 0),
(7, 0), (7, 1), (7, -1)]
# Maximum distance of connected components is 1, for 1.5 it may discover 3 clusters.
# For 3.2 it may discover only 2
#
# 1 5 8
# | | |
# 0 - 3 4 6 - 7
# | |
# 2 9
#
matrix = CondensedMatrix(pdist(points))
kmed_alg = KMedoidsAlgorithm(matrix, rand_seed=10)
# With a small cutoff we get all 3 connected components
numpy.testing.assert_array_equal(
kmed_alg.gromos_seeding(3, 1.4), [0, 7, 4]
) # if it's 1.5 we'll have 6 instead of 7 as medoid (as it is the first one to have 3 neighbours)
# With a bigger cutoff it has to try to find the minimum cutoff for 3 clusters, then 6 is returned instead of 7
numpy.testing.assert_array_equal(kmed_alg.gromos_seeding(3, 3.2),
[3, 6, 5])
# This one is regression
numpy.testing.assert_array_equal(kmed_alg.gromos_seeding(2, 3.2),
[4, 7])
# This one should return a random sequence, so is only testable because of the rand_seed
numpy.testing.assert_array_equal(kmed_alg.gromos_seeding(2, 0), [5, 3])
def __kmedoids_compression(self, clustering, matrix_handler):
"""
"""
representatives = []
for cluster in clustering.clusters:
# Guess 'correct' number of elements for this cluster
cluster_size = cluster.get_size()
expected_cluster_elements = cluster_size * (
float(self.parameters["final_number_of_frames"]) /
clustering.total_number_of_elements)
expected_cluster_elements = int(
math.ceil(expected_cluster_elements))
remapped_matrix = get_submatrix(matrix_handler.distance_matrix,
cluster.all_elements)
# Prepare and run kmedoids algorithm
kmedoids = KMedoidsAlgorithm(remapped_matrix)
# print "KMEDOIDS:: EXPECTED", expected_cluster_elements, cluster_size, clustering.total_number_of_elements, self.parameters["final_number_of_frames"]
new_clustering = kmedoids.perform_clustering({
"k":
expected_cluster_elements,
"seeding_type":
"EQUIDISTANT"
})
# print "NEW CLUSTERING SIZE clusters: %d elements: %d"%(len(new_clustering.clusters), new_clustering.total_number_of_elements)
# reverse the remapping and add it to representatives
remapped_representatives = new_clustering.get_medoids(
remapped_matrix)
fake_cluster = Cluster(None, remapped_representatives)
representatives.extend(
Refiner.redefine_cluster_with_map(cluster,
fake_cluster).all_elements)
return representatives
def perform_clustering(self, kwargs):
"""
Does the actual clustering by doing a k-medoids clustering of the first k eigenvector rows.
@param kwargs: Dictionary with this mandatory keys:
- 'k': Number of clusters to generate. Must be <= than max_clusters
@return: a Clustering instance with the clustered data.
"""
# Mandatory parameter
k = int(kwargs["k"])
if k > self.max_clusters:
print "[ERROR SpectralClusteringAlgorithm::perform_clustering] this algorithm was defined to generate at most %d clusters."%self.max_clusters,
algorithm_details = "Spectral algorithm with k = %d and sigma squared = %.3f" %(int(k), self.sigma_sq)
if self.use_k_medoids:
# The row vectors we have are in R^k (so k length)
eigen_distances = CondensedMatrix(pdist(self.eigenvectors[:,:k]))
k_medoids_args = {
"k":k,
"seeding_max_cutoff":-1,
"seeding_type": "RANDOM"
}
k_medoids_alg = KMedoidsAlgorithm(eigen_distances)
clustering = k_medoids_alg.perform_clustering(k_medoids_args)
clustering.details = algorithm_details
return k_medoids_alg.perform_clustering(k_medoids_args)
else:
centroid, labels = scipy.cluster.vq.kmeans2(self.eigenvectors[:,:k],
k,
iter = 1000,
minit = 'random')
del centroid
clusters = gen_clusters_from_class_list(labels)
return Clustering(clusters,details = algorithm_details)
def perform_clustering(self, kwargs):
"""
Does the actual clustering by doing a k-medoids clustering of the first k eigenvector rows.
@param kwargs: Dictionary with this mandatory keys:
- 'k': Number of clusters to generate. Must be <= than max_clusters
@return: a Clustering instance with the clustered data.
"""
# Mandatory parameter
k = int(kwargs["k"])
if k > self.max_clusters:
print "[ERROR SpectralClusteringAlgorithm::perform_clustering] this algorithm was defined to generate at most %d clusters." % self.max_clusters,
algorithm_details = "Spectral algorithm with k = %d and sigma squared = %.3f" % (
int(k), self.sigma_sq)
if self.use_k_medoids:
# The row vectors we have are in R^k (so k length)
eigen_distances = CondensedMatrix(pdist(self.eigenvectors[:, :k]))
k_medoids_args = {
"k": k,
"seeding_max_cutoff": -1,
"seeding_type": "RANDOM"
}
k_medoids_alg = KMedoidsAlgorithm(eigen_distances)
clustering = k_medoids_alg.perform_clustering(k_medoids_args)
clustering.details = algorithm_details
return k_medoids_alg.perform_clustering(k_medoids_args)
else:
centroid, labels = scipy.cluster.vq.kmeans2(
self.eigenvectors[:, :k], k, iter=1000, minit='random')
del centroid
clusters = gen_clusters_from_class_list(labels)
return Clustering(clusters, details=algorithm_details)
def setUpClass(cls):
cls.condensed_matrix = CondensedMatrix(
[1.0, 4.5, 7.2, 6.7, 8.5, 4.5, 3.6, 7.8, 2.2, 2.0])
cls.kmed_alg = KMedoidsAlgorithm(cls.condensed_matrix)