def test_get_submatrix(self):
old_matrix = CondensedMatrix([0.2, 1.0, 0.3, 4.0, 6.1, 0.5, 0.6, 0.7, 0.5, 0.9, 0.8, 0.3, 0.4, 1.4, 2.9])
expected = [0.9, 0.3, 1.4]
numpy.testing.assert_array_almost_equal(expected, get_submatrix(old_matrix, [2, 3, 5]).get_data(), 6)
data = [
1.0,
2.0,
3.0,
4.0,
5.0,
6.0,
7.0,
8.0,
9.0,
10.0,
11.0,
12.0,
13.0,
14.0,
15.0,
16.0,
17.0,
18.0,
19.0,
20.0,
21.0,
]
matrix = CondensedMatrix(data)
all_elements_map = [2, 4, 6]
new_matrix = get_submatrix(matrix, all_elements_map)
expected1 = [13, 15, 20]
numpy.testing.assert_array_equal(expected1, new_matrix.get_data())
def evaluate(self, clustering, condensed_distance_matrix):
"""
Returns the cohesion value of a cluster. The weight will be the number of elements
of each cluster.
The maximum value will be (in the case we have only one cluster) the sum of
all the pair distances. We can avoid the 2x factor, increasing the performance.
"""
clustered_elements = sorted(clustering.get_all_clustered_elements())
number_of_clustered_elements = len(clustered_elements)
if number_of_clustered_elements > 0:
distances = get_submatrix(condensed_distance_matrix, clustered_elements)
max_cohesion = numpy.sum(distances.get_data())/number_of_clustered_elements
total_cohesion = 0
for c in clustering.clusters:
size = c.get_size()
weight = 1. / size
cohesion = 0.
for i in range(size-1):
for j in range(i+1, size):
cohesion = cohesion + condensed_distance_matrix[c[i],c[j]]
total_cohesion += weight*cohesion
return 1 - total_cohesion / max_cohesion
else:
return 0.
def evaluate(self, clustering, condensed_distance_matrix):
"""
Returns the cohesion value of a cluster. The weight will be the number of elements
of each cluster.
The maximum value will be (in the case we have only one cluster) the sum of
all the pair distances. We can avoid the 2x factor, increasing the performance.
"""
clustered_elements = sorted(clustering.get_all_clustered_elements())
number_of_clustered_elements = len(clustered_elements)
if number_of_clustered_elements > 0:
distances = get_submatrix(condensed_distance_matrix,
clustered_elements)
max_cohesion = numpy.sum(
distances.get_data()) / number_of_clustered_elements
total_cohesion = 0
for c in clustering.clusters:
size = c.get_size()
weight = 1. / size
cohesion = 0.
for i in range(size - 1):
for j in range(i + 1, size):
cohesion = cohesion + condensed_distance_matrix[c[i],
c[j]]
total_cohesion += weight * cohesion
return 1 - total_cohesion / max_cohesion
else:
return 0.
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_get_submatrix(self):
old_matrix = CondensedMatrix([
0.2, 1., 0.3, 4.0, 6.1, 0.5, 0.6, 0.7, 0.5, 0.9, 0.8, 0.3, 0.4,
1.4, 2.9
])
expected = [0.9, 0.3, 1.4]
numpy.testing.assert_array_almost_equal(
expected,
get_submatrix(old_matrix, [2, 3, 5]).get_data(), 6)
data = [
1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0,
13.0, 14.0, 15.0, 16.0, 17.0, 18.0, 19.0, 20.0, 21.0
]
matrix = CondensedMatrix(data)
all_elements_map = [2, 4, 6]
new_matrix = get_submatrix(matrix, all_elements_map)
expected1 = [13, 15, 20]
numpy.testing.assert_array_equal(expected1, new_matrix.get_data())
def run (self, clustering):
"""
Refine a clustering recursively using a k-means over each cluster.
New clusters obtained from a cluster must have no noise and
"""
max_partitions = self.refinement_parameters["max_partitions"]
try_step = int(max(1, float(max_partitions) / self.refinement_parameters["tries_per_cluster"]))
matrix = self.matrixHandler.distance_matrix
new_clusters = []
for cluster in clustering.clusters:
base_id = cluster.id
# The initial clustering is added to the list of new clusters.
# With this 'trick' the initial cluster also enters the competition for the best clustering price.
clusterings = {base_id:{"type":"refined_base",
"clustering": Clustering([cluster]),
"parameters": {}}}
submatrix = get_submatrix(matrix, cluster.all_elements)
# Proceed with some K Medoids partitions
# TODO: Generate parameters with parameter generator
for k in range(2,max_partitions,try_step):
clustering = self.repartition_with_kmedoids(cluster, k, submatrix)
clusterings["%s_%d"%(base_id,k)] = {"type":"refined",
"clustering": clustering,
"parameters": {"k":k}}
# Evaluate all clusterings and pick the best one
AnalysisRunner(scheduling_tools.build_scheduler(
self.clustering_parameters["clustering"]["control"],
self.observer),
clusterings,
AnalysisPopulator(self.matrixHandler,
self.trajectoryHandler,
self.clustering_parameters)).evaluate()
best_clustering_id, all_scores = BestClusteringSelector(self.clustering_parameters).choose_best(clusterings) # @UnusedVariable
new_clusters.extend(clusterings[best_clustering_id]["clustering"].clusters)
# Convert all new clusters in the new clustering
return {"type":"refined_clustering",
"clustering": Clustering(new_clusters),
"parameters": self.refinement_parameters}
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
parameters["data"]["files"] = [sys.argv[1], sys.argv[2]]
frames_ini = get_number_of_frames(sys.argv[1])
frames_proto = get_number_of_frames(sys.argv[2])
print sys.argv[1],"->",frames_ini
print sys.argv[2],"->",frames_proto
try:
Driver(Observer()).run(parameters)
except SystemExit:
# Expected improductive search
# Load again the matrix
handler = MatrixHandler({
"method": "load",
"parameters": {
"path": parameters["workspace"]["base"]+"/matrix/matrix"
}
})
matrix = handler.create_matrix(None)
submatrix = get_submatrix(matrix, range(frames_ini,frames_ini+frames_proto))
matrixToImage(submatrix, parameters["workspace"]["base"] +"/submatrix.png")
print "Original mean:",get_submatrix(matrix, range(0,frames_ini)).calculateMean()
values = []
for i in range(0,frames_ini):
for j in range(frames_ini,frames_ini+frames_proto):
values.append((handler.distance_matrix[i,j],i,j-frames_ini))
for d,i,j in sorted(values):
print "%d %d %.2f"% (i,j,d)
print "Combined mean:", numpy.mean(values)
print sys.argv[1], "->", frames_ini
print sys.argv[2], "->", frames_proto
try:
Driver(Observer()).run(parameters)
except SystemExit:
# Expected improductive search
# Load again the matrix
handler = MatrixHandler({
"method": "load",
"parameters": {
"path": parameters["workspace"]["base"] + "/matrix/matrix"
}
})
matrix = handler.create_matrix(None)
submatrix = get_submatrix(matrix,
range(frames_ini, frames_ini + frames_proto))
matrixToImage(submatrix,
parameters["workspace"]["base"] + "/submatrix.png")
print "Original mean:", get_submatrix(matrix, range(
0, frames_ini)).calculateMean()
values = []
for i in range(0, frames_ini):
for j in range(frames_ini, frames_ini + frames_proto):
values.append((handler.distance_matrix[i,
j], i, j - frames_ini))
for d, i, j in sorted(values):
print "%d %d %.2f" % (i, j, d)
print "Combined mean:", numpy.mean(values)
def run(self, clustering):
"""
Refine a clustering recursively using a k-means over each cluster.
New clusters obtained from a cluster must have no noise and
"""
max_partitions = self.refinement_parameters["max_partitions"]
try_step = int(
max(
1,
float(max_partitions) /
self.refinement_parameters["tries_per_cluster"]))
matrix = self.matrixHandler.distance_matrix
new_clusters = []
for cluster in clustering.clusters:
base_id = cluster.id
# The initial clustering is added to the list of new clusters.
# With this 'trick' the initial cluster also enters the competition for the best clustering price.
clusterings = {
base_id: {
"type": "refined_base",
"clustering": Clustering([cluster]),
"parameters": {}
}
}
submatrix = get_submatrix(matrix, cluster.all_elements)
# Proceed with some K Medoids partitions
# TODO: Generate parameters with parameter generator
for k in range(2, max_partitions, try_step):
clustering = self.repartition_with_kmedoids(
cluster, k, submatrix)
clusterings["%s_%d" % (base_id, k)] = {
"type": "refined",
"clustering": clustering,
"parameters": {
"k": k
}
}
# Evaluate all clusterings and pick the best one
AnalysisRunner(
scheduling_tools.build_scheduler(
self.clustering_parameters["clustering"]["control"],
self.observer), clusterings,
AnalysisPopulator(self.matrixHandler, self.trajectoryHandler,
self.clustering_parameters)).evaluate()
best_clustering_id, all_scores = BestClusteringSelector(
self.clustering_parameters).choose_best(
clusterings) # @UnusedVariable
new_clusters.extend(
clusterings[best_clustering_id]["clustering"].clusters)
# Convert all new clusters in the new clustering
return {
"type": "refined_clustering",
"clustering": Clustering(new_clusters),
"parameters": self.refinement_parameters
}