def initialize(self, annotation, feature=None):
"""Initialize HAC with one cluster per label
Parameters
----------
annotation : Annotation
feature : Feature, optional
"""
# initialize annotation
self.annotation = annotation.copy()
# initialize history with original annotation
self.history = HACHistory(self.annotation)
# one cluster per label
clusters = self.annotation.labels()
# one model per cluster
self.models = self.hacModel.get_models(
clusters,
annotation=self.annotation, feature=feature
)
# cluster similarity matrix
self.matrix = self.hacModel.get_similarity_matrix(
clusters, models=self.models,
annotation=self.annotation, feature=feature)
# make sure diagonals are set to -np.inf
# -np.inf means "do not merge"
for c in clusters:
self.matrix[c, c] = -np.inf
# TODO: initialize constraints
# self.hacConstraint.initialize(
# annotation=self.annotation, models=self.models,
# matrix=self.matrix, history=self.history, feature=feature)
# initialize stopping criterion
self.hacStop.initialize(
annotation=self.annotation, models=self.models,
matrix=self.matrix, history=self.history, feature=feature)
class HierarchicalAgglomerativeClustering(object):
"""
Parameters
----------
model : HACModel
Model
stop : HACStop, optional
Stopping criterion
constraint : HACConstraint, optional
Constraint (not yet implemented)
debug : bool, optional
"""
def __init__(self, model, stop=None, constraint=None, debug=False):
super(HierarchicalAgglomerativeClustering, self).__init__()
assert isinstance(model, HACModel)
self.hacModel = model
assert isinstance(stop, HACStop)
self.hacStop = stop
# assert isinstance(constraint, HACConstraint)
# self.hacConstraint = constraint
self.debug = debug
def initialize(self, annotation, feature=None):
"""Initialize HAC with one cluster per label
Parameters
----------
annotation : Annotation
feature : Feature, optional
"""
# initialize annotation
self.annotation = annotation.copy()
# initialize history with original annotation
self.history = HACHistory(self.annotation)
# one cluster per label
clusters = self.annotation.labels()
# one model per cluster
self.models = self.hacModel.get_models(
clusters,
annotation=self.annotation, feature=feature
)
# cluster similarity matrix
self.matrix = self.hacModel.get_similarity_matrix(
clusters, models=self.models,
annotation=self.annotation, feature=feature)
# make sure diagonals are set to -np.inf
# -np.inf means "do not merge"
for c in clusters:
self.matrix[c, c] = -np.inf
# TODO: initialize constraints
# self.hacConstraint.initialize(
# annotation=self.annotation, models=self.models,
# matrix=self.matrix, history=self.history, feature=feature)
# initialize stopping criterion
self.hacStop.initialize(
annotation=self.annotation, models=self.models,
matrix=self.matrix, history=self.history, feature=feature)
def iterate(self, feature=None):
while True:
if len(self.models) <= 1:
break
# This second loop does not make sense for now.
# But it will, when we support constrained clustering in the future
while True:
# find two most similar clusters
# TODO: make this block overridable
# (e.g. one might want to merge more than 2 clusters at
# each iteration)
cluster1, cluster2 = self.matrix.argmax().popitem()
similarity = self.matrix[cluster1, cluster2]
if self.debug:
msg = (
"DEBUG > Next merging candidates "
"are %s and %s (s = %g).\n"
)
sys.stderr.write(msg % (cluster1, cluster2, similarity))
# if the best we can do is find clusters with -inf similarity,
# then stop here
if similarity == -np.inf:
break
# TODO: constrained clustering
# if mergeable(cluster1, cluster2)
# break
# self.matrix[cluster1, cluster2] = -np.inf
# self.matrix[cluster2, cluster1] = -np.inf
# if self.debug:
# msg = "DEBUG > Constraints prevented merging of %s and %s.\n"
# sys.stderr.write(msg % (cluster1, cluster2))
break
if similarity == -np.inf:
if self.debug:
msg = "DEBUG > Nothing left to merge.\n"
sys.stderr.write(msg)
break
# == update models
# (cluster1+cluster2 ==> cluster1)
self.models[cluster1] = self.hacModel.merge_models(
[cluster1, cluster2], annotation=self.annotation,
feature=feature, models=self.models,
matrix=self.matrix, history=self.history
)
# remove (now meaningless) cluster2's model
del self.models[cluster2]
# == update annotation (rename cluster2 into cluster1)
self.annotation = self.annotation % {cluster2: cluster1}
# == update history (keep track of this iteration)
self.history.add_iteration(
[cluster1, cluster2], similarity, cluster1)
# == update similarity matrix
# remove (now meaningless) cluster2's row and column
self.matrix.remove_row(cluster2)
self.matrix.remove_column(cluster2)
# update cluster1's row and column
for cluster in self.models:
if cluster == cluster1:
continue
# update matrix[cluster1, cluster]
s = self.hacModel.get_similarity(
cluster1, cluster, annotation=self.annotation,
models=self.models, matrix=self.matrix,
history=self.history, feature=feature
)
self.matrix[cluster1, cluster] = s
# update matrix[cluster, cluster1]
if not self.hacModel.is_symmetric():
s = self.hacModel.get_similarity(
cluster, cluster1, annotation=self.annotation,
models=self.models, matrix=self.matrix,
history=self.history, feature=feature
)
self.matrix[cluster, cluster1] = s
# TODO:
# == update constraints
# == update stopping criterion
# (most of the time, this does nothing)
self.hacStop.update(
[cluster1, cluster2], cluster1,
history=self.history, annotation=self.annotation,
models=self.models, matrix=self.matrix, feature=feature
)
# check if stopping criterion is reached
# and, if so, stop agglomerating...
if self.hacStop.reached(
history=self.history,
annotation=self.annotation, models=self.models,
matrix=self.matrix, feature=feature
):
if self.debug:
msg = "DEBUG > Reached stopping criterion.\n"
sys.stderr.write(msg)
break
yield self.annotation
def finalize(self, feature=None):
self.annotation = self.hacStop.finalize(
history=self.history, annotation=self.annotation,
models=self.models, matrix=self.matrix, feature=feature
)
return self.annotation
def __call__(self, annotation, feature=None):
"""
Parameters
----------
annotation : Annotation
feature : Feature, optional
"""
self.initialize(annotation, feature=feature)
for _ in self.iterate(feature=feature):
pass
return self.finalize(feature=feature)
