def _find_separability(self, training_classes):
# create a matrix list and give them indexes
vectors = []
training_classes_with_idx = {}
idx = 0
for name, points in training_classes.items():
this_class = training_classes_with_idx[name] = []
for point in points:
# give it an index
vector = point.tolist()
vector_with_idx = [idx] + vector
idx += 1
vectors.append(vector)
this_class.append(vector_with_idx)
training_classes_with_idx[name] = numpy.array(this_class)
vectors = numpy.array(vectors)
kernel = self.make_gram_matrix(vectors, self.gamma)
self.tools = Tools(kernel)
# calculate all the sqRadiuses
if self.verbose:
start_time = time.process_time()
sq_radiuses = {}
for name, points in training_classes_with_idx.items():
sq_radiuses[name] = self.tools.squared_radius(name, points)
if self.verbose:
print('sq_radiuses: %.4f' % (time.process_time() - start_time))
# separability section
# use the precalculated squared radiuses from above
def find_separability(name_a, name_b):
sq_ra = sq_radiuses[name_a]
sq_rb = sq_radiuses[name_b]
sq_dist = self.tools.squared_distance(
name_a,
training_classes_with_idx[name_a],
name_b,
training_classes_with_idx[name_b],
)
return sq_dist / (sq_ra + sq_rb)
# create mapping function from labels to integers and vice ver
class_cnt = len(training_classes.keys())
label_to_int = {}
int_to_label = [None for i in range(class_cnt)]
for i, label in enumerate(training_classes.keys()):
label_to_int[label] = i
int_to_label[i] = label
# 2d matrix showing separability of each
if self.verbose:
start_time = time.process_time()
separability = numpy.empty((class_cnt, class_cnt))
separability.fill(float('inf'))
for i, a in enumerate(training_classes.keys()):
int_a = label_to_int[a]
# should be no separability with itself
separability[int_a][int_a] = 0
for b in list(training_classes.keys())[i + 1:]:
int_b = label_to_int[b]
separability[int_a][int_b] = separability[int_b][
int_a] = find_separability(a, b)
if self.verbose:
print('separability: %.4f' % (time.process_time() - start_time))
return separability, label_to_int, int_to_label
def __init__(self, kernel):
self.kernel = kernel
self.groups = {}
self.similarity = {}
self.tools = Tools(kernel)
def _find_separability(self, training_classes):
# create a matrix list and give them indexes
vectors = []
training_classes_with_idx = {}
idx = 0
for name, points in training_classes.items():
this_class = training_classes_with_idx[name] = []
for point in points:
# give it an index
vector = point.tolist()
vector_with_idx = [idx] + vector
idx += 1
vectors.append(vector)
this_class.append(vector_with_idx)
training_classes_with_idx[name] = numpy.array(this_class)
vectors = numpy.array(vectors)
kernel = self.make_gram_matrix(vectors, self.gamma)
self.tools = Tools(kernel)
# find radius of each class
if self.verbose:
start_time = time.process_time()
sq_radiuses = {}
for name, points in training_classes_with_idx.items():
sq_radiuses[name] = self.tools.squared_radius(name, points)
if self.verbose:
print('train: %.4f' % (time.process_time() - start_time))
def find_separability(a, b):
sq_ra = sq_radiuses[a]
sq_rb = sq_radiuses[b]
sq_dist = self.tools.squared_distance(
a,
training_classes_with_idx[a],
b,
training_classes_with_idx[b],
)
return sq_dist / (sq_ra + sq_rb)
# relabelling
class_cnt = len(training_classes.keys())
label_to_int = {}
int_to_label = [None for i in range(class_cnt)]
for i, label in enumerate(training_classes.keys()):
label_to_int[label] = i
int_to_label[i] = label
# find separability of each pair
# default value is very high separability
if self.verbose:
start_time = time.process_time()
separability = numpy.empty((class_cnt, class_cnt))
separability.fill(float('inf'))
for i, a in enumerate(training_classes.keys()):
int_a = label_to_int[a]
# should be no separability with itself
separability[int_a][int_a] = 0
for b in list(training_classes.keys())[i + 1:]:
int_b = label_to_int[b]
separability[int_a][int_b] = separability[int_b][int_a] = find_separability(a, b)
if self.verbose:
print('train: %.4f' % (time.process_time() - start_time))
return separability, label_to_int, int_to_label
