def _get_metric_fct(self):
if self.metric == 'Mutual_information':
metric_fct = lambda TP, FP, TN, FN: BinaryClassificationDataset.mutual_information(
TN, FN, TP, FP)
elif self.metric == 'Normalized_mutual_information':
metric_fct = lambda TP, FP, TN, FN: BinaryClassificationDataset.normalized_mutual_information(
TN, FN, TP, FP)
elif self.metric == "Balanced_accuracy":
metric_fct = self.balanced_accuracy
elif '{TP}' in self.metric or '{FP}' in self.metric or '{TN}' in self.metric or '{FN}' in self.metric:
metric_fct = lambda TP, FP, TN, FN: eval(
self.metric.format(
TP=float(TP), FP=float(FP), TN=float(TN), FN=float(FN)))
elif self.inverse_metric:
metric_fct = lambda TP, FP, TN, FN: \
(-1.0)*BinaryClassificationDataset.calculate_confusion_metrics(
{"True_negatives": TN,
"True_positives": TP,
"False_positives": FP,
"False_negatives": FN},
weight=self.weight,)[self.metric]
else:
metric_fct = lambda TP, FP, TN, FN: \
BinaryClassificationDataset.calculate_confusion_metrics(
{"True_negatives": TN,
"True_positives": TP,
"False_positives": FP,
"False_negatives": FN},
weight=self.weight,)[self.metric]
return metric_fct
def _get_metric_fct(self):
if self.metric == 'Mutual_information':
metric_fct = lambda TP, FP, TN, FN: ClassificationCollection.mutual_information(TN, FN, TP, FP)
elif self.metric == 'Normalized_mutual_information':
metric_fct = lambda TP, FP, TN, FN: ClassificationCollection.normalized_mutual_information(TN, FN, TP, FP)
elif self.metric == "Balanced_accuracy":
metric_fct = self.balanced_accuracy
elif '{TP}' in self.metric or '{FP}' in self.metric or '{TN}' in self.metric or '{FN}' in self.metric:
metric_fct = lambda TP, FP, TN, FN: eval(self.metric.format(TP=float(TP),
FP=float(FP),
TN=float(TN),
FN=float(FN)))
elif self.inverse_metric:
metric_fct = lambda TP, FP, TN, FN: \
(-1.0)*ClassificationCollection.calculate_confusion_metrics(
{"True_negatives":TN,
"True_positives":TP,
"False_positives":FP,
"False_negatives":FN},
weight=self.weight,
)[self.metric]
else:
metric_fct = lambda TP, FP, TN, FN: \
ClassificationCollection.calculate_confusion_metrics(
{"True_negatives":TN,
"True_positives":TP,
"False_positives":FP,
"False_negatives":FN},
weight=self.weight,
)[self.metric]
return metric_fct
def __init__(self, erp_class_label, retained_channels=None):
super(SSNRSinkNode, self).__init__()
# We reuse the ClassificationCollection (maybe this should be renamed to
# MetricsDataset?)
self.set_permanent_attributes(# Object for handling SSNR related calculations
ssnr=SSNR(erp_class_label, retained_channels),
# Result collection
ssnr_collection=BinaryClassificationDataset(),
erp_class_label=erp_class_label)
def looCV(self):
""" Calculate leave one out metrics """
# remember original solution
optimal_w = copy.deepcopy(self.w)
optimal_b = copy.deepcopy(self.b)
optimal_dual_solution = copy.deepcopy(self.dual_solution)
# preparation of sorting
sort_dual = self.dual_solution
# sort indices --> zero weights do not need any changing and
# low weights are less relevant for changes
sorted_indices = map(list, [numpy.argsort(sort_dual)])[0]
sorted_indices.reverse()
prediction_vectors = []
using_initial_solution = True
for index in sorted_indices:
d_i = self.dual_solution[index]
# delete each index from the current observation
if d_i == 0 and using_initial_solution:
# no change in classifier necessary
pass
else:
# set weight to zero and track the corresponding changes
self.reduce_dual_weight(index)
# reiterate till convergence but skip current index
temp_iter = self.iterations
self.iteration_loop(self.M, reduced_indices=[index])
self.iterations += temp_iter
using_initial_solution = False
prediction_vectors.append((
self._execute(numpy.atleast_2d(self.samples[index])),
self.classes[self.labels[index]]))
self.loo_metrics = BinaryClassificationDataset.calculate_metrics(
prediction_vectors,
ir_class=self.classes[1],
sec_class=self.classes[0])
#undo changes
self.b = optimal_b
self.w = optimal_w
self.dual_solution = optimal_dual_solution
def looCV(self):
""" Calculate leave one out metrics """
# remember original solution
optimal_w = copy.deepcopy(self.w)
optimal_b = copy.deepcopy(self.b)
optimal_dual_solution = copy.deepcopy(self.dual_solution)
# preparation of sorting
sort_dual = self.dual_solution
# sort indices --> zero weights do not need any changing and
# low weights are less relevant for changes
sorted_indices = map(list, [numpy.argsort(sort_dual)])[0]
sorted_indices.reverse()
prediction_vectors = []
using_initial_solution = True
for index in sorted_indices:
d_i = self.dual_solution[index]
# delete each index from the current observation
if d_i == 0 and using_initial_solution:
# no change in classifier necessary
pass
else:
# set weight to zero and track the corresponding changes
self.reduce_dual_weight(index)
# reiterate till convergence but skip current index
temp_iter = self.iterations
self.iteration_loop(self.M, reduced_indices=[index])
self.iterations += temp_iter
using_initial_solution = False
prediction_vectors.append(
(self._execute(numpy.atleast_2d(self.samples[index])),
self.classes[self.labels[index]]))
self.loo_metrics = BinaryClassificationDataset.calculate_metrics(
prediction_vectors,
ir_class=self.classes[1],
sec_class=self.classes[0])
#undo changes
self.b = optimal_b
self.w = optimal_w
self.dual_solution = optimal_dual_solution
