def find_threshold(self, train, val, seed=None):
if seed:
random.seed(seed)
ratio = np.arange(0.2, 0.7, 0.05)
# Train
training_set = pc.join_list_df(train)
X_train = [x[:-1] for x in training_set]
y_train = [x[-1] for x in training_set]
if self.class_weight:
svc = SVC(kernel='linear',
random_state=SEED,
probability=True,
class_weight='balanced')
else:
svc = SVC(kernel='linear',
random_state=SEED,
probability=True,
class_weight=None)
svc.fit(X_train, y_train)
ratios_results = []
val_set = pc.join_list_df(val)
X_val = [x[:-1] for x in val_set]
y_val = [x[-1] for x in val_set]
border_amount = len(y_val) * 0.3
for r in ratio:
self.threshold = r
first_layer = self.first_layer_classifier(X_val, svc)
prediction = self.second_layer_classifier(X_val, first_layer, svc)
cm_d = np.sum(confusion_matrix(y_val, prediction).diagonal()) * 1.0
if cm_d / len(val) > border_amount:
ratios_results.append(accuracy_score(y_val, prediction))
else:
ratios_results.append(0.0)
return ratio[np.argmax(ratios_results)]
def train(self,training_set_list):
columns = training_set_list[0].columns
training_set = pd.DataFrame(pc.join_list_df(training_set_list))
training_set.columns = columns
if self.p1 <= -1:
sorted_phases = get_sorted_phases(training_set)
self.p1 = sorted_phases[-2]+1
self.p2 = sorted_phases[-1]+1
self.threshold = find_threshold(training_set_list,self.p1,self.p2)
def find_threshold(self, train, val, seed=None):
if seed:
random.seed(seed)
ratio = np.arange(0.0, 1.0, 0.1)
# Train
training_set = pc.join_list_df(train)
X_train = [x[:-1] for x in training_set]
y_train = [x[-1] for x in training_set]
class_weight = get_proportion(y_train)
if self.class_weight:
svc = SVC(kernel='linear',
random_state=SEED,
probability=True,
class_weight=class_weight)
else:
svc = SVC(kernel='linear',
random_state=SEED,
probability=True,
class_weight=None)
svc.fit(X_train, y_train)
ratios_results = []
val_set = pc.join_list_df(val)
X_val = [x[:-1] for x in val_set]
y_val = [x[-1] for x in val_set]
border_amount = len(y_val) * 0.3
_, coeff = pc.combination_before_after(X_val, y_val, svc)
self.coeff = coeff
for r in ratio:
self.threshold = r
first_layer = self.first_layer_classifier(X_val, svc)
second_layer = pc.second_layer_combination_test(X_val, coeff, svc)
prediction = []
for i, element in enumerate(first_layer):
if element == -1:
prediction.append(second_layer[i])
else:
prediction.append(first_layer[i])
cm_d = np.sum(confusion_matrix(y_val, prediction).diagonal()) * 1.0
if cm_d / len(val) > border_amount:
ratios_results.append(accuracy_score(y_val, prediction))
else:
ratios_results.append(0.0)
return ratio[np.argmax(ratios_results)]
def train(self, training_set_list):
train, val = pc.split_df_list(training_set_list, 0.2, SEED)
training_set = pc.join_list_df(train)
# find threshold to split using train and val set
self.t = self.find_threshold(train, val, SEED)
# train svm with all the data
training_set = pc.join_list_df(training_set_list)
X_train = [x[:-1] for x in training_set]
y_train = [x[-1] for x in training_set]
if self.class_weight:
svc = SVC(kernel='linear',
random_state=SEED,
probability=True,
class_weight='balanced')
else:
svc = SVC(kernel='linear',
random_state=SEED,
probability=True,
class_weight=None)
svc.fit(X_train, y_train)
self.svm = svc
def train(self, training_set_list):
training_set = pc.join_list_df(training_set_list)
X_train = [x[:-1] for x in training_set]
y_train = [x[-1] for x in training_set]
class_weight = get_proportion(y_train)
if self.class_weight:
svc = SVC(kernel='linear',
random_state=SEED,
probability=True,
class_weight=class_weight)
else:
svc = SVC(kernel='linear',
random_state=SEED,
probability=True,
class_weight=None)
svc.fit(X_train, y_train)
self.svm = svc
def train(self,training_set_list):
columns = training_set_list[0].columns
training_set = pd.DataFrame(pc.join_list_df(training_set_list))
training_set.columns = columns
self.biggest_phase = get_sorted_phases(training_set)[-1]+1