def trapezoidalExperimentOLSF(input_dataset, dataset_name):
print("Trapezoidal experiment with OLSF: " + str(dataset_name))
error_vector = np.zeros(len(input_dataset))
feature_summary = [
len(row) for row in preprocess2.removeDataTrapezoidal(
copy.deepcopy(input_dataset))
]
for i in range(parameters.rounds):
print("Round: " + str(i))
random.seed(parameters.seed)
random.shuffle(input_dataset)
current_dataset = preprocess2.removeDataTrapezoidal(
copy.deepcopy(input_dataset))
current_classifier = olsf.classifier(current_dataset, [])
classifier_summary, stream_error = current_classifier.train()
error_vector = np.add(error_vector, stream_error)
average_error_vector = np.divide(error_vector, parameters.rounds)
#misc.plotError(average_error_vector, dataset_name)
#misc.plotFeatures(feature_summary, dataset_name)
#misc.plotClassifierDimension(classifier_summary, dataset_name)
print(current_classifier.weight_dict)
return average_error_vector
def preprocessData(data, mode='variable'):
random.seed(50)
copydata= copy.deepcopy(data)
random.shuffle(copydata)
if mode=='trapezoidal': dataset=preprocess2.removeDataTrapezoidal(copydata)
if mode=='variable': dataset=preprocess2.removeRandomData(copydata)
all_keys = set().union(*(d.keys() for d in dataset))
X,y = [],[]
for row in dataset:
for key in all_keys:
if key not in row.keys() : row[key]=0
y.append(row['class_label'])
del row['class_label']
if 0 not in row.keys(): start=1
if 0 in row.keys(): start=0
for row in dataset:
X_row=[]
for i in range(start, len(row)):
X_row.append(row[i])
X.append(X_row)
return X,y
def train(self, training_set, dataset_name,
mode): #uses self.training_dataset
self.data = training_set
if mode == 'ASGD':
self.classifier = SGDClassifier(average=True,
max_iter=1,
penalty='elasticnet',
l1_ratio=0.5)
self.classifier2 = SGDClassifier(average=False,
max_iter=1,
penalty='elasticnet',
l1_ratio=0.5)
elif mode == 'Perceptron':
self.classifier = Perceptron(max_iter=1)
self.classifier2 = self.classifiers
elif mode == 'PA1':
self.classifier = PassiveAggressiveClassifier(loss='hinge',
C=-1.0,
max_iter=1)
self.classifier2 = self.classifier
elif mode == 'PA2':
self.classifier = PassiveAggressiveClassifier(loss='squared_hinge',
C=1.0,
max_iter=1)
self.classifier2 = self.classifier
init = np.zeros(len(self.data[0]) - 1).reshape(1, -1)
for i in range(0, parameters.rounds):
train_error_vector = []
iterations = 0
train_error = 0
copydata = copy.deepcopy(self.data)
random.shuffle(copydata)
self.data_preprocessor(
preprocess2.removeDataTrapezoidal(copydata)) #or trapezoidal
self.classifier = clone(self.classifier)
self.classifier2 = clone(self.classifier2)
self.classifier.partial_fit(init, [-self.y[0]], np.unique(self.y))
self.classifier2.partial_fit(init, [-self.y[0]], np.unique(self.y))
total_error_vector = np.zeros(len(self.y))
#c = list(zip(self.X, self.y))
#random.shuffle(c)
#self.X, self.y= zip(*c)
self.variance_vector = [np.ones(len(training_set[0]) - 1)]
self.average_vector = [np.zeros(len(training_set[0]) - 1)]
for i in range(0, len(self.y)):
#self.classifier.densify()
row = [self.X[i]]
label = self.y[i]
iterations = i + 1
old = self.classifier.coef_
self.classifier.coef_ = self.update_metadata(iterations)
result = self.classifier.predict(row)
self.classifier.coef_ = old
if result[0] != label:
train_error += 1
self.classifier.partial_fit(row, [self.y[i]],
np.unique(self.y))
self.classifier2.partial_fit(row, [self.y[i]],
np.unique(self.y))
#self.classifier.sparsify()
train_error_vector.append(train_error / iterations)
total_error_vector = np.add(train_error_vector, total_error_vector)
total_error_vector = np.divide(total_error_vector, parameters.rounds)
misc.plotError(train_error_vector[0::50], dataset_name)
return train_error_vector