def testSingle(self, original_row):
row = original_row.copy()
label = row.pop('class_label', None) #extract the label
commonKeys = misc.findCommonKeys(
row, self.mean_dict) #do the classifications based on shared keys
row_subset = misc.subsetDictionary(row, commonKeys)
mean_subset = misc.subsetDictionary(self.mean_dict, commonKeys)
row_vector = misc.dictToNumpyArray(row_subset)
mean_vector = misc.dictToNumpyArray(mean_subset)
covariance_vector = misc.dictToNumpyArray(
misc.subsetDictionary(
self.covariance_dict,
commonKeys)) #this is for alternative classification style
if len(covariance_vector) == 0:
return False
if np.average(covariance_vector) > np.average(
list(self.covariance_dict.values())) * 2.2:
return np.random.choice([True, False])
if np.sum(mean_vector) != 0:
#confidences= np.divide(np.reciprocal(covariance_vector), np.sum(np.reciprocal(covariance_vector)))
mean_vector = np.multiply(
np.divide(mean_vector, np.sum(np.abs(mean_vector))),
np.sum(np.abs(list(self.mean_dict.values()))))
#mean_vector= [a*b for a,b in zip(mean_vector,confidences)]
return self.algorithm_methods.classify(
row_vector, mean_vector, covariance_vector,
label) #check if it can classify correctly
def sparsify(self):
with warnings.catch_warnings(record=True) as w:
if parameters.sparsity_switch == 0: #dont touch
return
lamb = parameters.sparsity_regularizer
trunc = parameters.sparsity_B
mean_vector = misc.dictToNumpyArray(self.mean_dict)
covariance_vector = misc.dictToNumpyArray(self.covariance_dict)
value_vector = covariance_vector
if w:
print("mean vector: " + str(min(mean_vector)))
print("covariance vector: " + str(min(covariance_vector)))
warnings.simplefilter("error")
sys.exit()
num_elem_remove = int(len(value_vector) * (1 - trunc))
if (np.linalg.norm(value_vector, ord=1) >
trunc * len(value_vector)):
coefficient = np.minimum(
1, lamb / np.linalg.norm(value_vector, ord=1))
copy_mean_dict = self.mean_dict.copy()
#multiply by coefficient
for key, value in copy_mean_dict.items():
copy_mean_dict[key] = value * coefficient
#print("remove elements: "+str(num_elem_remove))
for i in range(0, num_elem_remove):
key_to_delete = max(copy_mean_dict,
key=lambda k: copy_mean_dict[k])
copy_mean_dict.pop(key_to_delete)
self.mean_dict.pop(key_to_delete)
self.covariance_dict.pop(key_to_delete)
def updateWeights(self, row, weights, label):
row_vec = misc.dictToNumpyArray(row)
weights_vec = misc.dictToNumpyArray(weights)
self.classifier.coef_ = np.array([weights_vec])
self.classifier.partial_fit([row_vec], [label], np.unique(self.y))
self.weight_dict = self.repackDict(self.classifier.coef_[0],
list(weights.keys()))
def testSingle(self, original_row):
row=original_row.copy()
label=row.pop('class_label', None) #extract the label
commonKeys=misc.findCommonKeys(row, self.weight_dict) #do the classifications based on shared keys
row_subset=misc.subsetDictionary(row, commonKeys)
weight_subset=misc.subsetDictionary(self.weight_dict, commonKeys)
row_vector=misc.dictToNumpyArray(row_subset)
weight_vector=misc.dictToNumpyArray(weight_subset)
loss=np.maximum(0, 1-label*(weight_vector.dot(row_vector)))
product=label*(weight_vector.dot(row_vector))
return loss, product #check if it can classify correctly
def testSingle(self, original_row):
row = original_row.copy()
label = row.pop('class_label', None) #extract the label
commonKeys = misc.findCommonKeys(
row, self.mean_dict) #do the classifications based on shared keys
row_subset = misc.subsetDictionary(row, commonKeys)
mean_subset = misc.subsetDictionary(self.mean_dict, commonKeys)
row_vector = misc.dictToNumpyArray(row_subset)
mean_vector = misc.dictToNumpyArray(mean_subset)
covariance_vector = misc.dictToNumpyArray(
misc.subsetDictionary(
self.covariance_dict,
commonKeys)) #this is for alternative classification style
return self.classify(row_vector, mean_vector, covariance_vector,
label) #check if it can classify correctly
def learnNew(self, new_attribute_dict, label):
#mean_new=example_new/(label*example_new*example_new)
new_attribute_vector = misc.dictToNumpyArray(new_attribute_dict)
new_partial_mean, new_covariance_dict = self.updateNewPart(
new_attribute_vector, label, new_attribute_dict)
new_mean_dict = misc.numpyArrayToDict(new_partial_mean,
list(new_attribute_dict.keys()))
return new_mean_dict, new_covariance_dict
def setParameter(self, loss, row):
row= misc.dictToNumpyArray(row)
if self.selector==0: #OLSF
return loss/((np.linalg.norm(row))*((np.linalg.norm(row))))
elif self.selector==1: #OLSF1
return np.minimum(self.C, loss/((np.linalg.norm(row))*((np.linalg.norm(row)))))
else: #OLSF2
return loss/((np.linalg.norm(row)*np.linalg.norm(row))+(1/(2*self.C)))
def learnCommon(self, mean_dict, covariance_dict, row_dict, label,
full_covariance_dict):
#order of the values in two dictionaries are the same
#transform dicts to matrix, vector for math operations
full_cov_vector = misc.dictToNumpyArray(full_covariance_dict)
mean_vector = misc.dictToNumpyArray(mean_dict)
row_vector = misc.dictToNumpyArray(row_dict)
covariance_matrix = misc.dictToUnitNumpyMatrix(covariance_dict)
if self.classify_train(row_vector, mean_vector, label):
return mean_dict, covariance_dict, 1 #large margin classified
else:
#recalculate mean and covariance
mean_vector, covariance_matrix = self.update(
mean_vector, covariance_matrix, row_vector, label,
full_cov_vector)
#transform everything back to labeled dictionary
new_mean_dict = misc.numpyArrayToDict(mean_vector,
list(mean_dict.keys()))
new_covariance_dict = misc.numpyMatrixToDict(
covariance_matrix, list(mean_dict.keys()))
return new_mean_dict, new_covariance_dict, 0
def sparsity_step(self):
weight_vector=misc.dictToNumpyArray(self.weight_dict)
#l1 projection
new_weight_vector= np.multiply(np.minimum(1, self.param_lambda/ np.linalg.norm(weight_vector, ord=1)), weight_vector)
#truncation
return_dict=self.weight_dict
if np.linalg.norm(new_weight_vector, ord=0) >= self.B*len(new_weight_vector):
num_elem_remove=int(len(new_weight_vector)- np.maximum(1, np.floor(self.B* len(new_weight_vector))))
copy_dict= copy.deepcopy(return_dict)
sorted_keys= sorted(copy_dict, key=lambda dict_key: abs(copy_dict[dict_key]))
for key in sorted_keys[:num_elem_remove]:
del return_dict[key]
return return_dict
def test(self): #returns average test accuracy
counter = 0 #correct counter
for original_row in self.test_dataset:
row = original_row.copy()
label = row.pop('class_label', None) #extract the label
commonKeys = misc.findCommonKeys(
row,
self.mean_dict) #do the classifications based on shared keys
row_subset = misc.subsetDictionary(row, commonKeys)
mean_subset = misc.subsetDictionary(self.mean_dict, commonKeys)
row_vector = misc.dictToNumpyArray(row_subset)
mean_vector = misc.dictToNumpyArray(mean_subset)
covariance_vector = misc.dictToNumpyArray(
misc.subsetDictionary(
self.covariance_dict,
commonKeys)) #this is for alternative classification style
if self.algorithm_methods.classify(
row_vector, mean_vector, covariance_vector,
label) == False: #check if it can classify correctly
counter += 1
return counter / len(
self.test_dataset
) #return number of false classification over all examples
def learnCommon(self, mean_dict, covariance_dict, row_dict,
label): #return 1 if correctly classified
mean_vector = misc.dictToNumpyArray(mean_dict)
row_vector = misc.dictToNumpyArray(row_dict)
covariance_matrix = misc.dictToUnitNumpyMatrix(covariance_dict)
covsum = np.sum(covariance_matrix)
#
V = np.dot(np.dot(row_vector, covariance_matrix), row_vector)
phi = global_phi / (V / covsum)
M = (row_vector.dot(mean_vector)) * label
part1 = -(1 + 2 * phi * M)
#print("Part 1:"+str(M))
#print("row vector: "+str(row_vector))
#print("covariance matrix: "+ str(covariance_matrix))
#print("v"+str(V))
part2 = np.sqrt(np.square(part1) - 8 * phi * (M - phi * V))
#print("part2: "+str(part2))
part3 = 4 * phi * V
gamma = (part1 + part2) / part3
#print("gamma: "+str(gamma))
alpha = np.max(gamma, 0)
mean_vector = mean_vector + np.dot(
np.multiply(alpha * label, covariance_matrix), row_vector)
#print(mean_vector)
covariance_matrix = covariance_matrix + np.multiply(
2 * alpha * phi, np.diag(row_vector))
#transform everything back to labeled dictionary
new_mean_dict = misc.numpyArrayToDict(mean_vector,
list(mean_dict.keys()))
new_covariance_dict = misc.numpyMatrixToDict(covariance_matrix,
list(mean_dict.keys()))
return new_mean_dict, new_covariance_dict, 0
def sparsity_step(self):
weight_vector = misc.dictToNumpyArray(self.mean_dict)
#l1 projection
new_weight_vector = np.multiply(
np.minimum(
1, olsf.param_lambda / np.linalg.norm(weight_vector, ord=1)),
weight_vector)
#truncation
return_dict = self.mean_dict
if np.linalg.norm(new_weight_vector,
ord=0) >= olsf.B * len(new_weight_vector):
num_elem_remove = int(
len(new_weight_vector) -
np.maximum(1, np.floor(olsf.B * len(new_weight_vector))))
#print(len(return_dict))
return_dict = dict(
sorted(return_dict.items(),
key=operator.itemgetter(1),
reverse=True)[:num_elem_remove])
#print(len(return_dict))
return return_dict
def predictSingle(self, row, weights, label):
row_vec = misc.dictToNumpyArray(row)
weights_vec = misc.dictToNumpyArray(weights)
#print(weights_vec)
self.classifier.coef_ = np.array([weights_vec])
return label != self.classifier.predict([row_vec])
def learnNew(self, row_dict, label, tao):
row_vector=misc.dictToNumpyArray(row_dict)
new_weight_vector= np.multiply(tao*label, row_vector)
new_weight_dict=misc.numpyArrayToDict(new_weight_vector, list(row_dict.keys()))
return new_weight_dict
def learnCommon(self, weight_dict, row_dict, label, tao):
weight_vector=misc.dictToNumpyArray(weight_dict)
row_vector=misc.dictToNumpyArray(row_dict)
new_weight_vector= weight_vector+ np.multiply(tao*label, row_vector)
new_weight_dict=misc.numpyArrayToDict(new_weight_vector, list(weight_dict.keys()))
return new_weight_dict