def ruleOfThumb(n=10**5, dim=2, C=1, seed=371986):
rng = np.random.RandomState(seed)
Cov = rng.normal(0, 1, (dim, dim))
XL, yL = data_gen.dataset_fixed_cov(n, dim, seed, Cov)
rng = np.random.RandomState(seed)
min_error = sys.maxsize
permutation = rng.permutation(len(ＸL))
#print(X , y)
XL, yL = XL[permutation], yL[permutation]
XL1 = XL
yL1 = yL
n_features = XL1.shape[1]
#print(n_features)
# split 50%-50% training sets and test sets
scaler = preprocessing.StandardScaler()
XL1 = scaler.fit_transform(XL1, yL1)
#test_X1 = scaler.fit_transform(test_X1,test_y1)
train_XL1, test_XL1, train_yL1, test_yL1 = train_test_split(
XL1, yL1, train_size=0.5, random_state=seed)
svm = my_svm(iter=100)
start_time = time.time()
w, obj, step = svm.my_gradient_descent(train_XL1,
train_yL1,
stop=stop.perfor,
C=C)
print("---Cost %s seconds for GD ---" % (time.time() - start_time))
error_GD = svm.predict(test_XL1, test_yL1, w=w[len(w) - 1])
svm = my_svm(iter=100)
start_time = time.time()
w, obj, sp = svm.my_sgd(train_XL1,
train_yL1,
seed=seed,
stop=stop.perfor,
step=0.11,
t0=2,
C=C)
print("---Cost %s seconds for SGD ---" % (time.time() - start_time))
error_SGD = svm.predict(test_XL1, test_yL1, w=w[len(w) - 1])
if (error_GD < error_SGD):
print(
"Use GD to run the data number = %f and dimention = %d , and C = %d , with error rate = %f "
% (n, dim, C, error_GD))
else:
print(
"Use SGD to run the data number = %f and dimention = %d , and C = %d , with error rate = %f"
% (n, dim, C, error_SGD))
def run(centroid , group_num , train_X , train_y ,test_X , test_y, method='knn' , n_nb = 2 , seed = 371986):
# will load data as the patch size defined , 3 means 3*3 = 9 for each patch, and will return the dictionary included:
# 'data' (one patch) , 'target' (the sample of this patch belongs to ) , 'filename' (the file comes from)
bofs = []
vlad = my_vlad(centroid[group_num])
for file in train_X:
mydata,y = myData.load_sig_data(file , 3)
bofs.append(vlad.get_vlad(mydata['data']).flatten() )
#knn_init = KNeighborsClassifier()
#parameters = {'n_neighbors':[ 5, 10 , 15]}
#knn = grid_search.GridSearchCV(knn_init, parameters)
bofs_test = []
for file in test_X:
mydata,y = myData.load_sig_data(file , 3)
bofs_test.append(vlad.get_vlad(mydata['data']).flatten() )
if(method == "knn"):
knn = KNeighborsClassifier(n_neighbors = n_nb)
knn.fit(bofs, train_y)
predicted = knn.predict(bofs_test)
score = knn.score(bofs_test,test_y)
if(method == "LinearSVM"):
svm = my_svm(iter = 100 )
w , obj , sp = svm.my_sgd(bofs,train_y, seed = seed , stop = stop.perfor , step = 0.11, t0 = 2, C = 1 )
score = 100-svm.predict(bofs_test,test_y)
return score
def ruleOfThumb(n = 10**5 , dim = 2 , C = 1 , seed = 371986 ):
rng = np.random.RandomState(seed)
Cov = rng.normal(0, 1, (dim, dim))
XL, yL = data_gen.dataset_fixed_cov(n, dim, seed , Cov)
rng = np.random.RandomState(seed)
min_error = sys.maxsize
permutation = rng.permutation(len(ＸL))
#print(X , y)
XL, yL = XL[permutation], yL[permutation]
XL1 = XL
yL1 = yL
n_features = XL1.shape[1]
#print(n_features)
# split 50%-50% training sets and test sets
scaler = preprocessing.StandardScaler()
XL1 = scaler.fit_transform(XL1,yL1)
#test_X1 = scaler.fit_transform(test_X1,test_y1)
train_XL1, test_XL1, train_yL1, test_yL1 = train_test_split(XL1, yL1, train_size=0.5, random_state=seed)
svm = my_svm(iter = 100 )
start_time = time.time()
w , obj , step= svm.my_gradient_descent(train_XL1,train_yL1, stop = stop.perfor , C = C )
print("---Cost %s seconds for GD ---" % (time.time() - start_time))
error_GD = svm.predict(test_XL1,test_yL1,w=w[len(w)-1])
svm = my_svm(iter = 100 )
start_time = time.time()
w , obj , sp = svm.my_sgd(train_XL1,train_yL1, seed = seed , stop = stop.perfor , step = 0.11, t0 = 2, C = C )
print("---Cost %s seconds for SGD ---" % (time.time() - start_time))
error_SGD = svm.predict(test_XL1,test_yL1,w=w[len(w)-1])
if(error_GD < error_SGD):
print("Use GD to run the data number = %f and dimention = %d , and C = %d , with error rate = %f " %(n , dim , C , error_GD))
else :
print("Use SGD to run the data number = %f and dimention = %d , and C = %d , with error rate = %f" %(n , dim , C , error_SGD))
def run(centroid,
group_num,
train_X,
train_y,
test_X,
test_y,
method='knn',
n_nb=2,
seed=371986):
# will load data as the patch size defined , 3 means 3*3 = 9 for each patch, and will return the dictionary included:
# 'data' (one patch) , 'target' (the sample of this patch belongs to ) , 'filename' (the file comes from)
bofs = []
vlad = my_vlad(centroid[group_num])
for file in train_X:
mydata, y = myData.load_sig_data(file, 3)
bofs.append(vlad.get_vlad(mydata['data']).flatten())
#knn_init = KNeighborsClassifier()
#parameters = {'n_neighbors':[ 5, 10 , 15]}
#knn = grid_search.GridSearchCV(knn_init, parameters)
bofs_test = []
for file in test_X:
mydata, y = myData.load_sig_data(file, 3)
bofs_test.append(vlad.get_vlad(mydata['data']).flatten())
if (method == "knn"):
knn = KNeighborsClassifier(n_neighbors=n_nb)
knn.fit(bofs, train_y)
predicted = knn.predict(bofs_test)
score = knn.score(bofs_test, test_y)
if (method == "LinearSVM"):
svm = my_svm(iter=100)
w, obj, sp = svm.my_sgd(bofs,
train_y,
seed=seed,
stop=stop.perfor,
step=0.11,
t0=2,
C=1)
score = 100 - svm.predict(bofs_test, test_y)
return score