def with_cross_validation(train_data_x, train_data_y, regularization=False):
c = 0.0005 # if regularization else NO_REGULARIZATION
pen = 'l2' if regularization else 'none'
percentage = [0.2, 0.4, 0.6, 0.8,
1] # will learn with 20%, 40% ,.... 100% of the data
acc_train = []
acc_vald = []
sampels_num = [int(x * len(train_data_x) * (4 / 5)) for x in percentage]
for k in sampels_num:
clf = LogisticRegression(C=c, max_iter=20000,
penalty=pen) # creat LR model
result = cross_validate(clf,
train_data_x[:k],
train_data_y[:k],
cv=5,
scoring='accuracy',
return_train_score=True) # 5-cross-validation
acc_vald.append(
round(np.sum(result['test_score']) / len(result['test_score']), 3))
acc_train.append(
round(
np.sum(result['train_score']) / len(result['train_score']), 3))
plot_train_vald(
acc_train,
acc_vald,
sampels_num,
x_label="Training set size (samples)",
y_label="Mean Accuracy (%)",
title="Mean Accuracy as function of training set size, cv (5), C={0}".
format(get_reg_title(c)))
def without_cross_validation(train_data_x, train_data_y, regularization=False):
c = 10 # if regularization else NO_REGULARIZATION
pen = 'l2' if regularization else 'none'
training_set_size = int((4 / 5) * len(train_data_x))
vald_x = train_data_x[training_set_size + 1:]
vald_y = train_data_y[training_set_size + 1:]
train_data_x = train_data_x[:
training_set_size] # separate to validation set
train_data_y = train_data_y[:training_set_size]
percentage = [0.2, 0.4, 0.6, 0.8,
1] # will learn with 20%, 40% ,.... 100% of the data
acc_train = []
acc_vald = []
sampels_num = [int(x * len(train_data_x)) for x in percentage]
for k in sampels_num:
print(int(len(train_data_x) * k))
clf = LogisticRegression(C=c, max_iter=20000, penalty=pen)
clf.fit(train_data_x[:k], train_data_y[:k]) # learn the data
y_hat = clf.predict(vald_x[:int(k /
5)]) # predict on the validation set
acc_vald.append(
np.sum(y_hat == vald_y[:int(k / 5)]) / len(vald_y[:int(k / 5)]))
y_hat = clf.predict(train_data_x[:k])
acc_train.append(
np.sum(y_hat == train_data_y[:k]) / len(train_data_y[:k]))
plot_train_vald(
acc_train,
acc_vald,
sampels_num,
x_label="Training set size (samples)",
y_label="Accuracy (%)",
title=
"Accuracy as function of training set size, single validation set, C={}"
.format(get_reg_title(c)))
def with_cross_validation(train_data_x, train_data_y, regularization=False):
alpha = 0.015 if regularization else NO_REGULARIZATION
percentage = [0.2, 0.4, 0.6, 0.8,
1] # will learn with 20%, 40% ,.... 100% of the data
acc_train = []
acc_vald = []
sampels_num = [int(x * len(train_data_x) * (4 / 5)) for x in percentage]
for k in sampels_num:
# creat ann with one hidden layer (3 neurons)
clf = MLPClassifier(hidden_layer_sizes=(3, ),
solver='sgd',
activation='relu',
alpha=alpha,
max_iter=20000)
result = cross_validate(clf,
train_data_x[:k],
train_data_y[:k],
cv=5,
scoring='accuracy',
return_train_score=True) # 5-cross-validation
acc_vald.append(
round(np.sum(result['test_score']) / len(result['test_score']), 3))
acc_train.append(
round(
np.sum(result['train_score']) / len(result['train_score']), 3))
plot_train_vald(
acc_train,
acc_vald,
sampels_num,
x_label="Training set size (samples)",
y_label="Mean Accuracy (%)",
title=
"Mean Accuracy as function of training set size, cv (5), $\\alpha$={0}"
.format(alpha))
def with_cross_validation(train_data_x, train_data_y, regularization=False):
c = 1 if regularization else NO_REGULARIZATION
percentage = [0.2, 0.4, 0.6, 0.8,
1] # will learn with 20%, 40% ,.... 100% of the data
acc_train = []
acc_vald = []
samples_num = [int(x * len(train_data_x) * (4 / 5)) for x in percentage]
for k in samples_num:
clf = clf = svm.SVC(C=c,
kernel='poly',
degree=1,
gamma=1,
coef0=0,
max_iter=2000000) # creat svm model
result = cross_validate(clf,
train_data_x[:k],
train_data_y[:k],
cv=5,
scoring='accuracy',
return_train_score=True
) # using a part of the data every iteration.
acc_vald.append(
round(np.sum(result['test_score']) / len(result['test_score']), 3))
acc_train.append(
round(
np.sum(result['train_score']) / len(result['train_score']), 3))
plot_train_vald(
acc_train,
acc_vald,
samples_num,
x_label="Training set size (samples)",
y_label="Mean Accuracy (%)",
title="Mean Accuracy as function of training set size, cv (5), C={0}".
format(get_reg_title(c)))
def without_cross_validation(train_data_x, train_data_y, regularization=False):
c = 1 if regularization else NO_REGULARIZATION
training_set_size = int((4 / 5) * len(train_data_x))
vald_x = train_data_x[training_set_size + 1:] # separate to validation set
vald_y = train_data_y[training_set_size + 1:]
train_data_x = train_data_x[:training_set_size]
train_data_y = train_data_y[:training_set_size]
percentage = [0.2, 0.4, 0.6, 0.8,
1] # will learn with 20%, 40% ,.... 100% of the data
acc_train = []
acc_vald = []
sampels_num = [int(x * len(train_data_x)) for x in percentage]
for k in sampels_num:
clf = svm.SVC(C=c,
kernel='poly',
degree=1,
gamma=1,
coef0=0,
max_iter=2000000) # creat svm model
clf.fit(train_data_x[:k], train_data_y[:k]) # train the model
y_hat = clf.predict(vald_x[:int(k /
5)]) # predict of the validation set
acc_vald.append(
np.sum(y_hat == vald_y[:int(k / 5)]) / len(vald_y[:int(k / 5)]))
y_hat = clf.predict(train_data_x[:k]) # predict of the training set
acc_train.append(
np.sum(y_hat == train_data_y[:k]) / len(train_data_y[:k]))
plot_train_vald(
acc_train,
acc_vald,
sampels_num,
x_label="Training set size (samples)",
y_label="Accuracy (%)",
title=
"Accuracy as function of training set size, single validation set, C={}"
.format(get_reg_title(c)))
def without_cross_validation(train_data_x, train_data_y, regularization=False):
alpha = 0.015 if regularization else NO_REGULARIZATION
training_set_size = int((4 / 5) * len(train_data_x))
vald_x = train_data_x[training_set_size + 1:] # separate to validation set
vald_y = train_data_y[training_set_size + 1:]
train_data_x = train_data_x[:training_set_size]
train_data_y = train_data_y[:training_set_size]
percentage = [0.2, 0.4, 0.6, 0.8,
1] # will learn with 20%, 40% ,.... 100% of the data
acc_train = []
acc_vald = []
sampels_num = [int(x * len(train_data_x)) for x in percentage]
for k in sampels_num:
clf = MLPClassifier(hidden_layer_sizes=(3, ),
solver='sgd',
activation='relu',
alpha=alpha,
max_iter=20000)
clf.fit(train_data_x[:k], train_data_y[:k]) # trainin g the model
y_hat = clf.predict(vald_x[:int(k / 5)])
acc_vald.append(
np.sum(y_hat == vald_y[:int(k / 5)]) / len(vald_y[:int(k / 5)]))
y_hat = clf.predict(train_data_x[:k]) # predict on the validation set
acc_train.append(
np.sum(y_hat == train_data_y[:k]) / len(train_data_y[:k]))
plot_train_vald(
acc_train,
acc_vald,
sampels_num,
x_label="Training set size (samples)",
y_label="Accuracy (%)",
title=
"Accuracy as function of training set size, single validation set, $\\alpha$={0}"
.format(alpha))