def test_bank_svmguide1():
print("========== Test BaNK on svmguide1 dataset ==========")
data_name = 'svmguide1'
(x_train, y_train), (x_test, y_test) = demo.load_svmguide1()
learner = BaNK(gamma=1.0 / 1.25727,
rf_dim=384,
inner_regularization=0.125,
outer_regularization=0.125,
alpha=10.0,
kappa=0.1,
inner_max_loop=1,
max_outer_loop=5,
batch_size=5)
learner.fit(x_train, y_train)
y_train_pred = learner.predict(x_train)
y_test_pred = learner.predict(x_test)
print("Dataset: {}".format(data_name))
print("Training error = %.4f" %
(1 - metrics.accuracy_score(y_train, y_train_pred)))
print("Test error = %.4f" %
(1 - metrics.accuracy_score(y_test, y_test_pred)))
def test_svmguide1(show_figure=False, block_figure_on_end=False):
print("========== Test SVRG on svmguide1 dataset ==========")
data_name = 'svmguide1'
(x_train, y_train), (x_test, y_test) = demo.load_svmguide1()
loss_display = Display(
freq=1,
dpi='auto',
show=show_figure,
block_on_end=block_figure_on_end,
monitor=[{'metrics': ['train_loss', 'obj_func'],
'type': 'line',
'title': "Learning losses",
'xlabel': "X1",
'ylabel': "X2",
}]
)
print(x_train.shape)
learner = SVRG(
mode='online',
regular_param=0.01,
learning_rate_scale=0.7,
gamma=2.0,
rf_dim=400,
num_epochs=1,
cache_size=50,
freq_update_full_model=100,
oracle=SVRG.COVERAGE,
core_max=100,
coverage_radius=0.9,
loss_func=SVRG.LOGISTIC,
smooth_hinge_theta=0.5,
smooth_hinge_tau=0.5,
callbacks=[loss_display],
metrics=['train_loss', 'obj_func'],
freq_calc_metrics=300,
random_state=random_seed()
)
learner.fit(x_train, y_train)
y_train_pred = learner.predict(x_train)
y_test_pred = learner.predict(x_test)
print("Dataset: {}".format(data_name))
print("Training error = %.4f" % (1 - metrics.accuracy_score(y_train, y_train_pred)))
print("Test error = %.4f" % (1 - metrics.accuracy_score(y_test, y_test_pred)))
def test_sgd_svmguide1_bin():
(x_train, y_train), (x_test, y_test) = demo.load_svmguide1()
learner = KSGD(lbd=0.1,
eps=0.01,
gamma=2,
kernel='gaussian',
loss='hinge',
batch_size=10,
avg_weight=False,
random_state=random_seed())
learner.fit(x_train, y_train)
train_err = 1.0 - learner.score(x_train, y_train)
test_err = 1.0 - learner.score(x_test, y_test)
print("Training error = %.4f" % train_err)
print("Testing error = %.4f" % test_err)
def test_svmguide1():
print("========== Test IGKOL on svmguide1 dataset ==========")
data_name = 'svmguide1'
(x_train, y_train), (x_test, y_test) = demo.load_svmguide1()
display = Display(
freq=1,
dpi='auto',
block_on_end=True,
monitor=[{'metrics': ['mistake_rate'],
'type': 'line',
'title': "Mistake Rate",
'xlabel': "data points",
'ylabel': "Error",
}]
)
learner = OnlineDualSVRG(
regular_param=0.01,
learning_rate_scale=1.0,
gamma=2.0,
rf_dim=400,
cache_size=50,
freq_update_full_model=10,
oracle='budget',
core_max=100,
coverage_radius=0.5,
loss_func='hinge',
smooth_hinge_theta=0.5,
smooth_hinge_tau=0.5,
callbacks=[display],
metrics=['mistake_rate'],
random_state=random_seed())
learner.fit(x_train, y_train)
print('OFFLINE Testing')
y_train_pred = learner.predict(x_train)
y_test_pred = learner.predict(x_test)
print("Dataset: {}".format(data_name))
print("Training error = %.4f" % (1 - metrics.accuracy_score(y_train, y_train_pred)))
print("Test error = %.4f" % (1 - metrics.accuracy_score(y_test, y_test_pred)))
def run_svmguide1_gridsearch():
print("========== Tune parameters for SVRG for classification ==========")
data_name = 'svmguide1'
(x_total, y_total), (x_test, y_test) = demo.load_svmguide1()
n_total = x_total.shape[0]
percent = 0.8
n_train = int(percent * n_total)
idx_train = np.random.permutation(n_total)[:n_train]
mask = np.zeros(n_total, dtype=bool)
mask[idx_train] = True
x_train = x_total[mask, :]
x_valid = x_total[~mask, :]
y_train = y_total[mask]
y_valid = y_total[~mask]
print("Number of training samples = {}".format(x_train.shape[0]))
print("Number of testing samples = {}".format(x_valid.shape[0]))
x = np.vstack((x_train, x_valid))
y = np.concatenate((y_train, y_valid))
params = {'regular_param': [0.0001, 0.00001],
'gamma': [0.25, 0.5]}
ps = PredefinedSplit(test_fold=[-1] * x_train.shape[0] + [1] * x_valid.shape[0])
clf = SVRG(
mode='batch',
regular_param=0.01,
learning_rate_scale=0.8,
gamma=2.0,
rf_dim=400,
num_epochs=1,
cache_size=50,
freq_update_full_model=100,
oracle=SVRG.COVERAGE,
core_max=100,
coverage_radius=0.9,
loss_func=SVRG.LOGISTIC,
smooth_hinge_theta=0.5,
smooth_hinge_tau=0.5,
random_state=random_seed()
)
gs = GridSearchCV(clf, params, cv=ps, n_jobs=-1, refit=False, verbose=True)
gs.fit(x, y)
print("Best error {} @ params {}".format(1.0 - gs.best_score_, gs.best_params_))
# best param prediction
print("Best param prediction")
best_clf = clone(clf).set_params(**gs.best_params_)
best_clf.fit(x_total, y_total)
y_total_pred = best_clf.predict(x_total)
y_test_pred = best_clf.predict(x_test)
total_err = 1 - metrics.accuracy_score(y_total, y_total_pred)
test_err = 1 - metrics.accuracy_score(y_test, y_test_pred)
print("Training error = %.4f" % total_err)
print("Testing error = %.4f" % test_err)