def test_l1_regularized_saga():
beta = 1e-3
pysaga = PySAGAClassifier(eta=1e-3, alpha=0.0, beta=beta, max_iter=10, penalty="l1", random_state=0)
saga = SAGAClassifier(eta=1e-3, alpha=0.0, beta=beta, max_iter=10, penalty="l1", random_state=0)
pysaga.fit(X_bin, y_bin)
saga.fit(X_bin, y_bin)
np.testing.assert_array_almost_equal(pysaga.coef_, saga.coef_)
def test_enet_regularized_saga():
X_sparse = sparse.rand(100, 50, density=.5, random_state=0).tocsr()
y_sparse = np.random.randint(0, high=2, size=100)
eta = 1e-3
for (X, y) in ((X_bin, y_bin), (X_sparse, y_sparse)):
for alpha in np.logspace(-3, 0, 5):
for beta in np.logspace(-3, 3, 5):
pysaga = PySAGAClassifier(eta=eta,
alpha=alpha,
beta=beta,
max_iter=5,
penalty='l1',
random_state=0)
saga = SAGAClassifier(eta=eta,
alpha=alpha,
beta=beta,
max_iter=5,
penalty='l1',
random_state=0,
tol=1e-24)
pysaga.fit(X, y)
saga.fit(X, y)
np.testing.assert_array_almost_equal(pysaga.coef_, saga.coef_)
def test_no_reg_saga():
# Using no regularisation at all
pysaga = PySAGAClassifier(eta=1e-3, alpha=0.0, beta=0.0, max_iter=10, penalty=None, random_state=0)
saga = SAGAClassifier(eta=1e-3, alpha=0.0, beta=0.0, max_iter=10, penalty=None, random_state=0)
pysaga.fit(X_bin, y_bin)
saga.fit(X_bin, y_bin)
np.testing.assert_array_almost_equal(pysaga.coef_, saga.coef_)
def test_elastic_saga():
ab = [1e-5, 1e-2, 1e-1, 1.0]
for alpha, beta in zip(ab, ab):
pysaga = PySAGAClassifier(eta=1e-3, alpha=alpha, beta=beta, max_iter=1, penalty="l1", random_state=0)
saga = SAGAClassifier(eta=1e-3, alpha=alpha, beta=beta, max_iter=1, penalty="l1", random_state=0, tol=0)
pysaga.fit(X_bin, y_bin)
saga.fit(X_bin, y_bin)
np.testing.assert_array_almost_equal(pysaga.coef_, saga.coef_)
def test_l2_regularized_saga():
pysaga = PySAGAClassifier(eta=1e-3, alpha=1.0, max_iter=10,
penalty=None, random_state=0)
saga = SAGAClassifier(eta=1e-3, alpha=1.0, max_iter=10,
penalty=None, random_state=0)
pysaga.fit(X_bin, y_bin)
saga.fit(X_bin, y_bin)
np.testing.assert_array_almost_equal(pysaga.coef_, saga.coef_)
def test_saga_score():
X, y = make_classification(1000, random_state=0)
pysaga = PySAGAClassifier(eta=1e-3, alpha=0.0, beta=0.0, max_iter=1, penalty=None, random_state=0)
saga = SAGAClassifier(eta=1e-3, alpha=0.0, beta=0.0, max_iter=1, penalty=None, random_state=0)
pysaga.fit(X, y)
saga.fit(X, y)
assert_equal(pysaga.score(X, y), saga.score(X, y))
def test_elastic_saga():
ab = [1e-5, 1e-2, 1e-1, 1.]
for alpha, beta in zip(ab, ab):
pysaga = PySAGAClassifier(eta=1e-3, alpha=alpha, beta=beta, max_iter=1,
penalty='l1', random_state=0)
saga = SAGAClassifier(eta=1e-3, alpha=alpha, beta=beta, max_iter=1,
penalty='l1', random_state=0, tol=0)
pysaga.fit(X_bin, y_bin)
saga.fit(X_bin, y_bin)
np.testing.assert_array_almost_equal(pysaga.coef_, saga.coef_)
def test_no_reg_saga():
# Using no regularisation at all
pysaga = PySAGAClassifier(eta=1e-3, alpha=0.0, beta=0.0, max_iter=10,
penalty=None, random_state=0)
saga = SAGAClassifier(eta=1e-3, alpha=0.0, beta=0.0, max_iter=10,
penalty=None, random_state=0)
pysaga.fit(X_bin, y_bin)
saga.fit(X_bin, y_bin)
np.testing.assert_array_almost_equal(pysaga.coef_, saga.coef_)
def test_saga_score():
X, y = make_classification(1000, random_state=0)
pysaga = PySAGAClassifier(eta=1e-3, alpha=0.0, beta=0.0, max_iter=1,
penalty=None, random_state=0)
saga = SAGAClassifier(eta=1e-3, alpha=0.0, beta=0.0, max_iter=1,
penalty=None, random_state=0)
pysaga.fit(X, y)
saga.fit(X, y)
assert_equal(pysaga.score(X, y), saga.score(X, y))
def test_enet_regularized_saga():
X_sparse = sparse.rand(100, 50, density=.5, random_state=0).tocsr()
y_sparse = np.random.randint(0, high=2, size=100)
eta = 1e-3
for (X, y) in ((X_bin, y_bin), (X_sparse, y_sparse)):
for alpha in np.logspace(-3, 0, 5):
for beta in np.logspace(-3, 3, 5):
pysaga = PySAGAClassifier(
eta=eta, alpha=alpha, beta=beta,
max_iter=5, penalty='l1', random_state=0)
saga = SAGAClassifier(
eta=eta, alpha=alpha, beta=beta, max_iter=5,
penalty='l1', random_state=0, tol=1e-24)
pysaga.fit(X, y)
saga.fit(X, y)
np.testing.assert_array_almost_equal(pysaga.coef_, saga.coef_)
def test_sag_sample_weights():
clf1 = SAGAClassifier(loss='log', max_iter=5, verbose=0, random_state=0)
clf2 = SAGAClassifier(loss='log', max_iter=5, verbose=0, random_state=0)
clf1.fit(X, y)
sample_weights = [1] * y.size
clf2.fit(X, y, sample_weight=sample_weights)
np.testing.assert_array_equal(clf1.coef_.ravel(), clf2.coef_.ravel())
# same thing but for a regression object
alpha = 1.0
clf1 = SAGARegressor(loss='squared',
alpha=alpha,
max_iter=5,
random_state=0)
clf1.fit(X, y)
sample_weights = [2] * y.size
# alpha needs to be multiplied accordingly
clf2 = SAGARegressor(loss='squared',
alpha=2 * alpha,
max_iter=5,
random_state=0)
clf2.fit(X, y, sample_weight=sample_weights)
np.testing.assert_array_equal(clf1.coef_.ravel(), clf2.coef_.ravel())
#
# check that samples with a zero weight do not have an influence on the
# resulting coefficients by adding noise to original samples
X2 = np.concatenate((X, np.random.randn(*X.shape)),
axis=0) # augment with noise
y2 = np.concatenate((y, y), axis=0)
sample_weights = np.ones(y2.size, dtype=float)
sample_weights[X.shape[0]:] = 0.
clf1 = SAGARegressor(loss='squared',
alpha=alpha,
max_iter=100,
random_state=0,
tol=1e-24)
clf1.fit(X, y)
clf2 = SAGARegressor(loss='squared',
alpha=0.5 * alpha,
max_iter=100,
random_state=0,
tol=1e-24)
clf2.fit(X2, y2, sample_weight=sample_weights)
np.testing.assert_array_almost_equal(clf1.coef_.ravel(),
clf2.coef_.ravel(),
decimal=6)
class BaselineStruct(BaseArgumentMixin):
def __init__(self, alpha_link, alpha_prop, l1_ratio):
self.alpha_link = alpha_link
self.alpha_prop = alpha_prop
self.l1_ratio = l1_ratio
self.compat_features = False
def initialize_labels(self, y_props_flat, y_links_flat):
self.prop_encoder_ = LabelEncoder().fit(y_props_flat)
self.link_encoder_ = LabelEncoder().fit(y_links_flat)
self.n_prop_states = len(self.prop_encoder_.classes_)
self.n_link_states = len(self.link_encoder_.classes_)
def fit(self, X_link, y_link, X_prop, y_prop):
self.initialize_labels(y_prop, y_link)
y_link = self.link_encoder_.transform(y_link)
y_prop = self.prop_encoder_.transform(y_prop)
self.link_clf_ = SAGAClassifier(loss='smooth_hinge',
penalty='l1',
tol=1e-4,
max_iter=500,
random_state=0,
verbose=0)
self.prop_clf_ = clone(self.link_clf_)
alpha_link = self.alpha_link * (1 - self.l1_ratio)
beta_link = self.alpha_link * self.l1_ratio
sw = compute_sample_weight('balanced', y_link)
self.link_clf_.set_params(alpha=alpha_link, beta=beta_link)
self.link_clf_.fit(X_link, y_link, sample_weight=sw)
alpha_prop = self.alpha_prop * (1 - self.l1_ratio)
beta_prop = self.alpha_prop * self.l1_ratio
self.prop_clf_.set_params(alpha=alpha_prop, beta=beta_prop)
self.prop_clf_.fit(X_prop, y_prop)
return self
def decision_function(self, X_link, X_prop, docs):
link_offsets = np.cumsum([len(doc.features) for doc in docs])
y_link_flat = self.link_clf_.decision_function(X_link)
y_link_marg = np.zeros(
(len(y_link_flat), len(self.link_encoder_.classes_)))
link_on, = self.link_encoder_.transform([True])
y_link_marg[:, link_on] = y_link_flat.ravel()
Y_link = [
y_link_marg[start:end]
for start, end in zip(np.append(0, link_offsets), link_offsets)
]
prop_offsets = np.cumsum([len(doc.prop_features) for doc in docs])
y_prop_marg = self.prop_clf_.decision_function(X_prop)
Y_prop = [
y_prop_marg[start:end]
for start, end in zip(np.append(0, prop_offsets), prop_offsets)
]
Y_pred = []
for y_link, y_prop in zip(Y_link, Y_prop):
Y_pred.append(DocLabel(y_prop, y_link))
assert len(Y_pred) == len(docs)
return Y_pred
def fit_single(solver, X, y, penalty='l2', single_target=True, C=1,
max_iter=10, skip_slow=False):
if skip_slow and solver == 'lightning' and penalty == 'l1':
print('skip_slowping l1 logistic regression with solver lightning.')
return
print('Solving %s logistic regression with penalty %s, solver %s.'
% ('binary' if single_target else 'multinomial',
penalty, solver))
if solver == 'lightning':
from lightning.classification import SAGAClassifier
if single_target or solver not in ['sag', 'saga']:
multi_class = 'ovr'
else:
multi_class = 'multinomial'
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42,
stratify=y)
n_samples = X_train.shape[0]
n_classes = np.unique(y_train).shape[0]
test_scores = [1]
train_scores = [1]
accuracies = [1 / n_classes]
times = [0]
if penalty == 'l2':
alpha = 1. / (C * n_samples)
beta = 0
lightning_penalty = None
else:
alpha = 0.
beta = 1. / (C * n_samples)
lightning_penalty = 'l1'
for this_max_iter in range(1, max_iter + 1, 2):
print('[%s, %s, %s] Max iter: %s' %
('binary' if single_target else 'multinomial',
penalty, solver, this_max_iter))
if solver == 'lightning':
lr = SAGAClassifier(loss='log', alpha=alpha, beta=beta,
penalty=lightning_penalty,
tol=-1, max_iter=this_max_iter)
else:
lr = LogisticRegression(solver=solver,
multi_class=multi_class,
C=C,
penalty=penalty,
fit_intercept=False, tol=1e-24,
max_iter=this_max_iter,
random_state=42,
)
t0 = time.clock()
lr.fit(X_train, y_train)
train_time = time.clock() - t0
scores = []
for (X, y) in [(X_train, y_train), (X_test, y_test)]:
try:
y_pred = lr.predict_proba(X)
except NotImplementedError:
# Lightning predict_proba is not implemented for n_classes > 2
y_pred = _predict_proba(lr, X)
score = log_loss(y, y_pred, normalize=False) / n_samples
score += (0.5 * alpha * np.sum(lr.coef_ ** 2) +
beta * np.sum(np.abs(lr.coef_)))
scores.append(score)
train_score, test_score = tuple(scores)
y_pred = lr.predict(X_test)
accuracy = np.sum(y_pred == y_test) / y_test.shape[0]
test_scores.append(test_score)
train_scores.append(train_score)
accuracies.append(accuracy)
times.append(train_time)
return lr, times, train_scores, test_scores, accuracies
def fit_single(
solver,
X,
y,
penalty="l2",
single_target=True,
C=1,
max_iter=10,
skip_slow=False,
dtype=np.float64,
):
if skip_slow and solver == "lightning" and penalty == "l1":
print("skip_slowping l1 logistic regression with solver lightning.")
return
print("Solving %s logistic regression with penalty %s, solver %s." %
("binary" if single_target else "multinomial", penalty, solver))
if solver == "lightning":
from lightning.classification import SAGAClassifier
if single_target or solver not in ["sag", "saga"]:
multi_class = "ovr"
else:
multi_class = "multinomial"
X = X.astype(dtype)
y = y.astype(dtype)
X_train, X_test, y_train, y_test = train_test_split(X,
y,
random_state=42,
stratify=y)
n_samples = X_train.shape[0]
n_classes = np.unique(y_train).shape[0]
test_scores = [1]
train_scores = [1]
accuracies = [1 / n_classes]
times = [0]
if penalty == "l2":
alpha = 1.0 / (C * n_samples)
beta = 0
lightning_penalty = None
else:
alpha = 0.0
beta = 1.0 / (C * n_samples)
lightning_penalty = "l1"
for this_max_iter in range(1, max_iter + 1, 2):
print("[%s, %s, %s] Max iter: %s" % (
"binary" if single_target else "multinomial",
penalty,
solver,
this_max_iter,
))
if solver == "lightning":
lr = SAGAClassifier(
loss="log",
alpha=alpha,
beta=beta,
penalty=lightning_penalty,
tol=-1,
max_iter=this_max_iter,
)
else:
lr = LogisticRegression(
solver=solver,
multi_class=multi_class,
C=C,
penalty=penalty,
fit_intercept=False,
tol=0,
max_iter=this_max_iter,
random_state=42,
)
# Makes cpu cache even for all fit calls
X_train.max()
t0 = time.clock()
lr.fit(X_train, y_train)
train_time = time.clock() - t0
scores = []
for (X, y) in [(X_train, y_train), (X_test, y_test)]:
try:
y_pred = lr.predict_proba(X)
except NotImplementedError:
# Lightning predict_proba is not implemented for n_classes > 2
y_pred = _predict_proba(lr, X)
score = log_loss(y, y_pred, normalize=False) / n_samples
score += 0.5 * alpha * np.sum(lr.coef_**2) + beta * np.sum(
np.abs(lr.coef_))
scores.append(score)
train_score, test_score = tuple(scores)
y_pred = lr.predict(X_test)
accuracy = np.sum(y_pred == y_test) / y_test.shape[0]
test_scores.append(test_score)
train_scores.append(train_score)
accuracies.append(accuracy)
times.append(train_time)
return lr, times, train_scores, test_scores, accuracies
class BaselineStruct(BaseArgumentMixin):
def __init__(self, alpha_link, alpha_prop, l1_ratio, exact_test=False):
self.alpha_link = alpha_link
self.alpha_prop = alpha_prop
self.l1_ratio = l1_ratio
self.compat_features = False
self.exact_test = exact_test
def initialize_labels(self, y_props_flat, y_links_flat):
self.prop_encoder_ = LabelEncoder().fit(y_props_flat)
self.link_encoder_ = LabelEncoder().fit(y_links_flat)
self.n_prop_states = len(self.prop_encoder_.classes_)
self.n_link_states = len(self.link_encoder_.classes_)
def fit(self, X_link, y_link, X_prop, y_prop):
self.initialize_labels(y_prop, y_link)
y_link = self.link_encoder_.transform(y_link)
y_prop = self.prop_encoder_.transform(y_prop)
self.link_clf_ = SAGAClassifier(loss='smooth_hinge',
penalty='l1',
tol=1e-4,
max_iter=500,
random_state=0,
verbose=0)
self.prop_clf_ = clone(self.link_clf_)
alpha_link = self.alpha_link * (1 - self.l1_ratio)
beta_link = self.alpha_link * self.l1_ratio
sw = compute_sample_weight('balanced', y_link)
self.link_clf_.set_params(alpha=alpha_link, beta=beta_link)
self.link_clf_.fit(X_link, y_link, sample_weight=sw)
alpha_prop = self.alpha_prop * (1 - self.l1_ratio)
beta_prop = self.alpha_prop * self.l1_ratio
self.prop_clf_.set_params(alpha=alpha_prop, beta=beta_prop)
self.prop_clf_.fit(X_prop, y_prop)
return self
def decision_function(self, X_link, X_prop, docs):
link_offsets = np.cumsum([len(doc.features) for doc in docs])
y_link_flat = self.link_clf_.decision_function(X_link)
y_link_marg = np.zeros(
(len(y_link_flat), len(self.link_encoder_.classes_)))
link_on, = self.link_encoder_.transform([True])
y_link_marg[:, link_on] = y_link_flat.ravel()
Y_link = [
y_link_marg[start:end]
for start, end in zip(np.append(0, link_offsets), link_offsets)
]
prop_offsets = np.cumsum([len(doc.prop_features) for doc in docs])
y_prop_marg = self.prop_clf_.decision_function(X_prop)
Y_prop = [
y_prop_marg[start:end]
for start, end in zip(np.append(0, prop_offsets), prop_offsets)
]
Y_pred = []
for y_link, y_prop in zip(Y_link, Y_prop):
Y_pred.append(DocLabel(y_prop, y_link))
assert len(Y_pred) == len(docs)
return Y_pred
def fast_decode(self, Y_marg, docs, constraints):
if constraints:
Y_pred = []
zero_compat = np.zeros(
(self.n_prop_states, self.n_prop_states, self.n_link_states))
for doc, y in zip(docs, Y_marg):
potentials = (y.nodes, y.links, zero_compat, [], [], [])
y_decoded, _ = self._inference(doc,
potentials,
relaxed=False,
exact=self.exact_test,
constraints=constraints)
Y_pred.append(y_decoded)
else:
Y_pred = [
self._round(y.nodes, y.links, inverse_transform=True)
for y in Y_marg
]
return Y_pred
def predict(self, X_link, X_prop, docs, constraints=""):
Y_marg = self.decision_function(X_link, X_prop, docs)
return self.fast_decode(Y_marg, docs, constraints)
def test_sag_sample_weights():
clf1 = SAGAClassifier(loss='log', max_iter=5, verbose=0, random_state=0)
clf2 = SAGAClassifier(loss='log', max_iter=5, verbose=0, random_state=0)
clf1.fit(X, y)
sample_weights = [1] * y.size
clf2.fit(X, y, sample_weight=sample_weights)
np.testing.assert_array_equal(clf1.coef_.ravel(), clf2.coef_.ravel())
# same thing but for a regression object
alpha = 1.0
clf1 = SAGARegressor(loss='squared', alpha=alpha, max_iter=5, random_state=0)
clf1.fit(X, y)
sample_weights = [2] * y.size
# alpha needs to be multiplied accordingly
clf2 = SAGARegressor(loss='squared', alpha=2 * alpha, max_iter=5, random_state=0)
clf2.fit(X, y, sample_weight=sample_weights)
np.testing.assert_array_equal(clf1.coef_.ravel(), clf2.coef_.ravel())
#
# check that samples with a zero weight do not have an influence on the
# resulting coefficients by adding noise to original samples
X2 = np.concatenate((X, np.random.randn(*X.shape)), axis=0) # augment with noise
y2 = np.concatenate((y, y), axis=0)
sample_weights = np.ones(y2.size, dtype=np.float)
sample_weights[X.shape[0]:] = 0.
clf1 = SAGARegressor(loss='squared', alpha=alpha, max_iter=100, random_state=0, tol=1e-24)
clf1.fit(X, y)
clf2 = SAGARegressor(loss='squared', alpha=0.5*alpha, max_iter=100, random_state=0, tol=1e-24)
clf2.fit(X2, y2, sample_weight=sample_weights)
np.testing.assert_array_almost_equal(clf1.coef_.ravel(), clf2.coef_.ravel(), decimal=6)
def saga_cv(which, alphas, l1_ratio):
if which == 'cdcp':
n_folds = 3
path = os.path.join("data", "process", "erule", "folds", "{}", "{}")
elif which == 'ukp':
n_folds = 5
path = os.path.join("data", "process", "ukp-essays", "folds", "{}",
"{}")
else:
raise ValueError
clf_link = SAGAClassifier(loss='smooth_hinge',
penalty='l1',
tol=1e-4,
max_iter=100,
random_state=0,
verbose=0)
clf_prop = clone(clf_link)
link_scores = np.zeros((n_folds, len(alphas)))
prop_scores = np.zeros_like(link_scores)
for k in range(n_folds):
X_tr_link, y_tr_link = load_csr(path.format(k, 'train.npz'),
return_y=True)
X_te_link, y_te_link = load_csr(path.format(k, 'val.npz'),
return_y=True)
X_tr_prop, y_tr_prop = load_csr(path.format(k, 'prop-train.npz'),
return_y=True)
X_te_prop, y_te_prop = load_csr(path.format(k, 'prop-val.npz'),
return_y=True)
le = LabelEncoder()
y_tr_prop_enc = le.fit_transform(y_tr_prop)
y_te_prop_enc = le.transform(y_te_prop)
link_sw = compute_sample_weight('balanced', y_tr_link)
for j, alpha in enumerate(alphas):
beta = alpha * l1_ratio
alpha *= 1 - l1_ratio
clf_link.set_params(alpha=alpha, beta=beta)
clf_prop.set_params(alpha=alpha, beta=beta)
clf_link.fit(X_tr_link, y_tr_link, sample_weight=link_sw)
y_pred_link = clf_link.predict(X_te_link)
clf_prop.fit(X_tr_prop, y_tr_prop_enc)
y_pred_prop = clf_prop.predict(X_te_prop)
with warnings.catch_warnings() as w:
warnings.simplefilter('ignore')
link_f = f1_score(y_te_link, y_pred_link, average='binary')
prop_f = f1_score(y_te_prop_enc, y_pred_prop, average='macro')
link_scores[k, j] = link_f
prop_scores[k, j] = prop_f
return link_scores, prop_scores