def test_overflow_averaged():
X = np.array([[np.finfo('d').max]])
Y = np.array([-1])
pcp = StructuredPerceptron(model=BinaryClf(n_features=1),
max_iter=2, average=True)
pcp.fit(X, Y)
assert_true(np.isfinite(pcp.w[0]))
def test_constraint_removal():
digits = load_digits()
X, y = digits.data, digits.target
y = 2 * (y % 2) - 1 # even vs odd as +1 vs -1
X = X / 16.
pbl = BinaryClf(n_features=X.shape[1])
clf_no_removal = OneSlackSSVM(model=pbl, max_iter=500, C=1,
inactive_window=0, tol=0.01)
clf_no_removal.fit(X, y)
clf = OneSlackSSVM(model=pbl, max_iter=500, C=1, tol=0.01,
inactive_threshold=1e-8)
clf.fit(X, y)
# check that we learned something
assert_greater(clf.score(X, y), .92)
# results are mostly equal
# if we decrease tol, they will get more similar
assert_less(np.mean(clf.predict(X) != clf_no_removal.predict(X)), 0.02)
# without removal, have as many constraints as iterations
assert_equal(len(clf_no_removal.objective_curve_),
len(clf_no_removal.constraints_))
# with removal, there are less constraints than iterations
assert_less(len(clf.constraints_),
len(clf.objective_curve_))
def train_cue_learner(sentence_dicts, C_value):
cue_lexicon, affixal_cue_lexicon = get_cue_lexicon(sentence_dicts)
cue_sentence_dicts, cue_instances, cue_labels = extract_features_cue(
sentence_dicts, cue_lexicon, affixal_cue_lexicon, 'training')
vectorizer = DictVectorizer()
fvs = vectorizer.fit_transform(cue_instances).toarray()
model = BinaryClf()
cue_ssvm = NSlackSSVM(model, C=C_value, batch_size=-1)
cue_ssvm.fit(fvs, np.asarray(cue_labels))
return cue_ssvm, vectorizer, cue_lexicon, affixal_cue_lexicon
def test_simple_1d_dataset_cutting_plane():
# 10 1d datapoints between 0 and 1
X = np.random.uniform(size=(30, 1))
# linearly separable labels
Y = 1 - 2 * (X.ravel() < .5)
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
pbl = BinaryClf(n_features=2)
svm = NSlackSSVM(pbl, check_constraints=True, C=1000)
svm.fit(X, Y)
assert_array_equal(Y, np.hstack(svm.predict(X)))
def test_blobs_2d_one_slack():
# make two gaussian blobs
X, Y = make_blobs(n_samples=80, centers=2, random_state=1)
Y = 2 * Y - 1
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
X_train, X_test, Y_train, Y_test = X[:40], X[40:], Y[:40], Y[40:]
pbl = BinaryClf(n_features=3)
svm = OneSlackSSVM(pbl, C=1000)
svm.fit(X_train, Y_train)
assert_array_equal(Y_test, np.hstack(svm.predict(X_test)))
def cue_trainer(filename, corenlp):
newfilename = process_data(filename, corenlp)
sentence_dicts = file_to_sentence_dict(newfilename)
cue_dict, affix_cue_dict = get_cue_dict(sentence_dicts)
sentence_dicts, cue_instances, cue_labels = extract_features_cue(
sentence_dicts, cue_dict, affix_cue_dict, 'training')
cue_vec = DictVectorizer()
model = cue_vec.fit_transform(cue_instances).toarray()
cue_ssvm = NSlackSSVM(BinaryClf(), C=0.2, batch_size=-1)
#cue_ssvm = SVC(C = 0.2)
cue_ssvm.fit(model, cue_labels)
return sentence_dicts, cue_ssvm, cue_vec, cue_dict, affix_cue_dict
"""pickle.dump(cue_ssvm, open("cue_model_%s.pkl" %filename, "wb"))
def test_partial_averaging():
"""Use XOR weight cycling to test partial averaging"""
X = np.array([[a, b, 1] for a in (-1, 1) for b in (-1, 1)], dtype=np.float)
Y = np.array([-1, 1, 1, -1])
pcp = StructuredPerceptron(model=BinaryClf(n_features=3), max_iter=5,
decay_exponent=1, decay_t0=1)
weight = {}
for average in (0, 1, 4, -1):
pcp.set_params(average=average)
pcp.fit(X, Y)
weight[average] = pcp.w
assert_array_equal(weight[4], weight[-1])
assert_array_almost_equal(weight[0], [1.5, 3, 0])
assert_array_almost_equal(weight[1], [1.75, 3.5, 0])
assert_array_almost_equal(weight[4], [2.5, 5, 0])
def test_averaging_early_stopping():
"""Test averaging over final epoch when early stopping"""
# we use logical OR, an easy problem solved after the second epoch
X = np.array([[a, b, 1] for a in (-1, 1) for b in (-1, 1)], dtype=np.float)
Y = np.array([-1, 1, 1, 1])
pcp = StructuredPerceptron(model=BinaryClf(n_features=3), max_iter=3,
average=-1)
pcp.fit(X, Y)
# The exact weight is used without the influence of the early iterations
assert_array_equal(pcp.w, [1, 1, 1])
# If we were expecting 3 iterations, we would end up with a zero vector
pcp.set_params(average=2)
pcp.fit(X, Y)
assert_array_equal(pcp.w, [0, 0, 0])
def test_blobs_batch():
# make two gaussian blobs
X, Y = make_blobs(n_samples=80, centers=2, random_state=1)
Y = 2 * Y - 1
pbl = BinaryClf(n_features=2)
# test psi
psi_mean = pbl.batch_psi(X, Y)
psi_mean2 = np.sum([pbl.psi(x, y) for x, y in zip(X, Y)], axis=0)
assert_array_equal(psi_mean, psi_mean2)
# test inference
w = np.random.uniform(-1, 1, size=pbl.size_psi)
Y_hat = pbl.batch_inference(X, w)
for i, (x, y_hat) in enumerate(zip(X, Y_hat)):
assert_array_equal(Y_hat[i], pbl.inference(x, w))
# test inference
Y_hat = pbl.batch_loss_augmented_inference(X, Y, w)
for i, (x, y, y_hat) in enumerate(zip(X, Y, Y_hat)):
assert_array_equal(Y_hat[i], pbl.loss_augmented_inference(x, y, w))
def test_model_1d():
# 10 1d datapoints between -1 and 1
np.random.seed(0)
X = np.random.uniform(size=(10, 1))
# linearly separable labels
Y = 1 - 2 * (X.ravel() < .5)
pbl = BinaryClf(n_features=2)
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
w = [1, -.5]
Y_pred = np.hstack([pbl.inference(x, w) for x in X])
assert_array_equal(Y, Y_pred)
# check that sign of psi and inference agree
for x, y in zip(X, Y):
assert_true(np.dot(w, pbl.psi(x, y)) > np.dot(w, pbl.psi(x, -y)))
# check that sign of psi and the sign of y correspond
for x, y in zip(X, Y):
assert_true(np.dot(w, pbl.psi(x, y)) == -np.dot(w, pbl.psi(x, -y)))
def test_blobs_batch():
# make two gaussian blobs
X, Y = make_blobs(n_samples=80, centers=2, random_state=1)
Y = 2 * Y - 1
pbl = BinaryClf(n_features=2)
# test psi
psi_mean = pbl.batch_psi(X, Y)
psi_mean2 = np.sum([pbl.psi(x, y) for x, y in zip(X, Y)], axis=0)
assert_array_equal(psi_mean, psi_mean2)
# test inference
w = np.random.uniform(-1, 1, size=pbl.size_psi)
Y_hat = pbl.batch_inference(X, w)
for i, (x, y_hat) in enumerate(zip(X, Y_hat)):
assert_array_equal(Y_hat[i], pbl.inference(x, w))
# test inference
Y_hat = pbl.batch_loss_augmented_inference(X, Y, w)
for i, (x, y, y_hat) in enumerate(zip(X, Y, Y_hat)):
assert_array_equal(Y_hat[i], pbl.loss_augmented_inference(x, y, w))
def test_xor():
"""Test perceptron behaviour against hand-computed values for XOR"""
X = np.array([[a, b, 1] for a in (-1, 1) for b in (-1, 1)], dtype=np.float)
Y = np.array([-1, 1, 1, -1])
# Should cycle weight vectors (1, 1, -1), (0, 2, 0), (1, 1, 1), (0, 0, 0)
# but this depends on how ties are settled. Maybe the test can be
# made robust to this
# Batch version should cycle (0, 0, -2), (0, 0, 0)
expected_predictions = [
np.array([1, 1, 1, 1]), # online, no average, w = (0, 0, 0, 0)
np.array([-1, 1, -1, 1]), # online, average, w ~= (0.5, 1, 0)
np.array([1, 1, 1, 1]), # batch, no average, w = (0, 0, 0)
np.array([-1, -1, -1, -1]) # batch, average, w ~= (0, 0, -2)
]
pcp = StructuredPerceptron(model=BinaryClf(n_features=3), max_iter=2)
for pred, (batch, average) in zip(expected_predictions,
product((False, True), (False, True))):
pcp.set_params(batch=batch, average=average)
pcp.fit(X, Y)
# We don't compare w explicitly but its prediction. As the perceptron
# is invariant to the scaling of w, this will allow the optimization of
# the underlying implementation
assert_array_equal(pcp.predict(X), pred)
def test_model_1d():
# 10 1d datapoints between -1 and 1
np.random.seed(0)
X = np.random.uniform(size=(10, 1))
# linearly separable labels
Y = 1 - 2 * (X.ravel() < .5)
pbl = BinaryClf(n_features=2)
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
w = [1, -.5]
Y_pred = np.hstack([pbl.inference(x, w) for x in X])
assert_array_equal(Y, Y_pred)
# check that sign of psi and inference agree
for x, y in zip(X, Y):
assert_true(np.dot(w, pbl.psi(x, y)) > np.dot(w, pbl.psi(x, -y)))
# check that sign of psi and the sign of y correspond
for x, y in zip(X, Y):
assert_true(np.dot(w, pbl.psi(x, y)) == -np.dot(w, pbl.psi(x, -y)))
from pystruct.models import BinaryClf
from pystruct.learners import (NSlackSSVM, OneSlackSSVM, SubgradientSSVM)
# do a binary digit classification
digits = load_digits()
X, y = digits.data, digits.target
# make binary task by doing odd vs even numers
y = y % 2
# code as +1 and -1
y = 2 * y - 1
X /= X.max()
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=0)
pbl = BinaryClf()
n_slack_svm = NSlackSSVM(pbl, C=10, batch_size=-1)
one_slack_svm = OneSlackSSVM(pbl, C=10, tol=0.1)
subgradient_svm = SubgradientSSVM(pbl,
C=10,
learning_rate=0.1,
max_iter=100,
batch_size=10)
# we add a constant 1 feature for the bias
X_train_bias = np.hstack([X_train, np.ones((X_train.shape[0], 1))])
X_test_bias = np.hstack([X_test, np.ones((X_test.shape[0], 1))])
# n-slack cutting plane ssvm
start = time()
n_slack_svm.fit(X_train_bias, y_train)
import numpy as np
import loader
import util
from sklearn import preprocessing
directory = "/Users/thijs/dev/boilerplate/src/main/resources/dataset/"
featureset = "features10"
print("Load files")
features, labels = \
loader.loadBinary(featureset+'.csv', 'labels.csv', directory)
# print("Shuffle results")
# features, labels = util.shuffle(features, labels)
print("Loaded")
# print(labels)
# features = preprocessing.scale(features)
from pystruct.models import BinaryClf
from pystruct.learners import (NSlackSSVM, OneSlackSSVM, SubgradientSSVM,
FrankWolfeSSVM)
clf = FrankWolfeSSVM(BinaryClf(), verbose=True)
# print(clf)
clf.fit(features, labels)
trscore = clf.score(features, labels)
# print("Training score: {0}".format(trscore))
print("Klaar")
def test_break_ties():
pbl = BinaryClf(n_features=2)
X = np.array([[-1., -1.], [-1., 1.], [1., 1.]])
w = np.array([1., 1.])
assert_array_equal(pbl.batch_inference(X, w), np.array([-1, 1, 1]))
def test_break_ties():
pbl = BinaryClf(n_features=2)
X = np.array([[-1., -1.], [-1., 1.], [1., 1.]])
w = np.array([1., 1.])
assert_array_equal(pbl.batch_inference(X, w), np.array([-1, 1, 1]))
