def test_pairs(self):
database = Database('test_annotations_10000_cleaned.csv', header_path='test_annotations_10000_cleaned_header.csv')
labels = fast_strong_cluster(database)
pair_seed = generate_pair_seed(database, labels, 0.5)
# x1_a, x2_a, m_a = _get_pairs(database, labels, 10, balancing=True)
# x1_b, x2_b, m_b = _get_pairs(database, labels, 10, balancing=True)
# self.assertNotEqual(x1_a, x1_b)
# self.assertNotEqual(x2_a, x2_b)
# self.assertNotEqual(m_a, m_b)
x1_a, x2_a, m_a = get_pairwise_features(database, labels, pair_seed)
x1_b, x2_b, m_b = get_pairwise_features(database, labels, pair_seed)
np.testing.assert_array_equal(x1_a, x1_b)
np.testing.assert_array_equal(x2_a, x2_b)
np.testing.assert_array_equal(m_a, m_b)
def test(self, database_test, labels_test, pair_seed):
"""
Get testing samples and test the surrogate match function. Evaluated with ROC curve
:param database_test: RecordDatabase object
:param labels_test: Corresponding labels, of dict form [record_id, label]
:param pair_seed: List of pairs, where each pair is a tuple of form (identifierA, identifierB)
:return RocCurve: An RocCurve object
"""
y_test, x_test, _ = get_pairwise_features(database_test, labels_test,
pair_seed)
self.class_balance_validation = float(sum(y_test)) / len(y_test)
x_bar_probability = self._classifier.predict_proba(x_test)[:, 1]
#output = np.column_stack((x1_test, x1_bar_probability))
#np.savetxt('roc_labels.csv', output, delimiter=",", header='label,probability', fmt='%.1i,%5.5f')
roc = RocCurve(y_test, x_bar_probability)
self.roc = roc
sorted_indices = np.argsort(-1 * x_bar_probability)
for sorted_index in sorted_indices:
x1_bar = x_bar_probability[sorted_index]
pair = pair_seed[sorted_index]
print 'Test pair P(match) = ', x1_bar
database_test.records[pair[0]].display(indent=' ')
print ' ----'
database_test.records[pair[1]].display(indent=' ')
return roc
def test(self, database_test, labels_test, pair_seed):
"""
Get testing samples and test the surrogate match function. Evaluated with ROC curve
:param database_test: RecordDatabase object
:param labels_test: Corresponding labels, of dict form [record_id, label]
:param pair_seed: List of pairs, where each pair is a tuple of form (identifierA, identifierB)
:return RocCurve: An RocCurve object
"""
y_test, x_test, _ = get_pairwise_features(database_test, labels_test, pair_seed)
self.class_balance_validation = float(sum(y_test))/len(y_test)
# Save to text file
# Run executable
# Load csv file
x_bar_probability = self._classifier.predict_proba(x_test)[:, 1]
#output = np.column_stack((x1_test, x1_bar_probability))
#np.savetxt('roc_labels.csv', output, delimiter=",", header='label,probability', fmt='%.1i,%5.5f')
roc = RocCurve(y_test, x_bar_probability)
self.roc = roc
sorted_indices = np.argsort(-1*x_bar_probability)
for sorted_index in sorted_indices:
x1_bar = x_bar_probability[sorted_index]
pair = pair_seed[sorted_index]
print 'Test pair P(match) = ', x1_bar
database_test.records[pair[0]].display(indent=' ')
print ' ----'
database_test.records[pair[1]].display(indent=' ')
return roc
def _train(self, database_train, labels_train, pair_seed):
"""
Gets training samples and trains the surrogate match function
:param database_train: Training database
:param labels_train: Dictionary of [identier, cluster label]
:param pair_seed: List of pairs to use
:return x2_mean: Mean feature values used in imputation (i.e. to fill in missing features)
"""
x1_train, x2_train, x2_mean = get_pairwise_features(
database_train, labels_train, pair_seed)
bounds = list()
for _ in range(x2_train.shape[1]):
bounds.append((
None,
0)) # restrict all feature weights to be negative (convex set)
bounds.append((None, None)) # no bounds for the constant offset term
print 'Training Logistic Regression pairwise match function...'
print ' Pos/Neg training sample class split: ', sum(
x1_train), '/', len(x1_train) - sum(x1_train)
print ' Enforcing weight vector bounds:'
for counter, bound in enumerate(bounds):
print ' Feature', counter, ': ', bounds
print ' (last term is intercept)'
self._logreg.bounds = bounds
self._logreg.fit(x2_train, x1_train)
print 'Model coefficients: ', self._logreg.coef_
print 'Intercept: ', self._logreg.intercept_
return x2_mean
def _train(self, database_train, labels_train, pair_seed):
"""
Gets training samples and trains the surrogate match function
:param database_train: Training database
:param labels_train: Dictionary of [identier, cluster label]
:param pair_seed: List of pairs to use
:return x_mean: Mean feature values used in imputation (i.e. to fill in missing features)
"""
y, x, x_mean = get_pairwise_features(database_train, labels_train, pair_seed)
print 'Training decision tree pairwise match function...'
print ' Pos/Neg training sample class split: ', sum(y), '/', len(y) - sum(y)
self._classifier.fit(x, y)
print 'Match function training complete.'
return x_mean
def _train(self, database_train, labels_train, pair_seed):
"""
Gets training samples and trains the surrogate match function
:param database_train: Training database
:param labels_train: Dictionary of [identier, cluster label]
:param pair_seed: List of pairs to use
:return x_mean: Mean feature values used in imputation (i.e. to fill in missing features)
"""
y, x, x_mean = get_pairwise_features(database_train, labels_train, pair_seed)
print 'Training decision tree pairwise match function...'
print ' Pos/Neg training sample class split: ', sum(y), '/', len(y) - sum(y)
self._classifier.fit(x, y)
print 'Match function training complete.'
return x_mean
def _train(self, database_train, labels_train, pair_seed):
"""
Gets training samples and trains the surrogate match function
:param database_train: Training database
:param labels_train: Dictionary of [identier, cluster label]
:param pair_seed: List of pairs to use
:return x_mean: Mean feature values used in imputation (i.e. to fill in missing features)
"""
y, x, _ = get_pairwise_features(database_train, labels_train, pair_seed, impute=False)
print 'Training decision forest pairwise match function...'
print ' Pos/Neg training sample class split: ', sum(y), '/', len(y) - sum(y)
# Write features to text file
np.savetxt('features.csv', x, delimiter=",", header=)
# Run executable
# save path to model as the classifier
self._classifier.fit(x, y)
print 'Match function training complete.'
def _train(self, database_train, labels_train, pair_seed):
"""
Gets training samples and trains the surrogate match function
:param database_train: Training database
:param labels_train: Dictionary of [identier, cluster label]
:param pair_seed: List of pairs to use
:return x2_mean: Mean feature values used in imputation (i.e. to fill in missing features)
"""
x1_train, x2_train, x2_mean = get_pairwise_features(database_train, labels_train, pair_seed)
bounds = list()
for _ in range(x2_train.shape[1]):
bounds.append((None, 0)) # restrict all feature weights to be negative (convex set)
bounds.append((None, None)) # no bounds for the constant offset term
print 'Training Logistic Regression pairwise match function...'
print ' Pos/Neg training sample class split: ', sum(x1_train), '/', len(x1_train) - sum(x1_train)
print ' Enforcing weight vector bounds:'
for counter, bound in enumerate(bounds):
print ' Feature', counter, ': ', bounds
print ' (last term is intercept)'
self._logreg.bounds = bounds
self._logreg.fit(x2_train, x1_train)
print 'Model coefficients: ', self._logreg.coef_
print 'Intercept: ', self._logreg.intercept_
return x2_mean
