def test_KNN_adwin(test_path, package_path):
test_file = os.path.join(package_path,
'src/skmultiflow/data/datasets/sea_big.csv')
stream = FileStream(test_file, -1, 1)
stream.prepare_for_use()
learner = KNNAdwin(n_neighbors=8, leaf_size=40, max_window_size=2000)
cnt = 0
max_samples = 5000
predictions = []
correct_predictions = 0
wait_samples = 100
while cnt < max_samples:
X, y = stream.next_sample()
# Test every n samples
if (cnt % wait_samples == 0) and (cnt != 0):
predictions.append(learner.predict(X)[0])
if y[0] == predictions[-1]:
correct_predictions += 1
learner.partial_fit(X, y)
cnt += 1
performance = correct_predictions / len(predictions)
expected_predictions = [
1, 0, 0, 1, 0, 0, 1, 1, 1, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 1, 1, 0, 0, 1,
1, 1, 1, 1, 0, 1, 0, 1, 0, 1, 1, 0, 1, 0, 1, 0, 1, 1, 0, 1, 0, 1, 1, 1,
1
]
expected_correct_predictions = 40
expected_performance = 0.8163265306122449
assert np.alltrue(predictions == expected_predictions)
assert np.isclose(expected_performance, performance)
assert correct_predictions == expected_correct_predictions
def __init__(self,
base_estimator=KNNAdwin(),
n_estimators=10,
sampling_rate=3,
algorithm=1,
drift_detection=True,
random_state=None):
super().__init__()
# default values
self.ensemble = None
self.n_estimators = None
self.classes = None
self.random_state = None
self._init_n_estimators = n_estimators
self._init_random_state = random_state
self.sampling_rate = sampling_rate
self.algorithm = algorithm
self.drift_detection = drift_detection
self.adwin_ensemble = None
self.lam_sc = None
self.lam_pos = None
self.lam_neg = None
self.lam_sw = None
self.epsilon = None
self.__configure(base_estimator)
def __init__(self,
base_estimator=KNNAdwin(),
n_estimators=10,
cost_positive=1,
cost_negative=0.1,
drift_detection=True,
random_state=None):
super().__init__()
# default values
self.ensemble = None
self.actual_n_estimators = None
self.classes = None
self._random_state = None
self.base_estimator = base_estimator
self.n_estimators = n_estimators
self.cost_positive = cost_positive
self.cost_negative = cost_negative
self.drift_detection = drift_detection
self.random_state = random_state
self.adwin_ensemble = None
self.lam_fn = None
self.lam_fp = None
self.lam_sum = None
self.lam_sw = None
self.werr = None
self.epsilon = None
self.__configure()
def __init__(self,
base_estimator=KNNAdwin(),
n_estimators=10,
sampling_rate=1,
drift_detection=True,
random_state=None):
super().__init__(base_estimator, n_estimators, sampling_rate,
drift_detection, random_state)
def __init__(self,
base_estimator=KNNAdwin(),
n_estimators=10,
random_state=None):
super().__init__(base_estimator, n_estimators, random_state)
# default values
self.adwin_ensemble = None
self.__configure()
def demo(instances=2000):
""" _test_comparison_prequential
This demo will test a prequential evaluation when more than one learner is
passed, which makes it a comparison task.
Parameters
----------
instances: int
The evaluation's maximum number of instances.
"""
# Stream setup
stream = FileStream("../data/datasets/covtype.csv", -1, 1)
# stream = SEAGenerator(classification_function=2, sample_seed=53432, balance_classes=False)
stream.prepare_for_use()
# Setup the classifier
clf = SGDClassifier()
# classifier = KNNAdwin(n_neighbors=8, max_window_size=2000,leaf_size=40, categorical_list=None)
# classifier = OzaBaggingAdwin(base_estimator=KNN(n_neighbors=8, max_window_size=2000, leaf_size=30, categorical_list=None))
clf_one = KNNAdwin(n_neighbors=8, max_window_size=1000, leaf_size=30)
# clf_two = KNN(n_neighbors=8, max_window_size=1000, leaf_size=30)
# clf_two = LeverageBagging(base_estimator=KNN(), n_estimators=2)
t_one = OneHotToCategorical([[10, 11, 12, 13],
[
14, 15, 16, 17, 18, 19, 20, 21, 22, 23,
24, 25, 26, 27, 28, 29, 30, 31, 32, 33,
34, 35, 36, 37, 38, 39, 40, 41, 42, 43,
44, 45, 46, 47, 48, 49, 50, 51, 52, 53
]])
# t_two = OneHotToCategorical([[10, 11, 12, 13],
# [14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35,
# 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53]])
pipe_one = Pipeline([('one_hot_to_categorical', t_one), ('KNN', clf_one)])
# pipe_two = Pipeline([('one_hot_to_categorical', t_two), ('KNN', clf_two)])
classifier = [clf, pipe_one]
# classifier = SGDRegressor()
# classifier = PerceptronMask()
# Setup the pipeline
# pipe = Pipeline([('Classifier', classifier)])
# Setup the evaluator
evaluator = EvaluatePrequential(
pretrain_size=2000,
output_file='test_comparison_prequential.csv',
max_samples=instances,
batch_size=1,
n_wait=200,
max_time=1000,
show_plot=True,
metrics=['performance', 'kappa_t'])
# Evaluate
evaluator.evaluate(stream=stream, model=classifier)
def __init__(self,
base_estimator=KNNAdwin(),
n_estimators=10,
cost_positive=1,
cost_negative=0.1,
drift_detection=True,
random_state=None):
super().__init__(base_estimator, n_estimators, cost_positive,
cost_negative, drift_detection, random_state)
def __init__(self, base_estimator=KNNAdwin(), n_estimators=10, random_state=None):
super().__init__()
# default values
self.ensemble = None
self.n_estimators = None
self.classes = None
self.random_state = None
self._init_n_estimators = n_estimators
self._init_random_state = random_state
self.__configure(base_estimator)
def __init__(self,
base_estimator=KNNAdwin(),
n_estimators=10,
random_state=None):
super().__init__()
# default values
self.ensemble = None
self.actual_n_estimators = None
self.classes = None
self._random_state = None # This is the actual random_state object used internally
self.base_estimator = base_estimator
self.n_estimators = n_estimators
self.random_state = random_state
self.__configure()
def demo():
""" _test_oza_bagging
This demo tests the OzaBagging classifier using KNNAdwin classifiers,
on samples given by a SEAGenerator.
The test computes the performance of the OzaBagging classifier as well
as the time to create the structure and classify max_samples (5000 by
default) instances.
"""
logging.basicConfig(format='%(message)s', level=logging.INFO)
warnings.filterwarnings("ignore", ".*Passing 1d.*")
stream = SEAGenerator(1, noise_percentage=.067, random_state=1)
stream.prepare_for_use()
clf = OzaBagging(base_estimator=KNNAdwin(n_neighbors=8, max_window_size=2000, leaf_size=30), n_estimators=2, random_state=1)
sample_count = 0
correctly_classified = 0
max_samples = 5000
train_size = 8
first = True
if train_size > 0:
X, y = stream.next_sample(train_size)
clf.partial_fit(X, y, classes=stream.target_values)
first = False
while sample_count < max_samples:
if sample_count % (max_samples/20) == 0:
logging.info('%s%%', str((sample_count//(max_samples/20)*5)))
X, y = stream.next_sample()
my_pred = clf.predict(X)
if first:
clf.partial_fit(X, y, classes=stream.target_values)
first = False
else:
clf.partial_fit(X, y)
if my_pred is not None:
if y[0] == my_pred[0]:
correctly_classified += 1
sample_count += 1
print(str(sample_count) + ' samples analyzed.')
print('My performance: ' + str(correctly_classified / sample_count))
def test_pipeline(test_path):
n_categories = 5
# Load test data generated using:
# RandomTreeGenerator(tree_random_state=1, sample_random_state=1,
# n_cat_features=n_categories, n_num_features=0)
test_file = os.path.join(test_path, 'data-one-hot.npz')
data = np.load(test_file)
X = data['X']
y = data['y']
stream = DataStream(data=X, y=y)
stream.prepare_for_use()
# Setup transformer
cat_att_idx = [[i + j for i in range(n_categories)]
for j in range(0, n_categories * n_categories, n_categories)
]
transformer = OneHotToCategorical(categorical_list=cat_att_idx)
# Set up the classifier
classifier = KNNAdwin(n_neighbors=2, max_window_size=50, leaf_size=40)
# Setup the pipeline
pipe = Pipeline([('one-hot', transformer), ('KNNAdwin', classifier)])
# Setup the evaluator
evaluator = EvaluatePrequential(show_plot=False,
pretrain_size=10,
max_samples=100)
# Evaluate
evaluator.evaluate(stream=stream, model=pipe)
metrics = evaluator.get_mean_measurements()
expected_accuracy = 0.5555555555555556
assert np.isclose(expected_accuracy, metrics[0].accuracy_score())
expected_kappa = 0.11111111111111116
assert np.isclose(expected_kappa, metrics[0].kappa_score())
print(pipe.get_info())
expected_info = "Pipeline:\n" \
"[OneHotToCategorical(categorical_list=[[0, 1, 2, 3, 4], [5, 6, 7, 8, 9],\n" \
" [10, 11, 12, 13, 14],\n" \
" [15, 16, 17, 18, 19],\n" \
" [20, 21, 22, 23, 24]])\n" \
"KNNAdwin(leaf_size=40, max_window_size=50, n_neighbors=2,\n" \
" nominal_attributes=None)]"
assert pipe.get_info() == expected_info
def __init__(self,
base_estimator=KNNAdwin(),
n_estimators=10,
sampling_rate=2,
drift_detection=True,
random_state=None):
super().__init__()
# default values
self.ensemble = None
self.n_estimators = None
self.classes = None
self.random_state = None
self.n_samples = None
self.drift_detection = drift_detection
self.adwin_ensemble = None
self._init_n_estimators = n_estimators
self._init_random_state = random_state
self.sampling_rate = sampling_rate
self.__configure(base_estimator)
def __init__(self,
base_estimator=KNNAdwin(),
n_estimators=10,
random_state=None):
super().__init__(base_estimator, n_estimators, random_state)
def demo():
""" _test_knn_adwin
This demo tests the KNNAdwin classifier on a file stream, which gives
instances coming from a SEA generator.
The test computes the performance of the KNNAdwin classifier as well as
the time to create the structure and classify max_samples (10000 by
default) instances.
"""
start = timer()
logging.basicConfig(format='%(message)s', level=logging.INFO)
# warnings.filterwarnings("ignore", ".*Passing 1d.*")
stream = FileStream('../data/datasets/sea_big.csv', -1, 1)
# stream = RandomRBFGeneratorDrift(change_speed=41.00, n_centroids=50, model_random_state=32523423,
# sample_seed=5435, n_classes=2, num_att=10, num_drift_centroids=50)
stream.prepare_for_use()
t = OneHotToCategorical([[10, 11, 12, 13],
[
14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24,
25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35,
36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46,
47, 48, 49, 50, 51, 52, 53
]])
t2 = OneHotToCategorical([[10, 11, 12, 13],
[
14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24,
25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35,
36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46,
47, 48, 49, 50, 51, 52, 53
]])
# knn = KNN(n_neighbors=8, max_window_size=2000, leaf_size=40)
knn = KNNAdwin(n_neighbors=8, leaf_size=40, max_window_size=2000)
# pipe = Pipeline([('one_hot_to_categorical', t), ('KNN', knn)])
compare = KNeighborsClassifier(n_neighbors=8,
algorithm='kd_tree',
leaf_size=40,
metric='euclidean')
# pipe2 = Pipeline([('one_hot_to_categorical', t2), ('KNN', compare)])
first = True
train = 200
if train > 0:
X, y = stream.next_sample(train)
# pipe.partial_fit(X, y, classes=stream.target_values)
# pipe.partial_fit(X, y, classes=stream.target_values)
# pipe2.fit(X, y)
knn.partial_fit(X, y, classes=stream.target_values)
compare.fit(X, y)
first = False
n_samples = 0
max_samples = 10000
my_corrects = 0
compare_corrects = 0
while n_samples < max_samples:
if n_samples % (max_samples / 20) == 0:
logging.info('%s%%', str((n_samples // (max_samples / 20) * 5)))
X, y = stream.next_sample()
# my_pred = pipe.predict(X)
my_pred = knn.predict(X)
# my_pred = [1]
if first:
# pipe.partial_fit(X, y, classes=stream.target_values)
# pipe.partial_fit(X, y, classes=stream.target_values)
knn.partial_fit(X, y, classes=stream.target_values)
first = False
else:
# pipe.partial_fit(X, y)
knn.partial_fit(X, y)
# compare_pred = pipe2.predict(X)
compare_pred = compare.predict(X)
if y[0] == my_pred[0]:
my_corrects += 1
if y[0] == compare_pred[0]:
compare_corrects += 1
n_samples += 1
end = timer()
print('Evaluation time: ' + str(end - start))
print(str(n_samples) + ' samples analyzed.')
print('My performance: ' + str(my_corrects / n_samples))
print('Compare performance: ' + str(compare_corrects / n_samples))
def test_knn_adwin():
stream = ConceptDriftStream(stream=SEAGenerator(random_state=1),
drift_stream=SEAGenerator(random_state=2, classification_function=2),
random_state=1, position=250, width=10)
stream.prepare_for_use()
learner = KNNAdwin(n_neighbors=8, leaf_size=40, max_window_size=200)
cnt = 0
max_samples = 1000
predictions = array('i')
correct_predictions = 0
wait_samples = 20
while cnt < max_samples:
X, y = stream.next_sample()
# Test every n samples
if (cnt % wait_samples == 0) and (cnt != 0):
predictions.append(learner.predict(X)[0])
if y[0] == predictions[-1]:
correct_predictions += 1
learner.partial_fit(X, y)
cnt += 1
expected_predictions = array('i', [1, 0, 1, 1, 1, 1, 0, 1, 1, 1,
0, 1, 1, 0, 1, 1, 0, 1, 0, 1,
1, 1, 1, 0, 1, 0, 0, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
0, 1, 1, 1, 0, 1, 0, 1, 1])
assert np.alltrue(predictions == expected_predictions)
expected_correct_predictions = 46
assert correct_predictions == expected_correct_predictions
learner.reset()
assert learner.window.n_samples == 0
expected_info = 'KNNAdwin: - n_neighbors: 8 - max_window_size: 200 - leaf_size: 40'
assert learner.get_info() == expected_info
stream.restart()
X, y = stream.next_sample(max_samples)
learner.fit(X[:950], y[:950])
predictions = learner.predict(X[951:])
correct_predictions = sum(np.array(predictions) == y[951:])
expected_correct_predictions = 47
assert correct_predictions == expected_correct_predictions
assert type(learner.predict(X)) == np.ndarray
assert type(learner.predict_proba(X)) == np.ndarray