def test_hat_mc(test_path):
stream = ConceptDriftStream(stream=SEAGenerator(random_state=1,
noise_percentage=0.05),
drift_stream=SEAGenerator(
random_state=2,
classification_function=2,
noise_percentage=0.05),
random_state=1,
position=250,
width=10)
stream.prepare_for_use()
learner = HAT(leaf_prediction='mc')
cnt = 0
max_samples = 1000
y_pred = array('i')
y_proba = []
wait_samples = 20
while cnt < max_samples:
X, y = stream.next_sample()
# Test every n samples
if (cnt % wait_samples == 0) and (cnt != 0):
y_pred.append(learner.predict(X)[0])
y_proba.append(learner.predict_proba(X)[0])
learner.partial_fit(X, y)
cnt += 1
expected_predictions = array('i', [
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1
])
assert np.alltrue(y_pred == expected_predictions)
test_file = os.path.join(test_path, 'test_hoeffding_adaptive_tree_mc.npy')
data = np.load(test_file)
assert np.allclose(y_proba, data)
expected_info = "HAT(binary_split=False, grace_period=200, leaf_prediction='mc',\n" \
" max_byte_size=33554432, memory_estimate_period=1000000, nb_threshold=0,\n" \
" no_preprune=False, nominal_attributes=None, remove_poor_atts=False,\n" \
" split_confidence=1e-07, split_criterion='info_gain',\n" \
" stop_mem_management=False, tie_threshold=0.05)"
assert learner.get_info() == expected_info
expected_model_1 = 'Leaf = Class 1.0 | {0.0: 398.0, 1.0: 1000.0}\n'
assert (learner.get_model_description() == expected_model_1)
assert type(learner.predict(X)) == np.ndarray
assert type(learner.predict_proba(X)) == np.ndarray
stream.restart()
X, y = stream.next_sample(5000)
learner = HAT(max_byte_size=30, leaf_prediction='mc', grace_period=10)
learner.partial_fit(X, y)
def test_hat_mc(test_path):
stream = ConceptDriftStream(stream=SEAGenerator(random_state=1,
noise_percentage=0.05),
drift_stream=SEAGenerator(
random_state=2,
classification_function=2,
noise_percentage=0.05),
random_state=1,
position=250,
width=10)
stream.prepare_for_use()
learner = HAT(leaf_prediction='mc')
cnt = 0
max_samples = 1000
y_pred = array('i')
y_proba = []
wait_samples = 20
while cnt < max_samples:
X, y = stream.next_sample()
# Test every n samples
if (cnt % wait_samples == 0) and (cnt != 0):
y_pred.append(learner.predict(X)[0])
y_proba.append(learner.predict_proba(X)[0])
learner.partial_fit(X, y)
cnt += 1
expected_predictions = array('i', [
1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 0, 1, 0, 1, 0,
1
])
assert np.alltrue(y_pred == expected_predictions)
test_file = os.path.join(test_path, 'test_hoeffding_adaptive_tree_mc.npy')
data = np.load(test_file)
assert np.allclose(y_proba, data)
expected_info = 'HAT: max_byte_size: 33554432 - memory_estimate_period: 1000000 - grace_period: 200' \
' - split_criterion: info_gain - split_confidence: 1e-07 - tie_threshold: 0.05' \
' - binary_split: False - stop_mem_management: False - remove_poor_atts: False' \
' - no_pre_prune: False - leaf_prediction: mc - nb_threshold: 0' \
' - nominal_attributes: [] - '
assert learner.get_info() == expected_info
expected_model_1 = 'Leaf = Class 1.0 | {0.0: 0.005295278636481529, 1.0: 1.9947047213635185}\n'
expected_model_2 = 'Leaf = Class 1.0 | {0.0: 0.0052952786364815294, 1.0: 1.9947047213635185}\n'
expected_model_3 = 'Leaf = Class 1.0 | {1.0: 1.9947047213635185, 0.0: 0.0052952786364815294}\n'
assert (learner.get_model_description() == expected_model_1) \
or (learner.get_model_description() == expected_model_2) \
or (learner.get_model_description() == expected_model_3)
stream.restart()
X, y = stream.next_sample(5000)
learner = HAT(max_byte_size=30, leaf_prediction='mc', grace_period=10)
learner.partial_fit(X, y)
def test_concept_drift_stream(test_path):
stream = ConceptDriftStream(random_state=1, position=20, width=5)
stream.prepare_for_use()
assert stream.n_remaining_samples() == -1
expected_names = [
"salary", "commission", "age", "elevel", "car", "zipcode", "hvalue",
"hyears", "loan"
]
assert stream.feature_names == expected_names
expected_targets = [0, 1]
assert stream.target_values == expected_targets
assert stream.target_names == ['target']
assert stream.n_features == 9
assert stream.n_cat_features == 3
assert stream.n_num_features == 6
assert stream.n_targets == 1
assert stream.get_info() == 'ConceptDriftStream: ' \
'First Stream: AGRAWALGenerator - ' \
'Drift Stream: AGRAWALGenerator - ' \
'alpha: 0.0 - position: 20 - width: 5'
assert stream.has_more_samples() is True
assert stream.is_restartable() is True
# Load test data corresponding to first 10 instances
test_file = os.path.join(test_path, 'concept_drift_stream.npz')
data = np.load(test_file)
X_expected = data['X']
y_expected = data['y']
X, y = stream.next_sample()
assert np.alltrue(X[0] == X_expected[0])
assert np.alltrue(y[0] == y_expected[0])
X, y = stream.last_sample()
assert np.alltrue(X[0] == X_expected[0])
assert np.alltrue(y[0] == y_expected[0])
stream.restart()
X, y = stream.next_sample(30)
assert np.alltrue(X == X_expected)
assert np.alltrue(y == y_expected)
assert stream.n_targets == np.array(y).ndim
assert stream.n_features == X.shape[1]
assert 'stream' == stream.get_class_type()
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 = KNNADWINClassifier(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 = 'KNNADWINClassifier(leaf_size=40, max_window_size=200, n_neighbors=8, nominal_attributes=None)'
info = " ".join([line.strip() for line in learner.get_info().split()])
assert 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
def test_hat_nb(test_path):
stream = ConceptDriftStream(stream=SEAGenerator(random_state=1,
noise_percentage=0.05),
drift_stream=SEAGenerator(
random_state=2,
classification_function=2,
noise_percentage=0.05),
random_state=1,
position=250,
width=10)
stream.prepare_for_use()
learner = HAT(leaf_prediction='nb')
cnt = 0
max_samples = 1000
y_pred = array('i')
y_proba = []
wait_samples = 20
while cnt < max_samples:
X, y = stream.next_sample()
# Test every n samples
if (cnt % wait_samples == 0) and (cnt != 0):
y_pred.append(learner.predict(X)[0])
y_proba.append(learner.predict_proba(X)[0])
learner.partial_fit(X, y)
cnt += 1
expected_predictions = array('i', [
1, 0, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 0, 1, 1, 1, 1, 0,
1, 0, 0, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1,
1
])
assert np.alltrue(y_pred == expected_predictions)
test_file = os.path.join(test_path, 'test_hoeffding_adaptive_tree_nb.npy')
data = np.load(test_file)
assert np.allclose(y_proba, data)
expected_info = 'HAT: max_byte_size: 33554432 - memory_estimate_period: 1000000 - grace_period: 200' \
' - split_criterion: info_gain - split_confidence: 1e-07 - tie_threshold: 0.05' \
' - binary_split: False - stop_mem_management: False - remove_poor_atts: False' \
' - no_pre_prune: False - leaf_prediction: nb - nb_threshold: 0' \
' - nominal_attributes: [] - '
assert learner.get_info() == expected_info
assert type(learner.predict(X)) == np.ndarray
assert type(learner.predict_proba(X)) == np.ndarray
def test_hoeffding_adaptive_tree_nb(test_path):
stream = ConceptDriftStream(stream=SEAGenerator(random_state=1, noise_percentage=0.05),
drift_stream=SEAGenerator(random_state=2, classification_function=2,
noise_percentage=0.05),
random_state=1, position=250, width=10)
stream.prepare_for_use()
learner = HoeffdingAdaptiveTreeClassifier(leaf_prediction='nb')
cnt = 0
max_samples = 1000
y_pred = array('i')
y_proba = []
wait_samples = 20
while cnt < max_samples:
X, y = stream.next_sample()
# Test every n samples
if (cnt % wait_samples == 0) and (cnt != 0):
y_pred.append(learner.predict(X)[0])
y_proba.append(learner.predict_proba(X)[0])
learner.partial_fit(X, y)
cnt += 1
expected_predictions = array('i', [1, 0, 1, 1, 1, 1, 0, 1, 1, 1,
0, 1, 1, 1, 1, 1, 0, 1, 0, 1,
1, 1, 1, 0, 1, 0, 0, 1, 1, 0,
1, 1, 1, 1, 1, 0, 1, 0, 1, 1,
0, 1, 1, 1, 1, 1, 0, 1, 1])
assert np.alltrue(y_pred == expected_predictions)
test_file = os.path.join(test_path, 'test_hoeffding_adaptive_tree_nb.npy')
data = np.load(test_file)
assert np.allclose(y_proba, data)
expected_info = "HoeffdingAdaptiveTreeClassifier(binary_split=False, bootstrap_sampling=True, grace_period=200, " \
"leaf_prediction='nb', max_byte_size=33554432, memory_estimate_period=1000000, nb_threshold=0, " \
"no_preprune=False, nominal_attributes=None, remove_poor_atts=False, split_confidence=1e-07, " \
"split_criterion='info_gain', stop_mem_management=False, tie_threshold=0.05)"
info = " ".join([line.strip() for line in learner.get_info().split()])
assert info == expected_info
assert type(learner.predict(X)) == np.ndarray
assert type(learner.predict_proba(X)) == np.ndarray
class RecurringConceptGradualStream(RecurringConceptStream):
""" A stream featuring gradual drift between given concepts.
Uses the scikit-multiflow concept drift stream to blend concepts over
a window.
Parameters
----------
rctype: RCStreamType
An enum describing the type of stream
num_samples: int
The number of samples in the stream
noise: float
The probability that noise will happen in the generation. At each
new sample generated, the sample with will perturbed by the amount of
perturbation.
Values go from 0.0 to 1.0.
concept_chain: list<int> or dict
A dict with key observation number and value
the concept begining at that observation
or
A list of concept ids. A dict will be generated
with each concept lasting its length given in desc
or uniform length.
window_size: int
The number of observations each gradual drift is
spread over.
seed: int
Random seed.
desc: dict<int><conceptOccurence>
A map of concept ID to options
boost_first_occurance: bool
If true, double the observations drawn from
the first occurence of a concept. Allows
a better model to be built and stored.
Examples
--------
>>> # An example stream using the STAGGER Generator.
>>> # Starts using generating function 0, then at
>>> # observation 5000 transitions to generating function
>>> # 1 then at 10000 transitions back to 0.
>>> from skika.data.synthetic.reccurring_concept_stream import RCStreamType, RecurringConceptGradualStream, conceptOccurence
>>> concept_chain = {0: 0, 5000: 1, 10000: 0}
>>> num_samples = 15000
>>> # init concept
>>> concept_0 = conceptOccurence(id = 0, difficulty = 2, noise = 0,
appearences = 2, examples_per_appearence = 5000)
>>> concept_1 = conceptOccurence(id = 1, difficulty = 3, noise = 0,
appearences = 1, examples_per_appearence = 5000)
>>> desc = {0: concept_0, 1: concept_1}
>>> datastream = RecurringConceptGradualStream(
rctype = RCStreamType.STAGGER,
num_samples =num_samples,
noise = 0,
concept_chain = concept_chain,
window_size = 1000,
seed = 42,
desc = desc,
boost_first_occurance = False)
>>> datastream.has_more_samples()
True
>>> datastream.get_drift_info()
{0: 0, 5000: 1, 10000: 0}
>>> datastream.n_remaining_samples()
15000
>>> datastream.get_stream_info()
{0: 0, 5000: 1, 10000: 0}
0 - 5000: STAGGERGenerator(balance_classes=False, classification_function=0,
random_state=42)
5000 - 10000: STAGGERGenerator(balance_classes=False, classification_function=1,
random_state=43)
10000 - 15000: STAGGERGenerator(balance_classes=False, classification_function=0,
random_state=42)
>>> datastream.get_moa_stream_info()
{0: 0, 5000: 1, 10000: 0}
'(ConceptDriftStream -s (generators.STAGGERGenerator -f 1 -i 42) -d (ConceptDriftStream -s (generators.STAGGERGenerator -f 2 -i 43) -d (generators.STAGGERGenerator -f 1 -i 42) -p 5000 -w 1) -p 5000 -w 1)'
>>> datastream.get_supplementary_info()
>>> datastream.next_sample()
(array([[2., 0., 2.]]), array([0]))
>>> datastream.n_remaining_samples()
14999
>>> datastream.next_sample()
(array([[2., 0., 0.]]), array([0]))
>>> datastream.n_remaining_samples()
14998
"""
def __init__(self,
rctype,
num_samples,
noise,
concept_chain,
window_size=1000,
seed=None,
desc=None,
boost_first_occurance=True):
self.in_drift = False
self.drift_switch = False
self.window_size = window_size
self.transition_stream = None
super().__init__(rctype,
num_samples,
noise,
concept_chain,
seed=seed,
desc=desc,
boost_first_occurance=boost_first_occurance)
def next_sample(self, batch_size=1):
if batch_size > 1:
print("Only batch size of 1 for now")
return None
if not self.in_drift:
samples = self.concepts[self.current_concept].next_sample(
batch_size)
else:
samples = self.transition_stream.next_sample(batch_size)
last_switch_point = 0 - self.window_size // 2
next_switch_point = self.num_samples + self.window_size
self.example_count += batch_size
for concept_switch_index in sorted(self.concept_chain.keys()):
if (concept_switch_index <= self.example_count):
last_switch_point = concept_switch_index
if concept_switch_index >= self.example_count:
next_switch_point = concept_switch_index
break
self.drifted = False
if not self.in_drift:
if self.example_count >= next_switch_point - self.window_size // 2:
self.in_drift = True
self.drift_switch = True
self.transition_stream = ConceptDriftStream(
stream=self.concepts[self.concept_chain[last_switch_point]]
.get_datastream(),
drift_stream=self.concepts[self.concept_chain[
next_switch_point]].get_datastream(),
position=self.window_size // 2,
width=self.window_size)
self.transition_stream.prepare_for_use()
else:
if self.example_count == next_switch_point:
self.current_concept = self.concept_chain[next_switch_point]
self.drifted = True
self.drift_switch = False
if self.example_count >= (last_switch_point + self.window_size //
2) and not self.drift_switch:
self.in_drift = False
return samples
class RecurringConceptGradualStream(RecurringConceptStream):
def __init__(self,
rctype,
num_samples,
noise,
concept_chain,
window_size=1000,
seed=None,
desc=None,
boost_first_occurance=True):
self.in_drift = False
self.drift_switch = False
self.window_size = window_size
self.transition_stream = None
super().__init__(rctype,
num_samples,
noise,
concept_chain,
seed=seed,
desc=desc,
boost_first_occurance=boost_first_occurance)
def next_sample(self, batch_size=1):
if batch_size > 1:
print("Only batch size of 1 for now")
return None
if not self.in_drift:
samples = self.concepts[self.current_concept].next_sample(
batch_size)
else:
samples = self.transition_stream.next_sample(batch_size)
last_switch_point = 0 - self.window_size // 2
next_switch_point = self.num_samples + self.window_size
self.example_count += batch_size
for concept_switch_index in sorted(self.concept_chain.keys()):
if (concept_switch_index <= self.example_count):
last_switch_point = concept_switch_index
if concept_switch_index >= self.example_count:
next_switch_point = concept_switch_index
break
self.drifted = False
if not self.in_drift:
# print(f"START GRADUAL DRIFT FROM {self.concept_chain[last_switch_point]} TO {self.concept_chain[next_switch_point]}")
if self.example_count >= next_switch_point - self.window_size // 2:
# print(f"{self.example_count}: START GRADUAL DRIFT FROM {self.concept_chain[last_switch_point]} TO {self.concept_chain[next_switch_point]}")
self.in_drift = True
self.drift_switch = True
self.transition_stream = ConceptDriftStream(
stream=self.concepts[self.concept_chain[last_switch_point]]
.get_datastream(),
drift_stream=self.concepts[self.concept_chain[
next_switch_point]].get_datastream(),
position=self.window_size // 2,
width=self.window_size)
self.transition_stream.prepare_for_use()
else:
if self.example_count == next_switch_point:
self.current_concept = self.concept_chain[next_switch_point]
self.drifted = True
self.drift_switch = False
# print(f"{self.example_count}: SWITCH POINT")
if self.example_count >= (last_switch_point + self.window_size //
2) and not self.drift_switch:
self.in_drift = False
# print(f"{self.example_count}: END GRADUAL DRIFT FROM {self.concept_chain[last_switch_point]} TO {self.concept_chain[next_switch_point]}")
return samples
random_state=None,
alpha=0.0)
stream_3 = ConceptDriftStream(
stream=AGRAWALGenerator(balance_classes=False,
classification_function=1,
perturbation=0.0,
random_state=11),
drift_stream=AGRAWALGenerator(balance_classes=False,
classification_function=2,
perturbation=0.0,
random_state=12),
position=6000,
width=500,
random_state=None,
alpha=0.0)
stream_1.prepare_for_use()
stream_2.prepare_for_use()
stream_3.prepare_for_use()
instances_num = 10000
instances_counter = 0
ENSEMBLE_TYPE = 'av'
### Arrays for storing accuracy values for Streams
accuracies_1 = []
accuracies_2 = []
accuracies_3_mv = []
accuracies_3_av = []
accuracies_3_goowe = []
num_features = stream_1.n_features