def test_adaptive_forest():
test_data_directory = os.path.join(TEST_DIRECTORY, 'data')
test_file = os.path.join(
test_data_directory,
'test_data/weather.csv'
)
raw_data = pd.read_csv(test_file)
stream1 = DataStream(raw_data, name='Test')
stream2 = DataStream(raw_data, name='Test')
# learner = ExtendedHoeffdingAdaptiveTree()
# learner1 = AdaptiveHoeffdingTreeEnsemble(n_estimators=4)
# stream1_learner = calculate_accuracy(learner, stream1, stream1.n_samples)
# stream2_learner = calculate_accuracy(learner1, stream2, stream2.n_samples)
# stream1_learner = calculate_accuracy(learner, stream1, stream1.n_samples)
learner3 = AdaptiveRandomForestClassifier(n_estimators=10)
stream3_learner = calculate_accuracy(learner3, stream1, stream1.n_samples)
# learner4 = StreamingRandomPatchesClassifier(n_estimators=3)
# stream4_learner = calculate_accuracy(learner4, stream1, stream1.n_samples)
# learner5 = DeepStreamLearner(classes=stream1.target_values)
# stream5_learner = calculate_accuracy(learner5, stream1, stream1.n_samples)
import pudb; pudb.set_trace() # XXX BREAKPOINT
assert 1 == 1
# print(stream2_learner.base_estimator.accuracy)
with open (
os.path.join(test_data_directory, 'test_data/adaptive_test_result.txt'),
'+w'
) as f:
f.write('stream2 average_accuracy:')
import pudb; pudb.set_trace() # XXX BREAKPOINT
assert 1 == 1
def partial_fit(self, data: pd.DataFrame) -> None:
""" The most basic working version for now.
TODO: improve, calculate accuracy, maybe add possibility to learn in batches? """
stream = DataStream(self.prepare_data(data))
n = stream.n_remaining_samples()
for i in range(n):
x, y = stream.next_sample()
if self.model.predict(x)[0] == y[0]:
self.correct_predictions += 1
self.model.partial_fit(x, y)
self.predictions += n
self.accuracy = self.correct_predictions / self.predictions
def test_hoeffding_tree_ensemble():
test_data_directory = os.path.join(TEST_DIRECTORY, 'data')
test_file = os.path.join(
test_data_directory,
'test_data/airlines.csv'
)
test_file2 = os.path.join(
test_data_directory,
'test_data/weather.csv'
)
raw_data = pd.read_csv(test_file)
raw_data2 = pd.read_csv(test_file2)
stream3 = DataStream(raw_data2, name='Test2')
stream1 = DataStream(raw_data, name='Test')
stream2 = RandomTreeGenerator(
tree_random_state=23, sample_random_state=12, n_classes=4,
n_cat_features=2, n_num_features=5, n_categories_per_cat_feature=5,
max_tree_depth=6, min_leaf_depth=3, fraction_leaves_per_level=0.15
)
# learner = HoeffdingTreeClassifier(
# leaf_prediction='nb',
# classes=stream.target_values
# )
# learner = HoeffdingTreeEnsemble(
# n_estimators=3,
# classes=stream.target_values)
# learner1 = DeepStreamLearner(classes=stream1.target_values)
# stream1_learner = calculate_accuracy(learner1, stream1, stream1.n_samples)
learner2 = DeepStreamLearner(classes=stream2.target_values)
stream2_learner = calculate_accuracy(learner2, stream2, 100000)
learner3 = DeepStreamLearner(classes=stream3.target_values)
stream3_learner = calculate_accuracy(learner3, stream3, stream3.n_samples)
with open (
os.path.join(test_data_directory, 'test_data/test_result.txt'),
'+w'
) as f:
# f.write('stream1 accuracy: {} \n'.format(stream1_learner.accuracy[-1]))
# f.write('stream1 first_layer_accuracy: {} \n'.format(stream1_learner.first_layer_cascade.accuracy_per_sample[-1]))
# f.write('stream1 average_accuracy: {} \n'.format(sum(stream1_learner.accuracy)/stream1_learner.number_of_samples))
# f.write('stream1 first_layer_average_accuracy: {} \n \n'.format(sum(stream1_learner.first_layer_cascade.accuracy_per_sample)/stream1_learner.number_of_samples))
# f.write('stream2 accuracy: {} \n'.format(stream2_learner.accuracy[-1]))
# f.write('stream2 first_layer_accuracy: {} \n'.format(stream2_learner.first_layer_cascade.accuracy_per_sample[-1]))
f.write('stream2 average_accuracy: {} \n'.format(sum(stream2_learner.accuracy)/stream2_learner.number_of_samples))
f.write('stream2 first_layer_average_accuracy: {} \n \n'.format(sum(stream2_learner.first_layer_cascade.accuracy_per_sample)/stream2_learner.number_of_samples))
f.write('stream3 accuracy: {} \n'.format(stream3_learner.accuracy[-1]))
f.write('stream3 first_layer_accuracy: {} \n'.format(stream3_learner.first_layer_cascade.accuracy_per_sample[-1]))
f.write('stream3 average_accuracy: {} \n'.format(sum(stream3_learner.accuracy)/stream3_learner.number_of_samples))
f.write('stream3 first_layer_average_accuracy: {} \n \n'.format(sum(stream3_learner.first_layer_cascade.accuracy_per_sample)/stream3_learner.number_of_samples))
import pudb; pudb.set_trace() # XXX BREAKPOINT
assert 1 == 1
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 get_error_hoeffdingtree(data, pre_train_size, **hf_kwargs):
orig_X = data[:, :-1]
orig_y = data[:, -1].astype(int)
stream = DataStream(orig_X, orig_y)
hf = HoeffdingTreeClassifier(**hf_kwargs)
pretrainX, pretrainy = stream.next_sample(pre_train_size)
# Pre-train
hf.partial_fit(pretrainX, pretrainy, classes=stream.target_values)
evaluations = []
while stream.has_more_samples():
X, y = stream.next_sample()
# Evaluation
y_hat = hf.predict(X)
evaluations.append(int(y_hat[0] == y[0]))
# Train
hf.partial_fit(X, y, classes=stream.target_values)
return evaluations
def train(name, clusters, window, normalize=False):
input_csv = '{}{}_clusters={}_window={}_prepared.csv'.format(
DATA_LOCATION, name, clusters, window)
data = pd.read_csv(input_csv, index_col=0)
if normalize:
states = data.filter(['current_state', 'next_state'])
sensors = data.drop(columns=['current_state', 'next_state'])
scaler = StandardScaler()
data = pd.DataFrame(data=scaler.fit_transform(X=sensors),
index=data.index,
columns=sensors.columns)
data = pd.concat([data, states], axis='columns')
stream = DataStream(data)
hf = HoeffdingTreeClassifier()
sgd = SGDClassifier()
evaluator = EvaluatePrequential()
evaluator.evaluate(stream=stream, model=[hf, sgd])
# print('---------------------------------------------')
# measurements = evaluator.get_mean_measurements()[0]
# print(measurements.confusion_matrix)
# print(measurements.accuracy_score())
data = []
for i, measurements in enumerate(evaluator.get_mean_measurements()):
data.append([
name, clusters, window, MODEL_NAMES[i], normalize,
measurements.accuracy_score(),
measurements.precision_score(),
measurements.recall_score(),
measurements.f1_score()
])
return pd.DataFrame(data=data,
columns=[
'name', 'clusters', 'window', 'model',
'normalized', 'accuracy', 'precision', 'recall',
'f1'
])
df = pd.read_csv(filename, comment='#')
y = df['Target'].values
anoms = (y == "'Anomaly'")
normal = (y == "'Normal'")
y[anoms] = 1
y[normal] = 0
X = df.drop(["Target"], axis=1)
return X, y
n_samples = 5000
X, y = generate_data(n_samples)
# data_file="data/mulcross.csv"
# X,y=read_data(data_file)
stream = DataStream(data=X, y=y)
n_clusters = 3
random_state = 3
# 2. Prepare for use
stream.prepare_for_use()
kmeans1 = KMeans(n_clusters=n_clusters,
random_state=random_state,
init="k-means++")
kmeans2 = KMeans(n_clusters=n_clusters,
random_state=random_state,
init="k-means++")
coreset_size = 100
precisions = []
recalls = []
f1s = []
tprs = []
aucs = []
mean_fpr = np.linspace(0, 1, 100)
for fold, split in enumerate(cross_validation.split(X_train, y_train)):
fold_train_indexes, fold_test_indexes = split
fold_X_train = X_train.iloc[fold_train_indexes]
fold_y_train = y_train.iloc[fold_train_indexes]
fold_X_test = X_train.iloc[fold_test_indexes]
fold_y_test = y_train.iloc[fold_test_indexes]
if (classifier_name == 'hoeffding'):
stream = DataStream(X, y.values.ravel())
stream.prepare_for_use()
evaluator = EvaluatePrequential(
show_plot=False, pretrain_size=200, metrics=['accuracy'])
model = evaluator.evaluate(
stream=stream, model=classifier)[0]
model.fit(fold_X_train, fold_y_train.values.ravel())
# elif (classifier_name == 'cn2'):
# model = CrossValidation(
# table_from_frame(data), [CN2Learner()], k=5)
else:
model = classifier.fit(
fold_X_train, fold_y_train.values.ravel())
y_pred = model.predict(fold_X_test)
# Global variable
TRAINING_SIZE = 1
grace = 1000
ignore = 0
elec_data = arff.load("elecNormNew.arff")
elec_df = pandas.DataFrame(elec_data)
elec_df.columns = ['date', 'day', 'period', 'nswprice', 'nswdemand', 'vicprice', 'vicdemand', 'transfer', 'class']
mapping = {"day":{"1":1, "2":2, "3":3, "4":4, "5":5, "6":6, "7":7}, "class": {"UP": 0, "DOWN": 1}}
elec_df = elec_df.replace(mapping)
elec_full_df = pandas.concat([elec_df] * 200)
STREAM_SIZE = elec_full_df.shape[0]
elec_stream = DataStream(elec_full_df, name="elec")
elec_stream.prepare_for_use()
X_train, y_train = elec_stream.next_sample(TRAINING_SIZE)
ht = HoeffdingTreeClassifier()
ht.partial_fit(X_train, y_train)
n_global = ignore + TRAINING_SIZE # Cumulative Number of observations
d_ddm = 0
w_ddm = 0
TP_ddm = []
FP_ddm = []
RT_ddm = []
DIST_ddm = []
from skmultiflow.evaluation import EvaluatePrequential
from forget_tree import ForgetHATT
from skmultiflow.data import DataStream
import pandas as pd
from skmultiflow.trees import HATT
data_filepath = "../datasets/transient_chess.data"
labels_filepath = "../datasets/transient_chess.labels"
data = pd.read_csv(data_filepath, delimiter=" ")
labels = pd.read_csv(labels_filepath, delimiter=" ")
labels['y'] = labels['y'].astype('category')
stream = DataStream(data=data, y=labels)
stream.prepare_for_use()
evaluator = EvaluatePrequential(output_file="log.log",
show_plot=False,
metrics=['accuracy'],
max_time=60)
models = [
ForgetHATT(data_filepath,
labels_filepath,
forget_percentage=0,
delimiter=" "),
ForgetHATT(data_filepath,
labels_filepath,
forget_percentage=0.1,
delimiter=" "),
ForgetHATT(data_filepath,
def start_run(options):
if not os.path.exists(options.experiment_directory):
print('No Directory')
return
name = '-'.join([options.moa_learner, str(options.concept_limit), 'py'])
print(name)
datastream_filename = None
datastream_pickle_filename = None
fns = glob.glob(os.sep.join([options.experiment_directory, "*.ARFF"]))
print(fns)
for fn in fns:
if fn.split('.')[-1] == 'ARFF':
actual_fn = fn.split(os.sep)[-1]
fn_path = os.sep.join(fn.split(os.sep)[:-1])
print(actual_fn)
print(fn_path)
pickle_fn = f"{actual_fn.split('.')[0]}_concept_chain.pickle"
pickle_full_fn = os.sep.join([fn_path, pickle_fn])
csv_fn = f"{name}.csv"
csv_full_fn = os.sep.join([fn_path, csv_fn])
print(csv_full_fn)
if os.path.exists(pickle_full_fn):
skip_file = False
if os.path.exists(csv_full_fn):
if os.path.getsize(csv_full_fn) > 2000:
skip_file = True
if not skip_file:
datastream_filename = fn
datastream_pickle_filename = pickle_full_fn
break
else:
print('csv exists')
if datastream_filename == None:
print('Not datastream file')
return
print(datastream_filename)
bat_filename = f"{options.experiment_directory}{os.sep}{name}.{'bat' if not options.using_linux else 'sh'}"
if not os.path.exists(bat_filename) or True:
with open(f'{datastream_pickle_filename}', 'rb') as f:
concept_chain = pickle.load(f)
print(concept_chain)
concepts = sorted(list(concept_chain.keys()))
num_examples = concepts[-1] + (concepts[-1] - concepts[-2])
stream_string = moaLink.get_moa_stream_from_filename(
os.sep.join(datastream_filename.split(os.sep)[:-1]),
datastream_filename.split(os.sep)[-1])
moa_string = moaLink.make_moa_command(stream_string,
options.moa_learner,
options.concept_limit,
'int',
num_examples,
config.report_window_length,
options.experiment_directory,
is_bat=not options.using_linux)
moaLink.save_moa_bat(moa_string, bat_filename, not options.using_linux)
# datastream = None
t_start = process_time()
command = f"{bat_filename} {options.moa_location}"
print(command)
print(options.moa_learner)
if options.moa_learner != 'arf':
if options.using_linux:
subprocess.run(['chmod', '+x', bat_filename])
subprocess.run([bat_filename, options.moa_location])
else:
subprocess.run(command)
else:
datastream_filename = f"{os.sep.join(datastream_filename.split(os.sep)[:-1])}{os.sep}{datastream_filename.split(os.sep)[-1]}"
data = arff.loadarff(datastream_filename)
df = pd.DataFrame(data[0], dtype='float64')
df['y0'] = df['y0'].astype('int64')
# df["y0"] = df["y0"].astype('category')
print(df.info())
datastream = DataStream(df)
datastream.prepare_for_use()
print(datastream.target_values)
learner = AdaptiveRandomForest(n_estimators=int(options.concept_limit))
right = 0
wrong = 0
overall_log = []
while datastream.has_more_samples():
X, y = datastream.next_sample()
prediction = learner.predict(X)
is_correct = prediction[0] == y[0]
if is_correct:
right += 1
else:
wrong += 1
learner.partial_fit(X, y)
if (right + wrong) > 0 and (right + wrong) % 200 == 0:
overall_log.append((right + wrong, right / (right + wrong)))
print(f'ex: {right + wrong}, Acc: {right / (right + wrong)}\r',
end="")
overall = pd.DataFrame(overall_log, columns=['ex', 'overall_accuracy'])
overall.to_csv(f"{options.experiment_directory}{os.sep}{name}.csv")
print("")
print(f'Accuracy: {right / (right + wrong)}')
#fsm, system_stats, concept_chain, ds, stream_examples = fsmsys.run_fsm(datastream, options, suppress = True, name = name, save_checkpoint=True)
t_stop = process_time()
print("")
print("Elapsed time during the whole program in seconds:",
t_stop - t_start)
def start_run(options):
if not os.path.exists(options.experiment_directory):
print('No Directory')
return
datastream_filename = None
datastream_pickle_filename = None
fns = glob.glob(os.sep.join([options.experiment_directory, "*.ARFF"]))
print(fns)
# mm_options = ['rA', 'age', 'LRU', 'acc'] if options.memory_management == 'all' else [options.memory_management]
for fn in fns:
save_mm = options.memory_management
mm_options = [options.memory_management
] if options.memory_management != 'all' else [
"score", "rA", 'auc', "age", "LRU", 'acc', 'div'
]
mm_options = mm_options if options.memory_management != 'mine' else [
'auc', "score", "rA"
]
for mm in mm_options:
print(mm)
options.memory_management = mm
sys_name = 'system'
if options.drift_detector != 'adwin':
sys_name += f"{options.drift_detector}"
name = '-'.join([
sys_name,
str(options.noise),
str(options.concept_limit),
str(options.memory_management),
str(options.sensitivity),
str(options.window),
str(options.optimal_selection),
str(options.learner_str),
str(options.poisson),
str(options.seed),
str(options.optimal_drift),
str(options.similarity_measure),
str(options.merge_strategy),
str(options.merge_similarity)
])
name_no_seed = '-'.join([
sys_name,
str(options.noise),
str(options.concept_limit),
str(options.memory_management),
str(options.sensitivity),
str(options.window),
str(options.optimal_selection),
str(options.learner_str),
str(options.poisson), "*",
str(options.optimal_drift),
str(options.similarity_measure),
str(options.merge_strategy),
str(options.merge_similarity)
])
print(name)
if fn.split('.')[-1] == 'ARFF':
actual_fn = fn.split(os.sep)[-1]
fn_path = os.sep.join(fn.split(os.sep)[:-1])
print(actual_fn)
print(fn_path)
pickle_fn = f"{actual_fn.split('.')[0]}_concept_chain.pickle"
pickle_full_fn = os.sep.join([fn_path, pickle_fn])
csv_fn = f"{name}.csv"
csv_full_fn = os.sep.join([fn_path, csv_fn])
print(f"checking {csv_full_fn}")
concept_chain_exists = os.path.exists(pickle_full_fn)
if not options.no_chain and not concept_chain_exists:
print("No concept chain pickle file")
continue
skip_file = False
existing_matches = glob.glob(
os.sep.join([fn_path, f"{name_no_seed}.csv"]))
if len(existing_matches):
if any(
[os.path.getsize(x) > 2000 for x in existing_matches]):
skip_file = True
if not skip_file:
datastream_filename = fn
datastream_pickle_filename = pickle_full_fn
else:
print(f'{csv_full_fn} exists')
if datastream_filename == None:
print('Not datastream file')
continue
print(datastream_filename)
if not options.no_chain:
with open(f'{datastream_pickle_filename}', 'rb') as f:
concept_chain = pickle.load(f)
else:
concept_chain = None
with open(f"{options.experiment_directory}{os.sep}{name}_info.txt",
"w") as f:
f.write(
json.dumps(options.__dict__,
default=lambda o: '<not serializable>'))
f.write(
f"\n Git Commit: {subprocess.check_output(['git', 'rev-parse', '--short', 'HEAD']).strip()}"
)
try:
data = arff.loadarff(datastream_filename)
df = pd.DataFrame(data[0])
except Exception as e:
print(e)
print("trying csv")
df = pd.read_csv(datastream_filename, header=None)
for c_i, c in enumerate(df.columns):
if pd.api.types.is_string_dtype(df[c]):
print(f"Factoizing {c}")
print(pd.factorize(df[c])[0].shape)
df[c] = pd.factorize(df[c])[0]
# print(f"{c_i}: {len(df.columns) - 1}")
# if c_i == len(df.columns) - 1:
# print(f"converting {c}")
# df[c] = df[c].astype('category')
print(df.info())
datastream = DataStream(df)
datastream.concept_chain = concept_chain
print(concept_chain)
datastream.prepare_for_use()
t_start = process_time()
print(options.__dict__)
classifier = FSMClassifier(
concept_limit=options.concept_limit,
memory_management=options.memory_management,
learner=options.learner,
window=options.window,
sensitivity=options.sensitivity,
concept_chain=concept_chain,
optimal_selection=options.optimal_selection,
optimal_drift=options.optimal_drift,
rand_weights=options.rand_weights,
poisson=options.poisson,
similarity_measure=options.similarity_measure,
merge_strategy=options.merge_strategy,
use_clean=options.use_clean,
merge_similarity=options.merge_similarity,
drift_detector=options.drift_detector)
avg_memory, max_memory = evaluate_prequential.evaluate_prequential(
datastream,
classifier,
directory=options.experiment_directory,
name=name,
noise=options.noise,
seed=options.seed)
t_stop = process_time()
print("")
print("Elapsed time during the whole program in seconds:",
t_stop - t_start)
with open(
f"{options.experiment_directory}{os.sep}{name}_timer.txt",
"w") as f:
f.write(
f"Elapsed time during the whole program in seconds: {t_stop-t_start}"
)
with open(
f"{options.experiment_directory}{os.sep}{name}_memory.txt",
"w") as f:
f.write(f"Average: {avg_memory}\n")
f.write(f"Max: {max_memory}")
options.memory_management = save_mm
options.memory_management = save_mm
X = tdf[["Pressure (millibars)", "Humidity",
"Wind Speed (km/h)"]].resample("6H").mean()
y = tdf[["Temperature (C)"]].resample("6H").max()
X.plot(subplots=True, layout=(1, 3))
y.plot()
#%%
reload(samknnreg)
from samknnreg import SAMKNNRegressor
sam = SAMKNNRegressor()
hat = RegressionHAT()
rht = RegressionHoeffdingTree()
ds = DataStream(X, y=y)
ds.prepare_for_use()
evaluator = EvaluatePrequential(
show_plot=True,
n_wait=730,
batch_size=28,
metrics=['mean_square_error', 'true_vs_predicted'])
#%%
evaluator.evaluate(stream=ds,
model=[sam, rht, hat],
model_names=[
"SAM", "Hoeffding Tree Regressor",
"Hoeffding Tree Regressor (Adaptive)"
])
def start_run(options):
if not os.path.exists(options.experiment_directory):
print('No Directory')
return
name = '-'.join([
options.moa_learner,
str(options.concept_limit), 'pyn',
str(options.seed)
])
print(name)
datastream_filename = None
datastream_pickle_filename = None
fns = glob.glob(os.sep.join([options.experiment_directory, "*.ARFF"]))
print(fns)
for fn in fns:
if fn.split('.')[-1] == 'ARFF':
actual_fn = fn.split(os.sep)[-1]
fn_path = os.sep.join(fn.split(os.sep)[:-1])
print(actual_fn)
print(fn_path)
csv_fn = f"{name}.csv"
csv_full_fn = os.sep.join([fn_path, csv_fn])
print(csv_full_fn)
skip_file = False
if os.path.exists(csv_full_fn):
if os.path.getsize(csv_full_fn) > 2000:
skip_file = True
if not skip_file:
datastream_filename = fn
break
else:
print('csv exists')
if datastream_filename == None:
print('Not datastream file')
return
print(datastream_filename)
datastream_filename = f"{os.sep.join(datastream_filename.split(os.sep)[:-1])}{os.sep}{datastream_filename.split(os.sep)[-1]}"
data = arff.loadarff(datastream_filename)
df = pd.DataFrame(data[0])
print(df.tail())
for c in df.columns:
print(f"Factoizing {c}")
if pd.api.types.is_string_dtype(df[c]):
print(pd.factorize(df[c])[0].shape)
df[c] = pd.factorize(df[c])[0]
print(df.tail())
bat_filename = f"{options.experiment_directory}{os.sep}{name}.{'bat' if not options.using_linux else 'sh'}"
if not os.path.exists(bat_filename) or True:
num_examples = df.shape[0]
stream_string = moaLink.get_moa_stream_from_filename(
os.sep.join(datastream_filename.split(os.sep)[:-1]),
datastream_filename.split(os.sep)[-1])
moa_string = moaLink.make_moa_command(stream_string,
options.moa_learner,
options.concept_limit,
'int',
num_examples,
config.report_window_length,
options.experiment_directory,
is_bat=not options.using_linux,
name=name,
num_features=len(df.columns) - 1)
moaLink.save_moa_bat(moa_string, bat_filename, not options.using_linux)
# datastream = None
t_start = process_time()
command = f'{bat_filename} "{options.moa_location}"'
print(command)
print(options.moa_learner)
if options.moa_learner != 'arf' or options.use_moa:
if options.using_linux:
subprocess.run(['chmod', '+x', bat_filename])
subprocess.run([bat_filename, options.moa_location])
else:
subprocess.run(command)
else:
# df['y0'] = df['y0'].astype('int64')
# df["y0"] = df["y0"].astype('category')
print(df.info())
datastream = DataStream(df)
datastream.prepare_for_use()
print(datastream.target_values)
learner = AdaptiveRandomForest(n_estimators=int(options.concept_limit))
avg_memory, max_memory = evaluate_prequential(
datastream=datastream,
classifier=learner,
directory=options.experiment_directory,
name=name)
# right = 0
# wrong = 0
# overall_log = []
# while datastream.has_more_samples():
# X,y = datastream.next_sample()
# prediction = learner.predict(X)
# is_correct = prediction[0] == y[0]
# if is_correct:
# right += 1
# else:
# wrong += 1
# learner.partial_fit(X, y)
# if (right + wrong) > 0 and (right + wrong) % 200 == 0:
# overall_log.append((right+ wrong, right / (right + wrong)))
# print(f'ex: {right + wrong}, Acc: {right / (right + wrong)}\r', end = "")
# overall = pd.DataFrame(overall_log, columns = ['ex', 'overall_accuracy'])
# overall.to_csv(f"{options.experiment_directory}{os.sep}{name}.csv")
# print("")
# print(f'Accuracy: {right / (right + wrong)}')
#fsm, system_stats, concept_chain, ds, stream_examples = fsmsys.run_fsm(datastream, options, suppress = True, name = name, save_checkpoint=True)
t_stop = process_time()
print("")
print("Elapsed time during the whole program in seconds:",
t_stop - t_start)
with open(f"{options.experiment_directory}{os.sep}{name}_timer.txt",
"w") as f:
f.write(
f"Elapsed time during the whole program in seconds: {t_stop-t_start}"
)
warnings.filterwarnings('ignore')
plt.style.use("seaborn-whitegrid")
# Global variable
TRAINING_SIZE = 1
grace = 1000
ignore = 0
weather_data = arff.load('weatherAUS.arff')
weather_df = pandas.DataFrame(weather_data)
weather_full_df = pandas.concat([weather_df] * 150)
STREAM_SIZE = weather_full_df.shape[0]
weather_stream = DataStream(weather_full_df, name="weather")
weather_stream.prepare_for_use()
X_train, y_train = weather_stream.next_sample(TRAINING_SIZE)
ht = HoeffdingTreeClassifier()
ht.partial_fit(X_train, y_train)
n_global = ignore + TRAINING_SIZE # Cumulative Number of observations
d_ddm = 0
w_ddm = 0
TP_ddm = []
FP_ddm = []
RT_ddm = []
DIST_ddm = []
def make_stream(path):
data = prepare_data(path)
stream = DataStream(data, y=None, target_idx=-1, n_targets=1, cat_features=None, name=None, allow_nan=False)
stream = stream.y
return stream
def test_regressor_chains():
X_reg, y_reg = make_regression(random_state=112,
n_targets=3,
n_samples=5150)
stream = DataStream(X_reg, y_reg)
estimator = SGDRegressor(random_state=112, max_iter=10)
learner = RegressorChain(base_estimator=estimator, random_state=112)
X, y = stream.next_sample(150)
learner.partial_fit(X, y)
cnt = 0
max_samples = 5000
predictions = []
true_labels = []
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(list(learner.predict(X)[0]))
true_labels.append(y[0])
learner.partial_fit(X, y)
cnt += 1
expected_predictions = [
[-21.932581119953333, 1265662295936.5574, 7.5406725414072326e+22],
[-97.17297744582125, 5438576501559.791, -1.1370581201037737e+24],
[-60.06308622605051, 26421144038311.047, 1.3207650552720094e+25],
[-285.32687352244847, 8881551118262.033, -1.1322856827798374e+24],
[-115.80322693771457, -24997431307818.508, 2.85747306174037e+24],
[-12.184193815918672, 3510562166726.0283, -4.8590562435597834e+23],
[-94.99008392491476, 4794062761133.606, -1.8849188211946465e+24],
[66.35576182871232, -8147485653396.883, -7.492944375995595e+23],
[-52.145505628056995, -1013810481101.9043, -4.5310283013446384e+23],
[16.715060622072958, 562391244392.6193, 3.3789644409962397e+22],
[96.32219400190282, -20397346086007.85, 1.558245298240083e+24],
[-281.8168065846582, 118681520215938.52, 4.815807486956294e+25],
[-135.62679760307105, 20260866750185.832, 1.605753540523006e+24],
[0.07932047636460954, -708539394047.3298, -3.61482684929158e+22],
[-292.1646176261883, -11162615183157.55, -8.674643964570704e+23],
[-176.92746747754094, -29231218161585.13, 1.411600743825668e+24],
[-348.0498644784687, -100615393132365.25, 9.759683002046948e+23],
[30.948974669258675, -1199287119275.6328, 2.0866927007519847e+23],
[214.0020659569134, -24437173206276.543, 9.450880718880671e+23],
[153.98931593720746, 32675842205528.723, -1.7246747286222668e+24],
[99.39074016354951, -11385065116243.611, 1.0770253102805811e+24],
[127.81660709796127, 16929726964275.697, 7.14820947257164e+24],
[40.45505653639006, -14311951591200.725, -9.33193290094133e+23],
[117.52219878440611, 17952367624051.36, 4.5651719663788677e+23],
[75.53942801239991, -9231543699137.594, 3.2317133158453914e+24],
[31.795193207760704, -4084783706153.4004, -4.188095047309216e+23],
[68.5318978502461, 5735810247065.921, 1.7284713503779943e+24],
[65.18438567482129, -13298743450357.943, -1.4367047198923567e+24],
[-116.63952028337805, -344127767223.9295, 2.3925104169428623e+22],
[-76.81599010889556, 8711205431447.733, -1.1575305916673031e+24],
[263.1077717649874, 32146618104196.434, -7.240279466740839e+24],
[-94.07597099457413, -8216681977657.527, 2.3785728690780553e+24],
[-175.78429788635424, -368856885004.46, -5.7200993095587195e+22],
[59.648477499483285, -1752783828320.242, 2.1429953624557326e+23],
[71.68447202426032, -27151271800666.492, 9.367463190825582e+24],
[-189.96629636835922, -27090727476080.18, -3.8659883994544866e+24],
[-240.7920206809074, 15406047062899.537, 2.0609123388035027e+24],
[-105.80996634043589, -1518636404558.1646, -1.4166487855869706e+23],
[-164.02527753963858, -61386039046571.125, -2.179071650432624e+25],
[52.451759456657975, -988509747123.6125, -7.334899319683594e+22],
[68.37044139814127, -7434200892467.581, -7.535677215142279e+23],
[164.9457843624521, -9474550940989.51, -1.3512944635293625e+24],
[189.34401690407307, -14349556896444.508, 1.0732760415617274e+24],
[0.8944005517286119, 463945767759.78735, -1.9938544157612443e+22],
[71.7856433565235, -9804063257174.584, 4.7874862540754335e+23],
[-5.450502769025279, 281585481223.33276, 2.1974700575843552e+22],
[248.00190755589915, -81874135462745.58, -2.6532557110860303e+25],
[-113.86249490223707, 2634310697909.643, 1.580428629322546e+23],
[-35.92856878407447, -5410985463428.589, 2.522168862637753e+23]
]
print(predictions)
assert np.allclose(
np.array(predictions).all(),
np.array(expected_predictions).all())
assert type(learner.predict(X)) == np.ndarray
expected_info = "RegressorChain(base_estimator=SGDRegressor(max_iter=10, random_state=112), " \
"order=None, random_state=112)"
info = " ".join([line.strip() for line in learner.get_info().split()])
assert info == expected_info
def increment_model(ht_regressor):
try:
start_time = time.time()
# val_df = pd.read_sql(engine.execute("select * from consumption where integrated = 0 limit 0,10").statement,session.bind)
logging.info("[ML - modIncrement] Loading data... Time: " +
str(round(time.time() - start_time, 2)))
val_df = pd.read_sql(
session.query(Consumption).filter(
Consumption.integrated == False).limit(2000000).statement,
session.bind)
logging.info("[ML - modIncrement] Data loaded... Time: " +
str(round(time.time() - start_time, 2)))
n_samples = 0
cnter = 0
client_ids = []
logging.info(
"[ML - modIncrement] Starting model incremental fitting... Time: "
+ str(round(time.time() - start_time, 2)))
client_id_max = max(val_df.client_id.unique())
client_id_min = min(val_df.client_id.unique())
df = val_df.drop(
columns=['id', 'client_id', 'year', 'month', 'integrated'])
stream = DataStream(data=df, target_idx=0)
plr = []
plprev_ht = []
while stream.has_more_samples():
X, y = stream.next_sample()
if (cnter % 7000 == 0):
y_prev = ht_regressor.predict(X)
plr.append(y)
plprev_ht.append(y_prev)
ht_regressor.partial_fit(X, y)
if (cnter % 10000 == 0):
logging.info("[ML - modIncrement] Extracting element #" +
str(cnter) + " Time: " +
str(round(time.time() - start_time, 2)))
n_samples += 1
cnter += 1
fig, ax = plt.subplots(figsize=(15, 6))
plt.plot(range(len(plr)), plr, 'b-', label='Real')
plt.plot(range(len(plprev_ht)),
plprev_ht,
'g--',
label='HoeffdingTreeRegressor')
plt.legend()
mse = mean_squared_error(plr, plprev_ht)
r2 = r2_score(plr, plprev_ht)
plt.suptitle(client_id_max, fontsize=12)
plt.title("R2: " + str(r2) + " MSE: " + str(mse))
filename = "images/predictionHT12F" + str(r2) + ".png"
plt.savefig(filename)
plt.close()
#Updating
logging.info("[ML - modIncrement] Execution %d --- %s seconds ---" %
(cnter, round(time.time() - start_time, 2)))
return ht_regressor, client_id_min, client_id_max
except:
logging.error("[ML - modIncrement] Stopping...")
