def test_adaptive_random_forests_nb():
stream = RandomTreeGenerator(tree_random_state=112,
sample_random_state=112,
n_classes=2)
stream.prepare_for_use()
learner = AdaptiveRandomForest(n_estimators=3,
random_state=112,
leaf_prediction='nb')
X, y = stream.next_sample(150)
learner.partial_fit(X, y)
cnt = 0
max_samples = 5000
predictions = []
true_labels = []
wait_samples = 100
correct_predictions = 0
while cnt < max_samples:
X, y = stream.next_sample()
# Test every n samples
if (cnt % wait_samples == 0) and (cnt != 0):
predictions.append(int(learner.predict(X)[0]))
true_labels.append(y[0])
if np.array_equal(y[0], predictions[-1]):
correct_predictions += 1
learner.partial_fit(X, y)
cnt += 1
last_version_predictions = [
1, 1, 0, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 0, 1, 1, 0, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 1, 0, 0, 0, 1, 1, 1, 0, 0,
1
]
# Performance below does not need to be guaranteed. This check is set up so that anything that changes
# to predictions are caught in the unit test. This helps prevent accidental changes.
assert type(learner.predict(X)) == np.ndarray
assert np.alltrue(predictions == last_version_predictions)
expected_info = "AdaptiveRandomForest(binary_split=False, disable_weighted_vote=False,\n" \
" drift_detection_method=ADWIN(delta=0.001), grace_period=50,\n" \
" lambda_value=6, leaf_prediction='nb',\n" \
" max_byte_size=33554432, max_features=5,\n" \
" memory_estimate_period=2000000, n_estimators=3,\n" \
" nb_threshold=0, no_preprune=False, nominal_attributes=None,\n" \
" performance_metric='acc', random_state=112,\n" \
" remove_poor_atts=False, split_confidence=0.01,\n" \
" split_criterion='info_gain', stop_mem_management=False,\n" \
" tie_threshold=0.05,\n" \
" warning_detection_method=ADWIN(delta=0.01))"
assert learner.get_info() == expected_info
def test_adaptive_random_forests():
stream = RandomTreeGenerator(tree_random_state=112, sample_random_state=112)
stream.prepare_for_use()
learner = AdaptiveRandomForest(n_estimators=3,
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
correct_predictions = 0
while cnt < max_samples:
X, y = stream.next_sample()
# Test every n samples
if (cnt % wait_samples == 0) and (cnt != 0):
predictions.append(int(learner.predict(X)[0]))
true_labels.append(y[0])
if np.array_equal(y[0], predictions[-1]):
correct_predictions += 1
learner.partial_fit(X, y)
cnt += 1
last_version_predictions = [1, 1, 1, 1, 1, 0, 0, 1, 0, 1, 0, 0, 1, 0, 1, 0,
1, 1, 1, 0, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0,
1, 1, 0, 1, 0, 1, 0, 0, 0, 0, 1, 0, 1, 1, 0, 0, 0]
# Performance below does not need to be guaranteed. This check is set up so that anything that changes
# to predictions are caught in the unit test. This helps prevent accidental changes.
# If these tests fail, make sure that what is worked on *should* change the predictions of ARF.
if sys.version_info.major == 3 and sys.version_info.minor >= 6:
# Temporary disable as pre-3.6 give different predictions than 3.6+
assert np.alltrue(predictions == last_version_predictions)
assert type(learner.predict(X)) == np.ndarray
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)
class ARFPredictor(AbstractPredictor):
def __init__(self, n_estimators=10, max_features='auto', grace_period=50):
super(ARFPredictor, self).__init__()
self._n_estimators = n_estimators
self._max_features = max_features
self._classifier = AdaptiveRandomForest(
n_estimators=self._n_estimators,
max_features=self._max_features,
grace_period=grace_period,
random_state=42)
self._trained_samples = 0
def predict(self, training_data, training_labels, todays_features,
training_years, trading_days_per_year) -> int:
self._classifier.partial_fit(training_data[self._trained_samples:],
training_labels[self._trained_samples:],
np.unique(training_labels))
self._trained_samples += len(training_labels[self._trained_samples:])
prediction = self._classifier.predict([todays_features])
return prediction[-1]
def tune(self,
stock_data,
symbols,
num_features,
measure='f1',
trading_frequency=10,
training_years=3,
trading_days_per_year=246):
gp_space = [10, 15, 20]
n_est_space = [50, 100]
sqrt_feat = sqrt(num_features)
max_f_space = [round(sqrt_feat)]
accuracies = []
f_scores = []
params = []
for gp in gp_space:
for n_estimators in n_est_space:
for max_features in max_f_space:
accs = []
fs = []
for stock in symbols:
print(stock)
self._n_estimators = n_estimators
self._max_features = max_features
self._trained_samples = 0
self._classifier = AdaptiveRandomForest(
n_estimators=self._n_estimators,
max_features=self._max_features,
grace_period=gp,
random_state=42)
preds = []
X = stock_data[stock][[
'I{}'.format(x)
for x in range(1, num_features + 1)
]]
y = stock_data[stock]['target']
first_trading_day = training_years * trading_days_per_year + trading_frequency
last_trading_day = first_trading_day - trading_frequency
for trading_day in range(first_trading_day, len(y)):
sys.stdout.write('\r%i/%i' % (trading_day, len(y)))
sys.stdout.flush()
preds.append(
self.predict(X.iloc[:last_trading_day].values,
y.iloc[:last_trading_day].values,
X.iloc[trading_day],
training_years,
trading_days_per_year))
last_trading_day = trading_day
print('')
preds = pd.Series(preds, name='Predictions')
accuracy = metrics.accuracy_score(
y.iloc[first_trading_day:], preds)
accs.append(accuracy)
f_score = metrics.f1_score(y.iloc[first_trading_day:],
preds)
fs.append(f_score)
mean_acc = np.array(accs).mean()
accuracies.append(mean_acc)
mean_f_score = np.array(fs).mean()
f_scores.append(mean_f_score)
params.append((n_estimators, max_features, gp))
print('\nARF n_estimators=%i max_features=%i gp=%i' %
(n_estimators, max_features, gp))
print('ARF Accuracy: %.3f' % mean_acc)
print('ARF F1 Score: %.3f' % mean_f_score)
accuracies = np.array(accuracies)
f_scores = np.array(f_scores)
print('\nARF FINAL RESULTS')
if measure == 'accuracy':
(n_estimators, max_features, gp) = params[accuracies.argmax()]
print(
'ARF Best result: accuracy %.3f (n_estimators=%i max_features=%i gp=%i)'
% (accuracies.max(), n_estimators, max_features, gp))
elif measure == 'f1':
(n_estimators, max_features, gp) = params[f_scores.argmax()]
print(
'ARF Best result: f score %.3f (n_estimators=%i max_features=%i gp=%i)'
% (f_scores.max(), n_estimators, max_features, gp))
else:
raise NotImplementedError
return [n_estimators, max_features]