def test_threshold_minmax(self):
sample_num = 10
feature_dim = 5
num_anomaly = 5
# actual value
y_test = np.zeros(sample_num * feature_dim)
gen_rand_indexes = [0, 7, 16, 33, 45]
y_test[gen_rand_indexes] = 10
y_test = y_test.reshape((sample_num, feature_dim))
td = ThresholdDetector()
td.set_params(threshold=(-1, 1))
td.fit(y_test)
anomaly_scores = td.score()
assert len(set(np.where(anomaly_scores > 0)[0])) == num_anomaly
anomaly_indexes = td.anomaly_indexes()
assert len(anomaly_indexes) == num_anomaly
def test_threshold_single(self):
sample_num = 10
feature_dim = 5
num_anomaly = 5
# predicted value
y_pred = np.full((sample_num, feature_dim), 0)
# actual value
y_test = np.full(sample_num * feature_dim, 0.2)
gen_rand_indexes = [0, 7, 16, 33, 45]
y_test[gen_rand_indexes] = 10
y_test = y_test.reshape((sample_num, feature_dim))
td = ThresholdDetector()
td.set_params(threshold=3)
td.fit(y_test, y_pred)
anomaly_scores = td.score()
assert len(set(np.where(anomaly_scores > 0)[0])) == num_anomaly
anomaly_indexes = td.anomaly_indexes()
assert len(anomaly_indexes) == num_anomaly
def test_mode_gaussian(self):
sample_num = 500
# actual value
y_test = np.full(sample_num, 2)
mu, sigma, ratio = 3, 0.1, 0.01
s = np.random.normal(mu, sigma, sample_num)
y_pred = y_test + s
td = ThresholdDetector()
td.set_params(mode="gaussian", ratio=ratio)
td.fit(y_test, y_pred)
# check estimated threshold
from scipy.stats import norm
assert abs(td.th - (norm.ppf(1 - ratio) * sigma + mu)) < 0.04
def test_fit_score(self):
look_back = 4
# generate dataframe
data = self.gen_data(feature_num=6, sample_num=100)
# split train and test dataframes
train_df, test_df = self.train_test_split(data, test_num=20, look_back=look_back)
# roll data to generate model input
x_train, y_train = self.roll_data(dataset=train_df, look_back=look_back,
target_col_indexes=[0])
x_test, y_test = self.roll_data(dataset=test_df, look_back=look_back,
target_col_indexes=[0])
# create model, train on train data and predict on test
lstm_config = {"lstm_units": [32] * 2, "lr": 0.001}
forecaster = LSTMForecaster(target_dim=1, feature_dim=x_train.shape[-1], **lstm_config)
forecaster.fit(x=x_train, y=y_train, batch_size=1024, epochs=50, distributed=False)
y_predict = forecaster.predict(x_test)
# find anomaly using a manual set threshold
td = ThresholdDetector()
td.set_params(threshold=10)
td.fit(y_test, y_predict)
anomaly_scores = td.score()
assert len(list(np.where(anomaly_scores > 0)[0])) == 0
anomaly_indexes = td.anomaly_indexes()
assert len(anomaly_indexes) == 0
# if threshold is not provided, ThresholdDetector can fit to the data
ratio = 0.1
td = ThresholdDetector()
td.set_params(ratio=ratio)
td.fit(y_test, y_predict)
fitted_anomaly_indexes = td.anomaly_indexes()
assert len(fitted_anomaly_indexes) == int(ratio * y_test.shape[0])
def test_corner_cases(self):
td = ThresholdDetector()
with pytest.raises(RuntimeError):
td.score()
with pytest.raises(RuntimeError):
td.anomaly_indexes()
time = np.arange(0, 1, 0.5)
y = np.sin(time)
td.set_params(mode="dummy")
with pytest.raises(ValueError):
td.fit(y, y)
td.set_params(mode="gaussian")
with pytest.raises(ValueError):
td.fit(y)
td.set_params(threshold="1")
with pytest.raises(ValueError):
td.fit(y)
td.set_params(threshold=(1, -1))
with pytest.raises(ValueError):
td.fit(y)
td.set_params(threshold=(np.array([-1]), np.array([-1])))
with pytest.raises(ValueError):
td.fit(y)
td.set_params(threshold=(np.array([1, 1]), np.array([-1, -1])))
with pytest.raises(ValueError):
td.fit(y)
def test_fit_score(self):
look_back = 4
# generate dataframe
data = self.gen_data(feature_num=6, sample_num=100)
# split train and test dataframes
train_df, test_df = self.train_test_split(data, test_num=20, look_back=look_back)
# roll data to generate model input
x_train, y_train = self.roll_data(dataset=train_df, look_back=look_back,
target_col_indexes=[0])
x_test, y_test = self.roll_data(dataset=test_df, look_back=look_back,
target_col_indexes=[0])
# create model, train on train data and predict on test
y_train = np.expand_dims(y_train, 1)
forecaster = LSTMForecaster(past_seq_len=look_back,
input_feature_num=x_train.shape[-1],
output_feature_num=1,
hidden_dim=32,
layer_num=2)
forecaster.fit(data=(x_train, y_train), batch_size=1024, epochs=50)
y_predict = forecaster.predict(x_test)
y_predict = np.squeeze(y_predict, axis=1)
# find anomaly using a manual set threshold
td = ThresholdDetector()
td.set_params(threshold=10)
td.fit(y_test, y_predict)
anomaly_scores = td.score()
assert len(list(np.where(anomaly_scores > 0)[0])) == 0
anomaly_indexes = td.anomaly_indexes()
assert len(anomaly_indexes) == 0
# if threshold is not provided, ThresholdDetector can fit to the data
ratio = 0.1
td = ThresholdDetector()
td.set_params(ratio=ratio)
td.fit(y_test, y_predict)
fitted_anomaly_indexes = td.anomaly_indexes()
assert len(fitted_anomaly_indexes) == int(ratio * y_test.shape[0])