def evaluate_sa_cnn2d(x_train: List, x_test: List, y_test: np.array) -> Dict:
sc = CustomMinMaxScaler()
x_test = sc.fit(x_train).transform(x_test)
training_stats = load_experiment('../../models/sa_cnn2d/experiments.json')
score = evaluate(x_test, y_test, training_stats['experiments'])
return score
def train_aetcnn(x_train: List, x_test: List, y_train: np.array,
y_test: np.array) -> Dict:
sc = CustomMinMaxScaler()
x_train = sc.fit_transform(x_train)
x_test = sc.transform(x_test)
model = AETCN()
n_experiments = 100
embeddings_dim = x_train[0].shape[1]
params = {
'epochs':
np.random.choice(np.arange(1, 10), size=n_experiments).tolist(),
'learning_rate':
np.random.choice(10**np.linspace(-4, -0.5),
size=n_experiments).tolist(),
'batch_size':
np.random.choice([2**i for i in range(3, 8)],
size=n_experiments).tolist(),
'input_shape': [embeddings_dim] * n_experiments,
'layers':
generate_layer_settings(embeddings_dim, n_experiments),
'kernel_size':
np.random.choice([2 * i + 1 for i in range(1, 6)],
size=n_experiments).tolist(),
'window':
np.random.randint(10, 100, size=n_experiments).tolist(),
'dropout':
np.random.uniform(0, 0.5, size=n_experiments).tolist()
}
evaluated_hyperparams = random_search(
(x_train[y_train == 0], x_test, None, y_test), model, params)
return evaluated_hyperparams
def get_extracted_features(x_train: List, x_val: List, x_test: List,
y_val: np.array):
sc = CustomMinMaxScaler()
x_train = sc.fit_transform(x_train)
x_val = sc.transform(x_val)
x_test = sc.transform(x_test)
model = torch.load(
'../../models/aetcn/5d9ad591-6d3c-428f-894f-02af96ca1930.pt')
y_pred = model.predict(x_val) # return reconstruction errors
train_features = model.extract_features(x_train).astype(dtype=np.float32)
val_features = model.extract_features(x_val).astype(dtype=np.float32)
test_features = model.extract_features(x_test).astype(dtype=np.float32)
theta, f1 = find_optimal_threshold(y_val, y_pred)
y_pred = classify(y_pred, theta)
metrics_report(y_val, y_pred)
return train_features, val_features, test_features
def train_sa_cnn1d(x_train: List, x_test: List, y_train: np.array,
y_test: np.array) -> Dict:
sc = CustomMinMaxScaler()
x_train = sc.fit_transform(x_train)
x_test = sc.transform(x_test)
model = SACNN1D()
n_experiments = 100
embeddings_dim = x_train[0].shape[1]
encoder_kernel_sizes = np.random.choice([2 * i + 1 for i in range(1, 4)],
size=n_experiments).tolist()
layers = generate_layer_settings(embeddings_dim, n_experiments)
params = {
'epochs':
np.random.choice(np.arange(1, 10), size=n_experiments).tolist(),
'learning_rate':
np.random.choice(10**np.linspace(-4, -0.5),
size=n_experiments).tolist(),
'batch_size':
np.random.choice([2**i for i in range(3, 8)],
size=n_experiments).tolist(),
'input_shape': [embeddings_dim] * n_experiments,
'layers':
layers,
'encoder_kernel_size':
encoder_kernel_sizes,
'decoder_kernel_size':
np.random.choice([2 * i + 1 for i in range(2, 7)],
size=n_experiments).tolist(),
'encoder_heads':
get_encoder_heads(layers),
'decoder_heads':
get_decoder_heads(layers),
'window':
get_1d_window_size(encoder_kernel_sizes, layers, get_encoder_size),
'dropout':
np.random.uniform(0, 0.3, size=n_experiments).tolist()
}
evaluated_hyperparams = random_search(
(x_train[y_train == 0], x_test, None, y_test), model, params)
return evaluated_hyperparams
def train_transformer(x_train: List, x_test: List, y_train: np.array,
y_test: np.array) -> Dict:
sc = CustomMinMaxScaler()
x_train = sc.fit_transform(x_train)
x_test = sc.transform(x_test)
model = TransformerAutoEncoder()
n_experiments = 100
embeddings_dim = x_train[0].shape[1]
divisors = get_all_divisors(embeddings_dim)
params = {
'epochs':
np.random.choice(np.arange(1, 5), size=n_experiments).tolist(),
'learning_rate':
np.random.choice(10**np.linspace(-4, -0.5),
size=n_experiments).tolist(),
'batch_size':
np.random.choice([2**i for i in range(3, 8)],
size=n_experiments).tolist(),
'input_dim': [embeddings_dim] * n_experiments,
'heads':
np.random.choice(divisors,
size=n_experiments,
p=get_normal_dist(divisors)).tolist(),
'n_encoders':
np.random.randint(1, 5, size=n_experiments).tolist(),
'n_decoders':
np.random.randint(1, 5, size=n_experiments).tolist(),
'dim_feedforward':
np.random.randint(100, 2000, size=n_experiments).tolist(),
'window':
np.random.randint(10, 100, size=n_experiments).tolist(),
'dropout':
np.random.uniform(0, 0.5, size=n_experiments).tolist()
}
evaluated_hyperparams = random_search(
(x_train[y_train == 0], x_test, None, y_test), model, params)
return evaluated_hyperparams
def train_window(x_train: List, x_test: List, y_train: np.array,
y_test: np.array) -> Dict:
sc = CustomMinMaxScaler()
x_train = sc.fit_transform(x_train)
x_test = sc.transform(x_test)
scores = []
for w in range(1, 50, 2):
print('Window:', w)
model = VanillaTCN(epochs=1, window=w)
model.fit(x_train[y_train == 0])
y_pred = model.predict(x_test) # return reconstruction errors
theta, f1 = find_optimal_threshold(y_test, y_pred)
y_pred = classify(y_pred, theta)
metrics_report(y_test, y_pred)
scores.append(
create_experiment_report(get_metrics(y_test, y_pred),
{'window': w}))
create_checkpoint(
{'experiments': scores},
'../../models/TCN-cropped-window-embeddings-HDFS1.json')
return {'experiments': scores}
# not used currently
"model_path": "../../models/aetcn/4f5f4682-1ca5-400a-a340-6243716690c0.pt",
"threshold": 0.00331703620031476
}
X = load_pickle_file(
'../../data/processed/HDFS1/X-val-HDFS1-cv1-1-block.npy')[:1000]
y = np.load('../../data/processed/HDFS1/y-val-HDFS1-cv1-1-block.npy')[:1000]
# list of matrices, a matrix == blk_id -> NumPy [[005515, ...], ...] (n_logs x 100)
# F1 = (2 * r * p) / (r + p)
n_examples = 700
sc = CustomMinMaxScaler() # range 0 -- 1
x_train = sc.fit_transform(X[:n_examples])
y_train = y[:n_examples]
x_test = sc.transform(X[n_examples:])
y_test = y[n_examples:]
model = AETCN()
model.set_params(**config['hyperparameters'])
model.fit(x_train[y_train == 0]) # 0 -> normal, 1 -> anomaly
y_pred = model.predict(x_test) # return reconstruction errors
theta, f1 = find_optimal_threshold(y_test, y_pred)
y_pred = classify(y_pred, theta)
metrics_report(y_test, y_pred)
confusion_matrix(y_test, y_pred)