def run_callable(run_config: dict):
experiment.print('Loading data')
bearing_dataset = load_data(run_config['data_set'])
x_train = bearing_dataset.train(column=config['data_column'], as_numpy=True)
experiment.print('Initializing optimizer')
experiment.print(f'max_iterations = {config["max_iterations"]}')
optimizer = create_optimizer(
run_config['optimizer'],
config_space=config_space,
model=KerasModel(
create_model_function=create_deep_easing_feed_forward_autoencoder,
evaluation_function=load_from_module(run_config['evaluation_function']),
),
x=x_train,
max_iterations=config['max_iterations'],
min_budget=config['min_budget'],
max_budget=config['max_budget'],
run_id=experiment.run_id,
validation_split=config['validation_split'],
)
try:
experiment.print('Optimizing')
optimization_result = optimizer.optimize()
except:
optimization_result = OptimizationResult(optimizer.evaluation_result_aggregator.get_evaluation_results())
experiment.log('search/error', traceback.format_exc())
experiment.print('Logging optimization results')
experiment.log_optimization_result(optimization_result)
log_model_configs = [
{'key': 'best', 'percentile_rank': 1.0},
{'key': 'average', 'percentile_rank': 0.5},
{'key': 'worst_10th', 'percentile_rank': 0.1},
{'key': 'worst', 'percentile_rank': 0.0},
]
for log_model_config in log_model_configs:
keras_model = cast(KerasModel, optimizer.refit_by_percentile_rank(log_model_config['percentile_rank']))
experiment.log_keras_model(keras_model.model, key=log_model_config['key'])
for threshold_percentile in config['threshold_percentiles']:
experiment.log_keras_predictions(
keras_model,
bearing_dataset.as_dict(config['data_column']),
key=f'{log_model_config["key"]}_{threshold_percentile}',
labels=bearing_dataset.as_dict(config['data_column'], labels_only=True),
has_multiple_features=True,
threshold_percentile=threshold_percentile,
)
def run_callable(run_config: dict):
def modifier(x):
return x[x.rpm > 0]
def pre_processing(data_frame):
if run_config['scaling'] == 'min_max':
samples = Scaler(MinMaxScaler, fit_mode=run_config['fit_mode']).fit_transform(data_frame.to_numpy())
else:
samples = data_frame.to_numpy()
return build_samples(samples, target_sample_length=config['input_size'], target_dimensions=3)
experiment.print('Building model')
model_function = load_from_module(run_config['model_function'])
model = model_function(config['input_size'])
experiment.log_keras_model(model)
experiment.print('Loading data')
bearing_dataset = load_data(run_config['data_set'])
data_frames = bearing_dataset.as_dict(column=config['data_column'], modifier=modifier)
train_samples = pre_processing(data_frames['train'])
experiment.print('Fitting model')
history = model.fit(
train_samples,
train_samples,
verbose=0,
epochs=config['epochs'],
callbacks=experiment.keras_callbacks(),
validation_split=config['validation_split'],
shuffle=config['shuffle'],
)
experiment.log_history(history)
experiment.log_keras_model(model)
experiment.log_keras_predictions(
model=model,
data_frames=data_frames,
pre_processing=pre_processing,
has_multiple_features=True,
)
def run_callable(run_config: dict):
experiment.print('Loading data')
bearing_dataset = load_data(run_config['data_set'])
train = bearing_dataset.train(column=config['data_column'],
as_numpy=True)
test = bearing_dataset.test(column=config['data_column'],
as_numpy=True)
test_labels = bearing_dataset.test(column=config['data_column'],
add_label=True)['label']
threshold_percentile = config['threshold_percentile']
x_train, x_valid, y_train, y_valid = train_test_split(
train,
train,
test_size=config['validation_split'],
shuffle=True,
)
model = KerasModel(
create_model_function=create_deep_easing_feed_forward_autoencoder,
evaluation_function=load_from_module(
run_config['evaluation_function']),
)
history = pd.DataFrame(columns=[
'cost', 'auc', 'accuracy', 'precision', 'recall', 'f_score',
'matthews_cc'
])
for i in range(1, config['num_evaluations']):
experiment.print(
f'Evaluating configuration {i} of {config["num_evaluations"]}')
current_config = dict(config_space.sample_configuration())
model.load_config(current_config)
evaluation_result = model.evaluate(x_train,
y_train,
x_valid,
y_valid,
budget=config['budget'])
train_reconstruction_error = compute_reconstruction_error(
y_train, model.predict(x_train))
train_z_scores = z_score(train_reconstruction_error)
anomaly_threshold = percentile(train_z_scores,
threshold_percentile)
test_prediction = model.predict(test)
test_reconstruction_error = compute_reconstruction_error(
test, test_prediction)
test_z_scores = z_score(
test_reconstruction_error,
given_median=median(train_reconstruction_error),
given_mad=mad(train_reconstruction_error))
if np.isnan(np.sum(test_reconstruction_error)):
experiment.print(
'Got a NaN value in test reconstruction error, skip this evaluation.'
)
continue
anomaly_prediction = (np.array(test_z_scores) >
anomaly_threshold).astype(int)
metrics = compute_classification_metrics(test_labels.values,
anomaly_prediction)
roc = compute_roc(test_labels.values, test_reconstruction_error)
history_record = {
'cost': evaluation_result.cost,
'auc': roc['auc'],
'accuracy': metrics['accuracy'],
'precision': metrics['precision'],
'recall': metrics['recall'],
'f_score': metrics['f_score'],
'matthews_cc': metrics['matthews_cc'],
**{f'info_{k}': v
for k, v in evaluation_result.info.items()},
**{f'config_{k}': v
for k, v in current_config.items()}
}
history = history.append(history_record, ignore_index=True)
experiment.log('history', history)
def create_deep_easing_feed_forward_autoencoder(
input_dimension,
latent_dimension,
easing='ease_linear',
number_of_hidden_layers=1,
hidden_layer_activations='relu',
output_layer_activation='relu',
optimizer='adam',
loss='mse',
activity_regularizer=None,
l1_activity_regularizer_factor=0.01,
l2_activity_regularizer_factor=0.01,
learning_rate=None,
learning_rate_decay=0,
sgd_momentum=None,
dropout_rate_input=0,
dropout_rate_hidden_layers=0,
dropout_rate_output=0,
dropout_rate_threshold=0.01,
**kwargs,
) -> Model:
if number_of_hidden_layers % 2 == 0:
raise ValueError(
f'Number of hidden layers must be odd, "{number_of_hidden_layers}" provided.'
)
if activity_regularizer == 'none':
activity_regularizer = None
elif activity_regularizer == 'l1':
activity_regularizer = regularizers.l1(l1_activity_regularizer_factor)
elif activity_regularizer == 'l2':
activity_regularizer = regularizers.l2(l2_activity_regularizer_factor)
number_of_encoding_layers = int((number_of_hidden_layers - 1) / 2)
encoding_layer_dimensions = []
input_layer = Input(shape=(input_dimension, ), name='input')
layer = input_layer
if dropout_rate_input > dropout_rate_threshold:
layer = Dropout(dropout_rate_input)(layer)
if isinstance(easing, str):
easing_function = load_from_module(f'a2e.utility.easing.{easing}')
for i in range(1, number_of_encoding_layers + 1):
encoding_layer_dimensions.append(
easing_function(input_dimension, latent_dimension, i,
number_of_encoding_layers))
# encoding
for i, layer_dimension in enumerate(encoding_layer_dimensions):
layer = Dense(layer_dimension,
activation=hidden_layer_activations,
name=f'hidden_encoding_layer_{i}')(layer)
if dropout_rate_hidden_layers > dropout_rate_threshold:
layer = Dropout(dropout_rate_hidden_layers)(layer)
encoded = Dense(latent_dimension,
activation=hidden_layer_activations,
activity_regularizer=activity_regularizer,
name='encoded')(layer)
layer = encoded
if len(encoding_layer_dimensions) > 0:
encoding_layer_dimensions.pop()
encoding_layer_dimensions.insert(0, input_dimension)
# decoding
for i, layer_dimension in enumerate(reversed(encoding_layer_dimensions)):
layer = Dense(layer_dimension,
activation=hidden_layer_activations,
name=f'hidden_decoding_layer_{i}')(layer)
if dropout_rate_hidden_layers > dropout_rate_threshold \
and (i + 1 < len(encoding_layer_dimensions) or dropout_rate_output < dropout_rate_threshold):
layer = Dropout(dropout_rate_hidden_layers)(layer)
if dropout_rate_output > dropout_rate_threshold:
layer = Dropout(dropout_rate_output)(layer)
output_layer = Dense(input_dimension,
activation=output_layer_activation,
name='output')(layer)
model = Model(input_layer, output_layer, name='a2e_deep_feed_forward')
if optimizer == 'adam':
optimizer = Adam(
learning_rate=0.001 if learning_rate is None else learning_rate,
decay=learning_rate_decay)
elif optimizer == 'sgd':
optimizer = SGD(
learning_rate=0.01 if learning_rate is None else learning_rate,
momentum=sgd_momentum,
decay=learning_rate_decay)
model.compile(optimizer=optimizer, loss=loss)
return model
def test_load_from_module(self):
loaded_function = load_from_module(
'a2e.utility.timestamp_to_date_time')
self.assertIsInstance(loaded_function, Callable)