def run_chain_from_automl(train_file_path: str,
test_file_path: str,
max_run_time: timedelta = timedelta(minutes=10)):
train_data = InputData.from_csv(train_file_path)
test_data = InputData.from_csv(test_file_path)
testing_target = test_data.target
chain = Chain()
node_tpot = PrimaryNode('tpot')
node_tpot.model.params = {'max_run_time_sec': max_run_time.seconds}
node_lda = PrimaryNode('lda')
node_rf = SecondaryNode('rf')
node_rf.nodes_from = [node_tpot, node_lda]
chain.add_node(node_rf)
chain.fit(train_data)
results = chain.predict(test_data)
roc_auc_value = roc_auc(y_true=testing_target, y_score=results.predict)
print(roc_auc_value)
return roc_auc_value
def compose_chain(self, data: InputData,
initial_chain: Optional[Chain],
composer_requirements: ComposerRequirements,
metrics: Optional[Callable],
optimiser_parameters=None,
is_visualise: bool = False) -> Chain:
new_chain = Chain()
if self.dummy_chain_type == DummyChainTypeEnum.hierarchical:
# (y1, y2) -> y
last_node = SecondaryNode(composer_requirements.secondary[0])
for requirement_model in composer_requirements.primary:
new_node = PrimaryNode(requirement_model)
new_chain.add_node(new_node)
last_node.nodes_from.append(new_node)
new_chain.add_node(last_node)
elif self.dummy_chain_type == DummyChainTypeEnum.flat:
# (y1) -> (y2) -> y
first_node = PrimaryNode(composer_requirements.primary[0])
new_chain.add_node(first_node)
prev_node = first_node
for requirement_model in composer_requirements.secondary:
new_node = SecondaryNode(requirement_model)
new_node.nodes_from = [prev_node]
prev_node = new_node
new_chain.add_node(new_node)
else:
raise NotImplementedError()
return new_chain
def get_simple_chain():
first = PrimaryNode(model_type='xgboost')
second = PrimaryNode(model_type='knn')
final = SecondaryNode(model_type='logit', nodes_from=[first, second])
chain = Chain(final)
return chain
def chain_with_incorrect_decomposition_structure():
first = PrimaryNode(model_type='trend_data_model')
second = PrimaryNode(model_type='residual_data_model')
final = SecondaryNode(model_type='linear', nodes_from=[first, second])
chain = Chain(final)
return chain
def chain_with_incorrect_task_type():
first = PrimaryNode(model_type='linear')
second = PrimaryNode(model_type='linear')
final = SecondaryNode(model_type='kmeans', nodes_from=[first, second])
chain = Chain(final)
return chain, Task(TaskTypesEnum.classification)
def chain_with_pca() -> Chain:
node_first = PrimaryNode('pca_data_model')
node_second = PrimaryNode('lda')
node_final = SecondaryNode('rf', nodes_from=[node_first, node_second])
chain = Chain(node_final)
return chain
def chain_simple() -> Chain:
node_first = PrimaryNode('svc')
node_second = PrimaryNode('lda')
node_final = SecondaryNode('rf', nodes_from=[node_first, node_second])
chain = Chain(node_final)
return chain
def chain_with_primary_composition_model():
first = PrimaryNode(model_type='additive_data_model')
second = PrimaryNode(model_type='residual_data_model')
final = SecondaryNode(model_type='additive_data_model',
nodes_from=[first, second])
chain = Chain(final)
return chain
def get_comp_chain():
node_first = PrimaryNode('lstm')
node_second = PrimaryNode('rfr')
node_final = SecondaryNode('linear',
nodes_from=[node_first, node_second],
manual_preprocessing_func=EmptyStrategy)
chain = Chain(node_final)
return chain
def chain_with_incorrect_root_model():
first = PrimaryNode(model_type='logit')
second = PrimaryNode(model_type='logit')
final = SecondaryNode(model_type='direct_data_model',
nodes_from=[first, second])
chain = Chain(final)
return chain
def get_class_chain():
# Chain composition
first = PrimaryNode(model_type='xgboost')
second = PrimaryNode(model_type='knn')
final = SecondaryNode(model_type='logit', nodes_from=[first, second])
chain = Chain()
chain.add_node(final)
return chain
def two_level_chain():
first = PrimaryNode(model_type='logit')
second = PrimaryNode(model_type='knn')
third = SecondaryNode(model_type='xgboost', nodes_from=[first, second])
chain = Chain()
for node in [first, second, third]:
chain.add_node(node)
return chain
def get_regr_chain():
# Chain composition
first = PrimaryNode(model_type='xgbreg')
second = PrimaryNode(model_type='knnreg')
final = SecondaryNode(model_type='linear', nodes_from=[first, second])
chain = Chain()
chain.add_node(final)
return chain
def compose_chain() -> Chain:
chain = Chain()
node_first = PrimaryNode('svc')
node_second = PrimaryNode('lda')
node_third = SecondaryNode('rf')
node_third.nodes_from.append(node_first)
node_third.nodes_from.append(node_second)
chain.add_node(node_third)
return chain
def test_secondary_nodes_is_invariant_to_inputs_order(data_setup):
data = data_setup
train, test = train_test_data_setup(data)
first = PrimaryNode(model_type='logit')
second = PrimaryNode(model_type='lda')
third = PrimaryNode(model_type='knn')
final = SecondaryNode(model_type='xgboost',
nodes_from=[first, second, third])
chain = Chain()
for node in [first, second, third, final]:
chain.add_node(node)
first = deepcopy(first)
second = deepcopy(second)
third = deepcopy(third)
final_shuffled = SecondaryNode(model_type='xgboost',
nodes_from=[third, first, second])
chain_shuffled = Chain()
# change order of nodes in list
for node in [final_shuffled, third, first, second]:
chain_shuffled.add_node(node)
train_predicted = chain.fit(input_data=train)
train_predicted_shuffled = chain_shuffled.fit(input_data=train)
# train results should be invariant
assert chain.root_node.descriptive_id == chain_shuffled.root_node.descriptive_id
assert np.equal(train_predicted.predict,
train_predicted_shuffled.predict).all()
test_predicted = chain.predict(input_data=test)
test_predicted_shuffled = chain_shuffled.predict(input_data=test)
# predict results should be invariant
assert np.equal(test_predicted.predict,
test_predicted_shuffled.predict).all()
# change parents order for the nodes fitted chain
nodes_for_change = chain.nodes[3].nodes_from
chain.nodes[3].nodes_from = [
nodes_for_change[2], nodes_for_change[0], nodes_for_change[1]
]
chain.nodes[3].cache.clear()
chain.fit(train)
test_predicted_re_shuffled = chain.predict(input_data=test)
# predict results should be invariant
assert np.equal(test_predicted.predict,
test_predicted_re_shuffled.predict).all()
def run_metocean_forecasting_problem(train_file_path,
test_file_path,
forecast_length=1,
max_window_size=64,
is_visualise=False):
# specify the task to solve
task_to_solve = Task(
TaskTypesEnum.ts_forecasting,
TsForecastingParams(forecast_length=forecast_length,
max_window_size=max_window_size))
full_path_train = os.path.join(str(project_root()), train_file_path)
dataset_to_train = InputData.from_csv(full_path_train,
task=task_to_solve,
data_type=DataTypesEnum.ts)
# a dataset for a final validation of the composed model
full_path_test = os.path.join(str(project_root()), test_file_path)
dataset_to_validate = InputData.from_csv(full_path_test,
task=task_to_solve,
data_type=DataTypesEnum.ts)
chain = get_composite_lstm_chain()
chain_simple = Chain()
node_single = PrimaryNode('ridge')
chain_simple.add_node(node_single)
chain_lstm = Chain()
node_lstm = PrimaryNode('lstm')
chain_lstm.add_node(node_lstm)
chain.fit(input_data=dataset_to_train, verbose=False)
rmse_on_valid = calculate_validation_metric(
chain.predict(dataset_to_validate), dataset_to_validate,
f'full-composite_{forecast_length}', is_visualise)
chain_lstm.fit(input_data=dataset_to_train, verbose=False)
rmse_on_valid_lstm_only = calculate_validation_metric(
chain_lstm.predict(dataset_to_validate), dataset_to_validate,
f'full-lstm-only_{forecast_length}', is_visualise)
chain_simple.fit(input_data=dataset_to_train, verbose=False)
rmse_on_valid_simple = calculate_validation_metric(
chain_simple.predict(dataset_to_validate), dataset_to_validate,
f'full-simple_{forecast_length}', is_visualise)
print(f'RMSE composite: {rmse_on_valid}')
print(f'RMSE simple: {rmse_on_valid_simple}')
print(f'RMSE LSTM only: {rmse_on_valid_lstm_only}')
return rmse_on_valid_simple
def get_composite_lstm_chain():
chain = Chain()
node_trend = PrimaryNode('trend_data_model')
node_lstm_trend = SecondaryNode('lasso', nodes_from=[node_trend])
node_residual = PrimaryNode('residual_data_model')
node_ridge_residual = SecondaryNode('ridge', nodes_from=[node_residual])
node_final = SecondaryNode(
'additive_data_model',
nodes_from=[node_ridge_residual, node_lstm_trend])
chain.add_node(node_final)
return chain
def test_eval_strategy_logreg(data_setup):
data_set = data_setup
train, test = train_test_data_setup(data=data_set)
test_skl_model = LogisticRegression(C=10., random_state=1,
solver='liblinear',
max_iter=10000, verbose=0)
test_skl_model.fit(train.features, train.target)
expected_result = test_skl_model.predict(test.features)
test_model_node = PrimaryNode(model_type='logit')
test_model_node.fit(input_data=train)
actual_result = test_model_node.predict(input_data=test)
assert len(actual_result.predict) == len(expected_result)
def test_regression_chain_with_datamodel_fit_correct():
data = get_synthetic_regression_data()
train_data, test_data = train_test_data_setup(data)
node_data = PrimaryNode('direct_data_model')
node_first = PrimaryNode('ridge')
node_second = SecondaryNode('lasso')
node_second.nodes_from = [node_first, node_data]
chain = Chain(node_second)
chain.fit(train_data)
results = chain.predict(test_data)
assert results.predict.shape == test_data.target.shape
def get_composite_lstm_chain():
chain = Chain()
node_trend = PrimaryNode('trend_data_model')
node_trend.labels = ["fixed"]
node_lstm_trend = SecondaryNode('linear', nodes_from=[node_trend])
node_trend.labels = ["fixed"]
node_residual = PrimaryNode('residual_data_model')
node_ridge_residual = SecondaryNode('linear', nodes_from=[node_residual])
node_final = SecondaryNode(
'additive_data_model',
nodes_from=[node_ridge_residual, node_lstm_trend])
node_final.labels = ["fixed"]
chain.add_node(node_final)
return chain
def get_decomposed_chain(model_trend='lstm', model_residual='ridge'):
chain = Chain()
node_trend = PrimaryNode('trend_data_model')
node_first_trend = SecondaryNode('lstm', nodes_from=[node_trend])
if model_trend == 'lstm':
# decrease the number of epochs to fit
node_first_trend.model.params = {'epochs': 1}
node_residual = PrimaryNode('residual_data_model')
node_model_residual = SecondaryNode(model_residual,
nodes_from=[node_residual])
node_final = SecondaryNode(
'additive_data_model',
nodes_from=[node_model_residual, node_first_trend])
chain.add_node(node_final)
return chain
def test_node_factory_log_reg_correct(data_setup):
model_type = 'logit'
node = PrimaryNode(model_type=model_type)
expected_model = Model(model_type=model_type).__class__
actual_model = node.model.__class__
assert node.__class__ == PrimaryNode
assert expected_model == actual_model
def test_chain_with_datamodel_fit_correct(data_setup):
data = data_setup
train_data, test_data = train_test_data_setup(data)
chain = Chain()
node_data = PrimaryNode('direct_data_model')
node_first = PrimaryNode('bernb')
node_second = SecondaryNode('rf')
node_second.nodes_from = [node_first, node_data]
chain.add_node(node_data)
chain.add_node(node_first)
chain.add_node(node_second)
chain.fit(train_data)
results = np.asarray(probs_to_labels(chain.predict(test_data).predict))
assert results.shape == test_data.target.shape
def chain_fourth():
# XG
# | \
# XG KNN
# | \ | \
# QDA KNN LR LDA
# | \ | \
# RF RF KNN KNN
chain = chain_first()
new_node = SecondaryNode('qda')
for model_type in ('rf', 'rf'):
new_node.nodes_from.append(PrimaryNode(model_type))
chain.replace_node_with_parents(chain.root_node.nodes_from[0].nodes_from[1], new_node)
new_node = SecondaryNode('knn')
for model_type in ('knn', 'knn'):
new_node.nodes_from.append(PrimaryNode(model_type))
chain.replace_node_with_parents(chain.root_node.nodes_from[0].nodes_from[0], new_node)
return chain
def default_valid_chain():
first = PrimaryNode(model_type='logit')
second = SecondaryNode(model_type='logit', nodes_from=[first])
third = SecondaryNode(model_type='logit', nodes_from=[first])
final = SecondaryNode(model_type='logit', nodes_from=[second, third])
chain = Chain(final)
return chain
def baseline_chain():
chain = Chain()
last_node = SecondaryNode(model_type='xgboost', nodes_from=[])
for requirement_model in ['knn', 'logit']:
new_node = PrimaryNode(requirement_model)
chain.add_node(new_node)
last_node.nodes_from.append(new_node)
chain.add_node(last_node)
return chain
def chain_with_cycle():
first = PrimaryNode(model_type='logit')
second = SecondaryNode(model_type='logit', nodes_from=[first])
third = SecondaryNode(model_type='logit', nodes_from=[second, first])
second.nodes_from.append(third)
chain = Chain()
for node in [first, second, third]:
chain.add_node(node)
return chain
def chain_with_multiple_roots():
first = PrimaryNode(model_type='logit')
root_first = SecondaryNode(model_type='logit', nodes_from=[first])
root_second = SecondaryNode(model_type='logit', nodes_from=[first])
chain = Chain()
for node in [first, root_first, root_second]:
chain.add_node(node)
return chain
def chain_with_self_cycle():
first = PrimaryNode(model_type='logit')
second = SecondaryNode(model_type='logit', nodes_from=[first])
second.nodes_from.append(second)
chain = Chain()
chain.add_node(first)
chain.add_node(second)
return chain
def chain_third():
# QDA
# | \
# RF RF
chain = Chain()
new_node = SecondaryNode('qda')
for model_type in ('rf', 'rf'):
new_node.nodes_from.append(PrimaryNode(model_type))
chain.add_node(new_node)
[chain.add_node(node_from) for node_from in new_node.nodes_from]
return chain