def chain_tuning(nodes_to_tune: str, chain: Chain, train_data: InputData,
test_data: InputData, local_iter: int,
tuner_iter_num: int = 50) -> (float, list):
several_iter_scores_test = []
if nodes_to_tune == 'primary':
print('primary_node_tuning')
chain_tune_strategy = chain.fine_tune_primary_nodes
elif nodes_to_tune == 'root':
print('root_node_tuning')
chain_tune_strategy = chain.fine_tune_all_nodes
else:
raise ValueError(f'Invalid type of nodes. Nodes must be primary or root')
for iteration in range(local_iter):
print(f'current local iteration {iteration}')
# Chain tuning
chain_tune_strategy(train_data, iterations=tuner_iter_num)
# After tuning prediction
chain.fit(train_data)
after_tuning_predicted = chain.predict(test_data)
# Metrics
aft_tun_roc_auc = roc_auc(y_true=test_data.target,
y_score=after_tuning_predicted.predict)
several_iter_scores_test.append(aft_tun_roc_auc)
return float(np.mean(several_iter_scores_test)), several_iter_scores_test
def test_save_load_fitted_atomized_chain_correctly():
chain = create_chain_with_several_nested_atomized_model()
train_data, test_data = create_data_for_train()
chain.fit(train_data)
json_actual = chain.save_chain(
'test_save_load_fitted_atomized_chain_correctly')
json_path_load = create_correct_path(
'test_save_load_fitted_atomized_chain_correctly')
chain_loaded = Chain()
chain_loaded.load_chain(json_path_load)
json_expected = chain_loaded.save_chain(
'test_save_load_fitted_atomized_chain_correctly_loaded')
assert chain.length == chain_loaded.length
assert json_actual == json_expected
before_save_predicted = chain.predict(test_data)
chain_loaded.fit(train_data)
after_save_predicted = chain_loaded.predict(test_data)
bfr_tun_mse = mean_squared_error(y_true=test_data.target,
y_pred=before_save_predicted.predict)
aft_tun_mse = mean_squared_error(y_true=test_data.target,
y_pred=after_save_predicted.predict)
assert aft_tun_mse <= bfr_tun_mse
def test_forecast_with_exog():
train_source_ts, predict_source_ts, train_exog_ts, predict_exog_ts, ts_test = synthetic_with_exogenous_ts(
)
# Source data for lagged node
node_lagged = PrimaryNode('lagged',
node_data={
'fit': train_source_ts,
'predict': predict_source_ts
})
# Set window size for lagged transformation
node_lagged.custom_params = {'window_size': window_size}
# Exogenous variable for exog node
node_exog = PrimaryNode('exog',
node_data={
'fit': train_exog_ts,
'predict': predict_exog_ts
})
node_final = SecondaryNode('linear', nodes_from=[node_lagged, node_exog])
chain = Chain(node_final)
chain.fit()
forecast = chain.predict()
prediction = np.ravel(np.array(forecast.predict))
assert tuple(prediction) == tuple(ts_test)
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 run_tpot_vs_fedot_example(train_file_path: str, test_file_path: str):
train_data = InputData.from_csv(train_file_path)
test_data = InputData.from_csv(test_file_path)
training_features = train_data.features
testing_features = test_data.features
training_target = train_data.target
testing_target = test_data.target
# Average CV score on the training set was: 0.93755
exported_pipeline = make_pipeline(
StackingEstimator(estimator=BernoulliNB()), RandomForestClassifier())
# Fix random state for all the steps in exported pipeline
set_param_recursive(exported_pipeline.steps, 'random_state', 1)
exported_pipeline.fit(training_features, training_target)
results = exported_pipeline.predict_proba(testing_features)[:, 1]
roc_auc_value = roc_auc(y_true=testing_target, y_score=results)
print(f'ROC AUC for TPOT: {roc_auc_value}')
node_scaling = PrimaryNode('scaling')
node_bernb = SecondaryNode('bernb', nodes_from=[node_scaling])
node_rf = SecondaryNode('rf', nodes_from=[node_bernb, node_scaling])
chain = Chain(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(f'ROC AUC for FEDOT: {roc_auc_value}')
return roc_auc_value
def run_chain_from_automl(train_file_path: str,
test_file_path: str,
max_run_time: timedelta = timedelta(minutes=10)):
""" Function run chain with Auto ML models in nodes
:param train_file_path: path to the csv file with data for train
:param test_file_path: path to the csv file with data for validation
:param max_run_time: maximum running time for customization of the "tpot" model
:return roc_auc_value: ROC AUC metric for chain
"""
train_data = InputData.from_csv(train_file_path)
test_data = InputData.from_csv(test_file_path)
testing_target = test_data.target
chain = Chain()
node_scaling = PrimaryNode('scaling')
node_tpot = PrimaryNode('tpot')
node_tpot.operation.params = {'max_run_time_sec': max_run_time.seconds}
node_lda = SecondaryNode('lda', nodes_from=[node_scaling])
node_rf = SecondaryNode('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 fine_tune_certain_node(self, model_id, input_data: InputData, iterations: int = 30,
max_lead_time: timedelta = timedelta(minutes=5)):
"""
Optimize hyperparameters of models in the certain node,
defined by model id
:param int model_id: number of the certain model in the chain.
Look for it in exported json file of your model.
:param input_data: data used for tuning
:param iterations: max number of iterations
:param max_lead_time: max time available for tuning process
:param verbose: flag used for status printing to console, default False
:return: updated chain object
"""
subchain = Chain()
new_root = extract_subtree_root(root_model_id=model_id,
chain_template=self.chain_template)
subchain.add_node(new_root)
subchain.fit(input_data=input_data, use_cache=False)
updated_subchain = Tune(subchain).fine_tune_root_node(input_data=input_data,
iterations=iterations,
max_lead_time=max_lead_time)
self._update_template(model_id=model_id,
updated_node=updated_subchain.root_node)
updated_chain = Chain()
self.chain_template.convert_to_chain(chain=updated_chain)
return updated_chain
def test_secondary_nodes_is_invariant_to_inputs_order(data_setup):
data = data_setup
train, test = train_test_data_setup(data)
first = PrimaryNode(operation_type='logit')
second = PrimaryNode(operation_type='lda')
third = PrimaryNode(operation_type='knn')
final = SecondaryNode(operation_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(operation_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].unfit()
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 metric_for_nodes(self, metric_function, train_data: InputData,
test_data: InputData, is_chain_shared: bool,
chain: Chain) -> float:
try:
validate(chain)
if is_chain_shared:
chain = SharedChain(base_chain=chain, shared_cache=self.shared_cache)
chain.fit(input_data=train_data)
return metric_function(chain, test_data)
except Exception as ex:
self.log.info(f'Error in chain assessment during composition: {ex}. Continue.')
return max_int_value
def test_arima_chain_fit_correct():
train_data, test_data = get_synthetic_ts_data_linear(forecast_length=12)
chain = Chain(PrimaryNode('arima'))
chain.fit(input_data=train_data)
rmse_on_train, rmse_on_test, _, _ = get_rmse_value(chain, train_data,
test_data)
rmse_threshold = _max_rmse_threshold_by_std(test_data.target)
assert rmse_on_train < rmse_threshold
def test_regression_chain_period_exog_forecast_multistep_correct():
train_data, test_data = get_synthetic_ts_data_period(forecast_length=2,
max_window_size=3)
chain = Chain(PrimaryNode('linear'))
chain.fit(input_data=train_data)
rmse_on_train, rmse_on_test, _, _ = get_rmse_value(chain, train_data,
test_data)
rmse_threshold = 1.5
assert rmse_on_train < rmse_threshold
assert rmse_on_test < rmse_threshold
def test_log_clustering_fit_correct(data_fixture, request):
data = request.getfixturevalue(data_fixture)
train_data, test_data = train_test_data_setup(data=data)
# Scaling chain. Fit predict it
scaling_chain = Chain(PrimaryNode('normalization'))
scaling_chain.fit(train_data)
scaled_data = scaling_chain.predict(train_data)
kmeans = Model(operation_type='kmeans')
_, train_predicted = kmeans.fit(data=scaled_data)
assert all(np.unique(train_predicted.predict) == [0, 1])
def test_regression_chain_forecast_multistep_correct():
train_data, test_data = get_synthetic_ts_data_period(forecast_length=20,
max_window_size=30)
chain = Chain(PrimaryNode('ridge'))
chain.fit(input_data=train_data)
_, rmse_on_test, _, _ = get_rmse_value(chain, train_data, test_data)
rmse_threshold = _max_rmse_threshold_by_std(test_data.target,
is_strict=False)
assert rmse_on_test < rmse_threshold
def test_import_custom_json_object_to_chain_and_fit_correctly_no_exception():
test_file_path = str(os.path.dirname(__file__))
file = '../../data/test_custom_json_template.json'
json_path_load = os.path.join(test_file_path, file)
train_file_path, test_file_path = get_scoring_case_data_paths()
train_data = InputData.from_csv(train_file_path)
chain = Chain()
chain.load_chain(json_path_load)
chain.fit(train_data)
chain.save_chain('test_import_custom_json_object_to_chain_and_fit_correctly_no_exception')
def execute_chain_for_text_problem(train_data, test_data):
node_text_clean = PrimaryNode('text_clean')
node_tfidf = SecondaryNode('tfidf', nodes_from=[node_text_clean])
model_node = SecondaryNode('multinb', nodes_from=[node_tfidf])
chain = Chain(model_node)
chain.fit(train_data)
predicted = chain.predict(test_data)
roc_auc_metric = roc_auc(y_true=test_data.target,
y_score=predicted.predict)
return roc_auc_metric
def test_chain_with_wrong_data():
chain = Chain(PrimaryNode('linear'))
data_seq = np.arange(0, 10)
task = Task(TaskTypesEnum.ts_forecasting,
TsForecastingParams(forecast_length=10))
data = InputData(idx=data_seq,
features=data_seq,
target=data_seq,
data_type=DataTypesEnum.ts,
task=task)
with pytest.raises(ValueError):
chain.fit(data)
def execute_chain_for_text_problem(train_data, test_data):
preproc_node = PrimaryNode(
'tfidf', manual_preprocessing_func=TextPreprocessingStrategy)
model_node = SecondaryNode('multinb',
nodes_from=[preproc_node],
manual_preprocessing_func=EmptyStrategy)
chain = Chain(nodes=[model_node, preproc_node])
chain.fit(train_data)
predicted = chain.predict(test_data)
roc_auc_metric = roc_auc(y_true=test_data.target,
y_score=predicted.predict)
return roc_auc_metric
def test_regression_chain_forecast_onestep_correct():
train_data, test_data = get_synthetic_ts_data_linear(forecast_length=1,
max_window_size=10)
chain = Chain(PrimaryNode('ridge'))
chain.fit(input_data=train_data)
rmse_on_train, rmse_on_test, _, _ = get_rmse_value(chain, train_data,
test_data)
rmse_threshold = _max_rmse_threshold_by_std(test_data.target,
is_strict=True)
assert rmse_on_train < rmse_threshold
assert rmse_on_test < rmse_threshold
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 test_pca_model_removes_redunant_features_correct():
n_informative = 5
data = classification_dataset_with_redunant_features(n_samples=1000, n_features=100,
n_informative=n_informative)
train_data, test_data = train_test_data_setup(data=data)
# Scaling chain. Fit predict it
scaling_chain = Chain(PrimaryNode('normalization'))
scaling_chain.fit(train_data)
scaled_data = scaling_chain.predict(train_data)
pca = DataOperation(operation_type='pca')
_, train_predicted = pca.fit(data=scaled_data)
transformed_features = train_predicted.predict
assert transformed_features.shape[1] < data.features.shape[1]
def test_fitted_chain_cache_correctness_after_export_and_import():
train_file_path, test_file_path = get_scoring_case_data_paths()
train_data = InputData.from_csv(train_file_path)
chain = Chain(PrimaryNode('logit'))
chain.fit(train_data)
chain.save_chain('test_fitted_chain_cache_correctness_after_export_and_import')
json_path_load = create_correct_path('test_fitted_chain_cache_correctness_after_export_and_import')
new_chain = Chain()
new_chain.load_chain(json_path_load)
results = new_chain.fit(train_data)
assert results is not None
def test_log_regression_fit_correct(classification_dataset):
data = classification_dataset
train_data, test_data = train_test_data_setup(data=data)
# Scaling chain. Fit predict it
scaling_chain = Chain(PrimaryNode('normalization'))
scaling_chain.fit(train_data)
scaled_data = scaling_chain.predict(train_data)
log_reg = Model(operation_type='logit')
_, train_predicted = log_reg.fit(data=scaled_data)
roc_on_train = get_roc_auc(valid_data=train_data,
predicted_data=train_predicted)
roc_threshold = 0.95
assert roc_on_train >= roc_threshold
def test_random_forest_fit_correct(data_fixture, request):
data = request.getfixturevalue(data_fixture)
train_data, test_data = train_test_data_setup(data=data)
# Scaling chain. Fit predict it
scaling_chain = Chain(PrimaryNode('normalization'))
scaling_chain.fit(train_data)
scaled_data = scaling_chain.predict(train_data)
random_forest = Model(operation_type='rf')
_, train_predicted = random_forest.fit(data=scaled_data)
roc_on_train = get_roc_auc(valid_data=train_data,
predicted_data=train_predicted)
roc_threshold = 0.95
assert roc_on_train >= roc_threshold
def test_chain_hierarchy_fit_correct(data_setup):
data = data_setup
train, _ = train_test_data_setup(data)
first = PrimaryNode(operation_type='logit')
second = SecondaryNode(operation_type='logit', nodes_from=[first])
third = SecondaryNode(operation_type='logit', nodes_from=[first])
final = SecondaryNode(operation_type='logit', nodes_from=[second, third])
chain = Chain()
for node in [first, second, third, final]:
chain.add_node(node)
chain.unfit()
train_predicted = chain.fit(input_data=train)
assert chain.root_node.descriptive_id == ('((/n_logit_default_params;)/'
'n_logit_default_params;;(/'
'n_logit_default_params;)/'
'n_logit_default_params;)/'
'n_logit_default_params')
assert chain.length == 4
assert chain.depth == 3
assert train_predicted.predict.shape[0] == train.target.shape[0]
assert final.fitted_operation is not None
def run_tpot_vs_fedot_example(train_file_path: str, test_file_path: str):
train_data = InputData.from_csv(train_file_path)
test_data = InputData.from_csv(test_file_path)
training_features = train_data.features
testing_features = test_data.features
training_target = train_data.target
testing_target = test_data.target
# Average CV score on the training set was: 0.93755
exported_pipeline = make_pipeline(
StackingEstimator(estimator=BernoulliNB()),
RandomForestClassifier()
)
# Fix random state for all the steps in exported pipeline
set_param_recursive(exported_pipeline.steps, 'random_state', 1)
exported_pipeline.fit(training_features, training_target)
results = exported_pipeline.predict_proba(testing_features)[:, 1]
roc_auc_value = roc_auc(y_true=testing_target,
y_score=results)
print(roc_auc_value)
chain = Chain()
node_first = PrimaryNode('direct_data_model')
node_second = PrimaryNode('bernb')
node_third = SecondaryNode('rf')
node_third.nodes_from.append(node_first)
node_third.nodes_from.append(node_second)
chain.add_node(node_third)
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 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('logit')
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 composer_metric(self, metrics, train_data: InputData,
test_data: InputData,
chain: Chain) -> Optional[Tuple[Any]]:
try:
validate(chain)
chain.log = self.log
if type(metrics) is not list:
metrics = [metrics]
if self.cache is not None:
# TODO improve cache
chain.fit_from_cache(self.cache)
if not chain.is_fitted:
self.log.debug(
f'Chain {chain.root_node.descriptive_id} fit started')
chain.fit(input_data=train_data,
time_constraint=self.composer_requirements.
max_chain_fit_time)
self.cache.save_chain(chain)
evaluated_metrics = ()
for metric in metrics:
if callable(metric):
metric_func = metric
else:
metric_func = MetricsRepository().metric_by_id(metric)
evaluated_metrics = evaluated_metrics + (metric_func(
chain, reference_data=test_data), )
self.log.debug(
f'Chain {chain.root_node.descriptive_id} with metrics: {list(evaluated_metrics)}'
)
except Exception as ex:
self.log.info(f'Chain assessment warning: {ex}. Continue.')
evaluated_metrics = None
return evaluated_metrics
def test_regression_chain_with_data_operation_fit_correct():
data = get_synthetic_regression_data()
train_data, test_data = train_test_data_setup(data)
# linear
# / \
# ridge |
# | |
# ransac_lin_reg lasso
# \ /
# scaling
node_scaling = PrimaryNode('scaling')
node_ransac = SecondaryNode('ransac_lin_reg', nodes_from=[node_scaling])
node_lasso = SecondaryNode('lasso', nodes_from=[node_scaling])
node_ridge = SecondaryNode('ridge', nodes_from=[node_ransac])
node_root = SecondaryNode('linear', nodes_from=[node_lasso, node_ridge])
chain = Chain(node_root)
chain.fit(train_data)
results = chain.predict(test_data)
assert results.predict.shape == test_data.target.shape
def test_chain_with_custom_params_for_model(data_setup):
data = data_setup
custom_params = dict(n_neighbors=1, weights='uniform', p=1)
first = PrimaryNode(operation_type='logit')
second = PrimaryNode(operation_type='lda')
final = SecondaryNode(operation_type='knn', nodes_from=[first, second])
chain = Chain()
chain.add_node(final)
chain_default_params = deepcopy(chain)
chain.root_node.custom_params = custom_params
chain_default_params.fit(data)
chain.fit(data)
custom_params_prediction = chain.predict(data).predict
default_params_prediction = chain_default_params.predict(data).predict
assert not np.array_equal(custom_params_prediction,
default_params_prediction)
def test_chain_sequential_fit_correct(data_setup):
data = data_setup
train, _ = train_test_data_setup(data)
first = PrimaryNode(model_type='logit')
second = SecondaryNode(model_type='logit', nodes_from=[first])
third = SecondaryNode(model_type='logit', nodes_from=[second])
final = SecondaryNode(model_type='logit', nodes_from=[third])
chain = Chain()
for node in [first, second, third, final]:
chain.add_node(node)
train_predicted = chain.fit(input_data=train, use_cache=False)
assert chain.root_node.descriptive_id == ('(((/n_logit_default_params;)/'
'n_logit_default_params;)/'
'n_logit_default_params;)/'
'n_logit_default_params')
assert chain.length == 4
assert chain.depth == 4
assert train_predicted.predict.shape[0] == train.target.shape[0]
assert final.cache.actual_cached_state is not None
