def test_choose_next(self):
configspace = ConfigurationSpace()
configspace.add_hyperparameter(UniformFloatHyperparameter('a', 0, 1))
configspace.add_hyperparameter(UniformFloatHyperparameter('b', 0, 1))
dataset_name = 'foo'
func_eval_time_limit = 15
total_walltime_limit = 15
memory_limit = 3000
auto = AutoMLSMBO(None, dataset_name, None, func_eval_time_limit,
total_walltime_limit, memory_limit, None)
auto.config_space = configspace
scenario = Scenario({'cs': configspace,
'cutoff-time': func_eval_time_limit,
'wallclock-limit': total_walltime_limit,
'memory-limit': memory_limit,
'run-obj': 'quality'})
smac = SMAC(scenario)
self.assertRaisesRegex(ValueError, 'Cannot use SMBO algorithm on '
'empty runhistory',
auto.choose_next, smac)
runhistory = smac.solver.runhistory
runhistory.add(config=Configuration(configspace,
values={'a': 0.1, 'b': 0.2}),
cost=0.5, time=0.5, status=StatusType.SUCCESS)
auto.choose_next(smac)
def test_choose_next(self):
configspace = ConfigurationSpace()
configspace.add_hyperparameter(UniformFloatHyperparameter('a', 0, 1))
configspace.add_hyperparameter(UniformFloatHyperparameter('b', 0, 1))
dataset_name = 'foo'
func_eval_time_limit = 15
total_walltime_limit = 15
memory_limit = 3072
auto = AutoMLSMBO(config_space=None,
dataset_name=dataset_name,
backend=None,
func_eval_time_limit=func_eval_time_limit,
total_walltime_limit=total_walltime_limit,
memory_limit=memory_limit,
watcher=None,
metric=accuracy)
auto.config_space = configspace
scenario = Scenario({
'cs': configspace,
'cutoff_time': func_eval_time_limit,
'wallclock_limit': total_walltime_limit,
'memory_limit': memory_limit,
'run_obj': 'quality',
})
smac = SMAC(scenario)
self.assertRaisesRegex(
ValueError, 'Cannot use SMBO algorithm on empty runhistory',
auto.choose_next, smac)
config = Configuration(configspace, values={'a': 0.1, 'b': 0.2})
# TODO make sure the incumbent is always set?
smac.solver.incumbent = config
runhistory = smac.solver.runhistory
runhistory.add(config=config,
cost=0.5,
time=0.5,
status=StatusType.SUCCESS)
auto.choose_next(smac)
def test_smbo_metalearning_configurations(backend, context, dask_client):
# Get the inputs to the optimizer
X_train, Y_train, X_test, Y_test = putil.get_dataset('iris')
config_space = AutoML(backend=backend,
metric=autosklearn.metrics.accuracy,
time_left_for_this_task=20,
per_run_time_limit=5).fit(
X_train,
Y_train,
task=BINARY_CLASSIFICATION,
only_return_configuration_space=True)
watcher = StopWatch()
# Create an optimizer
smbo = AutoMLSMBO(
config_space=config_space,
dataset_name='iris',
backend=backend,
total_walltime_limit=10,
func_eval_time_limit=5,
memory_limit=4096,
metric=autosklearn.metrics.accuracy,
watcher=watcher,
n_jobs=1,
dask_client=dask_client,
port=logging.handlers.DEFAULT_TCP_LOGGING_PORT,
start_num_run=1,
data_memory_limit=None,
num_metalearning_cfgs=25,
pynisher_context=context,
)
assert smbo.pynisher_context == context
# Create the inputs to metalearning
datamanager = XYDataManager(
X_train,
Y_train,
X_test,
Y_test,
task=BINARY_CLASSIFICATION,
dataset_name='iris',
feat_type={i: 'numerical'
for i in range(X_train.shape[1])},
)
backend.save_datamanager(datamanager)
smbo.task = BINARY_CLASSIFICATION
smbo.reset_data_manager()
metalearning_configurations = smbo.get_metalearning_suggestions()
# We should have 25 metalearning configurations
assert len(metalearning_configurations) == 25
assert [
isinstance(config, Configuration)
for config in metalearning_configurations
]
def get_meta_learning_configs(X,
y,
task_type,
dataset_name='default',
metric='accuracy',
num_cfgs=5):
if X is None or y is None:
X, y, _ = load_data(dataset_name)
backend = create(temporary_directory=None,
output_directory=None,
delete_tmp_folder_after_terminate=False,
delete_output_folder_after_terminate=False,
shared_mode=True)
dm = XYDataManager(X, y, None, None, task_type, None, dataset_name)
configuration_space = pipeline.get_configuration_space(
dm.info,
include_estimators=None,
exclude_estimators=None,
include_preprocessors=None,
exclude_preprocessors=None)
watcher = StopWatch()
name = os.path.basename(dm.name)
watcher.start_task(name)
def reset_data_manager(max_mem=None):
pass
automlsmbo = AutoMLSMBO(
config_space=configuration_space,
dataset_name=dataset_name,
backend=backend,
total_walltime_limit=1e5,
func_eval_time_limit=1e5,
memory_limit=1e5,
metric=metric,
watcher=watcher,
metadata_directory='components/meta_learning/meta_resource',
num_metalearning_cfgs=num_cfgs)
automlsmbo.reset_data_manager = reset_data_manager
automlsmbo.task = task_type
automlsmbo.datamanager = dm
configs = automlsmbo.get_metalearning_suggestions()
return configs
def _fit(self, datamanager, metric):
# Reset learnt stuff
self.models_ = None
self.ensemble_ = None
# Check arguments prior to doing anything!
if not isinstance(self._disable_evaluator_output, (bool, list)):
raise ValueError('disable_evaluator_output must be of type bool '
'or list.')
if isinstance(self._disable_evaluator_output, list):
allowed_elements = ['model', 'y_optimization']
for element in self._disable_evaluator_output:
if element not in allowed_elements:
raise ValueError("List member '%s' for argument "
"'disable_evaluator_output' must be one "
"of " + str(allowed_elements))
if self._resampling_strategy not in [
'holdout', 'holdout-iterative-fit', 'cv', 'partial-cv',
'partial-cv-iterative-fit'
]:
raise ValueError('Illegal resampling strategy: %s' %
self._resampling_strategy)
if self._resampling_strategy in ['partial-cv', 'partial-cv-iterative-fit'] \
and self._ensemble_size != 0:
raise ValueError("Resampling strategy %s cannot be used "
"together with ensembles." %
self._resampling_strategy)
if self._resampling_strategy in ['partial-cv', 'cv',
'partial-cv-iterative-fit'] and \
not 'folds' in self._resampling_strategy_arguments:
self._resampling_strategy_arguments['folds'] = 5
self._backend._make_internals_directory()
if self._keep_models:
try:
os.makedirs(self._backend.get_model_dir())
except (OSError, FileExistsError) as e:
if not self._shared_mode:
raise
self._metric = metric
self._task = datamanager.info['task']
self._label_num = datamanager.info['label_num']
# == Pickle the data manager to speed up loading
data_manager_path = self._backend.save_datamanager(datamanager)
time_for_load_data = self._stopwatch.wall_elapsed(self._dataset_name)
if self._debug_mode:
self._print_load_time(self._dataset_name, self._time_for_task,
time_for_load_data, self._logger)
# == Perform dummy predictions
num_run = 1
#if self._resampling_strategy in ['holdout', 'holdout-iterative-fit']:
num_run = self._do_dummy_prediction(datamanager, num_run)
# = Create a searchspace
# Do this before One Hot Encoding to make sure that it creates a
# search space for a dense classifier even if one hot encoding would
# make it sparse (tradeoff; if one hot encoding would make it sparse,
# densifier and truncatedSVD would probably lead to a MemoryError,
# like this we can't use some of the preprocessing methods in case
# the data became sparse)
self.configuration_space, configspace_path = self._create_search_space(
self._backend.temporary_directory,
self._backend,
datamanager,
include_estimators=self._include_estimators,
exclude_estimators=self._exclude_estimators,
include_preprocessors=self._include_preprocessors,
exclude_preprocessors=self._exclude_preprocessors)
# == RUN ensemble builder
# Do this before calculating the meta-features to make sure that the
# dummy predictions are actually included in the ensemble even if
# calculating the meta-features takes very long
ensemble_task_name = 'runEnsemble'
self._stopwatch.start_task(ensemble_task_name)
time_left_for_ensembles = max(0,self._time_for_task \
- self._stopwatch.wall_elapsed(self._dataset_name))
if self._logger:
self._logger.info('Start Ensemble with %5.2fsec time left' %
time_left_for_ensembles)
if time_left_for_ensembles <= 0:
self._logger.warning("Not starting ensemble builder because there "
"is no time left!")
self._proc_ensemble = None
else:
self._proc_ensemble = self._get_ensemble_process(
time_left_for_ensembles)
if self._ensemble_size > 0:
self._proc_ensemble.start()
else:
self._logger.info('Not starting ensemble builder because '
'ensemble size is <= 0.')
self._stopwatch.stop_task(ensemble_task_name)
# kill the datamanager as it will be re-loaded anyways from sub processes
try:
del self._datamanager
except Exception:
pass
# => RUN SMAC
smac_task_name = 'runSMAC'
self._stopwatch.start_task(smac_task_name)
time_left_for_smac = max(
0, self._time_for_task -
self._stopwatch.wall_elapsed(self._dataset_name))
if self._logger:
self._logger.info('Start SMAC with %5.2fsec time left' %
time_left_for_smac)
if time_left_for_smac <= 0:
self._logger.warning("Not starting SMAC because there is no time "
"left.")
_proc_smac = None
else:
if self._per_run_time_limit is None or \
self._per_run_time_limit > time_left_for_smac:
print('Time limit for a single run is higher than total time '
'limit. Capping the limit for a single run to the total '
'time given to SMAC (%f)' % time_left_for_smac)
per_run_time_limit = time_left_for_smac
else:
per_run_time_limit = self._per_run_time_limit
_proc_smac = AutoMLSMBO(
config_space=self.configuration_space,
dataset_name=self._dataset_name,
backend=self._backend,
total_walltime_limit=time_left_for_smac,
func_eval_time_limit=per_run_time_limit,
memory_limit=self._ml_memory_limit,
data_memory_limit=self._data_memory_limit,
watcher=self._stopwatch,
start_num_run=num_run,
num_metalearning_cfgs=self.
_initial_configurations_via_metalearning,
config_file=configspace_path,
seed=self._seed,
metadata_directory=self._metadata_directory,
metric=self._metric,
resampling_strategy=self._resampling_strategy,
resampling_strategy_args=self._resampling_strategy_arguments,
shared_mode=self._shared_mode,
include_estimators=self._include_estimators,
exclude_estimators=self._exclude_estimators,
include_preprocessors=self._include_preprocessors,
exclude_preprocessors=self._exclude_preprocessors,
disable_file_output=self._disable_evaluator_output,
get_smac_object_callback=self._get_smac_object_callback,
smac_scenario_args=self._smac_scenario_args,
)
self.runhistory_, self.trajectory_ = \
_proc_smac.run_smbo()
trajectory_filename = os.path.join(
self._backend.get_smac_output_directory_for_run(self._seed),
'trajectory.json')
saveable_trajectory = \
[list(entry[:2]) + [entry[2].get_dictionary()] + list(entry[3:])
for entry in self.trajectory_]
with open(trajectory_filename, 'w') as fh:
json.dump(saveable_trajectory, fh)
# Wait until the ensemble process is finished to avoid shutting down
# while the ensemble builder tries to access the data
if self._proc_ensemble is not None and self._ensemble_size > 0:
self._proc_ensemble.join()
self._proc_ensemble = None
self._load_models()
return self
class AutoML(BaseEstimator, multiprocessing.Process):
def __init__(self,
tmp_dir,
output_dir,
time_left_for_this_task,
per_run_time_limit,
log_dir=None,
initial_configurations_via_metalearning=25,
ensemble_size=1,
ensemble_nbest=1,
seed=1,
ml_memory_limit=3000,
metadata_directory=None,
queue=None,
keep_models=True,
debug_mode=False,
include_estimators=None,
include_preprocessors=None,
resampling_strategy='holdout-iterative-fit',
resampling_strategy_arguments=None,
delete_tmp_folder_after_terminate=False,
delete_output_folder_after_terminate=False,
shared_mode=False,
precision=32,
max_iter_smac=None,
acquisition_function='EI'):
super(AutoML, self).__init__()
self._tmp_dir = tmp_dir
self._output_dir = output_dir
self._time_for_task = time_left_for_this_task
self._per_run_time_limit = per_run_time_limit
self._log_dir = log_dir if log_dir is not None else self._tmp_dir
self._initial_configurations_via_metalearning = \
initial_configurations_via_metalearning
self._ensemble_size = ensemble_size
self._ensemble_nbest = ensemble_nbest
self._seed = seed
self._ml_memory_limit = ml_memory_limit
self._data_memory_limit = None
self._metadata_directory = metadata_directory
self._queue = queue
self._keep_models = keep_models
self._include_estimators = include_estimators
self._include_preprocessors = include_preprocessors
self._resampling_strategy = resampling_strategy
self._resampling_strategy_arguments = resampling_strategy_arguments
self._max_iter_smac = max_iter_smac
self.delete_tmp_folder_after_terminate = \
delete_tmp_folder_after_terminate
self.delete_output_folder_after_terminate = \
delete_output_folder_after_terminate
self._shared_mode = shared_mode
self.precision = precision
self.acquisition_function = acquisition_function
self._datamanager = None
self._dataset_name = None
self._stopwatch = StopWatch()
self._logger = None
self._task = None
self._metric = None
self._label_num = None
self._parser = None
self.models_ = None
self.ensemble_ = None
self._can_predict = False
self._debug_mode = debug_mode
if not isinstance(self._time_for_task, int):
raise ValueError("time_left_for_this_task not of type integer, "
"but %s" % str(type(self._time_for_task)))
if not isinstance(self._per_run_time_limit, int):
raise ValueError("per_run_time_limit not of type integer, but %s" %
str(type(self._per_run_time_limit)))
# After assignging and checking variables...
self._backend = Backend(self._output_dir, self._tmp_dir)
def start_automl(self, parser):
self._parser = parser
self.start()
def start(self):
if self._parser is None:
raise ValueError('You must invoke start() only via start_automl()')
super(AutoML, self).start()
def run(self):
if self._parser is None:
raise ValueError('You must invoke run() only via start_automl()')
self._backend.save_start_time(self._seed)
self._stopwatch = StopWatch()
datamanager = get_data_manager(namespace=self._parser)
self._stopwatch.start_task(datamanager.name)
self._logger = self._get_logger(datamanager.name)
self._datamanager = datamanager
self._dataset_name = datamanager.name
self._fit(self._datamanager)
def fit(self,
X,
y,
task=MULTICLASS_CLASSIFICATION,
metric='acc_metric',
feat_type=None,
dataset_name=None):
if dataset_name is None:
m = hashlib.md5()
m.update(X.data)
dataset_name = m.hexdigest()
self._backend.save_start_time(self._seed)
self._stopwatch = StopWatch()
self._dataset_name = dataset_name
self._stopwatch.start_task(self._dataset_name)
self._logger = self._get_logger(dataset_name)
if isinstance(metric, str):
metric = STRING_TO_METRIC[metric]
if feat_type is not None and len(feat_type) != X.shape[1]:
raise ValueError('Array feat_type does not have same number of '
'variables as X has features. %d vs %d.' %
(len(feat_type), X.shape[1]))
if feat_type is not None and not all(
[isinstance(f, str) for f in feat_type]):
raise ValueError('Array feat_type must only contain strings.')
if feat_type is not None:
for ft in feat_type:
if ft.lower() not in ['categorical', 'numerical']:
raise ValueError('Only `Categorical` and `Numerical` are '
'valid feature types, you passed `%s`' %
ft)
self._data_memory_limit = None
loaded_data_manager = XYDataManager(X,
y,
task=task,
metric=metric,
feat_type=feat_type,
dataset_name=dataset_name,
encode_labels=False)
return self._fit(loaded_data_manager)
def fit_automl_dataset(self, dataset):
self._stopwatch = StopWatch()
self._backend.save_start_time(self._seed)
name = os.path.basename(dataset)
self._stopwatch.start_task(name)
self._start_task(self._stopwatch, name)
self._dataset_name = name
self._logger = self._get_logger(name)
self._logger.debug('======== Reading and converting data ==========')
# Encoding the labels will be done after the metafeature calculation!
self._data_memory_limit = float(self._ml_memory_limit) / 3
loaded_data_manager = CompetitionDataManager(
dataset,
encode_labels=False,
max_memory_in_mb=self._data_memory_limit)
loaded_data_manager_str = str(loaded_data_manager).split('\n')
for part in loaded_data_manager_str:
self._logger.debug(part)
return self._fit(loaded_data_manager)
def _get_logger(self, name):
logger_name = 'AutoML(%d):%s' % (self._seed, name)
setup_logger(os.path.join(self._tmp_dir, '%s.log' % str(logger_name)))
return get_logger(logger_name)
@staticmethod
def _start_task(watcher, task_name):
watcher.start_task(task_name)
@staticmethod
def _stop_task(watcher, task_name):
watcher.stop_task(task_name)
@staticmethod
def _print_load_time(basename, time_left_for_this_task, time_for_load_data,
logger):
time_left_after_reading = max(
0, time_left_for_this_task - time_for_load_data)
logger.info('Remaining time after reading %s %5.2f sec' %
(basename, time_left_after_reading))
return time_for_load_data
def _do_dummy_prediction(self, datamanager, num_run):
self._logger.info("Starting to create dummy predictions.")
time_limit = int(self._time_for_task / 6.)
memory_limit = int(self._ml_memory_limit)
_info = eval_with_limits(datamanager, self._tmp_dir, 1, self._seed,
num_run, self._resampling_strategy,
self._resampling_strategy_arguments,
memory_limit, time_limit)
if _info[4] == StatusType.SUCCESS:
self._logger.info("Finished creating dummy prediction 1/2.")
else:
self._logger.error('Error creating dummy prediction 1/2:%s ',
_info[3])
num_run += 1
_info = eval_with_limits(datamanager, self._tmp_dir, 2, self._seed,
num_run, self._resampling_strategy,
self._resampling_strategy_arguments,
memory_limit, time_limit)
if _info[4] == StatusType.SUCCESS:
self._logger.info("Finished creating dummy prediction 2/2.")
else:
self._logger.error('Error creating dummy prediction 2/2 %s',
_info[3])
num_run += 1
return num_run
def _fit(self, datamanager):
# Reset learnt stuff
self.models_ = None
self.ensemble_ = None
# Check arguments prior to doing anything!
if self._resampling_strategy not in [
'holdout', 'holdout-iterative-fit', 'cv', 'nested-cv',
'partial-cv'
]:
raise ValueError('Illegal resampling strategy: %s' %
self._resampling_strategy)
if self._resampling_strategy == 'partial-cv' and \
self._ensemble_size != 0:
raise ValueError("Resampling strategy partial-cv cannot be used "
"together with ensembles.")
acquisition_functions = ['EI', 'EIPS']
if self.acquisition_function not in acquisition_functions:
raise ValueError(
'Illegal acquisition %s: Must be one of %s.' %
(self.acquisition_function, acquisition_functions))
self._backend._make_internals_directory()
if self._keep_models:
try:
os.mkdir(self._backend.get_model_dir())
except OSError:
self._logger.warning("model directory already exists")
if not self._shared_mode:
raise
self._metric = datamanager.info['metric']
self._task = datamanager.info['task']
self._label_num = datamanager.info['label_num']
# == Pickle the data manager to speed up loading
data_manager_path = self._backend.save_datamanager(datamanager)
self._save_ensemble_data(datamanager.data['X_train'],
datamanager.data['Y_train'])
time_for_load_data = self._stopwatch.wall_elapsed(self._dataset_name)
if self._debug_mode:
self._print_load_time(self._dataset_name, self._time_for_task,
time_for_load_data, self._logger)
# == Perform dummy predictions
num_run = 1
#if self._resampling_strategy in ['holdout', 'holdout-iterative-fit']:
num_run = self._do_dummy_prediction(datamanager, num_run)
# = Create a searchspace
# Do this before One Hot Encoding to make sure that it creates a
# search space for a dense classifier even if one hot encoding would
# make it sparse (tradeoff; if one hot encoding would make it sparse,
# densifier and truncatedSVD would probably lead to a MemoryError,
# like this we can't use some of the preprocessing methods in case
# the data became sparse)
self.configuration_space, configspace_path = self._create_search_space(
self._tmp_dir, self._backend, datamanager,
self._include_estimators, self._include_preprocessors)
# == RUN ensemble builder
# Do this before calculating the meta-features to make sure that the
# dummy predictions are actually included in the ensemble even if
# calculating the meta-features takes very long
ensemble_task_name = 'runEnsemble'
self._stopwatch.start_task(ensemble_task_name)
time_left_for_ensembles = max(0,self._time_for_task \
- self._stopwatch.wall_elapsed(self._dataset_name))
if self._logger:
self._logger.info('Start Ensemble with %5.2fsec time left' %
time_left_for_ensembles)
if time_left_for_ensembles <= 0:
self._logger.warning("Not starting ensemble builder because there "
"is no time left!")
self._proc_ensemble = None
else:
self._proc_ensemble = self._get_ensemble_process(
time_left_for_ensembles)
self._proc_ensemble.start()
self._stopwatch.stop_task(ensemble_task_name)
# == RUN SMBO
# kill the datamanager as it will be re-loaded anyways from sub processes
try:
del self._datamanager
except Exception:
pass
# => RUN SMAC
smac_task_name = 'runSMAC'
self._stopwatch.start_task(smac_task_name)
time_left_for_smac = max(
0, self._time_for_task -
self._stopwatch.wall_elapsed(self._dataset_name))
if self._logger:
self._logger.info('Start SMAC with %5.2fsec time left' %
time_left_for_smac)
if time_left_for_smac <= 0:
self._logger.warning("Not starting SMAC because there is no time "
"left.")
self._procsmac = None
else:
self._proc_smac = AutoMLSMBO(
config_space=self.configuration_space,
dataset_name=self._dataset_name,
tmp_dir=self._tmp_dir,
output_dir=self._output_dir,
total_walltime_limit=time_left_for_smac,
func_eval_time_limit=self._per_run_time_limit,
memory_limit=self._ml_memory_limit,
data_memory_limit=self._data_memory_limit,
watcher=self._stopwatch,
start_num_run=num_run,
num_metalearning_cfgs=self.
_initial_configurations_via_metalearning,
config_file=configspace_path,
smac_iters=self._max_iter_smac,
seed=self._seed,
metadata_directory=self._metadata_directory,
resampling_strategy=self._resampling_strategy,
resampling_strategy_args=self._resampling_strategy_arguments,
acquisition_function=self.acquisition_function,
shared_mode=self._shared_mode)
self._proc_smac.start()
psutil_procs = []
procs = []
if self._proc_smac is not None:
proc_smac = psutil.Process(self._proc_smac.pid)
psutil_procs.append(proc_smac)
procs.append(self._proc_smac)
if self._proc_ensemble is not None:
proc_ensemble = psutil.Process(self._proc_ensemble.pid)
psutil_procs.append(proc_ensemble)
procs.append(self._proc_ensemble)
if self._queue is not None:
self._queue.put(
[time_for_load_data, data_manager_path, psutil_procs])
else:
for proc in procs:
proc.join()
if self._queue is None:
self._load_models()
return self
def refit(self, X, y):
if self._keep_models is not True:
raise ValueError(
"Predict can only be called if 'keep_models==True'")
if self.models_ is None or len(self.models_) == 0 or \
self.ensemble_ is None:
self._load_models()
for identifier in self.models_:
if identifier in self.ensemble_.get_model_identifiers():
model = self.models_[identifier]
# this updates the model inplace, it can then later be used in
# predict method
model.fit(X.copy(), y.copy())
self._can_predict = True
def predict(self, X):
return np.argmax(self.predict_proba(X), axis=1)
def predict_proba(self, X):
if self._keep_models is not True:
raise ValueError(
"Predict can only be called if 'keep_models==True'")
if not self._can_predict and \
self._resampling_strategy not in \
['holdout', 'holdout-iterative-fit']:
raise NotImplementedError(
'Predict is currently only implemented for resampling '
'strategy holdout.')
if self.models_ is None or len(self.models_) == 0 or \
self.ensemble_ is None:
self._load_models()
all_predictions = []
for identifier in self.ensemble_.get_model_identifiers():
model = self.models_[identifier]
X_ = X.copy()
if self._task in REGRESSION_TASKS:
prediction = model.predict(X_)
else:
prediction = model.predict_proba(X_)
if len(prediction.shape) < 1 or len(X_.shape) < 1 or \
X_.shape[0] < 1 or prediction.shape[0] != X_.shape[0]:
self._logger.warning(
"Prediction shape for model %s is %s "
"while X_.shape is %s" %
(model, str(prediction.shape), str(X_.shape)))
all_predictions.append(prediction)
if len(all_predictions) == 0:
raise ValueError(
'Something went wrong generating the predictions. '
'The ensemble should consist of the following '
'models: %s, the following models were loaded: '
'%s' % (str(list(self.ensemble_indices_.keys())),
str(list(self.models_.keys()))))
predictions = self.ensemble_.predict(all_predictions)
return predictions
def fit_ensemble(self,
task=None,
metric=None,
precision='32',
dataset_name=None,
ensemble_nbest=None,
ensemble_size=None):
if self._logger is None:
self._logger = self._get_logger(dataset_name)
self._proc_ensemble = self._get_ensemble_process(
1,
task,
metric,
precision,
dataset_name,
max_iterations=1,
ensemble_nbest=ensemble_nbest,
ensemble_size=ensemble_size)
self._proc_ensemble.main()
def _get_ensemble_process(self,
time_left_for_ensembles,
task=None,
metric=None,
precision=None,
dataset_name=None,
max_iterations=-1,
ensemble_nbest=None,
ensemble_size=None):
if task is None:
task = self._task
if metric is None:
metric = self._metric
if precision is None:
precision = self.precision
if dataset_name is None:
dataset_name = self._dataset_name
if ensemble_nbest is None:
ensemble_nbest = self._ensemble_nbest
if ensemble_size is None:
ensemble_size = self._ensemble_size
return EnsembleBuilder(autosklearn_tmp_dir=self._tmp_dir,
dataset_name=dataset_name,
task_type=task,
metric=metric,
limit=time_left_for_ensembles,
output_dir=self._output_dir,
ensemble_size=ensemble_size,
ensemble_nbest=ensemble_nbest,
seed=self._seed,
shared_mode=self._shared_mode,
precision=precision,
max_iterations=max_iterations)
def _load_models(self):
if self._shared_mode:
seed = -1
else:
seed = self._seed
self.ensemble_ = self._backend.load_ensemble(seed)
if self.ensemble_:
identifiers = self.ensemble_.identifiers_
self.models_ = self._backend.load_models_by_identifiers(
identifiers)
else:
self.models_ = self._backend.load_all_models(seed)
if len(self.models_) == 0:
raise ValueError('No models fitted!')
def score(self, X, y):
# fix: Consider only index 1 of second dimension
# Don't know if the reshaping should be done there or in calculate_score
prediction = self.predict_proba(X)
return calculate_score(y,
prediction,
self._task,
self._metric,
self._label_num,
logger=self._logger)
def show_models(self):
if self.models_ is None or len(self.models_) == 0 or \
self.ensemble_ is None:
self._load_models()
return self.ensemble_.pprint_ensemble_string(self.models_)
def _save_ensemble_data(self, X, y):
"""Split dataset and store Data for the ensemble script.
:param X:
:param y:
:return:
"""
task_name = 'LoadData'
self._start_task(self._stopwatch, task_name)
_, _, _, y_ensemble = resampling.split_data(X, y)
self._backend.save_targets_ensemble(y_ensemble)
self._stop_task(self._stopwatch, task_name)
def _create_search_space(self,
tmp_dir,
backend,
datamanager,
include_estimators=None,
include_preprocessors=None):
task_name = 'CreateConfigSpace'
self._stopwatch.start_task(task_name)
configspace_path = os.path.join(tmp_dir, 'space.pcs')
configuration_space = pipeline.get_configuration_space(
datamanager.info,
include_estimators=include_estimators,
include_preprocessors=include_preprocessors)
configuration_space = self.configuration_space_created_hook(
datamanager, configuration_space)
sp_string = pcs.write(configuration_space)
backend.write_txt_file(configspace_path, sp_string,
'Configuration space')
self._stopwatch.stop_task(task_name)
return configuration_space, configspace_path
def configuration_space_created_hook(self, datamanager,
configuration_space):
return configuration_space
def get_params(self, deep=True):
raise NotImplementedError('auto-sklearn does not implement '
'get_params() because it is not intended to '
'be optimized.')
def set_params(self, deep=True):
raise NotImplementedError('auto-sklearn does not implement '
'set_params() because it is not intended to '
'be optimized.')
def __del__(self):
self._delete_output_directories()
def _delete_output_directories(self):
if self.delete_output_folder_after_terminate:
try:
shutil.rmtree(self._output_dir)
except Exception:
if self._logger is not None:
self._logger.warning("Could not delete output dir: %s" %
self._output_dir)
else:
print("Could not delete output dir: %s" % self._output_dir)
if self.delete_tmp_folder_after_terminate:
try:
shutil.rmtree(self._tmp_dir)
except Exception:
if self._logger is not None:
self._logger.warning("Could not delete tmp dir: %s" %
self._tmp_dir)
pass
else:
print("Could not delete tmp dir: %s" % self._tmp_dir)
def _fit(self, datamanager):
# Reset learnt stuff
self.models_ = None
self.ensemble_ = None
# Check arguments prior to doing anything!
if self._resampling_strategy not in [
'holdout', 'holdout-iterative-fit', 'cv', 'nested-cv',
'partial-cv'
]:
raise ValueError('Illegal resampling strategy: %s' %
self._resampling_strategy)
if self._resampling_strategy == 'partial-cv' and \
self._ensemble_size != 0:
raise ValueError("Resampling strategy partial-cv cannot be used "
"together with ensembles.")
acquisition_functions = ['EI', 'EIPS']
if self.acquisition_function not in acquisition_functions:
raise ValueError(
'Illegal acquisition %s: Must be one of %s.' %
(self.acquisition_function, acquisition_functions))
self._backend._make_internals_directory()
if self._keep_models:
try:
os.mkdir(self._backend.get_model_dir())
except OSError:
self._logger.warning("model directory already exists")
if not self._shared_mode:
raise
self._metric = datamanager.info['metric']
self._task = datamanager.info['task']
self._label_num = datamanager.info['label_num']
# == Pickle the data manager to speed up loading
data_manager_path = self._backend.save_datamanager(datamanager)
self._save_ensemble_data(datamanager.data['X_train'],
datamanager.data['Y_train'])
time_for_load_data = self._stopwatch.wall_elapsed(self._dataset_name)
if self._debug_mode:
self._print_load_time(self._dataset_name, self._time_for_task,
time_for_load_data, self._logger)
# == Perform dummy predictions
num_run = 1
#if self._resampling_strategy in ['holdout', 'holdout-iterative-fit']:
num_run = self._do_dummy_prediction(datamanager, num_run)
# = Create a searchspace
# Do this before One Hot Encoding to make sure that it creates a
# search space for a dense classifier even if one hot encoding would
# make it sparse (tradeoff; if one hot encoding would make it sparse,
# densifier and truncatedSVD would probably lead to a MemoryError,
# like this we can't use some of the preprocessing methods in case
# the data became sparse)
self.configuration_space, configspace_path = self._create_search_space(
self._tmp_dir, self._backend, datamanager,
self._include_estimators, self._include_preprocessors)
# == RUN ensemble builder
# Do this before calculating the meta-features to make sure that the
# dummy predictions are actually included in the ensemble even if
# calculating the meta-features takes very long
ensemble_task_name = 'runEnsemble'
self._stopwatch.start_task(ensemble_task_name)
time_left_for_ensembles = max(0,self._time_for_task \
- self._stopwatch.wall_elapsed(self._dataset_name))
if self._logger:
self._logger.info('Start Ensemble with %5.2fsec time left' %
time_left_for_ensembles)
if time_left_for_ensembles <= 0:
self._logger.warning("Not starting ensemble builder because there "
"is no time left!")
self._proc_ensemble = None
else:
self._proc_ensemble = self._get_ensemble_process(
time_left_for_ensembles)
self._proc_ensemble.start()
self._stopwatch.stop_task(ensemble_task_name)
# == RUN SMBO
# kill the datamanager as it will be re-loaded anyways from sub processes
try:
del self._datamanager
except Exception:
pass
# => RUN SMAC
smac_task_name = 'runSMAC'
self._stopwatch.start_task(smac_task_name)
time_left_for_smac = max(
0, self._time_for_task -
self._stopwatch.wall_elapsed(self._dataset_name))
if self._logger:
self._logger.info('Start SMAC with %5.2fsec time left' %
time_left_for_smac)
if time_left_for_smac <= 0:
self._logger.warning("Not starting SMAC because there is no time "
"left.")
self._procsmac = None
else:
self._proc_smac = AutoMLSMBO(
config_space=self.configuration_space,
dataset_name=self._dataset_name,
tmp_dir=self._tmp_dir,
output_dir=self._output_dir,
total_walltime_limit=time_left_for_smac,
func_eval_time_limit=self._per_run_time_limit,
memory_limit=self._ml_memory_limit,
data_memory_limit=self._data_memory_limit,
watcher=self._stopwatch,
start_num_run=num_run,
num_metalearning_cfgs=self.
_initial_configurations_via_metalearning,
config_file=configspace_path,
smac_iters=self._max_iter_smac,
seed=self._seed,
metadata_directory=self._metadata_directory,
resampling_strategy=self._resampling_strategy,
resampling_strategy_args=self._resampling_strategy_arguments,
acquisition_function=self.acquisition_function,
shared_mode=self._shared_mode)
self._proc_smac.start()
psutil_procs = []
procs = []
if self._proc_smac is not None:
proc_smac = psutil.Process(self._proc_smac.pid)
psutil_procs.append(proc_smac)
procs.append(self._proc_smac)
if self._proc_ensemble is not None:
proc_ensemble = psutil.Process(self._proc_ensemble.pid)
psutil_procs.append(proc_ensemble)
procs.append(self._proc_ensemble)
if self._queue is not None:
self._queue.put(
[time_for_load_data, data_manager_path, psutil_procs])
else:
for proc in procs:
proc.join()
if self._queue is None:
self._load_models()
return self
class AutoML(BaseEstimator, multiprocessing.Process):
def __init__(self,
tmp_dir,
output_dir,
time_left_for_this_task,
per_run_time_limit,
log_dir=None,
initial_configurations_via_metalearning=25,
ensemble_size=1,
ensemble_nbest=1,
seed=1,
ml_memory_limit=3000,
metadata_directory=None,
queue=None,
keep_models=True,
debug_mode=False,
include_estimators=None,
include_preprocessors=None,
resampling_strategy='holdout-iterative-fit',
resampling_strategy_arguments=None,
delete_tmp_folder_after_terminate=False,
delete_output_folder_after_terminate=False,
shared_mode=False,
precision=32,
max_iter_smac=None,
acquisition_function='EI'):
super(AutoML, self).__init__()
self._tmp_dir = tmp_dir
self._output_dir = output_dir
self._time_for_task = time_left_for_this_task
self._per_run_time_limit = per_run_time_limit
self._log_dir = log_dir if log_dir is not None else self._tmp_dir
self._initial_configurations_via_metalearning = \
initial_configurations_via_metalearning
self._ensemble_size = ensemble_size
self._ensemble_nbest = ensemble_nbest
self._seed = seed
self._ml_memory_limit = ml_memory_limit
self._data_memory_limit = None
self._metadata_directory = metadata_directory
self._queue = queue
self._keep_models = keep_models
self._include_estimators = include_estimators
self._include_preprocessors = include_preprocessors
self._resampling_strategy = resampling_strategy
self._resampling_strategy_arguments = resampling_strategy_arguments
self._max_iter_smac = max_iter_smac
self.delete_tmp_folder_after_terminate = \
delete_tmp_folder_after_terminate
self.delete_output_folder_after_terminate = \
delete_output_folder_after_terminate
self._shared_mode = shared_mode
self.precision = precision
self.acquisition_function = acquisition_function
self._datamanager = None
self._dataset_name = None
self._stopwatch = StopWatch()
self._logger = None
self._task = None
self._metric = None
self._label_num = None
self._parser = None
self.models_ = None
self.ensemble_ = None
self._can_predict = False
self._debug_mode = debug_mode
if not isinstance(self._time_for_task, int):
raise ValueError("time_left_for_this_task not of type integer, "
"but %s" % str(type(self._time_for_task)))
if not isinstance(self._per_run_time_limit, int):
raise ValueError("per_run_time_limit not of type integer, but %s" %
str(type(self._per_run_time_limit)))
# After assignging and checking variables...
self._backend = Backend(self._output_dir, self._tmp_dir)
def start_automl(self, parser):
self._parser = parser
self.start()
def start(self):
if self._parser is None:
raise ValueError('You must invoke start() only via start_automl()')
super(AutoML, self).start()
def run(self):
if self._parser is None:
raise ValueError('You must invoke run() only via start_automl()')
self._backend.save_start_time(self._seed)
self._stopwatch = StopWatch()
datamanager = get_data_manager(namespace=self._parser)
self._stopwatch.start_task(datamanager.name)
self._logger = self._get_logger(datamanager.name)
self._datamanager = datamanager
self._dataset_name = datamanager.name
self._fit(self._datamanager)
def fit(self, X, y,
task=MULTICLASS_CLASSIFICATION,
metric='acc_metric',
feat_type=None,
dataset_name=None):
if dataset_name is None:
m = hashlib.md5()
m.update(X.data)
dataset_name = m.hexdigest()
self._backend.save_start_time(self._seed)
self._stopwatch = StopWatch()
self._dataset_name = dataset_name
self._stopwatch.start_task(self._dataset_name)
self._logger = self._get_logger(dataset_name)
if isinstance(metric, str):
metric = STRING_TO_METRIC[metric]
if feat_type is not None and len(feat_type) != X.shape[1]:
raise ValueError('Array feat_type does not have same number of '
'variables as X has features. %d vs %d.' %
(len(feat_type), X.shape[1]))
if feat_type is not None and not all([isinstance(f, str)
for f in feat_type]):
raise ValueError('Array feat_type must only contain strings.')
if feat_type is not None:
for ft in feat_type:
if ft.lower() not in ['categorical', 'numerical']:
raise ValueError('Only `Categorical` and `Numerical` are '
'valid feature types, you passed `%s`' % ft)
self._data_memory_limit = None
loaded_data_manager = XYDataManager(X, y,
task=task,
metric=metric,
feat_type=feat_type,
dataset_name=dataset_name,
encode_labels=False)
return self._fit(loaded_data_manager)
def fit_automl_dataset(self, dataset):
self._stopwatch = StopWatch()
self._backend.save_start_time(self._seed)
name = os.path.basename(dataset)
self._stopwatch.start_task(name)
self._start_task(self._stopwatch, name)
self._dataset_name = name
self._logger = self._get_logger(name)
self._logger.debug('======== Reading and converting data ==========')
# Encoding the labels will be done after the metafeature calculation!
self._data_memory_limit = float(self._ml_memory_limit) / 3
loaded_data_manager = CompetitionDataManager(
dataset, encode_labels=False,
max_memory_in_mb=self._data_memory_limit)
loaded_data_manager_str = str(loaded_data_manager).split('\n')
for part in loaded_data_manager_str:
self._logger.debug(part)
return self._fit(loaded_data_manager)
def _get_logger(self, name):
logger_name = 'AutoML(%d):%s' % (self._seed, name)
setup_logger(os.path.join(self._tmp_dir, '%s.log' % str(logger_name)))
return get_logger(logger_name)
@staticmethod
def _start_task(watcher, task_name):
watcher.start_task(task_name)
@staticmethod
def _stop_task(watcher, task_name):
watcher.stop_task(task_name)
@staticmethod
def _print_load_time(basename, time_left_for_this_task,
time_for_load_data, logger):
time_left_after_reading = max(
0, time_left_for_this_task - time_for_load_data)
logger.info('Remaining time after reading %s %5.2f sec' %
(basename, time_left_after_reading))
return time_for_load_data
def _do_dummy_prediction(self, datamanager, num_run):
self._logger.info("Starting to create dummy predictions.")
time_limit = int(self._time_for_task / 6.)
memory_limit = int(self._ml_memory_limit)
_info = eval_with_limits(datamanager, self._tmp_dir, 1,
self._seed, num_run,
self._resampling_strategy,
self._resampling_strategy_arguments,
memory_limit, time_limit)
if _info[4] == StatusType.SUCCESS:
self._logger.info("Finished creating dummy prediction 1/2.")
else:
self._logger.error('Error creating dummy prediction 1/2:%s ',
_info[3])
num_run += 1
_info = eval_with_limits(datamanager, self._tmp_dir, 2,
self._seed, num_run,
self._resampling_strategy,
self._resampling_strategy_arguments,
memory_limit, time_limit)
if _info[4] == StatusType.SUCCESS:
self._logger.info("Finished creating dummy prediction 2/2.")
else:
self._logger.error('Error creating dummy prediction 2/2 %s',
_info[3])
num_run += 1
return num_run
def _fit(self, datamanager):
# Reset learnt stuff
self.models_ = None
self.ensemble_ = None
# Check arguments prior to doing anything!
if self._resampling_strategy not in ['holdout', 'holdout-iterative-fit',
'cv', 'nested-cv', 'partial-cv']:
raise ValueError('Illegal resampling strategy: %s' %
self._resampling_strategy)
if self._resampling_strategy == 'partial-cv' and \
self._ensemble_size != 0:
raise ValueError("Resampling strategy partial-cv cannot be used "
"together with ensembles.")
acquisition_functions = ['EI', 'EIPS']
if self.acquisition_function not in acquisition_functions:
raise ValueError('Illegal acquisition %s: Must be one of %s.' %
(self.acquisition_function, acquisition_functions))
self._backend._make_internals_directory()
if self._keep_models:
try:
os.mkdir(self._backend.get_model_dir())
except OSError:
self._logger.warning("model directory already exists")
if not self._shared_mode:
raise
self._metric = datamanager.info['metric']
self._task = datamanager.info['task']
self._label_num = datamanager.info['label_num']
# == Pickle the data manager to speed up loading
data_manager_path = self._backend.save_datamanager(datamanager)
self._save_ensemble_data(
datamanager.data['X_train'],
datamanager.data['Y_train'])
time_for_load_data = self._stopwatch.wall_elapsed(self._dataset_name)
if self._debug_mode:
self._print_load_time(
self._dataset_name,
self._time_for_task,
time_for_load_data,
self._logger)
# == Perform dummy predictions
num_run = 1
#if self._resampling_strategy in ['holdout', 'holdout-iterative-fit']:
num_run = self._do_dummy_prediction(datamanager, num_run)
# = Create a searchspace
# Do this before One Hot Encoding to make sure that it creates a
# search space for a dense classifier even if one hot encoding would
# make it sparse (tradeoff; if one hot encoding would make it sparse,
# densifier and truncatedSVD would probably lead to a MemoryError,
# like this we can't use some of the preprocessing methods in case
# the data became sparse)
self.configuration_space, configspace_path = self._create_search_space(
self._tmp_dir,
self._backend,
datamanager,
self._include_estimators,
self._include_preprocessors)
# == RUN ensemble builder
# Do this before calculating the meta-features to make sure that the
# dummy predictions are actually included in the ensemble even if
# calculating the meta-features takes very long
ensemble_task_name = 'runEnsemble'
self._stopwatch.start_task(ensemble_task_name)
time_left_for_ensembles = max(0,self._time_for_task \
- self._stopwatch.wall_elapsed(self._dataset_name))
if self._logger:
self._logger.info(
'Start Ensemble with %5.2fsec time left' % time_left_for_ensembles)
if time_left_for_ensembles <= 0:
self._logger.warning("Not starting ensemble builder because there "
"is no time left!")
self._proc_ensemble = None
else:
self._proc_ensemble = self._get_ensemble_process(time_left_for_ensembles)
self._proc_ensemble.start()
self._stopwatch.stop_task(ensemble_task_name)
# == RUN SMBO
# kill the datamanager as it will be re-loaded anyways from sub processes
try:
del self._datamanager
except Exception:
pass
# => RUN SMAC
smac_task_name = 'runSMAC'
self._stopwatch.start_task(smac_task_name)
time_left_for_smac = max(0,
self._time_for_task - self._stopwatch.wall_elapsed(
self._dataset_name))
if self._logger:
self._logger.info(
'Start SMAC with %5.2fsec time left' % time_left_for_smac)
if time_left_for_smac <= 0:
self._logger.warning("Not starting SMAC because there is no time "
"left.")
self._procsmac = None
else:
self._proc_smac = AutoMLSMBO(config_space=self.configuration_space,
dataset_name=self._dataset_name,
tmp_dir=self._tmp_dir,
output_dir=self._output_dir,
total_walltime_limit=time_left_for_smac,
func_eval_time_limit=self._per_run_time_limit,
memory_limit=self._ml_memory_limit,
data_memory_limit=self._data_memory_limit,
watcher=self._stopwatch,
start_num_run=num_run,
num_metalearning_cfgs=self._initial_configurations_via_metalearning,
config_file=configspace_path,
smac_iters=self._max_iter_smac,
seed=self._seed,
metadata_directory=self._metadata_directory,
resampling_strategy=self._resampling_strategy,
resampling_strategy_args=self._resampling_strategy_arguments,
acquisition_function=self.acquisition_function,
shared_mode=self._shared_mode)
self._proc_smac.start()
psutil_procs = []
procs = []
if self._proc_smac is not None:
proc_smac = psutil.Process(self._proc_smac.pid)
psutil_procs.append(proc_smac)
procs.append(self._proc_smac)
if self._proc_ensemble is not None:
proc_ensemble = psutil.Process(self._proc_ensemble.pid)
psutil_procs.append(proc_ensemble)
procs.append(self._proc_ensemble)
if self._queue is not None:
self._queue.put([time_for_load_data, data_manager_path, psutil_procs])
else:
for proc in procs:
proc.join()
if self._queue is None:
self._load_models()
return self
def refit(self, X, y):
if self._keep_models is not True:
raise ValueError(
"Predict can only be called if 'keep_models==True'")
if self.models_ is None or len(self.models_) == 0 or \
self.ensemble_ is None:
self._load_models()
for identifier in self.models_:
if identifier in self.ensemble_.get_model_identifiers():
model = self.models_[identifier]
# this updates the model inplace, it can then later be used in
# predict method
model.fit(X.copy(), y.copy())
self._can_predict = True
def predict(self, X):
return np.argmax(self.predict_proba(X), axis=1)
def predict_proba(self, X):
if self._keep_models is not True:
raise ValueError(
"Predict can only be called if 'keep_models==True'")
if not self._can_predict and \
self._resampling_strategy not in \
['holdout', 'holdout-iterative-fit']:
raise NotImplementedError(
'Predict is currently only implemented for resampling '
'strategy holdout.')
if self.models_ is None or len(self.models_) == 0 or \
self.ensemble_ is None:
self._load_models()
all_predictions = []
for identifier in self.ensemble_.get_model_identifiers():
model = self.models_[identifier]
X_ = X.copy()
if self._task in REGRESSION_TASKS:
prediction = model.predict(X_)
else:
prediction = model.predict_proba(X_)
if len(prediction.shape) < 1 or len(X_.shape) < 1 or \
X_.shape[0] < 1 or prediction.shape[0] != X_.shape[0]:
self._logger.warning("Prediction shape for model %s is %s "
"while X_.shape is %s" %
(model, str(prediction.shape),
str(X_.shape)))
all_predictions.append(prediction)
if len(all_predictions) == 0:
raise ValueError('Something went wrong generating the predictions. '
'The ensemble should consist of the following '
'models: %s, the following models were loaded: '
'%s' % (str(list(self.ensemble_indices_.keys())),
str(list(self.models_.keys()))))
predictions = self.ensemble_.predict(all_predictions)
return predictions
def fit_ensemble(self, task=None, metric=None, precision='32',
dataset_name=None, ensemble_nbest=None,
ensemble_size=None):
if self._logger is None:
self._logger = self._get_logger(dataset_name)
self._proc_ensemble = self._get_ensemble_process(
1, task, metric, precision, dataset_name, max_iterations=1,
ensemble_nbest=ensemble_nbest, ensemble_size=ensemble_size)
self._proc_ensemble.main()
def _get_ensemble_process(self, time_left_for_ensembles,
task=None, metric=None, precision=None,
dataset_name=None, max_iterations=-1,
ensemble_nbest=None, ensemble_size=None):
if task is None:
task = self._task
if metric is None:
metric = self._metric
if precision is None:
precision = self.precision
if dataset_name is None:
dataset_name = self._dataset_name
if ensemble_nbest is None:
ensemble_nbest = self._ensemble_nbest
if ensemble_size is None:
ensemble_size = self._ensemble_size
return EnsembleBuilder(autosklearn_tmp_dir=self._tmp_dir,
dataset_name=dataset_name,
task_type=task,
metric=metric,
limit=time_left_for_ensembles,
output_dir=self._output_dir,
ensemble_size=ensemble_size,
ensemble_nbest=ensemble_nbest,
seed=self._seed,
shared_mode=self._shared_mode,
precision=precision,
max_iterations=max_iterations)
def _load_models(self):
if self._shared_mode:
seed = -1
else:
seed = self._seed
self.ensemble_ = self._backend.load_ensemble(seed)
if self.ensemble_:
identifiers = self.ensemble_.identifiers_
self.models_ = self._backend.load_models_by_identifiers(identifiers)
else:
self.models_ = self._backend.load_all_models(seed)
if len(self.models_) == 0:
raise ValueError('No models fitted!')
def score(self, X, y):
# fix: Consider only index 1 of second dimension
# Don't know if the reshaping should be done there or in calculate_score
prediction = self.predict_proba(X)
return calculate_score(y, prediction, self._task,
self._metric, self._label_num,
logger=self._logger)
def show_models(self):
if self.models_ is None or len(self.models_) == 0 or \
self.ensemble_ is None:
self._load_models()
return self.ensemble_.pprint_ensemble_string(self.models_)
def _save_ensemble_data(self, X, y):
"""Split dataset and store Data for the ensemble script.
:param X:
:param y:
:return:
"""
task_name = 'LoadData'
self._start_task(self._stopwatch, task_name)
_, _, _, y_ensemble = resampling.split_data(X, y)
self._backend.save_targets_ensemble(y_ensemble)
self._stop_task(self._stopwatch, task_name)
def _create_search_space(self, tmp_dir, backend, datamanager,
include_estimators=None,
include_preprocessors=None):
task_name = 'CreateConfigSpace'
self._stopwatch.start_task(task_name)
configspace_path = os.path.join(tmp_dir, 'space.pcs')
configuration_space = pipeline.get_configuration_space(
datamanager.info,
include_estimators=include_estimators,
include_preprocessors=include_preprocessors)
configuration_space = self.configuration_space_created_hook(
datamanager, configuration_space)
sp_string = pcs.write(configuration_space)
backend.write_txt_file(configspace_path, sp_string,
'Configuration space')
self._stopwatch.stop_task(task_name)
return configuration_space, configspace_path
def configuration_space_created_hook(self, datamanager, configuration_space):
return configuration_space
def get_params(self, deep=True):
raise NotImplementedError('auto-sklearn does not implement '
'get_params() because it is not intended to '
'be optimized.')
def set_params(self, deep=True):
raise NotImplementedError('auto-sklearn does not implement '
'set_params() because it is not intended to '
'be optimized.')
def __del__(self):
self._delete_output_directories()
def _delete_output_directories(self):
if self.delete_output_folder_after_terminate:
try:
shutil.rmtree(self._output_dir)
except Exception:
if self._logger is not None:
self._logger.warning("Could not delete output dir: %s" %
self._output_dir)
else:
print("Could not delete output dir: %s" %
self._output_dir)
if self.delete_tmp_folder_after_terminate:
try:
shutil.rmtree(self._tmp_dir)
except Exception:
if self._logger is not None:
self._logger.warning("Could not delete tmp dir: %s" %
self._tmp_dir)
pass
else:
print("Could not delete tmp dir: %s" %
self._tmp_dir)
def _fit(self, datamanager):
# Reset learnt stuff
self.models_ = None
self.ensemble_ = None
# Check arguments prior to doing anything!
if self._resampling_strategy not in ['holdout', 'holdout-iterative-fit',
'cv', 'nested-cv', 'partial-cv']:
raise ValueError('Illegal resampling strategy: %s' %
self._resampling_strategy)
if self._resampling_strategy == 'partial-cv' and \
self._ensemble_size != 0:
raise ValueError("Resampling strategy partial-cv cannot be used "
"together with ensembles.")
acquisition_functions = ['EI', 'EIPS']
if self.acquisition_function not in acquisition_functions:
raise ValueError('Illegal acquisition %s: Must be one of %s.' %
(self.acquisition_function, acquisition_functions))
self._backend._make_internals_directory()
if self._keep_models:
try:
os.mkdir(self._backend.get_model_dir())
except OSError:
self._logger.warning("model directory already exists")
if not self._shared_mode:
raise
self._metric = datamanager.info['metric']
self._task = datamanager.info['task']
self._label_num = datamanager.info['label_num']
# == Pickle the data manager to speed up loading
data_manager_path = self._backend.save_datamanager(datamanager)
self._save_ensemble_data(
datamanager.data['X_train'],
datamanager.data['Y_train'])
time_for_load_data = self._stopwatch.wall_elapsed(self._dataset_name)
if self._debug_mode:
self._print_load_time(
self._dataset_name,
self._time_for_task,
time_for_load_data,
self._logger)
# == Perform dummy predictions
num_run = 1
#if self._resampling_strategy in ['holdout', 'holdout-iterative-fit']:
num_run = self._do_dummy_prediction(datamanager, num_run)
# = Create a searchspace
# Do this before One Hot Encoding to make sure that it creates a
# search space for a dense classifier even if one hot encoding would
# make it sparse (tradeoff; if one hot encoding would make it sparse,
# densifier and truncatedSVD would probably lead to a MemoryError,
# like this we can't use some of the preprocessing methods in case
# the data became sparse)
self.configuration_space, configspace_path = self._create_search_space(
self._tmp_dir,
self._backend,
datamanager,
self._include_estimators,
self._include_preprocessors)
# == RUN ensemble builder
# Do this before calculating the meta-features to make sure that the
# dummy predictions are actually included in the ensemble even if
# calculating the meta-features takes very long
ensemble_task_name = 'runEnsemble'
self._stopwatch.start_task(ensemble_task_name)
time_left_for_ensembles = max(0,self._time_for_task \
- self._stopwatch.wall_elapsed(self._dataset_name))
if self._logger:
self._logger.info(
'Start Ensemble with %5.2fsec time left' % time_left_for_ensembles)
if time_left_for_ensembles <= 0:
self._logger.warning("Not starting ensemble builder because there "
"is no time left!")
self._proc_ensemble = None
else:
self._proc_ensemble = self._get_ensemble_process(time_left_for_ensembles)
self._proc_ensemble.start()
self._stopwatch.stop_task(ensemble_task_name)
# == RUN SMBO
# kill the datamanager as it will be re-loaded anyways from sub processes
try:
del self._datamanager
except Exception:
pass
# => RUN SMAC
smac_task_name = 'runSMAC'
self._stopwatch.start_task(smac_task_name)
time_left_for_smac = max(0,
self._time_for_task - self._stopwatch.wall_elapsed(
self._dataset_name))
if self._logger:
self._logger.info(
'Start SMAC with %5.2fsec time left' % time_left_for_smac)
if time_left_for_smac <= 0:
self._logger.warning("Not starting SMAC because there is no time "
"left.")
self._procsmac = None
else:
self._proc_smac = AutoMLSMBO(config_space=self.configuration_space,
dataset_name=self._dataset_name,
tmp_dir=self._tmp_dir,
output_dir=self._output_dir,
total_walltime_limit=time_left_for_smac,
func_eval_time_limit=self._per_run_time_limit,
memory_limit=self._ml_memory_limit,
data_memory_limit=self._data_memory_limit,
watcher=self._stopwatch,
start_num_run=num_run,
num_metalearning_cfgs=self._initial_configurations_via_metalearning,
config_file=configspace_path,
smac_iters=self._max_iter_smac,
seed=self._seed,
metadata_directory=self._metadata_directory,
resampling_strategy=self._resampling_strategy,
resampling_strategy_args=self._resampling_strategy_arguments,
acquisition_function=self.acquisition_function,
shared_mode=self._shared_mode)
self._proc_smac.start()
psutil_procs = []
procs = []
if self._proc_smac is not None:
proc_smac = psutil.Process(self._proc_smac.pid)
psutil_procs.append(proc_smac)
procs.append(self._proc_smac)
if self._proc_ensemble is not None:
proc_ensemble = psutil.Process(self._proc_ensemble.pid)
psutil_procs.append(proc_ensemble)
procs.append(self._proc_ensemble)
if self._queue is not None:
self._queue.put([time_for_load_data, data_manager_path, psutil_procs])
else:
for proc in procs:
proc.join()
if self._queue is None:
self._load_models()
return self
class AutoML(BaseEstimator):
def __init__(self,
backend,
time_left_for_this_task,
per_run_time_limit,
log_dir=None,
initial_configurations_via_metalearning=25,
ensemble_size=1,
ensemble_nbest=1,
seed=1,
ml_memory_limit=3000,
metadata_directory=None,
keep_models=True,
debug_mode=False,
include_estimators=None,
include_preprocessors=None,
resampling_strategy='holdout-iterative-fit',
resampling_strategy_arguments=None,
delete_tmp_folder_after_terminate=False,
delete_output_folder_after_terminate=False,
shared_mode=False,
precision=32,
max_iter_smac=None,
acquisition_function='EI'):
super(AutoML, self).__init__()
self._backend = backend
#self._tmp_dir = tmp_dir
#self._output_dir = output_dir
self._time_for_task = time_left_for_this_task
self._per_run_time_limit = per_run_time_limit
#self._log_dir = log_dir if log_dir is not None else self._tmp_dir
self._initial_configurations_via_metalearning = \
initial_configurations_via_metalearning
self._ensemble_size = ensemble_size
self._ensemble_nbest = ensemble_nbest
self._seed = seed
self._ml_memory_limit = ml_memory_limit
self._data_memory_limit = None
self._metadata_directory = metadata_directory
self._keep_models = keep_models
self._include_estimators = include_estimators
self._include_preprocessors = include_preprocessors
self._resampling_strategy = resampling_strategy
self._resampling_strategy_arguments = resampling_strategy_arguments
self._max_iter_smac = max_iter_smac
#self.delete_tmp_folder_after_terminate = \
# delete_tmp_folder_after_terminate
#self.delete_output_folder_after_terminate = \
# delete_output_folder_after_terminate
self._shared_mode = shared_mode
self.precision = precision
self.acquisition_function = acquisition_function
self._datamanager = None
self._dataset_name = None
self._stopwatch = StopWatch()
self._logger = None
self._task = None
self._metric = None
self._label_num = None
self._parser = None
self.models_ = None
self.ensemble_ = None
self._can_predict = False
self._debug_mode = debug_mode
if not isinstance(self._time_for_task, int):
raise ValueError("time_left_for_this_task not of type integer, "
"but %s" % str(type(self._time_for_task)))
if not isinstance(self._per_run_time_limit, int):
raise ValueError("per_run_time_limit not of type integer, but %s" %
str(type(self._per_run_time_limit)))
# After assignging and checking variables...
#self._backend = Backend(self._output_dir, self._tmp_dir)
def start_automl(self, parser):
self._parser = parser
self.start()
def start(self):
if self._parser is None:
raise ValueError('You must invoke start() only via start_automl()')
super(AutoML, self).start()
def run(self):
if self._parser is None:
raise ValueError('You must invoke run() only via start_automl()')
self._backend.save_start_time(self._seed)
self._stopwatch = StopWatch()
datamanager = get_data_manager(namespace=self._parser)
self._stopwatch.start_task(datamanager.name)
self._logger = self._get_logger(datamanager.name)
self._datamanager = datamanager
self._dataset_name = datamanager.name
self._fit(self._datamanager)
def fit(self, X, y,
task=MULTICLASS_CLASSIFICATION,
metric='acc_metric',
feat_type=None,
dataset_name=None):
if not self._shared_mode:
self._backend.context.delete_directories()
else:
# If this fails, it's likely that this is the first call to get
# the data manager
try:
D = self._backend.load_datamanager()
dataset_name = D.name
except IOError:
pass
self._backend.context.create_directories()
if dataset_name is None:
m = hashlib.md5()
m.update(X.data)
dataset_name = m.hexdigest()
self._backend.save_start_time(self._seed)
self._stopwatch = StopWatch()
self._dataset_name = dataset_name
self._stopwatch.start_task(self._dataset_name)
self._logger = self._get_logger(dataset_name)
if isinstance(metric, str):
metric = STRING_TO_METRIC[metric]
if feat_type is not None and len(feat_type) != X.shape[1]:
raise ValueError('Array feat_type does not have same number of '
'variables as X has features. %d vs %d.' %
(len(feat_type), X.shape[1]))
if feat_type is not None and not all([isinstance(f, str)
for f in feat_type]):
raise ValueError('Array feat_type must only contain strings.')
if feat_type is not None:
for ft in feat_type:
if ft.lower() not in ['categorical', 'numerical']:
raise ValueError('Only `Categorical` and `Numerical` are '
'valid feature types, you passed `%s`' % ft)
self._data_memory_limit = None
loaded_data_manager = XYDataManager(X, y,
task=task,
metric=metric,
feat_type=feat_type,
dataset_name=dataset_name,
encode_labels=False)
return self._fit(loaded_data_manager)
def fit_automl_dataset(self, dataset):
self._stopwatch = StopWatch()
self._backend.save_start_time(self._seed)
name = os.path.basename(dataset)
self._stopwatch.start_task(name)
self._start_task(self._stopwatch, name)
self._dataset_name = name
self._logger = self._get_logger(name)
self._logger.debug('======== Reading and converting data ==========')
# Encoding the labels will be done after the metafeature calculation!
self._data_memory_limit = float(self._ml_memory_limit) / 3
loaded_data_manager = CompetitionDataManager(
dataset, encode_labels=False,
max_memory_in_mb=self._data_memory_limit)
loaded_data_manager_str = str(loaded_data_manager).split('\n')
for part in loaded_data_manager_str:
self._logger.debug(part)
return self._fit(loaded_data_manager)
def _get_logger(self, name):
logger_name = 'AutoML(%d):%s' % (self._seed, name)
setup_logger(os.path.join(self._backend.temporary_directory, '%s.log' % str(logger_name)))
return get_logger(logger_name)
@staticmethod
def _start_task(watcher, task_name):
watcher.start_task(task_name)
@staticmethod
def _stop_task(watcher, task_name):
watcher.stop_task(task_name)
@staticmethod
def _print_load_time(basename, time_left_for_this_task,
time_for_load_data, logger):
time_left_after_reading = max(
0, time_left_for_this_task - time_for_load_data)
logger.info('Remaining time after reading %s %5.2f sec' %
(basename, time_left_after_reading))
return time_for_load_data
def _do_dummy_prediction(self, datamanager, num_run):
self._logger.info("Starting to create dummy predictions.")
time_limit = int(self._time_for_task / 6.)
memory_limit = int(self._ml_memory_limit)
_info = eval_with_limits(datamanager, self._backend, 1,
self._seed, num_run,
self._resampling_strategy,
self._resampling_strategy_arguments,
memory_limit, time_limit,
logger=self._logger)
if _info[4] == StatusType.SUCCESS:
self._logger.info("Finished creating dummy prediction 1/2.")
else:
self._logger.error('Error creating dummy prediction 1/2:%s ',
_info[3])
num_run += 1
_info = eval_with_limits(datamanager, self._backend, 2,
self._seed, num_run,
self._resampling_strategy,
self._resampling_strategy_arguments,
memory_limit, time_limit,
logger=self._logger)
if _info[4] == StatusType.SUCCESS:
self._logger.info("Finished creating dummy prediction 2/2.")
else:
self._logger.error('Error creating dummy prediction 2/2 %s',
_info[3])
num_run += 1
return num_run
def _fit(self, datamanager):
# Reset learnt stuff
self.models_ = None
self.ensemble_ = None
# Check arguments prior to doing anything!
if self._resampling_strategy not in ['holdout', 'holdout-iterative-fit',
'cv', 'nested-cv', 'partial-cv']:
raise ValueError('Illegal resampling strategy: %s' %
self._resampling_strategy)
if self._resampling_strategy == 'partial-cv' and \
self._ensemble_size != 0:
raise ValueError("Resampling strategy partial-cv cannot be used "
"together with ensembles.")
acquisition_functions = ['EI', 'EIPS']
if self.acquisition_function not in acquisition_functions:
raise ValueError('Illegal acquisition %s: Must be one of %s.' %
(self.acquisition_function, acquisition_functions))
self._backend._make_internals_directory()
if self._keep_models:
try:
os.mkdir(self._backend.get_model_dir())
except OSError:
if not self._shared_mode:
raise
self._metric = datamanager.info['metric']
self._task = datamanager.info['task']
self._label_num = datamanager.info['label_num']
# == Pickle the data manager to speed up loading
data_manager_path = self._backend.save_datamanager(datamanager)
self._save_ensemble_data(
datamanager.data['X_train'],
datamanager.data['Y_train'])
time_for_load_data = self._stopwatch.wall_elapsed(self._dataset_name)
if self._debug_mode:
self._print_load_time(
self._dataset_name,
self._time_for_task,
time_for_load_data,
self._logger)
# == Perform dummy predictions
num_run = 1
#if self._resampling_strategy in ['holdout', 'holdout-iterative-fit']:
num_run = self._do_dummy_prediction(datamanager, num_run)
# = Create a searchspace
# Do this before One Hot Encoding to make sure that it creates a
# search space for a dense classifier even if one hot encoding would
# make it sparse (tradeoff; if one hot encoding would make it sparse,
# densifier and truncatedSVD would probably lead to a MemoryError,
# like this we can't use some of the preprocessing methods in case
# the data became sparse)
self.configuration_space, configspace_path = self._create_search_space(
self._backend.temporary_directory,
self._backend,
datamanager,
self._include_estimators,
self._include_preprocessors)
# == RUN ensemble builder
# Do this before calculating the meta-features to make sure that the
# dummy predictions are actually included in the ensemble even if
# calculating the meta-features takes very long
ensemble_task_name = 'runEnsemble'
self._stopwatch.start_task(ensemble_task_name)
time_left_for_ensembles = max(0,self._time_for_task \
- self._stopwatch.wall_elapsed(self._dataset_name))
if self._logger:
self._logger.info(
'Start Ensemble with %5.2fsec time left' % time_left_for_ensembles)
if time_left_for_ensembles <= 0:
self._logger.warning("Not starting ensemble builder because there "
"is no time left!")
self._proc_ensemble = None
else:
self._proc_ensemble = self._get_ensemble_process(time_left_for_ensembles)
if self._ensemble_size > 0:
self._proc_ensemble.start()
else:
self._logger.info('Not starting ensemble builder because '
'ensemble size is <= 0.')
self._stopwatch.stop_task(ensemble_task_name)
# kill the datamanager as it will be re-loaded anyways from sub processes
try:
del self._datamanager
except Exception:
pass
# => RUN SMAC
smac_task_name = 'runSMAC'
self._stopwatch.start_task(smac_task_name)
time_left_for_smac = max(0,
self._time_for_task - self._stopwatch.wall_elapsed(
self._dataset_name))
if self._logger:
self._logger.info(
'Start SMAC with %5.2fsec time left' % time_left_for_smac)
if time_left_for_smac <= 0:
self._logger.warning("Not starting SMAC because there is no time "
"left.")
self._proc_smac = None
else:
if self._per_run_time_limit is None or \
self._per_run_time_limit > time_left_for_smac:
print('Time limit for a single run is higher than total time '
'limit. Capping the limit for a single run to the total '
'time given to SMAC (%f)' % time_left_for_smac)
per_run_time_limit = time_left_for_smac
else:
per_run_time_limit = self._per_run_time_limit
self._proc_smac = AutoMLSMBO(config_space=self.configuration_space,
dataset_name=self._dataset_name,
backend=self._backend,
total_walltime_limit=time_left_for_smac,
func_eval_time_limit=per_run_time_limit,
memory_limit=self._ml_memory_limit,
data_memory_limit=self._data_memory_limit,
watcher=self._stopwatch,
start_num_run=num_run,
num_metalearning_cfgs=self._initial_configurations_via_metalearning,
config_file=configspace_path,
smac_iters=self._max_iter_smac,
seed=self._seed,
metadata_directory=self._metadata_directory,
resampling_strategy=self._resampling_strategy,
resampling_strategy_args=self._resampling_strategy_arguments,
acquisition_function=self.acquisition_function,
shared_mode=self._shared_mode)
self._proc_smac.run_smbo()
self._proc_ensemble = None
self._load_models()
return self
def refit(self, X, y):
if self._keep_models is not True:
raise ValueError(
"Predict can only be called if 'keep_models==True'")
if self.models_ is None or len(self.models_) == 0 or \
self.ensemble_ is None:
self._load_models()
for identifier in self.models_:
if identifier in self.ensemble_.get_model_identifiers():
model = self.models_[identifier]
# this updates the model inplace, it can then later be used in
# predict method
model.fit(X.copy(), y.copy())
self._can_predict = True
return self
def predict(self, X):
if self._keep_models is not True:
raise ValueError(
"Predict can only be called if 'keep_models==True'")
if not self._can_predict and \
self._resampling_strategy not in \
['holdout', 'holdout-iterative-fit']:
raise NotImplementedError(
'Predict is currently only implemented for resampling '
'strategy holdout.')
if self.models_ is None or len(self.models_) == 0 or \
self.ensemble_ is None:
self._load_models()
all_predictions = []
for identifier in self.ensemble_.get_model_identifiers():
model = self.models_[identifier]
X_ = X.copy()
if self._task in REGRESSION_TASKS:
prediction = model.predict(X_)
else:
prediction = model.predict_proba(X_)
if len(prediction.shape) < 1 or len(X_.shape) < 1 or \
X_.shape[0] < 1 or prediction.shape[0] != X_.shape[0]:
self._logger.warning("Prediction shape for model %s is %s "
"while X_.shape is %s" %
(model, str(prediction.shape),
str(X_.shape)))
all_predictions.append(prediction)
if len(all_predictions) == 0:
raise ValueError('Something went wrong generating the predictions. '
'The ensemble should consist of the following '
'models: %s, the following models were loaded: '
'%s' % (str(list(self.ensemble_indices_.keys())),
str(list(self.models_.keys()))))
predictions = self.ensemble_.predict(all_predictions)
return predictions
def fit_ensemble(self, y, task=None, metric=None, precision='32',
dataset_name=None, ensemble_nbest=None,
ensemble_size=None):
if self._logger is None:
self._logger = self._get_logger(dataset_name)
self._proc_ensemble = self._get_ensemble_process(
1, task, metric, precision, dataset_name, max_iterations=1,
ensemble_nbest=ensemble_nbest, ensemble_size=ensemble_size)
self._proc_ensemble.main()
return self
def _get_ensemble_process(self, time_left_for_ensembles,
task=None, metric=None, precision=None,
dataset_name=None, max_iterations=-1,
ensemble_nbest=None, ensemble_size=None):
if task is None:
task = self._task
else:
self._task = task
if metric is None:
metric = self._metric
else:
self._metric = metric
if precision is None:
precision = self.precision
else:
self.precision = precision
if dataset_name is None:
dataset_name = self._dataset_name
else:
self._dataset_name = dataset_name
if ensemble_nbest is None:
ensemble_nbest = self._ensemble_nbest
else:
self._ensemble_nbest = ensemble_nbest
if ensemble_size is None:
ensemble_size = self._ensemble_size
else:
self._ensemble_size = ensemble_size
return EnsembleBuilder(backend=self._backend,
dataset_name=dataset_name,
task_type=task,
metric=metric,
limit=time_left_for_ensembles,
ensemble_size=ensemble_size,
ensemble_nbest=ensemble_nbest,
seed=self._seed,
shared_mode=self._shared_mode,
precision=precision,
max_iterations=max_iterations)
def _load_models(self):
if self._shared_mode:
seed = -1
else:
seed = self._seed
self.ensemble_ = self._backend.load_ensemble(seed)
if self.ensemble_:
identifiers = self.ensemble_.identifiers_
self.models_ = self._backend.load_models_by_identifiers(identifiers)
else:
self.models_ = self._backend.load_all_models(seed)
if len(self.models_) == 0:
raise ValueError('No models fitted!')
def score(self, X, y):
# fix: Consider only index 1 of second dimension
# Don't know if the reshaping should be done there or in calculate_score
prediction = self.predict(X)
return calculate_score(y, prediction, self._task,
self._metric, self._label_num,
logger=self._logger)
@property
def grid_scores_(self):
grid_scores = list()
scores_per_config = defaultdict(list)
config_list = list()
for run_key in self._proc_smac.runhistory.data:
run_value = self._proc_smac.runhistory.data[run_key]
config_id = run_key.config_id
cost = run_value.cost
if config_id not in config_list:
config_list.append(config_id)
scores_per_config[config_id].append(cost)
for config_id in config_list:
scores = [1 - score for score in scores_per_config[config_id]]
mean_score = np.mean(scores)
config = self._proc_smac.runhistory.ids_config[config_id]
grid_score = _CVScoreTuple(config.get_dictionary(), mean_score,
scores)
grid_scores.append(grid_score)
return grid_scores
@property
def cv_results_(self):
results = dict()
# Missing in contrast to scikit-learn
# splitX_test_score - auto-sklearn does not store the scores on a split
# basis
# std_test_score - auto-sklearn does not store the scores on a split
# basis
# splitX_train_score - auto-sklearn does not compute train scores, add
# flag to compute the train scores
# mean_train_score - auto-sklearn does not store the train scores
# std_train_score - auto-sklearn does not store the train scores
# std_fit_time - auto-sklearn does not store the fit times per split
# mean_score_time - auto-sklearn does not store the score time
# std_score_time - auto-sklearn does not store the score time
# TODO: add those arguments
parameter_dictionaries = dict()
masks = dict()
hp_names = []
# Set up dictionary for parameter values
for hp in self.configuration_space.get_hyperparameters():
name = hp.name
parameter_dictionaries[name] = []
masks[name] = []
hp_names.append(name)
mean_test_score = []
mean_fit_time = []
params = []
status = []
for run_key in self._proc_smac.runhistory.data:
run_value = self._proc_smac.runhistory.data[run_key]
config_id = run_key.config_id
config = self._proc_smac.runhistory.ids_config[config_id]
param_dict = config.get_dictionary()
params.append(param_dict)
mean_test_score.append(1 - run_value.cost)
mean_fit_time.append(run_value.time)
s = run_value.status
if s == 1:
status.append('Success')
elif s == 2:
status.append('Timeout')
elif s == 3:
status.append('Crash')
elif s == 4:
status.append('Abort')
elif s == 5:
status.append('Memout')
else:
status.append('Unknown')
for hp_name in hp_names:
if hp_name in param_dict:
hp_value = param_dict[hp_name]
mask_value = False
else:
hp_value = np.NaN
mask_value = True
parameter_dictionaries[hp_name].append(hp_value)
masks[hp_name].append(mask_value)
results['mean_test_score'] = np.array(mean_test_score)
results['mean_fit_time'] = np.array(mean_fit_time)
results['params'] = params
results['rank_test_scores'] = scipy.stats.rankdata(1 - results['mean_test_score'],
method='min')
results['status'] = status
for hp_name in hp_names:
masked_array = ma.MaskedArray(parameter_dictionaries[hp_name],
masks[hp_name])
results['param_%s' % hp_name] = masked_array
return results
def sprint_statistics(self):
cv_results = self.cv_results_
sio = io.StringIO()
sio.write('auto-sklearn results:\n')
sio.write(' Dataset name: %s\n' % self._dataset_name)
sio.write(' Metric: %s\n' % METRIC_TO_STRING[self._metric])
idx_best_run = np.argmax(cv_results['mean_test_score'])
best_score = cv_results['mean_test_score'][idx_best_run]
sio.write(' Best validation score: %f\n' % best_score)
num_runs = len(cv_results['status'])
sio.write(' Number of target algorithm runs: %d\n' % num_runs)
num_success = sum([s == 'Success' for s in cv_results['status']])
sio.write(' Number of successful target algorithm runs: %d\n' % num_success)
num_crash = sum([s == 'Crash' for s in cv_results['status']])
sio.write(' Number of crashed target algorithm runs: %d\n' % num_crash)
num_timeout = sum([s == 'Timeout' for s in cv_results['status']])
sio.write(' Number of target algorithms that exceeded the memory '
'limit: %d\n' % num_timeout)
num_memout = sum([s == 'Memout' for s in cv_results['status']])
sio.write(' Number of target algorithms that exceeded the time '
'limit: %d\n' % num_memout)
return sio.getvalue()
def show_models(self):
if self.models_ is None or len(self.models_) == 0 or \
self.ensemble_ is None:
self._load_models()
return self.ensemble_.pprint_ensemble_string(self.models_)
def _save_ensemble_data(self, X, y):
"""Split dataset and store Data for the ensemble script.
:param X:
:param y:
:return:
"""
task_name = 'LoadData'
self._start_task(self._stopwatch, task_name)
_, _, _, y_ensemble = resampling.split_data(X, y)
self._backend.save_targets_ensemble(y_ensemble)
self._stop_task(self._stopwatch, task_name)
def _create_search_space(self, tmp_dir, backend, datamanager,
include_estimators=None,
include_preprocessors=None):
task_name = 'CreateConfigSpace'
self._stopwatch.start_task(task_name)
configspace_path = os.path.join(tmp_dir, 'space.pcs')
configuration_space = pipeline.get_configuration_space(
datamanager.info,
include_estimators=include_estimators,
include_preprocessors=include_preprocessors)
configuration_space = self.configuration_space_created_hook(
datamanager, configuration_space)
sp_string = pcs.write(configuration_space)
backend.write_txt_file(configspace_path, sp_string,
'Configuration space')
self._stopwatch.stop_task(task_name)
return configuration_space, configspace_path
def configuration_space_created_hook(self, datamanager, configuration_space):
return configuration_space
def fit(
self,
X: np.ndarray,
y: np.ndarray,
task: int,
X_test: Optional[np.ndarray] = None,
y_test: Optional[np.ndarray] = None,
feat_type: Optional[List[str]] = None,
dataset_name: Optional[str] = None,
only_return_configuration_space: Optional[bool] = False,
load_models: bool = True,
):
# Reset learnt stuff
self.models_ = None
self.cv_models_ = None
self.ensemble_ = None
# The metric must exist as of this point
# It can be provided in the constructor, or automatically
# defined in the estimator fit call
if self._metric is None:
raise ValueError('No metric given.')
if not isinstance(self._metric, Scorer):
raise ValueError('Metric must be instance of '
'autosklearn.metrics.Scorer.')
if self._shared_mode:
# If this fails, it's likely that this is the first call to get
# the data manager
try:
D = self._backend.load_datamanager()
dataset_name = D.name
except IOError:
pass
if dataset_name is None:
dataset_name = hash_array_or_matrix(X)
self._backend.save_start_time(self._seed)
self._stopwatch = StopWatch()
self._dataset_name = dataset_name
self._stopwatch.start_task(self._dataset_name)
self._logger = self._get_logger(dataset_name)
if feat_type is not None and len(feat_type) != X.shape[1]:
raise ValueError('Array feat_type does not have same number of '
'variables as X has features. %d vs %d.' %
(len(feat_type), X.shape[1]))
if feat_type is not None and not all(
[isinstance(f, str) for f in feat_type]):
raise ValueError('Array feat_type must only contain strings.')
if feat_type is not None:
for ft in feat_type:
if ft.lower() not in ['categorical', 'numerical']:
raise ValueError('Only `Categorical` and `Numerical` are '
'valid feature types, you passed `%s`' %
ft)
datamanager = XYDataManager(
X,
y,
X_test=X_test,
y_test=y_test,
task=task,
feat_type=feat_type,
dataset_name=dataset_name,
)
self._backend._make_internals_directory()
try:
os.makedirs(self._backend.get_model_dir())
except (OSError, FileExistsError):
if not self._shared_mode:
raise
try:
os.makedirs(self._backend.get_cv_model_dir())
except (OSError, FileExistsError):
if not self._shared_mode:
raise
self._task = datamanager.info['task']
self._label_num = datamanager.info['label_num']
# == Pickle the data manager to speed up loading
self._backend.save_datamanager(datamanager)
time_for_load_data = self._stopwatch.wall_elapsed(self._dataset_name)
if self._debug_mode:
self._print_load_time(self._dataset_name, self._time_for_task,
time_for_load_data, self._logger)
# == Perform dummy predictions
num_run = 1
# if self._resampling_strategy in ['holdout', 'holdout-iterative-fit']:
num_run = self._do_dummy_prediction(datamanager, num_run)
# = Create a searchspace
# Do this before One Hot Encoding to make sure that it creates a
# search space for a dense classifier even if one hot encoding would
# make it sparse (tradeoff; if one hot encoding would make it sparse,
# densifier and truncatedSVD would probably lead to a MemoryError,
# like this we can't use some of the preprocessing methods in case
# the data became sparse)
self.configuration_space, configspace_path = self._create_search_space(
self._backend.temporary_directory,
self._backend,
datamanager,
include_estimators=self._include_estimators,
exclude_estimators=self._exclude_estimators,
include_preprocessors=self._include_preprocessors,
exclude_preprocessors=self._exclude_preprocessors)
if only_return_configuration_space:
return self.configuration_space
# == RUN ensemble builder
# Do this before calculating the meta-features to make sure that the
# dummy predictions are actually included in the ensemble even if
# calculating the meta-features takes very long
ensemble_task_name = 'runEnsemble'
self._stopwatch.start_task(ensemble_task_name)
elapsed_time = self._stopwatch.wall_elapsed(self._dataset_name)
time_left_for_ensembles = max(0, self._time_for_task - elapsed_time)
if time_left_for_ensembles <= 0:
self._proc_ensemble = None
# Fit only raises error when ensemble_size is not zero but
# time_left_for_ensembles is zero.
if self._ensemble_size > 0:
raise ValueError("Not starting ensemble builder because there "
"is no time left. Try increasing the value "
"of time_left_for_this_task.")
elif self._ensemble_size <= 0:
self._proc_ensemble = None
self._logger.info('Not starting ensemble builder because '
'ensemble size is <= 0.')
else:
self._logger.info('Start Ensemble with %5.2fsec time left' %
time_left_for_ensembles)
self._proc_ensemble = self._get_ensemble_process(
time_left_for_ensembles)
self._proc_ensemble.start()
self._stopwatch.stop_task(ensemble_task_name)
# kill the datamanager as it will be re-loaded anyways from sub processes
try:
del self._datamanager
except Exception:
pass
# => RUN SMAC
smac_task_name = 'runSMAC'
self._stopwatch.start_task(smac_task_name)
elapsed_time = self._stopwatch.wall_elapsed(self._dataset_name)
time_left_for_smac = max(0, self._time_for_task - elapsed_time)
if self._logger:
self._logger.info('Start SMAC with %5.2fsec time left' %
time_left_for_smac)
if time_left_for_smac <= 0:
self._logger.warning("Not starting SMAC because there is no time "
"left.")
_proc_smac = None
self._budget_type = None
else:
if self._per_run_time_limit is None or \
self._per_run_time_limit > time_left_for_smac:
self._logger.warning(
'Time limit for a single run is higher than total time '
'limit. Capping the limit for a single run to the total '
'time given to SMAC (%f)' % time_left_for_smac)
per_run_time_limit = time_left_for_smac
else:
per_run_time_limit = self._per_run_time_limit
# Make sure that at least 2 models are created for the ensemble process
num_models = time_left_for_smac // per_run_time_limit
if num_models < 2:
per_run_time_limit = time_left_for_smac // 2
self._logger.warning(
"Capping the per_run_time_limit to {} to have "
"time for a least 2 models in each process.".format(
per_run_time_limit))
_proc_smac = AutoMLSMBO(
config_space=self.configuration_space,
dataset_name=self._dataset_name,
backend=self._backend,
total_walltime_limit=time_left_for_smac,
func_eval_time_limit=per_run_time_limit,
memory_limit=self._ml_memory_limit,
data_memory_limit=self._data_memory_limit,
watcher=self._stopwatch,
start_num_run=num_run,
num_metalearning_cfgs=self.
_initial_configurations_via_metalearning,
config_file=configspace_path,
seed=self._seed,
metadata_directory=self._metadata_directory,
metric=self._metric,
resampling_strategy=self._resampling_strategy,
resampling_strategy_args=self._resampling_strategy_arguments,
shared_mode=self._shared_mode,
include_estimators=self._include_estimators,
exclude_estimators=self._exclude_estimators,
include_preprocessors=self._include_preprocessors,
exclude_preprocessors=self._exclude_preprocessors,
disable_file_output=self._disable_evaluator_output,
get_smac_object_callback=self._get_smac_object_callback,
smac_scenario_args=self._smac_scenario_args,
)
try:
self.runhistory_, self.trajectory_, self._budget_type = \
_proc_smac.run_smbo()
trajectory_filename = os.path.join(
self._backend.get_smac_output_directory_for_run(
self._seed), 'trajectory.json')
saveable_trajectory = \
[list(entry[:2]) + [entry[2].get_dictionary()] + list(entry[3:])
for entry in self.trajectory_]
with open(trajectory_filename, 'w') as fh:
json.dump(saveable_trajectory, fh)
except Exception as e:
self._logger.exception(e)
raise
# Wait until the ensemble process is finished to avoid shutting down
# while the ensemble builder tries to access the data
if self._proc_ensemble is not None and self._ensemble_size > 0:
self._proc_ensemble.join()
self._proc_ensemble = None
if load_models:
self._load_models()
return self
def _fit(self, datamanager, metric, only_return_configuration_space=False):
# Reset learnt stuff
self.models_ = None
self.ensemble_ = None
# Check arguments prior to doing anything!
if not isinstance(self._disable_evaluator_output, (bool, list)):
raise ValueError('disable_evaluator_output must be of type bool '
'or list.')
if isinstance(self._disable_evaluator_output, list):
allowed_elements = ['model', 'y_optimization']
for element in self._disable_evaluator_output:
if element not in allowed_elements:
raise ValueError("List member '%s' for argument "
"'disable_evaluator_output' must be one "
"of " + str(allowed_elements))
if self._resampling_strategy not in [
'holdout', 'holdout-iterative-fit',
'cv', 'partial-cv',
'partial-cv-iterative-fit'] \
and not issubclass(self._resampling_strategy, BaseCrossValidator)\
and not issubclass(self._resampling_strategy, _RepeatedSplits)\
and not issubclass(self._resampling_strategy, BaseShuffleSplit):
raise ValueError('Illegal resampling strategy: %s' %
self._resampling_strategy)
if self._resampling_strategy in ['partial-cv', 'partial-cv-iterative-fit'] \
and self._ensemble_size != 0:
raise ValueError("Resampling strategy %s cannot be used "
"together with ensembles." % self._resampling_strategy)
if self._resampling_strategy in ['partial-cv', 'cv',
'partial-cv-iterative-fit'] and \
not 'folds' in self._resampling_strategy_arguments:
self._resampling_strategy_arguments['folds'] = 5
self._backend._make_internals_directory()
if self._keep_models:
try:
os.makedirs(self._backend.get_model_dir())
except (OSError, FileExistsError) as e:
if not self._shared_mode:
raise
self._metric = metric
self._task = datamanager.info['task']
self._label_num = datamanager.info['label_num']
# == Pickle the data manager to speed up loading
data_manager_path = self._backend.save_datamanager(datamanager)
time_for_load_data = self._stopwatch.wall_elapsed(self._dataset_name)
if self._debug_mode:
self._print_load_time(
self._dataset_name,
self._time_for_task,
time_for_load_data,
self._logger)
# == Perform dummy predictions
num_run = 1
#if self._resampling_strategy in ['holdout', 'holdout-iterative-fit']:
num_run = self._do_dummy_prediction(datamanager, num_run)
# = Create a searchspace
# Do this before One Hot Encoding to make sure that it creates a
# search space for a dense classifier even if one hot encoding would
# make it sparse (tradeoff; if one hot encoding would make it sparse,
# densifier and truncatedSVD would probably lead to a MemoryError,
# like this we can't use some of the preprocessing methods in case
# the data became sparse)
self.configuration_space, configspace_path = self._create_search_space(
self._backend.temporary_directory,
self._backend,
datamanager,
include_estimators=self._include_estimators,
exclude_estimators=self._exclude_estimators,
include_preprocessors=self._include_preprocessors,
exclude_preprocessors=self._exclude_preprocessors)
if only_return_configuration_space:
return self.configuration_space
# == RUN ensemble builder
# Do this before calculating the meta-features to make sure that the
# dummy predictions are actually included in the ensemble even if
# calculating the meta-features takes very long
ensemble_task_name = 'runEnsemble'
self._stopwatch.start_task(ensemble_task_name)
time_left_for_ensembles = max(0,self._time_for_task \
- self._stopwatch.wall_elapsed(self._dataset_name))
if self._logger:
self._logger.info(
'Start Ensemble with %5.2fsec time left' % time_left_for_ensembles)
if time_left_for_ensembles <= 0:
self._proc_ensemble = None
# Fit only raises error when ensemble_size is not zero but
# time_left_for_ensembles is zero.
if self._ensemble_size > 0:
raise ValueError("Not starting ensemble builder because there "
"is no time left. Try increasing the value "
"of time_left_for_this_task.")
else:
self._proc_ensemble = self._get_ensemble_process(time_left_for_ensembles)
if self._ensemble_size > 0:
self._proc_ensemble.start()
else:
self._logger.info('Not starting ensemble builder because '
'ensemble size is <= 0.')
self._stopwatch.stop_task(ensemble_task_name)
# kill the datamanager as it will be re-loaded anyways from sub processes
try:
del self._datamanager
except Exception:
pass
# => RUN SMAC
smac_task_name = 'runSMAC'
self._stopwatch.start_task(smac_task_name)
time_left_for_smac = max(0,
self._time_for_task - self._stopwatch.wall_elapsed(
self._dataset_name))
if self._logger:
self._logger.info(
'Start SMAC with %5.2fsec time left' % time_left_for_smac)
if time_left_for_smac <= 0:
self._logger.warning("Not starting SMAC because there is no time "
"left.")
_proc_smac = None
else:
if self._per_run_time_limit is None or \
self._per_run_time_limit > time_left_for_smac:
print('Time limit for a single run is higher than total time '
'limit. Capping the limit for a single run to the total '
'time given to SMAC (%f)' % time_left_for_smac)
per_run_time_limit = time_left_for_smac
else:
per_run_time_limit = self._per_run_time_limit
_proc_smac = AutoMLSMBO(
config_space=self.configuration_space,
dataset_name=self._dataset_name,
backend=self._backend,
total_walltime_limit=time_left_for_smac,
func_eval_time_limit=per_run_time_limit,
memory_limit=self._ml_memory_limit,
data_memory_limit=self._data_memory_limit,
watcher=self._stopwatch,
start_num_run=num_run,
num_metalearning_cfgs=self._initial_configurations_via_metalearning,
config_file=configspace_path,
seed=self._seed,
metadata_directory=self._metadata_directory,
metric=self._metric,
resampling_strategy=self._resampling_strategy,
resampling_strategy_args=self._resampling_strategy_arguments,
shared_mode=self._shared_mode,
include_estimators=self._include_estimators,
exclude_estimators=self._exclude_estimators,
include_preprocessors=self._include_preprocessors,
exclude_preprocessors=self._exclude_preprocessors,
disable_file_output=self._disable_evaluator_output,
get_smac_object_callback=self._get_smac_object_callback,
smac_scenario_args=self._smac_scenario_args,
)
self.runhistory_, self.trajectory_ = \
_proc_smac.run_smbo()
trajectory_filename = os.path.join(
self._backend.get_smac_output_directory_for_run(self._seed),
'trajectory.json')
saveable_trajectory = \
[list(entry[:2]) + [entry[2].get_dictionary()] + list(entry[3:])
for entry in self.trajectory_]
with open(trajectory_filename, 'w') as fh:
json.dump(saveable_trajectory, fh)
# Wait until the ensemble process is finished to avoid shutting down
# while the ensemble builder tries to access the data
if self._proc_ensemble is not None and self._ensemble_size > 0:
self._proc_ensemble.join()
self._proc_ensemble = None
self._load_models()
return self