def fit_on_datamanager(self, datamanager, metric, load_models=True):
self._stopwatch = StopWatch()
self._backend.save_start_time(self._seed)
name = os.path.basename(datamanager.name)
self._stopwatch.start_task(name)
self._start_task(self._stopwatch, name)
self._dataset_name = name
self._logger = self._get_logger(name)
self._fit(
datamanager=datamanager,
metric=metric,
load_models=load_models,
)
def calculate_metafeatures(task_id):
X_train, y_train, X_test, y_test, cat, task_type, dataset_name = load_task(task_id)
watch = StopWatch()
if task_type == 'classification':
if len(np.unique(y_train)) == 2:
task_type = BINARY_CLASSIFICATION
else:
task_type = MULTICLASS_CLASSIFICATION
else:
task_type = REGRESSION
_metafeatures_labels = _calculate_metafeatures(
x_train=X_train, y_train=y_train, data_feat_type=cat,
data_info_task=task_type, basename=dataset_name, logger_=logger,
watcher=watch,
)
_metafeatures_encoded_labels = _calculate_metafeatures_encoded(
x_train=X_train, y_train=y_train, data_feat_type=cat,
task=task_type, basename=dataset_name, logger_=logger,
watcher=watch,
)
mf = _metafeatures_labels
mf.metafeature_values.update(
_metafeatures_encoded_labels.metafeature_values)
return mf
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 test_stopwatch_overhead(self):
# Wall Overhead
start = time.time()
cpu_start = time.process_time()
watch = StopWatch()
for i in range(1, 1000):
watch.start_task('task_%d' % i)
watch.stop_task('task_%d' % i)
cpu_stop = time.process_time()
stop = time.time()
dur = stop - start
cpu_dur = cpu_stop - cpu_start
cpu_overhead = cpu_dur - watch.cpu_sum()
wall_overhead = dur - watch.wall_sum()
self.assertLess(cpu_overhead, 1)
self.assertLess(wall_overhead, 1)
self.assertLess(watch.cpu_sum(), 2 * watch.wall_sum())
def fit_automl_dataset(self, dataset, metric, load_models=True):
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, 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(
datamanager=loaded_data_manager,
metric=metric,
load_models=load_models,
)
def main(predictions_dir,
basename,
task_type,
metric,
limit,
output_dir,
ensemble_size=None,
seed=1,
indices_output_dir="."):
watch = StopWatch()
watch.start_task("ensemble_builder")
task_type = STRING_TO_TASK_TYPES[task_type]
used_time = 0
time_iter = 0
index_run = 0
current_num_models = 0
logging.basicConfig(filename=os.path.join(predictions_dir,
"ensemble_%d.log" % seed),
level=logging.DEBUG)
while used_time < limit:
logging.debug("Time left: %f", limit - used_time)
logging.debug("Time last iteration: %f", time_iter)
# Load the true labels of the validation data
true_labels = np.load(
os.path.join(predictions_dir, "true_labels_ensemble.npy"))
# Load the predictions from the models
dir_ensemble = os.path.join(predictions_dir,
"predictions_ensemble_%s/" % seed)
dir_valid = os.path.join(predictions_dir,
"predictions_valid_%s/" % seed)
dir_test = os.path.join(predictions_dir, "predictions_test_%s/" % seed)
paths_ = [dir_ensemble, dir_valid, dir_test]
exists = [os.path.isdir(dir_) for dir_ in paths_]
if not exists[0]: #all(exists):
logging.debug("Prediction directory %s does not exist!" %
dir_ensemble)
time.sleep(2)
used_time = watch.wall_elapsed("ensemble_builder")
continue
dir_ensemble_list = sorted(os.listdir(dir_ensemble))
dir_valid_list = sorted(os.listdir(dir_valid)) if exists[1] else []
dir_test_list = sorted(os.listdir(dir_test)) if exists[2] else []
if len(dir_ensemble_list) == 0:
logging.debug("Directories are empty")
time.sleep(2)
used_time = watch.wall_elapsed("ensemble_builder")
continue
if len(dir_ensemble_list) <= current_num_models:
logging.debug("Nothing has changed since the last time")
time.sleep(2)
used_time = watch.wall_elapsed("ensemble_builder")
continue
watch.start_task("ensemble_iter_" + str(index_run))
# List of num_runs (which are in the filename) which will be included
# later
include_num_runs = []
re_num_run = re.compile(r'_([0-9]*)\.npy$')
if ensemble_size is not None:
# Keeps track of the single scores of each model in our ensemble
scores_nbest = []
# The indices of the model that are currently in our ensemble
indices_nbest = []
# The names of the models
model_names = []
# The num run of the models
num_runs = []
model_names_to_scores = dict()
model_idx = 0
for model_name in dir_ensemble_list:
predictions = np.load(os.path.join(dir_ensemble, model_name))
score = evaluator.calculate_score(true_labels, predictions,
task_type, metric,
predictions.shape[1])
model_names_to_scores[model_name] = score
num_run = int(re_num_run.search(model_name).group(1))
if ensemble_size is not None:
if score <= 0.001:
# include_num_runs.append(True)
logging.error("Model only predicts at random: " +
model_name + " has score: " + str(score))
# If we have less models in our ensemble than ensemble_size add the current model if it is better than random
elif len(scores_nbest) < ensemble_size:
scores_nbest.append(score)
indices_nbest.append(model_idx)
include_num_runs.append(num_run)
model_names.append(model_name)
num_runs.append(num_run)
else:
# Take the worst performing model in our ensemble so far
idx = np.argmin(np.array([scores_nbest]))
# If the current model is better than the worst model in our ensemble replace it by the current model
if (scores_nbest[idx] < score):
logging.debug(
"Worst model in our ensemble: %s with "
"score %f will be replaced by model %s "
"with score %f", model_names[idx],
scores_nbest[idx], model_name, score)
# Exclude the old model
del scores_nbest[idx]
scores_nbest.append(score)
del include_num_runs[idx]
del indices_nbest[idx]
indices_nbest.append(model_idx)
include_num_runs.append(num_run)
del model_names[idx]
model_names.append(model_name)
del num_runs[idx]
num_runs.append(num_run)
# Otherwise exclude the current model from the ensemble
else:
#include_num_runs.append(True)
pass
else:
# Load all predictions that are better than random
if score <= 0.001:
#include_num_runs.append(True)
logging.error("Model only predicts at random: " +
model_name + " has score: " + str(score))
else:
include_num_runs.append(num_run)
model_idx += 1
indices_to_model_names = dict()
indices_to_run_num = dict()
for i, model_name in enumerate(dir_ensemble_list):
num_run = int(re_num_run.search(model_name).group(1))
if num_run in include_num_runs:
num_indices = len(indices_to_model_names)
indices_to_model_names[num_indices] = model_name
indices_to_run_num[num_indices] = num_run
#logging.info("Indices to model names:")
#logging.info(indices_to_model_names)
#for i, item in enumerate(sorted(model_names_to_scores.items(),
# key=lambda t: t[1])):
# logging.info("%d: %s", i, item)
include_num_runs = set(include_num_runs)
all_predictions_train = []
for i, model_name in enumerate(dir_ensemble_list):
num_run = int(re_num_run.search(model_name).group(1))
if num_run in include_num_runs:
predictions = np.load(os.path.join(dir_ensemble, model_name))
all_predictions_train.append(predictions)
all_predictions_valid = []
for i, model_name in enumerate(dir_valid_list):
num_run = int(re_num_run.search(model_name).group(1))
if num_run in include_num_runs:
predictions = np.load(os.path.join(dir_valid, model_name))
all_predictions_valid.append(predictions)
all_predictions_test = []
for i, model_name in enumerate(dir_test_list):
num_run = int(re_num_run.search(model_name).group(1))
if num_run in include_num_runs:
predictions = np.load(os.path.join(dir_test, model_name))
all_predictions_test.append(predictions)
if len(all_predictions_train) == len(all_predictions_test) == len(
all_predictions_valid) == 0:
logging.error("All models do just random guessing")
time.sleep(2)
continue
elif len(all_predictions_train) == 1:
logging.debug("Only one model so far we just copy its predictions")
ensemble_members_run_numbers = {0: 1.0}
# Output the score
logging.info("Training performance: %f" %
np.max(model_names_to_scores.values()))
else:
try:
indices, trajectory = ensemble_selection(
np.array(all_predictions_train), true_labels,
ensemble_size, task_type, metric)
logging.info("Trajectory and indices!")
logging.info(trajectory)
logging.info(indices)
except ValueError as e:
logging.error("Caught ValueError: " + str(e))
used_time = watch.wall_elapsed("ensemble_builder")
continue
except Exception as e:
logging.error("Caught error! %s", e.message)
used_time = watch.wall_elapsed("ensemble_builder")
continue
# Output the score
logging.info("Training performance: %f" % trajectory[-1])
# Print the ensemble members:
ensemble_members_run_numbers = dict()
ensemble_members = Counter(indices).most_common()
ensemble_members_string = "Ensemble members:\n"
logging.info(ensemble_members)
for ensemble_member in ensemble_members:
weight = float(ensemble_member[1]) / len(indices)
ensemble_members_string += \
(" %s; weight: %10f; performance: %10f\n" %
(indices_to_model_names[ensemble_member[0]],
weight,
model_names_to_scores[indices_to_model_names[ensemble_member[0]]]))
ensemble_members_run_numbers[indices_to_run_num[
ensemble_member[0]]] = weight
logging.info(ensemble_members_string)
# Save the ensemble indices for later use!
filename_indices = os.path.join(indices_output_dir,
str(index_run).zfill(5) + ".indices")
logging.info(ensemble_members_run_numbers)
with open(filename_indices, "w") as fh:
pickle.dump(ensemble_members_run_numbers, fh)
# Save predictions for valid and test data set
if len(dir_valid_list) == len(dir_ensemble_list):
ensemble_predictions_valid = np.mean(
all_predictions_valid[indices.astype(int)], axis=0)
filename_test = os.path.join(
output_dir,
basename + '_valid_' + str(index_run).zfill(3) + '.predict')
data_util.save_predictions(
os.path.join(predictions_dir, filename_test),
ensemble_predictions_valid)
else:
logging.info("Could not find as many validation set predictions "
"as ensemble predictions!.")
if len(dir_test_list) == len(dir_ensemble_list):
ensemble_predictions_test = np.mean(
all_predictions_test[indices.astype(int)], axis=0)
filename_test = os.path.join(
output_dir,
basename + '_test_' + str(index_run).zfill(3) + '.predict')
data_util.save_predictions(
os.path.join(predictions_dir, filename_test),
ensemble_predictions_test)
else:
logging.info("Could not find as many test set predictions as "
"ensemble predictions!")
current_num_models = len(dir_ensemble_list)
watch.stop_task("ensemble_iter_" + str(index_run))
time_iter = watch.get_wall_dur("ensemble_iter_" + str(index_run))
used_time = watch.wall_elapsed("ensemble_builder")
index_run += 1
return
def __init__(
self,
backend,
time_left_for_this_task,
per_run_time_limit,
initial_configurations_via_metalearning=25,
ensemble_size=1,
ensemble_nbest=1,
max_models_on_disc=1,
ensemble_memory_limit=1000,
seed=1,
ml_memory_limit=3072,
metadata_directory=None,
keep_models=True,
debug_mode=False,
include_estimators=None,
exclude_estimators=None,
include_preprocessors=None,
exclude_preprocessors=None,
resampling_strategy='holdout-iterative-fit',
resampling_strategy_arguments=None,
shared_mode=False,
precision=32,
disable_evaluator_output=False,
get_smac_object_callback=None,
smac_scenario_args=None,
logging_config=None,
):
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._initial_configurations_via_metalearning = \
initial_configurations_via_metalearning
self._ensemble_size = ensemble_size
self._ensemble_nbest = ensemble_nbest
self._max_models_on_disc = max_models_on_disc
self._ensemble_memory_limit = ensemble_memory_limit
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._exclude_estimators = exclude_estimators
self._include_preprocessors = include_preprocessors
self._exclude_preprocessors = exclude_preprocessors
self._resampling_strategy = resampling_strategy
self._resampling_strategy_arguments = resampling_strategy_arguments \
if resampling_strategy_arguments is not None else {}
self._shared_mode = shared_mode
self.precision = precision
self._disable_evaluator_output = disable_evaluator_output
self._get_smac_object_callback = get_smac_object_callback
self._smac_scenario_args = smac_scenario_args
self.logging_config = logging_config
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)))
class AutoML(BaseEstimator):
def __init__(
self,
backend,
time_left_for_this_task,
per_run_time_limit,
initial_configurations_via_metalearning=25,
ensemble_size=1,
ensemble_nbest=1,
max_models_on_disc=1,
ensemble_memory_limit=1000,
seed=1,
ml_memory_limit=3072,
metadata_directory=None,
keep_models=True,
debug_mode=False,
include_estimators=None,
exclude_estimators=None,
include_preprocessors=None,
exclude_preprocessors=None,
resampling_strategy='holdout-iterative-fit',
resampling_strategy_arguments=None,
shared_mode=False,
precision=32,
disable_evaluator_output=False,
get_smac_object_callback=None,
smac_scenario_args=None,
logging_config=None,
):
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._initial_configurations_via_metalearning = \
initial_configurations_via_metalearning
self._ensemble_size = ensemble_size
self._ensemble_nbest = ensemble_nbest
self._max_models_on_disc = max_models_on_disc
self._ensemble_memory_limit = ensemble_memory_limit
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._exclude_estimators = exclude_estimators
self._include_preprocessors = include_preprocessors
self._exclude_preprocessors = exclude_preprocessors
self._resampling_strategy = resampling_strategy
self._resampling_strategy_arguments = resampling_strategy_arguments \
if resampling_strategy_arguments is not None else {}
self._shared_mode = shared_mode
self.precision = precision
self._disable_evaluator_output = disable_evaluator_output
self._get_smac_object_callback = get_smac_object_callback
self._smac_scenario_args = smac_scenario_args
self.logging_config = logging_config
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 assigning and checking variables...
# self._backend = Backend(self._output_dir, self._tmp_dir)
def fit(
self,
X: np.ndarray,
y: np.ndarray,
task: int,
metric: Scorer,
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,
):
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 metric is None:
raise ValueError('No metric given.')
if not isinstance(metric, Scorer):
raise ValueError('Metric must be instance of '
'autosklearn.metrics.Scorer.')
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,
X_test=X_test,
y_test=y_test,
task=task,
feat_type=feat_type,
dataset_name=dataset_name,
)
return self._fit(
datamanager=loaded_data_manager,
metric=metric,
load_models=load_models,
only_return_configuration_space=only_return_configuration_space,
)
# TODO this is very old code which can be dropped!
def fit_automl_dataset(self, dataset, metric, load_models=True):
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, 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(
datamanager=loaded_data_manager,
metric=metric,
load_models=load_models,
)
def fit_on_datamanager(self, datamanager, metric, load_models=True):
self._stopwatch = StopWatch()
self._backend.save_start_time(self._seed)
name = os.path.basename(datamanager.name)
self._stopwatch.start_task(name)
self._start_task(self._stopwatch, name)
self._dataset_name = name
self._logger = self._get_logger(name)
self._fit(
datamanager=datamanager,
metric=metric,
load_models=load_models,
)
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)),
self.logging_config,
)
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):
# When using partial-cv it makes no sense to do dummy predictions
if self._resampling_strategy in [
'partial-cv', 'partial-cv-iterative-fit'
]:
return num_run
self._logger.info("Starting to create dummy predictions.")
memory_limit = int(self._ml_memory_limit)
scenario_mock = unittest.mock.Mock()
scenario_mock.wallclock_limit = self._time_for_task
# This stats object is a hack - maybe the SMAC stats object should
# already be generated here!
stats = Stats(scenario_mock)
stats.start_timing()
ta = ExecuteTaFuncWithQueue(
backend=self._backend,
autosklearn_seed=self._seed,
resampling_strategy=self._resampling_strategy,
initial_num_run=num_run,
logger=self._logger,
stats=stats,
metric=self._metric,
memory_limit=memory_limit,
disable_file_output=self._disable_evaluator_output,
**self._resampling_strategy_arguments)
status, cost, runtime, additional_info = \
ta.run(1, cutoff=self._time_for_task)
if status == StatusType.SUCCESS:
self._logger.info("Finished creating dummy predictions.")
else:
self._logger.error('Error creating dummy predictions: %s ',
str(additional_info))
# Fail if dummy prediction fails.
raise ValueError("Dummy prediction failed: %s " %
str(additional_info))
return ta.num_run
def _fit(
self,
datamanager: AbstractDataManager,
metric: Scorer,
load_models: bool,
only_return_configuration_space: bool = 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',
'cv-iterative-fit',
'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',
'cv-iterative-fit',
'partial-cv-iterative-fit',
]\
and 'folds' not 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):
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
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:
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_, 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)
# 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 refit(self, X, y):
if self._keep_models is not True:
raise ValueError("Refit 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()
# Refit is not applicable when ensemble_size is set to zero.
if self.ensemble_ is None:
raise ValueError(
"Refit can only be called if 'ensemble_size != 0'")
random_state = np.random.RandomState(self._seed)
for identifier in self.models_:
if identifier in self.ensemble_.get_selected_model_identifiers():
model = self.models_[identifier]
# this updates the model inplace, it can then later be used in
# predict method
# try to fit the model. If it fails, shuffle the data. This
# could alleviate the problem in algorithms that depend on
# the ordering of the data.
for i in range(10):
try:
if self._budget_type is None:
_fit_and_suppress_warnings(self._logger, model, X,
y)
else:
_fit_with_budget(
X_train=X,
Y_train=y,
budget=identifier[2],
budget_type=self._budget_type,
logger=self._logger,
model=model,
train_indices=np.arange(len(X), dtype=int),
task_type=self._task,
)
break
except ValueError as e:
indices = list(range(X.shape[0]))
random_state.shuffle(indices)
X = X[indices]
y = y[indices]
if i == 9:
raise e
self._can_predict = True
return self
def predict(self, X, batch_size=None, n_jobs=1):
"""predict.
Parameters
----------
X: array-like, shape = (n_samples, n_features)
batch_size: int or None, defaults to None
batch_size controls whether the pipelines will be
called on small chunks of the data. Useful when calling the
predict method on the whole array X results in a MemoryError.
n_jobs: int, defaults to 1
Parallelize the predictions across the models with n_jobs
processes.
"""
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 not implemented for resampling '
'strategy %s, please call refit().' %
self._resampling_strategy)
if self.models_ is None or len(self.models_) == 0 or \
self.ensemble_ is None:
self._load_models()
# If self.ensemble_ is None, it means that ensemble_size is set to zero.
# In such cases, raise error because predict and predict_proba cannot
# be called.
if self.ensemble_ is None:
raise ValueError("Predict and predict_proba can only be called "
"if 'ensemble_size != 0'")
# Parallelize predictions across models with n_jobs processes.
# Each process computes predictions in chunks of batch_size rows.
all_predictions = joblib.Parallel(n_jobs=n_jobs)(
joblib.delayed(_model_predict)(self, X, batch_size, identifier)
for identifier in self.ensemble_.get_selected_model_identifiers())
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._resampling_strategy in [
'partial-cv', 'partial-cv-iterative-fit'
]:
raise ValueError('Cannot call fit_ensemble with resampling '
'strategy %s.' % self._resampling_strategy)
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()
self._proc_ensemble = None
self._load_models()
return self
def _get_ensemble_process(self,
time_left_for_ensembles,
task=None,
metric=None,
precision=None,
dataset_name=None,
max_iterations=None,
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,
max_models_on_disc=self._max_models_on_disc,
seed=self._seed,
shared_mode=self._shared_mode,
precision=precision,
max_iterations=max_iterations,
read_at_most=np.inf,
memory_limit=self._ensemble_memory_limit,
random_state=self._seed,
)
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)
if len(self.models_) == 0 and self._resampling_strategy not in \
['partial-cv', 'partial-cv-iterative-fit']:
raise ValueError('No models fitted!')
elif self._disable_evaluator_output is False or \
(isinstance(self._disable_evaluator_output, list) and
'model' not in self._disable_evaluator_output):
model_names = self._backend.list_all_models(seed)
if len(model_names) == 0 and self._resampling_strategy not in \
['partial-cv', 'partial-cv-iterative-fit']:
raise ValueError('No models fitted!')
self.models_ = []
else:
self.models_ = []
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(solution=y,
prediction=prediction,
task_type=self._task,
metric=self._metric,
all_scoring_functions=False)
@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
# TODO remove this restriction!
if self._resampling_strategy in [
'partial-cv', 'partial-cv-iterative-fit'
]:
raise ValueError('Cannot call cv_results when using partial-cv!')
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 = []
budgets = []
for run_key in self.runhistory_.data:
run_value = self.runhistory_.data[run_key]
config_id = run_key.config_id
config = self.runhistory_.ids_config[config_id]
param_dict = config.get_dictionary()
params.append(param_dict)
mean_test_score.append(self._metric._optimum -
(self._metric._sign * run_value.cost))
mean_fit_time.append(run_value.time)
budgets.append(run_key.budget)
s = run_value.status
if s == StatusType.SUCCESS:
status.append('Success')
elif s == StatusType.DONOTADVANCE:
status.append('Success (but do not advance to higher budget)')
elif s == StatusType.TIMEOUT:
status.append('Timeout')
elif s == StatusType.CRASHED:
status.append('Crash')
elif s == StatusType.ABORT:
status.append('Abort')
elif s == StatusType.MEMOUT:
status.append('Memout')
else:
raise NotImplementedError(s)
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
results['budgets'] = budgets
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' % self._metric)
idx_success = np.where(
np.array([
status in [
'Success', 'Success (but do not advance to higher budget)'
] for status in cv_results['status']
]))[0]
if len(idx_success) > 0:
if not self._metric._optimum:
idx_best_run = np.argmin(
cv_results['mean_test_score'][idx_success])
else:
idx_best_run = np.argmax(
cv_results['mean_test_score'][idx_success])
best_score = cv_results['mean_test_score'][idx_success][
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 in ['Success', 'Success (but do not advance to higher budget)']
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 time '
'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 memory '
'limit: %d\n' % num_memout)
return sio.getvalue()
def get_models_with_weights(self):
if self.models_ is None or len(self.models_) == 0 or \
self.ensemble_ is None:
self._load_models()
return self.ensemble_.get_models_with_weights(self.models_)
def show_models(self):
models_with_weights = self.get_models_with_weights()
with io.StringIO() as sio:
sio.write("[")
for weight, model in models_with_weights:
sio.write("(%f, %s),\n" % (weight, model))
sio.write("]")
return sio.getvalue()
def _create_search_space(self,
tmp_dir,
backend,
datamanager,
include_estimators=None,
exclude_estimators=None,
include_preprocessors=None,
exclude_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,
exclude_estimators=exclude_estimators,
include_preprocessors=include_preprocessors,
exclude_preprocessors=exclude_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,
metric: Scorer,
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,
):
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 metric is None:
raise ValueError('No metric given.')
if not isinstance(metric, Scorer):
raise ValueError('Metric must be instance of '
'autosklearn.metrics.Scorer.')
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,
X_test=X_test,
y_test=y_test,
task=task,
feat_type=feat_type,
dataset_name=dataset_name,
)
return self._fit(
datamanager=loaded_data_manager,
metric=metric,
load_models=load_models,
only_return_configuration_space=only_return_configuration_space,
)
def main(predictions_dir, basename, task_type, metric, limit, output_dir,
ensemble_size=None, seed=1, indices_output_dir="."):
watch = StopWatch()
watch.start_task("ensemble_builder")
task_type = STRING_TO_TASK_TYPES[task_type]
used_time = 0
time_iter = 0
index_run = 0
current_num_models = 0
logging.basicConfig(filename=os.path.join(predictions_dir, "ensemble_%d.log" % seed), level=logging.DEBUG)
while used_time < limit:
logging.debug("Time left: %f", limit - used_time)
logging.debug("Time last iteration: %f", time_iter)
# Load the true labels of the validation data
true_labels = np.load(os.path.join(predictions_dir, "true_labels_ensemble.npy"))
# Load the predictions from the models
dir_ensemble = os.path.join(predictions_dir,
"predictions_ensemble_%s/" % seed)
dir_valid = os.path.join(predictions_dir,
"predictions_valid_%s/" % seed)
dir_test = os.path.join(predictions_dir,
"predictions_test_%s/" % seed)
paths_ = [dir_ensemble, dir_valid, dir_test]
exists = [os.path.isdir(dir_) for dir_ in paths_]
if not exists[0]: #all(exists):
logging.debug("Prediction directory %s does not exist!" %
dir_ensemble)
time.sleep(2)
used_time = watch.wall_elapsed("ensemble_builder")
continue
dir_ensemble_list = sorted(os.listdir(dir_ensemble))
dir_valid_list = sorted(os.listdir(dir_valid)) if exists[1] else []
dir_test_list = sorted(os.listdir(dir_test)) if exists[2] else []
if len(dir_ensemble_list) == 0:
logging.debug("Directories are empty")
time.sleep(2)
used_time = watch.wall_elapsed("ensemble_builder")
continue
if len(dir_ensemble_list) <= current_num_models:
logging.debug("Nothing has changed since the last time")
time.sleep(2)
used_time = watch.wall_elapsed("ensemble_builder")
continue
watch.start_task("ensemble_iter_" + str(index_run))
# List of num_runs (which are in the filename) which will be included
# later
include_num_runs = []
re_num_run = re.compile(r'_([0-9]*)\.npy$')
if ensemble_size is not None:
# Keeps track of the single scores of each model in our ensemble
scores_nbest = []
# The indices of the model that are currently in our ensemble
indices_nbest = []
# The names of the models
model_names = []
# The num run of the models
num_runs = []
model_names_to_scores = dict()
model_idx = 0
for model_name in dir_ensemble_list:
predictions = np.load(os.path.join(dir_ensemble, model_name))
score = evaluator.calculate_score(true_labels, predictions,
task_type, metric,
predictions.shape[1])
model_names_to_scores[model_name] = score
num_run = int(re_num_run.search(model_name).group(1))
if ensemble_size is not None:
if score <= 0.001:
# include_num_runs.append(True)
logging.error("Model only predicts at random: " + model_name + " has score: " + str(score))
# If we have less models in our ensemble than ensemble_size add the current model if it is better than random
elif len(scores_nbest) < ensemble_size:
scores_nbest.append(score)
indices_nbest.append(model_idx)
include_num_runs.append(num_run)
model_names.append(model_name)
num_runs.append(num_run)
else:
# Take the worst performing model in our ensemble so far
idx = np.argmin(np.array([scores_nbest]))
# If the current model is better than the worst model in our ensemble replace it by the current model
if(scores_nbest[idx] < score):
logging.debug("Worst model in our ensemble: %s with "
"score %f will be replaced by model %s "
"with score %f",
model_names[idx], scores_nbest[idx],
model_name, score)
# Exclude the old model
del scores_nbest[idx]
scores_nbest.append(score)
del include_num_runs[idx]
del indices_nbest[idx]
indices_nbest.append(model_idx)
include_num_runs.append(num_run)
del model_names[idx]
model_names.append(model_name)
del num_runs[idx]
num_runs.append(num_run)
# Otherwise exclude the current model from the ensemble
else:
#include_num_runs.append(True)
pass
else:
# Load all predictions that are better than random
if score <= 0.001:
#include_num_runs.append(True)
logging.error("Model only predicts at random: " + model_name + " has score: " + str(score))
else:
include_num_runs.append(num_run)
model_idx += 1
indices_to_model_names = dict()
indices_to_run_num = dict()
for i, model_name in enumerate(dir_ensemble_list):
num_run = int(re_num_run.search(model_name).group(1))
if num_run in include_num_runs:
num_indices = len(indices_to_model_names)
indices_to_model_names[num_indices] = model_name
indices_to_run_num[num_indices] = num_run
#logging.info("Indices to model names:")
#logging.info(indices_to_model_names)
#for i, item in enumerate(sorted(model_names_to_scores.items(),
# key=lambda t: t[1])):
# logging.info("%d: %s", i, item)
include_num_runs = set(include_num_runs)
all_predictions_train = []
for i, model_name in enumerate(dir_ensemble_list):
num_run = int(re_num_run.search(model_name).group(1))
if num_run in include_num_runs:
predictions = np.load(os.path.join(dir_ensemble, model_name))
all_predictions_train.append(predictions)
all_predictions_valid = []
for i, model_name in enumerate(dir_valid_list):
num_run = int(re_num_run.search(model_name).group(1))
if num_run in include_num_runs:
predictions = np.load(os.path.join(dir_valid, model_name))
all_predictions_valid.append(predictions)
all_predictions_test = []
for i, model_name in enumerate(dir_test_list):
num_run = int(re_num_run.search(model_name).group(1))
if num_run in include_num_runs:
predictions = np.load(os.path.join(dir_test, model_name))
all_predictions_test.append(predictions)
if len(all_predictions_train) == len(all_predictions_test) == len(all_predictions_valid) == 0:
logging.error("All models do just random guessing")
time.sleep(2)
continue
elif len(all_predictions_train) == 1:
logging.debug("Only one model so far we just copy its predictions")
ensemble_members_run_numbers = {0: 1.0}
# Output the score
logging.info("Training performance: %f" % np.max(
model_names_to_scores.values()))
else:
try:
indices, trajectory = ensemble_selection(
np.array(all_predictions_train), true_labels,
ensemble_size, task_type, metric)
logging.info("Trajectory and indices!")
logging.info(trajectory)
logging.info(indices)
except ValueError as e:
logging.error("Caught ValueError: " + str(e))
used_time = watch.wall_elapsed("ensemble_builder")
continue
except Exception as e:
logging.error("Caught error! %s", e.message)
used_time = watch.wall_elapsed("ensemble_builder")
continue
# Output the score
logging.info("Training performance: %f" % trajectory[-1])
# Print the ensemble members:
ensemble_members_run_numbers = dict()
ensemble_members = Counter(indices).most_common()
ensemble_members_string = "Ensemble members:\n"
logging.info(ensemble_members)
for ensemble_member in ensemble_members:
weight = float(ensemble_member[1]) / len(indices)
ensemble_members_string += \
(" %s; weight: %10f; performance: %10f\n" %
(indices_to_model_names[ensemble_member[0]],
weight,
model_names_to_scores[indices_to_model_names[ensemble_member[0]]]))
ensemble_members_run_numbers[indices_to_run_num[
ensemble_member[0]]] = weight
logging.info(ensemble_members_string)
# Save the ensemble indices for later use!
filename_indices = os.path.join(indices_output_dir,
str(index_run).zfill(5) + ".indices")
logging.info(ensemble_members_run_numbers)
with open(filename_indices, "w") as fh:
pickle.dump(ensemble_members_run_numbers, fh)
# Save predictions for valid and test data set
if len(dir_valid_list) == len(dir_ensemble_list):
ensemble_predictions_valid = np.mean(
all_predictions_valid[indices.astype(int)], axis=0)
filename_test = os.path.join(output_dir, basename + '_valid_' + str(index_run).zfill(3) + '.predict')
data_util.save_predictions(os.path.join(predictions_dir, filename_test),
ensemble_predictions_valid)
else:
logging.info("Could not find as many validation set predictions "
"as ensemble predictions!.")
if len(dir_test_list) == len(dir_ensemble_list):
ensemble_predictions_test = np.mean(
all_predictions_test[indices.astype(int)], axis=0)
filename_test = os.path.join(output_dir, basename + '_test_' + str(index_run).zfill(3) + '.predict')
data_util.save_predictions(os.path.join(predictions_dir, filename_test),
ensemble_predictions_test)
else:
logging.info("Could not find as many test set predictions as "
"ensemble predictions!")
current_num_models = len(dir_ensemble_list)
watch.stop_task("ensemble_iter_" + str(index_run))
time_iter = watch.get_wall_dur("ensemble_iter_" + str(index_run))
used_time = watch.wall_elapsed("ensemble_builder")
index_run += 1
return
def __init__(
self,
backend,
time_left_for_this_task,
per_run_time_limit,
initial_configurations_via_metalearning=25,
ensemble_size=1,
ensemble_nbest=1,
max_models_on_disc=1,
ensemble_memory_limit=1000,
seed=1,
ml_memory_limit=3072,
metadata_directory=None,
debug_mode=False,
include_estimators=None,
exclude_estimators=None,
include_preprocessors=None,
exclude_preprocessors=None,
resampling_strategy='holdout-iterative-fit',
resampling_strategy_arguments=None,
shared_mode=False,
precision=32,
disable_evaluator_output=False,
get_smac_object_callback=None,
smac_scenario_args=None,
logging_config=None,
metric=None,
):
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._initial_configurations_via_metalearning = \
initial_configurations_via_metalearning
self._ensemble_size = ensemble_size
self._ensemble_nbest = ensemble_nbest
self._max_models_on_disc = max_models_on_disc
self._ensemble_memory_limit = ensemble_memory_limit
self._seed = seed
self._ml_memory_limit = ml_memory_limit
self._data_memory_limit = None
self._metadata_directory = metadata_directory
self._include_estimators = include_estimators
self._exclude_estimators = exclude_estimators
self._include_preprocessors = include_preprocessors
self._exclude_preprocessors = exclude_preprocessors
self._resampling_strategy = resampling_strategy
self._resampling_strategy_arguments = resampling_strategy_arguments \
if resampling_strategy_arguments is not None else {}
if self._resampling_strategy not in ['holdout',
'holdout-iterative-fit',
'cv',
'cv-iterative-fit',
'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',
'cv-iterative-fit',
'partial-cv-iterative-fit',
]\
and 'folds' not in self._resampling_strategy_arguments:
self._resampling_strategy_arguments['folds'] = 5
self._shared_mode = shared_mode
self.precision = precision
self._disable_evaluator_output = disable_evaluator_output
# 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))
self._get_smac_object_callback = get_smac_object_callback
self._smac_scenario_args = smac_scenario_args
self.logging_config = logging_config
self._datamanager = None
self._dataset_name = None
self._stopwatch = StopWatch()
self._logger = None
self._task = None
self._metric = metric
self._label_num = None
self._parser = None
self.models_ = None
self.cv_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)))
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
