def get_metalearning_suggestions(self):
# == METALEARNING suggestions
# we start by evaluating the defaults on the full dataset again
# and add the suggestions from metalearning behind it
if self.num_metalearning_cfgs > 0:
if self.metadata_directory is None:
metalearning_directory = os.path.dirname(
autosklearn.metalearning.__file__)
# There is no multilabel data in OpenML
if self.task == MULTILABEL_CLASSIFICATION:
meta_task = BINARY_CLASSIFICATION
else:
meta_task = self.task
metadata_directory = os.path.join(
metalearning_directory, 'files',
'%s_%s_%s' % (self.metric, TASK_TYPES_TO_STRING[meta_task],
'sparse' if self.datamanager.info['is_sparse']
else 'dense'))
self.metadata_directory = metadata_directory
if os.path.exists(self.metadata_directory):
self.logger.info('Metadata directory: %s',
self.metadata_directory)
meta_base = MetaBase(self.config_space, self.metadata_directory)
try:
meta_base.remove_dataset(self.dataset_name)
except:
pass
metafeature_calculation_time_limit = int(
self.total_walltime_limit / 4)
metafeature_calculation_start_time = time.time()
meta_features = self._calculate_metafeatures_with_limits(
metafeature_calculation_time_limit)
metafeature_calculation_end_time = time.time()
metafeature_calculation_time_limit = \
metafeature_calculation_time_limit - (
metafeature_calculation_end_time -
metafeature_calculation_start_time)
if metafeature_calculation_time_limit < 1:
self.logger.warning(
'Time limit for metafeature calculation less '
'than 1 seconds (%f). Skipping calculation '
'of metafeatures for encoded dataset.',
metafeature_calculation_time_limit)
meta_features_encoded = None
else:
with warnings.catch_warnings():
warnings.showwarning = self._send_warnings_to_log
self.datamanager.perform1HotEncoding()
meta_features_encoded = \
self._calculate_metafeatures_encoded_with_limits(
metafeature_calculation_time_limit)
# In case there is a problem calculating the encoded meta-features
if meta_features is None:
if meta_features_encoded is not None:
meta_features = meta_features_encoded
else:
if meta_features_encoded is not None:
meta_features.metafeature_values.update(
meta_features_encoded.metafeature_values)
if meta_features is not None:
meta_base.add_dataset(self.dataset_name, meta_features)
# Do mean imputation of the meta-features - should be done specific
# for each prediction model!
all_metafeatures = meta_base.get_metafeatures(
features=list(meta_features.keys()))
all_metafeatures.fillna(all_metafeatures.mean(),
inplace=True)
with warnings.catch_warnings():
warnings.showwarning = self._send_warnings_to_log
metalearning_configurations = self.collect_metalearning_suggestions(
meta_base)
if metalearning_configurations is None:
metalearning_configurations = []
self.reset_data_manager()
self.logger.info('%s', meta_features)
# Convert meta-features into a dictionary because the scenario
# expects a dictionary
meta_features_dict = {}
for dataset, series in all_metafeatures.iterrows():
meta_features_dict[dataset] = series.values
meta_features_list = []
for meta_feature_name in all_metafeatures.columns:
meta_features_list.append(
meta_features[meta_feature_name].value)
meta_features_list = np.array(meta_features_list).reshape(
(1, -1))
self.logger.info(list(meta_features_dict.keys()))
else:
meta_features = None
self.logger.warning('Could not find meta-data directory %s' %
metadata_directory)
else:
meta_features = None
if meta_features is None:
meta_features_list = []
metalearning_configurations = []
return metalearning_configurations
class MetaBaseTest(unittest.TestCase):
_multiprocess_can_split_ = True
def setUp(self):
self.cwd = os.getcwd()
data_dir = os.path.dirname(__file__)
data_dir = os.path.join(data_dir, 'test_meta_base_data')
os.chdir(data_dir)
cs = autosklearn.pipeline.classification.SimpleClassificationPipeline()\
.get_hyperparameter_search_space()
self.base = MetaBase(cs, data_dir)
def tearDown(self):
os.chdir(self.cwd)
def test_get_all_runs(self):
runs = self.base.get_all_runs()
self.assertIsInstance(runs, pd.DataFrame)
# TODO update this ASAP
self.assertEqual((134, 24), runs.shape)
def test_get_runs(self):
runs = self.base.get_runs('38_acc')
# TODO update this ASAP
self.assertEqual(24, len(runs))
self.assertIsInstance(runs, pd.Series)
def test_get_metafeatures_single_dataset(self):
mf = self.base.get_metafeatures('38_acc')
self.assertIsInstance(mf, pd.Series)
self.assertEqual(mf.name, u'38_acc')
self.assertEqual(mf.loc['NumberOfInstances'], 2527.0)
def test_get_metafeatures_single_feature(self):
mf = self.base.get_metafeatures(features='NumberOfInstances')
self.assertIsInstance(mf, pd.Series)
self.assertEqual(mf.shape, (140, ))
def test_get_metafeatures_single_dataset_and_single_feature(self):
mf = self.base.get_metafeatures('38_acc', features='NumberOfInstances')
self.assertEqual(mf.shape, ())
def test_get_metafeatures_multiple_datasets(self):
mf = self.base.get_metafeatures(['38_acc', '24_acc'])
self.assertIsInstance(mf, pd.DataFrame)
self.assertEqual(mf.shape, (2, 46))
def test_get_metafeatures_multiple_features(self):
mf = self.base.get_metafeatures(features=['NumberOfInstances',
'NumberOfClasses'])
self.assertIsInstance(mf, pd.DataFrame)
self.assertEqual(mf.shape, (140, 2))
def test_remove_dataset(self):
name = "1000_acc"
for key in self.base.algorithm_runs:
self.assertIn(name, self.base.algorithm_runs[key].index)
self.assertIn(name, self.base.metafeatures.index)
metafeatures_shape = self.base.metafeatures.shape
self.base.remove_dataset(name)
for key in self.base.algorithm_runs:
self.assertNotIn(name, self.base.algorithm_runs[key].index)
self.assertNotIn(name, self.base.metafeatures.index)
# Check that only one thing was removed
self.assertEqual(self.base.metafeatures.shape,
(metafeatures_shape[0] - 1, metafeatures_shape[1]))
def run_smbo(self):
self.watcher.start_task('SMBO')
# == first things first: load the datamanager
self.reset_data_manager()
# == Initialize non-SMBO stuff
# first create a scenario
seed = self.seed
self.config_space.seed(seed)
num_params = len(self.config_space.get_hyperparameters())
# allocate a run history
num_run = self.start_num_run
# Initialize some SMAC dependencies
runhistory = RunHistory(aggregate_func=average_cost)
# meta_runhistory = RunHistory(aggregate_func=average_cost)
# meta_runs_dataset_indices = {}
# == METALEARNING suggestions
# we start by evaluating the defaults on the full dataset again
# and add the suggestions from metalearning behind it
if self.num_metalearning_cfgs > 0:
if self.metadata_directory is None:
metalearning_directory = os.path.dirname(
autosklearn.metalearning.__file__)
# There is no multilabel data in OpenML
if self.task == MULTILABEL_CLASSIFICATION:
meta_task = BINARY_CLASSIFICATION
else:
meta_task = self.task
metadata_directory = os.path.join(
metalearning_directory, 'files', '%s_%s_%s' %
(self.metric, TASK_TYPES_TO_STRING[meta_task], 'sparse'
if self.datamanager.info['is_sparse'] else 'dense'))
self.metadata_directory = metadata_directory
if os.path.exists(self.metadata_directory):
self.logger.info('Metadata directory: %s',
self.metadata_directory)
meta_base = MetaBase(self.config_space,
self.metadata_directory)
try:
meta_base.remove_dataset(self.dataset_name)
except:
pass
metafeature_calculation_time_limit = int(
self.total_walltime_limit / 4)
metafeature_calculation_start_time = time.time()
meta_features = self._calculate_metafeatures_with_limits(
metafeature_calculation_time_limit)
metafeature_calculation_end_time = time.time()
metafeature_calculation_time_limit = \
metafeature_calculation_time_limit - (
metafeature_calculation_end_time -
metafeature_calculation_start_time)
if metafeature_calculation_time_limit < 1:
self.logger.warning(
'Time limit for metafeature calculation less '
'than 1 seconds (%f). Skipping calculation '
'of metafeatures for encoded dataset.',
metafeature_calculation_time_limit)
meta_features_encoded = None
else:
with warnings.catch_warnings():
warnings.showwarning = self._send_warnings_to_log
self.datamanager.perform1HotEncoding()
meta_features_encoded = \
self._calculate_metafeatures_encoded_with_limits(
metafeature_calculation_time_limit)
# In case there is a problem calculating the encoded meta-features
if meta_features is None:
if meta_features_encoded is not None:
meta_features = meta_features_encoded
else:
if meta_features_encoded is not None:
meta_features.metafeature_values.update(
meta_features_encoded.metafeature_values)
if meta_features is not None:
meta_base.add_dataset(self.dataset_name, meta_features)
# Do mean imputation of the meta-features - should be done specific
# for each prediction model!
all_metafeatures = meta_base.get_metafeatures(
features=list(meta_features.keys()))
all_metafeatures.fillna(all_metafeatures.mean(),
inplace=True)
with warnings.catch_warnings():
warnings.showwarning = self._send_warnings_to_log
metalearning_configurations = self.collect_metalearning_suggestions(
meta_base)
if metalearning_configurations is None:
metalearning_configurations = []
self.reset_data_manager()
self.logger.info('%s', meta_features)
# Convert meta-features into a dictionary because the scenario
# expects a dictionary
meta_features_dict = {}
for dataset, series in all_metafeatures.iterrows():
meta_features_dict[dataset] = series.values
meta_features_list = []
for meta_feature_name in all_metafeatures.columns:
meta_features_list.append(
meta_features[meta_feature_name].value)
meta_features_list = np.array(meta_features_list).reshape(
(1, -1))
self.logger.info(list(meta_features_dict.keys()))
# meta_runs = meta_base.get_all_runs(METRIC_TO_STRING[self.metric])
# meta_runs_index = 0
# try:
# meta_durations = meta_base.get_all_runs('runtime')
# read_runtime_data = True
# except KeyError:
# read_runtime_data = False
# self.logger.critical('Cannot read runtime data.')
# if self.acquisition_function == 'EIPS':
# self.logger.critical('Reverting to acquisition function EI!')
# self.acquisition_function = 'EI'
# for meta_dataset in meta_runs.index:
# meta_dataset_start_index = meta_runs_index
# for meta_configuration in meta_runs.columns:
# if np.isfinite(meta_runs.loc[meta_dataset, meta_configuration]):
# try:
# config = meta_base.get_configuration_from_algorithm_index(
# meta_configuration)
# cost = meta_runs.loc[meta_dataset, meta_configuration]
# if read_runtime_data:
# runtime = meta_durations.loc[meta_dataset,
# meta_configuration]
# else:
# runtime = 1
# # TODO read out other status types!
# meta_runhistory.add(config, cost, runtime,
# StatusType.SUCCESS,
# instance_id=meta_dataset)
# meta_runs_index += 1
# except:
# # TODO maybe add warning
# pass
#
# meta_runs_dataset_indices[meta_dataset] = (
# meta_dataset_start_index, meta_runs_index)
else:
meta_features = None
self.logger.warning('Could not find meta-data directory %s' %
metadata_directory)
else:
meta_features = None
if meta_features is None:
if self.acquisition_function == 'EIPS':
self.logger.critical('Reverting to acquisition function EI!')
self.acquisition_function = 'EI'
meta_features_list = []
meta_features_dict = {}
metalearning_configurations = []
if self.resampling_strategy in [
'partial-cv', 'partial-cv-iterative-fit'
]:
num_folds = self.resampling_strategy_args['folds']
instances = [[
json.dumps({
'task_id': self.dataset_name,
'fold': fold_number
})
] for fold_number in range(num_folds)]
else:
instances = [[json.dumps({'task_id': self.dataset_name})]]
startup_time = self.watcher.wall_elapsed(self.dataset_name)
total_walltime_limit = self.total_walltime_limit - startup_time - 5
scenario_dict = {
'cs': self.config_space,
'cutoff-time': self.func_eval_time_limit,
'memory-limit': self.memory_limit,
'wallclock-limit': total_walltime_limit,
'output-dir': self.backend.get_smac_output_directory(self.seed),
'shared-model': self.shared_mode,
'run-obj': 'quality',
'deterministic': 'true',
'instances': instances
}
if self.configuration_mode == 'RANDOM':
scenario_dict['minR'] = len(
instances) if instances is not None else 1
scenario_dict['initial_incumbent'] = 'RANDOM'
self.scenario = Scenario(scenario_dict)
# TODO rebuild target algorithm to be it's own target algorithm
# evaluator, which takes into account that a run can be killed prior
# to the model being fully fitted; thus putting intermediate results
# into a queue and querying them once the time is over
exclude = dict()
include = dict()
if self.include_preprocessors is not None and \
self.exclude_preprocessors is not None:
raise ValueError('Cannot specify include_preprocessors and '
'exclude_preprocessors.')
elif self.include_preprocessors is not None:
include['preprocessor'] = self.include_preprocessors
elif self.exclude_preprocessors is not None:
exclude['preprocessor'] = self.exclude_preprocessors
if self.include_estimators is not None and \
self.exclude_preprocessors is not None:
raise ValueError('Cannot specify include_estimators and '
'exclude_estimators.')
elif self.include_estimators is not None:
if self.task in CLASSIFICATION_TASKS:
include['classifier'] = self.include_estimators
elif self.task in REGRESSION_TASKS:
include['regressor'] = self.include_estimators
else:
raise ValueError(self.task)
elif self.exclude_estimators is not None:
if self.task in CLASSIFICATION_TASKS:
exclude['classifier'] = self.exclude_estimators
elif self.task in REGRESSION_TASKS:
exclude['regressor'] = self.exclude_estimators
else:
raise ValueError(self.task)
ta = ExecuteTaFuncWithQueue(
backend=self.backend,
autosklearn_seed=seed,
resampling_strategy=self.resampling_strategy,
initial_num_run=num_run,
logger=self.logger,
include=include,
exclude=exclude,
metric=self.metric,
memory_limit=self.memory_limit,
disable_file_output=self.disable_file_output,
**self.resampling_strategy_args)
types, bounds = get_types(self.config_space,
self.scenario.feature_array)
# TODO extract generation of SMAC object into it's own function for
# testing
if self.acquisition_function == 'EI':
model = RandomForestWithInstances(
types=types,
bounds=bounds,
#instance_features=meta_features_list,
seed=1,
num_trees=10)
rh2EPM = RunHistory2EPM4Cost(num_params=num_params,
scenario=self.scenario,
success_states=[
StatusType.SUCCESS,
StatusType.MEMOUT,
StatusType.TIMEOUT
],
impute_censored_data=False,
impute_state=None)
_smac_arguments = dict(scenario=self.scenario,
model=model,
rng=seed,
runhistory2epm=rh2EPM,
tae_runner=ta,
runhistory=runhistory)
elif self.acquisition_function == 'EIPS':
rh2EPM = RunHistory2EPM4EIPS(num_params=num_params,
scenario=self.scenario,
success_states=[
StatusType.SUCCESS,
StatusType.MEMOUT,
StatusType.TIMEOUT
],
impute_censored_data=False,
impute_state=None)
model = UncorrelatedMultiObjectiveRandomForestWithInstances(
['cost', 'runtime'],
types=types,
bounds=bounds,
num_trees=10,
instance_features=meta_features_list,
seed=1)
acquisition_function = EIPS(model)
_smac_arguments = dict(scenario=self.scenario,
model=model,
rng=seed,
tae_runner=ta,
runhistory2epm=rh2EPM,
runhistory=runhistory,
acquisition_function=acquisition_function)
else:
raise ValueError('Unknown acquisition function value %s!' %
self.acquisition_function)
if self.configuration_mode == 'SMAC':
smac = SMAC(**_smac_arguments)
elif self.configuration_mode in ['ROAR', 'RANDOM']:
for not_in_roar in ['runhistory2epm', 'model']:
if not_in_roar in _smac_arguments:
del _smac_arguments[not_in_roar]
smac = ROAR(**_smac_arguments)
else:
raise ValueError(self.configuration_mode)
# Build a runtime model
# runtime_rf = RandomForestWithInstances(types,
# instance_features=meta_features_list,
# seed=1, num_trees=10)
# runtime_rh2EPM = RunHistory2EPM4EIPS(num_params=num_params,
# scenario=self.scenario,
# success_states=None,
# impute_censored_data=False,
# impute_state=None)
# X_runtime, y_runtime = runtime_rh2EPM.transform(meta_runhistory)
# runtime_rf.train(X_runtime, y_runtime[:, 1].flatten())
# X_meta, Y_meta = rh2EPM.transform(meta_runhistory)
# # Transform Y_meta on a per-dataset base
# for meta_dataset in meta_runs_dataset_indices:
# start_index, end_index = meta_runs_dataset_indices[meta_dataset]
# end_index += 1 # Python indexing
# Y_meta[start_index:end_index, 0]\
# [Y_meta[start_index:end_index, 0] >2.0] = 2.0
# dataset_minimum = np.min(Y_meta[start_index:end_index, 0])
# Y_meta[start_index:end_index, 0] = 1 - (
# (1. - Y_meta[start_index:end_index, 0]) /
# (1. - dataset_minimum))
# Y_meta[start_index:end_index, 0]\
# [Y_meta[start_index:end_index, 0] > 2] = 2
smac.solver.stats.start_timing()
# == first, evaluate all metelearning and default configurations
smac.solver.incumbent = smac.solver.initial_design.run()
for challenger in metalearning_configurations:
smac.solver.incumbent, inc_perf = smac.solver.intensifier.intensify(
challengers=[challenger],
incumbent=smac.solver.incumbent,
run_history=smac.solver.runhistory,
aggregate_func=smac.solver.aggregate_func,
time_bound=self.total_walltime_limit)
if smac.solver.scenario.shared_model:
pSMAC.write(run_history=smac.solver.runhistory,
output_directory=smac.solver.scenario.output_dir,
num_run=self.seed)
if smac.solver.stats.is_budget_exhausted():
break
# == after metalearning run SMAC loop
while True:
if smac.solver.scenario.shared_model:
pSMAC.read(run_history=smac.solver.runhistory,
output_dirs=glob.glob(
self.backend.get_smac_output_glob()),
configuration_space=self.config_space,
logger=self.logger)
choose_next_start_time = time.time()
try:
challengers = self.choose_next(smac)
except Exception as e:
self.logger.error(e)
self.logger.error("Error in getting next configurations "
"with SMAC. Using random configuration!")
next_config = self.config_space.sample_configuration()
challengers = [next_config]
time_for_choose_next = time.time() - choose_next_start_time
self.logger.info('Used %g seconds to find next '
'configurations' % (time_for_choose_next))
time_for_choose_next = max(time_for_choose_next, 1.0)
smac.solver.incumbent, inc_perf = smac.solver.intensifier.intensify(
challengers=challengers,
incumbent=smac.solver.incumbent,
run_history=smac.solver.runhistory,
aggregate_func=smac.solver.aggregate_func,
time_bound=time_for_choose_next)
if smac.solver.scenario.shared_model:
pSMAC.write(run_history=smac.solver.runhistory,
output_directory=smac.solver.scenario.output_dir,
num_run=self.seed)
if smac.solver.stats.is_budget_exhausted():
break
self.runhistory = smac.solver.runhistory
self.trajectory = smac.solver.intensifier.traj_logger.trajectory
smac.runhistory = self.runhistory
self.fANOVA_input = smac.get_X_y()
return self.runhistory, self.trajectory, self.fANOVA_input
