class SMBO(BaseSolver):
def __init__(self,
scenario,
tae_runner=None,
acquisition_function=None,
model=None,
runhistory2epm=None,
stats=None,
rng=None):
'''
Interface that contains the main Bayesian optimization loop
Parameters
----------
scenario: smac.scenario.scenario.Scenario
Scenario object
tae_runner: object
object that implements the following method to call the target
algorithm (or any other arbitrary function):
run(self, config)
If not set, it will be initialized with the tae.ExecuteTARunOld()
acquisition_function : AcquisitionFunction
Object that implements the AbstractAcquisitionFunction. Will use
EI if not set.
model : object
Model that implements train() and predict(). Will use a
RandomForest if not set.
runhistory2epm : RunHistory2EMP
Object that implements the AbstractRunHistory2EPM. If None,
will use RunHistory2EPM4Cost if objective is cost or
RunHistory2EPM4LogCost if objective is runtime.
stats: Stats
optional stats object
rng: numpy.random.RandomState
Random number generator
'''
if stats:
self.stats = stats
else:
self.stats = Stats(scenario)
self.runhistory = RunHistory()
self.logger = logging.getLogger("smbo")
if rng is None:
self.num_run = np.random.randint(1234567980)
self.rng = np.random.RandomState(seed=self.num_run)
elif isinstance(rng, int):
self.num_run = rng
self.rng = np.random.RandomState(seed=rng)
elif isinstance(rng, np.random.RandomState):
self.num_run = rng.randint(1234567980)
self.rng = rng
else:
raise TypeError('Unknown type %s for argument rng. Only accepts '
'None, int or np.random.RandomState' %
str(type(rng)))
self.scenario = scenario
self.config_space = scenario.cs
self.traj_logger = TrajLogger(output_dir=self.scenario.output_dir,
stats=self.stats)
self.types = get_types(self.config_space, scenario.feature_array)
if model is None:
self.model = RandomForestWithInstances(
self.types,
scenario.feature_array,
seed=self.rng.randint(1234567980))
else:
self.model = model
if acquisition_function is None:
self.acquisition_func = EI(self.model)
else:
self.acquisition_func = acquisition_function
self.local_search = LocalSearch(self.acquisition_func,
self.config_space)
self.incumbent = None
if tae_runner is None:
self.executor = ExecuteTARunOld(ta=scenario.ta,
stats=self.stats,
run_obj=scenario.run_obj,
par_factor=scenario.par_factor)
else:
self.executor = tae_runner
self.inten = Intensifier(
executor=self.executor,
stats=self.stats,
traj_logger=self.traj_logger,
instances=self.scenario.train_insts,
cutoff=self.scenario.cutoff,
deterministic=self.scenario.deterministic,
run_obj_time=self.scenario.run_obj == "runtime",
instance_specifics=self.scenario.instance_specific)
num_params = len(self.config_space.get_hyperparameters())
self.objective = average_cost
if self.scenario.run_obj == "runtime":
if runhistory2epm is None:
# if we log the performance data,
# the RFRImputator will already get
# log transform data from the runhistory
cutoff = np.log10(self.scenario.cutoff)
threshold = np.log10(self.scenario.cutoff *
self.scenario.par_factor)
imputor = RFRImputator(cs=self.config_space,
rs=self.rng,
cutoff=cutoff,
threshold=threshold,
model=self.model,
change_threshold=0.01,
max_iter=10)
self.rh2EPM = RunHistory2EPM4LogCost(scenario=self.scenario,
num_params=num_params,
success_states=[
StatusType.SUCCESS,
],
impute_censored_data=True,
impute_state=[
StatusType.TIMEOUT,
],
imputor=imputor)
else:
self.rh2EPM = runhistory2epm
elif self.scenario.run_obj == 'quality':
if runhistory2epm is None:
self.rh2EPM = RunHistory2EPM4Cost\
(scenario=self.scenario, num_params=num_params,
success_states=[StatusType.SUCCESS, ],
impute_censored_data=False, impute_state=None)
else:
self.rh2EPM = runhistory2epm
else:
raise ValueError('Unknown run objective: %s. Should be either '
'quality or runtime.' % self.scenario.run_obj)
def run_initial_design(self):
'''
runs algorithm runs for a initial design;
default implementation: running the default configuration on
a random instance-seed pair
Side effect: adds runs to self.runhistory
Returns
-------
incumbent: Configuration()
initial incumbent configuration
'''
default_conf = self.config_space.get_default_configuration()
self.incumbent = default_conf
# add this incumbent right away to have an entry to time point 0
self.traj_logger.add_entry(train_perf=2**31,
incumbent_id=1,
incumbent=self.incumbent)
rand_inst_id = self.rng.randint(0, len(self.scenario.train_insts))
# ignore instance specific values
rand_inst = self.scenario.train_insts[rand_inst_id]
if self.scenario.deterministic:
initial_seed = 0
else:
initial_seed = random.randint(0, MAXINT)
status, cost, runtime, additional_info = self.executor.start(
default_conf,
instance=rand_inst,
cutoff=self.scenario.cutoff,
seed=initial_seed,
instance_specific=self.scenario.instance_specific.get(
rand_inst, "0"))
if status in [StatusType.CRASHED or StatusType.ABORT]:
self.logger.critical("First run crashed -- Abort")
sys.exit(1)
self.runhistory.add(config=default_conf,
cost=cost,
time=runtime,
status=status,
instance_id=rand_inst,
seed=initial_seed,
additional_info=additional_info)
defaul_inst_seeds = set(
self.runhistory.get_runs_for_config(default_conf))
default_perf = self.objective(default_conf, self.runhistory,
defaul_inst_seeds)
self.runhistory.update_cost(default_conf, default_perf)
self.stats.inc_changed += 1 # first incumbent
self.traj_logger.add_entry(train_perf=default_perf,
incumbent_id=self.stats.inc_changed,
incumbent=self.incumbent)
return default_conf
def run(self, max_iters=10):
'''
Runs the Bayesian optimization loop for max_iters iterations
Parameters
----------
max_iters: int
The maximum number of iterations
Returns
----------
incumbent: np.array(1, H)
The best found configuration
'''
self.stats.start_timing()
#self.runhistory = RunHisory()
self.incumbent = self.run_initial_design()
# Main BO loop
iteration = 1
while True:
if self.scenario.shared_model:
pSMAC.read(run_history=self.runhistory,
output_directory=self.scenario.output_dir,
configuration_space=self.config_space,
logger=self.logger)
start_time = time.time()
X, Y = self.rh2EPM.transform(self.runhistory)
self.logger.debug("Search for next configuration")
# get all found configurations sorted according to acq
challengers = self.choose_next(X, Y)
time_spend = time.time() - start_time
logging.debug(
"Time spend to choose next configurations: %.2f sec" %
(time_spend))
self.logger.debug("Intensify")
self.incumbent, inc_perf = self.inten.intensify(
challengers=challengers,
incumbent=self.incumbent,
run_history=self.runhistory,
objective=self.objective,
time_bound=max(0.01, time_spend))
# TODO: Write run history into database
if self.scenario.shared_model:
pSMAC.write(run_history=self.runhistory,
output_directory=self.scenario.output_dir,
num_run=self.num_run)
if iteration == max_iters:
break
iteration += 1
logging.debug(
"Remaining budget: %f (wallclock), %f (ta costs), %f (target runs)"
% (self.stats.get_remaing_time_budget(),
self.stats.get_remaining_ta_budget(),
self.stats.get_remaining_ta_runs()))
if self.stats.is_budget_exhausted():
break
self.stats.print_stats(debug_out=True)
return self.incumbent
def choose_next(self,
X,
Y,
num_interleaved_random=1010,
num_configurations_by_random_search_sorted=1000,
num_configurations_by_local_search=10):
"""Choose next candidate solution with Bayesian optimization.
Parameters
----------
X : (N, D) numpy array
Each row contains a configuration and one set of
instance features.
Y : (N, O) numpy array
The function values for each configuration instance pair.
Returns
-------
list
List of 2020 suggested configurations to evaluate.
"""
self.model.train(X, Y)
if self.runhistory.empty():
incumbent_value = 0.0
elif self.incumbent is None:
# TODO try to calculate an incumbent from the runhistory!
incumbent_value = 0.0
else:
incumbent_value = self.runhistory.get_cost(self.incumbent)
self.acquisition_func.update(model=self.model, eta=incumbent_value)
# Remove dummy acquisition function value
next_configs_by_random_search = [
x[1] for x in self._get_next_by_random_search(
num_points=num_interleaved_random)
]
# Get configurations sorted by EI
next_configs_by_random_search_sorted = \
self._get_next_by_random_search(
num_configurations_by_random_search_sorted, _sorted=True)
next_configs_by_local_search = \
self._get_next_by_local_search(num_configurations_by_local_search)
next_configs_by_acq_value = next_configs_by_random_search_sorted + \
next_configs_by_local_search
next_configs_by_acq_value.sort(reverse=True, key=lambda x: x[0])
self.logger.debug(
"First 10 acq func values of selected configurations: %s" %
(str([_[0] for _ in next_configs_by_acq_value[:10]])))
next_configs_by_acq_value = [_[1] for _ in next_configs_by_acq_value]
challengers = list(
itertools.chain(*zip(next_configs_by_acq_value,
next_configs_by_random_search)))
return challengers
def _get_next_by_random_search(self, num_points=1000, _sorted=False):
"""Get candidate solutions via local search.
Parameters
----------
num_points : int, optional (default=10)
Number of local searches and returned values.
_sorted : bool, optional (default=True)
Whether to sort the candidate solutions by acquisition function
value.
Returns
-------
list : (acquisition value, Candidate solutions)
"""
rand_configs = self.config_space.sample_configuration(size=num_points)
if _sorted:
imputed_rand_configs = map(ConfigSpace.util.impute_inactive_values,
rand_configs)
imputed_rand_configs = [
x.get_array() for x in imputed_rand_configs
]
imputed_rand_configs = np.array(imputed_rand_configs,
dtype=np.float64)
acq_values = self.acquisition_func(imputed_rand_configs)
# From here
# http://stackoverflow.com/questions/20197990/how-to-make-argsort-result-to-be-random-between-equal-values
random = self.rng.rand(len(acq_values))
# Last column is primary sort key!
indices = np.lexsort((random.flatten(), acq_values.flatten()))
for i in range(len(rand_configs)):
rand_configs[i].origin = 'Random Search (sorted)'
# Cannot use zip here because the indices array cannot index the
# rand_configs list, because the second is a pure python list
return [(acq_values[ind][0], rand_configs[ind])
for ind in indices[::-1]]
else:
for i in range(len(rand_configs)):
rand_configs[i].origin = 'Random Search'
return [(0, rand_configs[i]) for i in range(len(rand_configs))]
def _get_next_by_local_search(self, num_points=10):
"""Get candidate solutions via local search.
In case acquisition function values tie, these will be broken randomly.
Parameters
----------
num_points : int, optional (default=10)
Number of local searches and returned values.
Returns
-------
list : (acquisition value, Candidate solutions),
ordered by their acquisition function value
"""
configs_acq = []
# Start N local search from different random start points
for i in range(num_points):
if i == 0 and self.incumbent is not None:
start_point = self.incumbent
else:
start_point = self.config_space.sample_configuration()
configuration, acq_val = self.local_search.maximize(start_point)
configuration.origin = 'Local Search'
configs_acq.append((acq_val[0][0], configuration))
# shuffle for random tie-break
random.shuffle(configs_acq, self.rng.rand)
# sort according to acq value
# and return n best configurations
configs_acq.sort(reverse=True, key=lambda x: x[0])
return configs_acq
def run_smbo(self, max_iters=1000):
global evaluator
# == first things first: load the datamanager
self.reset_data_manager()
# == Initialize SMBO stuff
# first create a scenario
seed = self.seed # TODO
num_params = len(self.config_space.get_hyperparameters())
# allocate a run history
run_history = RunHistory()
meta_runhistory = RunHistory()
meta_runs_dataset_indices = {}
num_run = self.start_num_run
instance_id = self.dataset_name + SENTINEL
# == Train on subset
# before doing anything, let us run the default_cfg
# on a subset of the available data to ensure that
# we at least have some models
# we will try three different ratios of decreasing magnitude
# in the hope that at least on the last one we will be able
# to get a model
n_data = self.datamanager.data['X_train'].shape[0]
subset_ratio = 10000. / n_data
if subset_ratio >= 0.5:
subset_ratio = 0.33
subset_ratios = [subset_ratio, subset_ratio * 0.10]
else:
subset_ratios = [subset_ratio, 500. / n_data]
self.logger.info("Training default configurations on a subset of "
"%d/%d data points." %
(int(n_data * subset_ratio), n_data))
# the time limit for these function evaluations is rigorously
# set to only 1/2 of a full function evaluation
subset_time_limit = max(5, int(self.func_eval_time_limit / 2))
# the configs we want to run on the data subset are:
# 1) the default configs
# 2) a set of configs we selected for training on a subset
subset_configs = [self.config_space.get_default_configuration()] \
+ self.collect_additional_subset_defaults()
subset_config_succesful = [False] * len(subset_configs)
for subset_config_id, next_config in enumerate(subset_configs):
for i, ratio in enumerate(subset_ratios):
self.reset_data_manager()
n_data_subsample = int(n_data * ratio)
# run the config, but throw away the result afterwards
# since this cfg was evaluated only on a subset
# and we don't want to confuse SMAC
self.logger.info(
"Starting to evaluate %d on SUBSET "
"with size %d and time limit %ds.", num_run,
n_data_subsample, subset_time_limit)
self.logger.info(next_config)
_info = eval_with_limits(self.datamanager, self.tmp_dir,
next_config, seed, num_run,
self.resampling_strategy,
self.resampling_strategy_args,
self.memory_limit, subset_time_limit,
n_data_subsample)
(duration, result, _, additional_run_info, status) = _info
self.logger.info(
"Finished evaluating %d. configuration on SUBSET. "
"Duration %f; loss %f; status %s; additional run "
"info: %s ", num_run, duration, result, str(status),
additional_run_info)
num_run += 1
if i < len(subset_ratios) - 1:
if status != StatusType.SUCCESS:
# Do not increase num_run here, because we will try
# the same configuration with less data
self.logger.info(
"A CONFIG did not finish "
" for subset ratio %f -> going smaller", ratio)
continue
else:
self.logger.info(
"Finished SUBSET training sucessfully "
"with ratio %f", ratio)
subset_config_succesful[subset_config_id] = True
break
else:
if status != StatusType.SUCCESS:
self.logger.info(
"A CONFIG did not finish "
" for subset ratio %f.", ratio)
continue
else:
self.logger.info(
"Finished SUBSET training sucessfully "
"with ratio %f", ratio)
subset_config_succesful[subset_config_id] = True
break
# Use the first non-failing configuration from the subsets as the new
# default configuration -> this guards us against the random forest
# failing on large, sparse datasets
default_cfg = None
for subset_config_id, next_config in enumerate(subset_configs):
if subset_config_succesful[subset_config_id]:
default_cfg = next_config
break
if default_cfg is None:
default_cfg = self.config_space.get_default_configuration()
# == METALEARNING suggestions
# we start by evaluating the defaults on the full dataset again
# and add the suggestions from metalearning behind it
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' %
(METRIC_TO_STRING[self.metric],
TASK_TYPES_TO_STRING[meta_task],
'sparse' if self.datamanager.info['is_sparse'] else 'dense'))
self.metadata_directory = metadata_directory
self.logger.info('Metadata directory: %s', self.metadata_directory)
meta_base = MetaBase(self.config_space, self.metadata_directory)
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:
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(instance_id, 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)
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:
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 = []
self.scenario = AutoMLScenario(self.config_space,
self.total_walltime_limit,
self.func_eval_time_limit,
meta_features_dict, self.tmp_dir,
self.shared_mode)
types = get_types(self.config_space, self.scenario.feature_array)
if self.acquisition_function == 'EI':
rh2EPM = RunHistory2EPM4Cost(num_params=num_params,
scenario=self.scenario,
success_states=None,
impute_censored_data=False,
impute_state=None)
model = RandomForestWithInstances(
types,
instance_features=meta_features_list,
seed=1,
num_trees=10)
smac = SMBO(self.scenario, model=model, rng=seed)
elif self.acquisition_function == 'EIPS':
rh2EPM = RunHistory2EPM4EIPS(num_params=num_params,
scenario=self.scenario,
success_states=None,
impute_censored_data=False,
impute_state=None)
model = UncorrelatedMultiObjectiveRandomForestWithInstances(
['cost', 'runtime'],
types,
num_trees=10,
instance_features=meta_features_list,
seed=1)
acquisition_function = EIPS(model)
smac = SMBO(self.scenario,
acquisition_function=acquisition_function,
model=model,
runhistory2epm=rh2EPM,
rng=seed)
else:
raise ValueError('Unknown acquisition function value %s!' %
self.acquisition_function)
# 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
# == first, evaluate all metelearning and default configurations
for i, next_config in enumerate(
([default_cfg] + metalearning_configurations)):
# Do not evaluate default configurations more than once
if i >= len([default_cfg]) and next_config in [default_cfg]:
continue
config_name = 'meta-learning' if i >= len([default_cfg]) \
else 'default'
self.logger.info(
"Starting to evaluate %d. configuration "
"(%s configuration) with time limit %ds.", num_run,
config_name, self.func_eval_time_limit)
self.logger.info(next_config)
self.reset_data_manager()
info = eval_with_limits(self.datamanager, self.tmp_dir,
next_config, seed, num_run,
self.resampling_strategy,
self.resampling_strategy_args,
self.memory_limit,
self.func_eval_time_limit)
(duration, result, _, additional_run_info, status) = info
run_history.add(config=next_config,
cost=result,
time=duration,
status=status,
instance_id=instance_id,
seed=seed)
run_history.update_cost(next_config, result)
self.logger.info(
"Finished evaluating %d. configuration. "
"Duration %f; loss %f; status %s; additional run "
"info: %s ", num_run, duration, result, str(status),
additional_run_info)
num_run += 1
if smac.incumbent is None:
smac.incumbent = next_config
elif result < run_history.get_cost(smac.incumbent):
smac.incumbent = next_config
if self.scenario.shared_model:
pSMAC.write(run_history=run_history,
output_directory=self.scenario.output_dir,
num_run=self.seed)
# == after metalearning run SMAC loop
smac.runhistory = run_history
smac_iter = 0
finished = False
while not finished:
if self.scenario.shared_model:
pSMAC.read(run_history=run_history,
output_directory=self.scenario.output_dir,
configuration_space=self.config_space,
logger=self.logger)
next_configs = []
time_for_choose_next = -1
try:
X_cfg, Y_cfg = rh2EPM.transform(run_history)
if not run_history.empty():
# Update costs by normalization
dataset_minimum = np.min(Y_cfg[:, 0])
Y_cfg[:, 0] = 1 - ((1. - Y_cfg[:, 0]) /
(1. - dataset_minimum))
Y_cfg[:, 0][Y_cfg[:, 0] > 2] = 2
if len(X_meta) > 0 and len(X_cfg) > 0:
pass
#X_cfg = np.concatenate((X_meta, X_cfg))
#Y_cfg = np.concatenate((Y_meta, Y_cfg))
elif len(X_meta) > 0:
X_cfg = X_meta.copy()
Y_cfg = Y_meta.copy()
elif len(X_cfg) > 0:
X_cfg = X_cfg.copy()
Y_cfg = Y_cfg.copy()
else:
raise ValueError(
'No training data for SMAC random forest!')
self.logger.info('Using %d training points for SMAC.' %
X_cfg.shape[0])
choose_next_start_time = time.time()
next_configs_tmp = smac.choose_next(
X_cfg,
Y_cfg,
num_interleaved_random=110,
num_configurations_by_local_search=10,
num_configurations_by_random_search_sorted=100)
time_for_choose_next = time.time() - choose_next_start_time
self.logger.info('Used %g seconds to find next '
'configurations' % (time_for_choose_next))
next_configs.extend(next_configs_tmp)
# TODO put Exception here!
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()
next_configs.append(next_config)
models_fitted_this_iteration = 0
start_time_this_iteration = time.time()
for next_config in next_configs:
x_runtime = impute_inactive_values(next_config)
x_runtime = impute_inactive_values(x_runtime).get_array()
# predicted_runtime = runtime_rf.predict_marginalized_over_instances(
# x_runtime.reshape((1, -1)))
# predicted_runtime = np.exp(predicted_runtime[0][0][0]) - 1
self.logger.info(
"Starting to evaluate %d. configuration (from "
"SMAC) with time limit %ds.", num_run,
self.func_eval_time_limit)
self.logger.info(next_config)
self.reset_data_manager()
info = eval_with_limits(self.datamanager, self.tmp_dir,
next_config, seed, num_run,
self.resampling_strategy,
self.resampling_strategy_args,
self.memory_limit,
self.func_eval_time_limit)
(duration, result, _, additional_run_info, status) = info
run_history.add(config=next_config,
cost=result,
time=duration,
status=status,
instance_id=instance_id,
seed=seed)
run_history.update_cost(next_config, result)
#self.logger.info('Predicted runtime %g, true runtime %g',
# predicted_runtime, duration)
# TODO add unittest to make sure everything works fine and
# this does not get outdated!
if smac.incumbent is None:
smac.incumbent = next_config
elif result < run_history.get_cost(smac.incumbent):
smac.incumbent = next_config
self.logger.info(
"Finished evaluating %d. configuration. "
"Duration: %f; loss: %f; status %s; additional "
"run info: %s ", num_run, duration, result, str(status),
additional_run_info)
smac_iter += 1
num_run += 1
models_fitted_this_iteration += 1
time_used_this_iteration = time.time(
) - start_time_this_iteration
if models_fitted_this_iteration >= 2 and \
time_for_choose_next > 0 and \
time_used_this_iteration > time_for_choose_next:
break
elif time_for_choose_next <= 0 and \
models_fitted_this_iteration >= 1:
break
elif models_fitted_this_iteration >= 50:
break
if max_iters is not None:
finished = (smac_iter < max_iters)
if self.scenario.shared_model:
pSMAC.write(run_history=run_history,
output_directory=self.scenario.output_dir,
num_run=self.seed)
def run_smbo(self, max_iters=1000):
global evaluator
# == first things first: load the datamanager
self.reset_data_manager()
# == Initialize SMBO stuff
# first create a scenario
seed = self.seed # TODO
num_params = len(self.config_space.get_hyperparameters())
# allocate a run history
run_history = RunHistory()
meta_runhistory = RunHistory()
meta_runs_dataset_indices = {}
num_run = self.start_num_run
instance_id = self.dataset_name + SENTINEL
# == Train on subset
# before doing anything, let us run the default_cfg
# on a subset of the available data to ensure that
# we at least have some models
# we will try three different ratios of decreasing magnitude
# in the hope that at least on the last one we will be able
# to get a model
n_data = self.datamanager.data['X_train'].shape[0]
subset_ratio = 10000. / n_data
if subset_ratio >= 0.5:
subset_ratio = 0.33
subset_ratios = [subset_ratio, subset_ratio * 0.10]
else:
subset_ratios = [subset_ratio, 500. / n_data]
self.logger.info("Training default configurations on a subset of "
"%d/%d data points." %
(int(n_data * subset_ratio), n_data))
# the time limit for these function evaluations is rigorously
# set to only 1/2 of a full function evaluation
subset_time_limit = max(5, int(self.func_eval_time_limit / 2))
# the configs we want to run on the data subset are:
# 1) the default configs
# 2) a set of configs we selected for training on a subset
subset_configs = [self.config_space.get_default_configuration()] \
+ self.collect_additional_subset_defaults()
subset_config_succesful = [False] * len(subset_configs)
for subset_config_id, next_config in enumerate(subset_configs):
for i, ratio in enumerate(subset_ratios):
self.reset_data_manager()
n_data_subsample = int(n_data * ratio)
# run the config, but throw away the result afterwards
# since this cfg was evaluated only on a subset
# and we don't want to confuse SMAC
self.logger.info("Starting to evaluate %d on SUBSET "
"with size %d and time limit %ds.",
num_run, n_data_subsample,
subset_time_limit)
self.logger.info(next_config)
_info = eval_with_limits(
self.datamanager, self.tmp_dir, next_config,
seed, num_run,
self.resampling_strategy,
self.resampling_strategy_args,
self.memory_limit,
subset_time_limit, n_data_subsample)
(duration, result, _, additional_run_info, status) = _info
self.logger.info("Finished evaluating %d. configuration on SUBSET. "
"Duration %f; loss %f; status %s; additional run "
"info: %s ", num_run, duration, result,
str(status), additional_run_info)
num_run += 1
if i < len(subset_ratios) - 1:
if status != StatusType.SUCCESS:
# Do not increase num_run here, because we will try
# the same configuration with less data
self.logger.info("A CONFIG did not finish "
" for subset ratio %f -> going smaller",
ratio)
continue
else:
self.logger.info("Finished SUBSET training sucessfully "
"with ratio %f", ratio)
subset_config_succesful[subset_config_id] = True
break
else:
if status != StatusType.SUCCESS:
self.logger.info("A CONFIG did not finish "
" for subset ratio %f.",
ratio)
continue
else:
self.logger.info("Finished SUBSET training sucessfully "
"with ratio %f", ratio)
subset_config_succesful[subset_config_id] = True
break
# Use the first non-failing configuration from the subsets as the new
# default configuration -> this guards us against the random forest
# failing on large, sparse datasets
default_cfg = None
for subset_config_id, next_config in enumerate(subset_configs):
if subset_config_succesful[subset_config_id]:
default_cfg = next_config
break
if default_cfg is None:
default_cfg = self.config_space.get_default_configuration()
# == METALEARNING suggestions
# we start by evaluating the defaults on the full dataset again
# and add the suggestions from metalearning behind it
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' % (METRIC_TO_STRING[self.metric],
TASK_TYPES_TO_STRING[meta_task],
'sparse' if self.datamanager.info['is_sparse']
else 'dense'))
self.metadata_directory = metadata_directory
self.logger.info('Metadata directory: %s', self.metadata_directory)
meta_base = MetaBase(self.config_space, self.metadata_directory)
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:
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(instance_id, 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)
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:
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 = []
self.scenario = AutoMLScenario(self.config_space,
self.total_walltime_limit,
self.func_eval_time_limit,
meta_features_dict,
self.tmp_dir,
self.shared_mode)
types = get_types(self.config_space, self.scenario.feature_array)
if self.acquisition_function == 'EI':
rh2EPM = RunHistory2EPM4Cost(num_params=num_params,
scenario=self.scenario,
success_states=None,
impute_censored_data=False,
impute_state=None)
model = RandomForestWithInstances(types,
instance_features=meta_features_list,
seed=1, num_trees=10)
smac = SMBO(self.scenario, model=model,
rng=seed)
elif self.acquisition_function == 'EIPS':
rh2EPM = RunHistory2EPM4EIPS(num_params=num_params,
scenario=self.scenario,
success_states=None,
impute_censored_data=False,
impute_state=None)
model = UncorrelatedMultiObjectiveRandomForestWithInstances(
['cost', 'runtime'], types, num_trees = 10,
instance_features=meta_features_list, seed=1)
acquisition_function = EIPS(model)
smac = SMBO(self.scenario,
acquisition_function=acquisition_function,
model=model, runhistory2epm=rh2EPM, rng=seed)
else:
raise ValueError('Unknown acquisition function value %s!' %
self.acquisition_function)
# 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
# == first, evaluate all metelearning and default configurations
for i, next_config in enumerate(([default_cfg] +
metalearning_configurations)):
# Do not evaluate default configurations more than once
if i >= len([default_cfg]) and next_config in [default_cfg]:
continue
config_name = 'meta-learning' if i >= len([default_cfg]) \
else 'default'
self.logger.info("Starting to evaluate %d. configuration "
"(%s configuration) with time limit %ds.",
num_run, config_name, self.func_eval_time_limit)
self.logger.info(next_config)
self.reset_data_manager()
info = eval_with_limits(self.datamanager, self.tmp_dir, next_config,
seed, num_run,
self.resampling_strategy,
self.resampling_strategy_args,
self.memory_limit,
self.func_eval_time_limit)
(duration, result, _, additional_run_info, status) = info
run_history.add(config=next_config, cost=result,
time=duration , status=status,
instance_id=instance_id, seed=seed)
run_history.update_cost(next_config, result)
self.logger.info("Finished evaluating %d. configuration. "
"Duration %f; loss %f; status %s; additional run "
"info: %s ", num_run, duration, result,
str(status), additional_run_info)
num_run += 1
if smac.incumbent is None:
smac.incumbent = next_config
elif result < run_history.get_cost(smac.incumbent):
smac.incumbent = next_config
if self.scenario.shared_model:
pSMAC.write(run_history=run_history,
output_directory=self.scenario.output_dir,
num_run=self.seed)
# == after metalearning run SMAC loop
smac.runhistory = run_history
smac_iter = 0
finished = False
while not finished:
if self.scenario.shared_model:
pSMAC.read(run_history=run_history,
output_directory=self.scenario.output_dir,
configuration_space=self.config_space,
logger=self.logger)
next_configs = []
time_for_choose_next = -1
try:
X_cfg, Y_cfg = rh2EPM.transform(run_history)
if not run_history.empty():
# Update costs by normalization
dataset_minimum = np.min(Y_cfg[:, 0])
Y_cfg[:, 0] = 1 - ((1. - Y_cfg[:, 0]) /
(1. - dataset_minimum))
Y_cfg[:, 0][Y_cfg[:, 0] > 2] = 2
if len(X_meta) > 0 and len(X_cfg) > 0:
pass
#X_cfg = np.concatenate((X_meta, X_cfg))
#Y_cfg = np.concatenate((Y_meta, Y_cfg))
elif len(X_meta) > 0:
X_cfg = X_meta.copy()
Y_cfg = Y_meta.copy()
elif len(X_cfg) > 0:
X_cfg = X_cfg.copy()
Y_cfg = Y_cfg.copy()
else:
raise ValueError('No training data for SMAC random forest!')
self.logger.info('Using %d training points for SMAC.' %
X_cfg.shape[0])
choose_next_start_time = time.time()
next_configs_tmp = smac.choose_next(X_cfg, Y_cfg,
num_interleaved_random=110,
num_configurations_by_local_search=10,
num_configurations_by_random_search_sorted=100)
time_for_choose_next = time.time() - choose_next_start_time
self.logger.info('Used %g seconds to find next '
'configurations' % (time_for_choose_next))
next_configs.extend(next_configs_tmp)
# TODO put Exception here!
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()
next_configs.append(next_config)
models_fitted_this_iteration = 0
start_time_this_iteration = time.time()
for next_config in next_configs:
x_runtime = impute_inactive_values(next_config)
x_runtime = impute_inactive_values(x_runtime).get_array()
# predicted_runtime = runtime_rf.predict_marginalized_over_instances(
# x_runtime.reshape((1, -1)))
# predicted_runtime = np.exp(predicted_runtime[0][0][0]) - 1
self.logger.info("Starting to evaluate %d. configuration (from "
"SMAC) with time limit %ds.", num_run,
self.func_eval_time_limit)
self.logger.info(next_config)
self.reset_data_manager()
info = eval_with_limits(self.datamanager, self.tmp_dir, next_config,
seed, num_run,
self.resampling_strategy,
self.resampling_strategy_args,
self.memory_limit,
self.func_eval_time_limit)
(duration, result, _, additional_run_info, status) = info
run_history.add(config=next_config, cost=result,
time=duration , status=status,
instance_id=instance_id, seed=seed)
run_history.update_cost(next_config, result)
#self.logger.info('Predicted runtime %g, true runtime %g',
# predicted_runtime, duration)
# TODO add unittest to make sure everything works fine and
# this does not get outdated!
if smac.incumbent is None:
smac.incumbent = next_config
elif result < run_history.get_cost(smac.incumbent):
smac.incumbent = next_config
self.logger.info("Finished evaluating %d. configuration. "
"Duration: %f; loss: %f; status %s; additional "
"run info: %s ", num_run, duration, result,
str(status), additional_run_info)
smac_iter += 1
num_run += 1
models_fitted_this_iteration += 1
time_used_this_iteration = time.time() - start_time_this_iteration
if models_fitted_this_iteration >= 2 and \
time_for_choose_next > 0 and \
time_used_this_iteration > time_for_choose_next:
break
elif time_for_choose_next <= 0 and \
models_fitted_this_iteration >= 1:
break
elif models_fitted_this_iteration >= 50:
break
if max_iters is not None:
finished = (smac_iter < max_iters)
if self.scenario.shared_model:
pSMAC.write(run_history=run_history,
output_directory=self.scenario.output_dir,
num_run=self.seed)
