def optimize(self, max_expr_times=1):
"""Method for auto tuning."""
target_info = self._config_info.get('target')
params_info = self._config_info.get('parameters')
command = self._config_info.get('command')
tuner = self._config_info.get('tuner')
for _ in range(max_expr_times):
self._update_from_lineage()
suggestion, user_defined_info = self._suggest(
self._lineage_table, params_info, target_info, tuner)
hyper_config = {
'params':
suggestion,
'summary_dir':
os.path.join(self._summary_base_dir,
f'{self._dir_prefix}_{str(uuid.uuid1())}'),
'custom_lineage_data':
user_defined_info
}
logger.info("Suggest values are: %s.", suggestion)
os.environ[HYPER_CONFIG_ENV_NAME] = json.dumps(hyper_config)
s = subprocess.Popen(shlex.split(command))
s.wait()
if s.returncode != _OK:
logger.error(
"An error occurred during execution, the auto tuning will be terminated."
)
raise OptimizerTerminateError(
"An error occurred during execution, the auto tuning was terminated."
)
def suggest(self, params, target, params_info: dict):
"""Get suggest values."""
bounds = []
for param_info in params_info.values():
bound = param_info[HyperParamKey.BOUND.value] if HyperParamKey.BOUND.value in param_info \
else param_info['choice']
bounds.append([min(bound), max(bound)])
bounds = np.array(bounds)
min_lineage_rows = 2
if not np.array(params).any() or params.shape[0] < min_lineage_rows:
logger.info(
"Without valid histories or the rows of lineages < %s, "
"parameters will be recommended randomly.", min_lineage_rows)
suggestion = generate_arrays(params_info)
else:
self._gp.fit(params, target)
suggestion = self._acq_max(gp=self._gp,
y_max=target.max(),
bounds=bounds,
params_info=params_info)
suggestion, user_defined_info = Transformer.transform_list_to_dict(
params_info, suggestion)
return suggestion, user_defined_info
def suggest(self, params, target, params_info: dict):
"""Get suggest values."""
bounds = []
for param_info in params_info.values():
bound = param_info[HyperParamKey.BOUND.value] if HyperParamKey.BOUND.value in param_info \
else param_info['choice']
bounds.append([min(bound), max(bound)])
bounds = np.array(bounds)
min_lineage_rows = 2
if not np.array(params).any() or params.shape[0] < min_lineage_rows:
logger.info(
"Without valid histories or the rows of lineages < %s, "
"parameters will be recommended randomly.", min_lineage_rows)
suggestion = generate_arrays(params_info)
# in scikit-learn<0.24.2, if np.std(target) include 0, errors would occur.
elif (np.std(target, axis=0) == 0).any():
logger.warning(
"The std of target: %s include zero. To avoid errors, "
"parameters will be recommended randomly.", target)
suggestion = generate_arrays(params_info)
else:
self._gp.fit(params, target)
suggestion = self._acq_max(gp=self._gp,
y_max=target.max(),
bounds=bounds,
params_info=params_info)
suggestion, user_defined_info = Transformer.transform_list_to_dict(
params_info, suggestion)
return suggestion, user_defined_info
def _update_from_lineage(self):
"""Update lineage from lineagemgr."""
try:
lineage_table = get_lineage_table(
summary_base_dir=self._summary_base_dir)
except MindInsightException as err:
logger.info("Can not query lineage. Detail: %s", str(err))
lineage_table = None
self._lineage_table = lineage_table
def _make_summary_base_dir(self, summary_base_dir):
"""Check and make summary_base_dir."""
if not os.path.exists(summary_base_dir):
permissions = os.R_OK | os.W_OK | os.X_OK
os.umask(permissions << 3 | permissions)
mode = permissions << 6
try:
logger.info(
"The summary_base_dir is generated automatically, path is %s.",
summary_base_dir)
os.makedirs(summary_base_dir, mode=mode, exist_ok=True)
except OSError as exc:
raise UnknownError(
"Can not make the summary base directory. Detail: %s." %
str(exc))