def load_selector(self):
"""
Load and initialize the BTB class which will be responsible for
selecting hyperpartitions.
"""
# selector will either be a key into SELECTORS_MAP or a path to
# a file that defines a class called CustomSelector.
if self.datarun.selector in SELECTORS_MAP:
Selector = SELECTORS_MAP[self.datarun.selector]
else:
path, classname = re.match(CUSTOM_CLASS_REGEX,
self.datarun.selector).groups()
mod = imp.load_source('btb.selection.custom', path)
Selector = getattr(mod, classname)
logger.info('Selector: %s' % Selector)
# generate the arguments we need to initialize the selector
hyperpartitions = self.db.get_hyperpartitions(
datarun_id=self.datarun.id)
hp_by_method = defaultdict(list)
for hp in hyperpartitions:
hp_by_method[hp.method].append(hp.id)
hyperpartition_ids = [hp.id for hp in hyperpartitions]
# Selector classes support passing in redundant arguments
self.selector = get_instance(Selector,
choices=hyperpartition_ids,
k=self.datarun.k_window,
by_algorithm=dict(hp_by_method))
def test_make_selector():
kwargs = {
'choices': [1, 2, 3],
'k': 3,
'by_algorithm': {'svm': [1, 2], 'rf': [3, 4]}
}
for selector_class in SELECTORS.values():
selector = utilities.get_instance(selector_class, **kwargs)
assert isinstance(selector, Selector)
def tune_hyperparameters(self, hyperpartition):
"""
Use the hyperparameter tuning method specified by our datarun to choose
a set of hyperparameters from the potential space.
"""
# Get parameter metadata for this hyperpartition
tunables = hyperpartition.tunables
# If there aren't any tunable parameters, we're done. Return the vector
# of values in the hyperpartition and mark the set as finished.
if not len(tunables):
logger.warning('No tunables for hyperpartition %d' %
hyperpartition.id)
self.db.mark_hyperpartition_gridding_done(hyperpartition.id)
return update_params(params=[],
tunables=tunables,
categoricals=hyperpartition.categoricals,
constants=hyperpartition.constants)
# Get previously-used parameters: every classifier should either be
# completed or have thrown an error
all_clfs = self.db.get_classifiers(hyperpartition_id=hyperpartition.id)
classifiers = [
c for c in all_clfs if c.status == ClassifierStatus.COMPLETE
]
X = [c.hyperparameter_values for c in classifiers]
y = np.array([
float(getattr(c, self.datarun.score_target)) for c in classifiers
])
# Initialize the tuner and propose a new set of parameters
# this has to be initialized with information from the hyperpartition, so we
# need to do it fresh for each classifier (not in load_tuner)
tuner = get_instance(self.Tuner,
tunables=tunables,
gridding=self.datarun.gridding,
r_minimum=self.datarun.r_minimum)
if len(X) > 0:
tuner.add(X, y)
params = tuner.propose()
if params is None and self.datarun.gridding:
logger.info('Gridding done for hyperpartition %d' %
hyperpartition.id)
self.db.mark_hyperpartition_gridding_done(hyperpartition.id)
return None
# Append categorical and constants to the params.
return update_params(params=params,
categoricals=hyperpartition.categoricals,
constants=hyperpartition.constants)