def __setstate__(self, state):
import tempfile
with tempfile.NamedTemporaryFile(delete=False) as tf:
tf.write(state["so_f"])
self._n_features = state["n_features"]
self._so_f = tf.name
self._evaluator = _compiled.CompiledPredictor(
tf.name.encode("ascii"), cg.EVALUATE_FN_NAME.encode("ascii"))
def __setstate__(self, state):
import tempfile
self._so_f_object = tempfile.NamedTemporaryFile(mode='w+b', prefix='compiledtrees_', suffix='.so', delete=True)
self._so_f_object.write(state["so_f"])
self._so_f_object.flush()
self._so_f = self._so_f_object.name
self._n_features = state["n_features"]
self._evaluator = _compiled.CompiledPredictor(
self._so_f_object.name.encode("ascii"),
cg.EVALUATE_FN_NAME.encode("ascii"))
def _build(cls, clf, n_jobs=1):
if not cls.compilable(clf):
raise ValueError("Predictor {} cannot be compiled".format(
clf.__class__.__name__))
files = None
n_features = None
if isinstance(clf, DecisionTreeRegressor):
n_features = clf.n_features_
files = cg.code_gen_regressor_tree(tree=clf.tree_)
if isinstance(clf, GradientBoostingRegressor):
n_features = clf.n_features_
# hack to get the initial (prior) on the decision tree.
if hasattr(clf, '_raw_predict_init'):
initial_value = clf._raw_predict_init(
np.zeros(shape=(1, n_features))).item((0, 0))
else:
# older scikit-learn
initial_value = clf._init_decision_function(
np.zeros(shape=(1, n_features))).item((0, 0))
files = cg.code_gen_ensemble_regressor(
trees=[e.tree_ for e in clf.estimators_.flat],
individual_learner_weight=clf.learning_rate,
initial_value=initial_value,
n_jobs=n_jobs)
if isinstance(clf, ForestRegressor):
n_features = clf.n_features_
files = cg.code_gen_ensemble_regressor(
trees=[e.tree_ for e in clf.estimators_],
individual_learner_weight=1.0 / clf.n_estimators,
initial_value=0.0,
n_jobs=n_jobs)
assert n_features is not None
assert files is not None
so_f = cg.compile_code_to_object(files, n_jobs=n_jobs)
evaluator = _compiled.CompiledPredictor(
so_f.name.encode("ascii"), cg.EVALUATE_FN_NAME.encode("ascii"))
return n_features, evaluator, so_f
def __setstate__(self, state):
import tempfile
self._so_f_object = tempfile.NamedTemporaryFile(
mode='w+b',
prefix='compiledtrees_',
suffix='.so',
delete=delete_files)
if isinstance(state["so_f"], six.text_type):
state["so_f"] = state["so_f"].encode('latin1')
self._so_f_object.write(state["so_f"])
self._so_f_object.flush()
self._so_f = self._so_f_object.name
if platform.system() == 'Windows':
self._so_f_object.close()
self._n_features = state["n_features"]
self._evaluator = _compiled.CompiledPredictor(
self._so_f_object.name.encode("ascii"),
cg.EVALUATE_FN_NAME.encode("ascii"))
def _build(cls, clf):
if not cls.compilable(clf):
raise ValueError("Predictor {} cannot be compiled".format(
clf.__class__.__name__))
lines = None
n_features = None
if isinstance(clf, DecisionTreeRegressor):
n_features = clf.n_features_
lines = cg.code_gen_tree(tree=clf.tree_)
if isinstance(clf, GradientBoostingRegressor):
n_features = clf.n_features
# hack to get the initial (prior) on the decision tree.
initial_value = clf._init_decision_function(
np.zeros(shape=(1, n_features))).item((0, 0))
lines = cg.code_gen_ensemble(
trees=[e.tree_ for e in clf.estimators_.flat],
individual_learner_weight=clf.learning_rate,
initial_value=initial_value)
if isinstance(clf, ForestRegressor):
n_features = clf.n_features_
lines = cg.code_gen_ensemble(
trees=[e.tree_ for e in clf.estimators_],
individual_learner_weight=1.0 / clf.n_estimators,
initial_value=0.0)
assert n_features is not None
assert lines is not None
so_f = cg.compile_code_to_object("\n".join(lines))
return n_features, _compiled.CompiledPredictor(
so_f.encode("ascii"), cg.EVALUATE_FN_NAME.encode("ascii"))
def __setstate__(self, state):
super(CompiledRegressionPredictor, self).__setstate__(state)
self._evaluator = _compiled.CompiledPredictor(
self._so_f_object.name.encode("ascii"),
cg.EVALUATE_FN_NAME.encode("ascii"))