def __init__(self, configs=MethodConfigs()):
super(FuseTransfer, self).__init__(configs)
self.use_oracle = False
#self.target_weight_scale = None
self.target_weight_scale = .75
self.label_transform = NanLabelBinarizer()
def __init__(self, configs=None):
super(FuseTransfer, self).__init__(configs)
self.use_oracle = False
# self.target_weight_scale = None
self.target_weight_scale = 0.75
self.label_transform = NanLabelBinarizer()
class FuseTransfer(TargetTranfer):
def __init__(self, configs=MethodConfigs()):
super(FuseTransfer, self).__init__(configs)
self.use_oracle = False
#self.target_weight_scale = None
self.target_weight_scale = .75
self.label_transform = NanLabelBinarizer()
def train(self, data):
is_labeled_train = data.is_labeled & data.is_train
n_labeled_target = (data.is_train & is_labeled_train).sum()
n_labeled_source = (data.is_train & is_labeled_train).sum()
data.instance_weights = np.ones(data.n)
if self.target_weight_scale is not None:
assert 0 <= self.target_weight_scale <= 1
data.instance_weights[data.is_source] /= n_labeled_source
data.instance_weights[data.is_target] /= n_labeled_target
data.instance_weights[data.is_target] *= self.target_weight_scale
data.instance_weights[data.is_source] *= (1 -
self.target_weight_scale)
y_old = data.y
if self.label_transform is not None:
data.y = self.label_transform.fit_transform(data.y)
super(FuseTransfer, self).train(data)
data.y = y_old
def _prepare_data(self, data, include_unlabeled=True):
source_labels = self.configs.source_labels
target_labels = self.configs.target_labels
data_copy = copy.deepcopy(data)
if data.data_set_ids is not None:
#assert source_labels is None
#assert target_labels is None
#data_copy.type[data_copy.data_set_ids > 0] = data_lib.TYPE_SOURCE
for i in source_labels:
data_copy.type[data_copy.data_set_ids ==
i] = data_lib.TYPE_SOURCE
return data_copy
#source_inds = array_functions.find_set(data_copy.true_y,source_labels)
if self.use_oracle:
oracle_labels = self.configs.oracle_labels
data_copy = data_copy.get_transfer_subset(np.concatenate(
(oracle_labels.ravel(), target_labels.ravel())),
include_unlabeled=True)
data_copy.data_set_ids = np.zeros(data_copy.n)
for i, s in enumerate(source_labels):
source_inds = data_copy.get_transfer_inds(s)
if not data_copy.is_regression:
data_copy.change_labels(s, target_labels)
data_copy.type[source_inds] = data_lib.TYPE_SOURCE
data_copy.is_train[source_inds] = True
data_copy.data_set_ids[source_inds] = i + 1
if getattr(self, 'use_all_source', False):
data_copy.reveal_labels(data_copy.is_source)
return data_copy
@property
def prefix(self):
s = 'FuseTransfer+' + self.base_learner.prefix
if 'target_weight_scale' in self.__dict__ and self.target_weight_scale is not None:
s += '-tws=' + str(self.target_weight_scale)
if 'use_oracle' in self.__dict__ and self.use_oracle:
s += '-Oracle'
return s
class FuseTransfer(TargetTranfer):
def __init__(self, configs=None):
super(FuseTransfer, self).__init__(configs)
self.use_oracle = False
# self.target_weight_scale = None
self.target_weight_scale = 0.75
self.label_transform = NanLabelBinarizer()
def train(self, data):
is_labeled_train = data.is_labeled & data.is_train
n_labeled_target = (data.is_train & is_labeled_train).sum()
n_labeled_source = (data.is_train & is_labeled_train).sum()
data.instance_weights = np.ones(data.n)
if self.target_weight_scale is not None:
assert 0 <= self.target_weight_scale <= 1
data.instance_weights[data.is_source] /= n_labeled_source
data.instance_weights[data.is_target] /= n_labeled_target
data.instance_weights[data.is_target] *= self.target_weight_scale
data.instance_weights[data.is_source] *= 1 - self.target_weight_scale
y_old = data.y
if self.label_transform is not None:
data.y = self.label_transform.fit_transform(data.y)
super(FuseTransfer, self).train(data)
data.y = y_old
def _prepare_data(self, data, include_unlabeled=True):
source_labels = self.configs.source_labels
target_labels = self.configs.target_labels
data_copy = copy.deepcopy(data)
if data.data_set_ids is not None:
# assert source_labels is None
# assert target_labels is None
# data_copy.type[data_copy.data_set_ids > 0] = data_lib.TYPE_SOURCE
for i in source_labels:
data_copy.type[data_copy.data_set_ids == i] = data_lib.TYPE_SOURCE
return data_copy
# source_inds = array_functions.find_set(data_copy.true_y,source_labels)
if self.use_oracle:
oracle_labels = self.configs.oracle_labels
data_copy = data_copy.get_transfer_subset(
np.concatenate((oracle_labels.ravel(), target_labels.ravel())), include_unlabeled=True
)
data_copy.data_set_ids = np.zeros(data_copy.n)
for i, s in enumerate(source_labels):
source_inds = data_copy.get_transfer_inds(s)
if not data_copy.is_regression:
data_copy.change_labels(s, target_labels)
data_copy.type[source_inds] = data_lib.TYPE_SOURCE
data_copy.is_train[source_inds] = True
data_copy.data_set_ids[source_inds] = i + 1
data_copy.reveal_labels(data_copy.is_source)
return data_copy
@property
def prefix(self):
s = "FuseTransfer+" + self.base_learner.prefix
if "target_weight_scale" in self.__dict__ and self.target_weight_scale is not None:
s += "-tws=" + str(self.target_weight_scale)
if "use_oracle" in self.__dict__ and self.use_oracle:
s += "-Oracle"
return s
