def setup(self, stage=None):
# MiraBest test samples #
mb_test = MB_nohybrids(
path_dict["mb"], train=False, transform=test_transforms["mb"]
)
self.test = mb_test
# Labelled MiraBest samples #
mb_l = MB_nohybrids(path_dict["mb"], train=True, transform=transforms["mb"])
if self.fraction != 1:
mb_l, _ = d_split(mb_l, self.fraction)
self.train_l = mb_l
# Cut unlabelled RGZ samples and define unlabelled training set #
mb_u = RGZ20k(path_dict["rgz"], train=True, transform=transforms["mb"])
size_cut(config["data"]["cut_threshold"], mb_u)
mb_cut(mb_u)
self.train_u = mb_u
# Combine labelled and unlabelled datasets #
self.train = torch.utils.data.ConcatDataset([mb_u, mb_l])
# Flip a number of targets randomly #
if config["train"]["flip"]:
mb_l = flip_targets(mb_l, config["train"]["flip"])
self.save_hparams()
def setup(self, stage=None):
mb_train = MB_nohybrids(self.path,
train=True,
transform=transforms["mb"])
mb_test = MB_nohybrids(self.path,
train=False,
transform=test_transforms["mb"])
self.test_dataset = mb_test
if self.fraction != 1:
mb_train, _ = d_split(mb_train, self.fraction)
self.train_dataset = mb_train
mb_l, mb_u = d_split(mb_train, self.split)
# Flip a number of targets randomly
if config["train"]["flip"]:
mb_u = flip_targets(mb_u, config["train"]["flip"])
self.train_dataset_u = mb_u
self.train_dataset_l = mb_l
self.f_u = label_fraction(mb_u, 0)
self.f_l = label_fraction(mb_l, 0)
def prepare_data(self):
MB_nohybrids(path_dict["mb"], train=True, download=True)
MB_nohybrids(path_dict["mb"], train=False, download=True)
RGZ20k(path_dict["rgz"], train=True, download=True)
def prepare_data(self):
MB_nohybrids(self.path, train=True, download=True)
MB_nohybrids(self.path, train=False, download=True)