def prepare_train_data(self):
label_dset = NLVR2Dataset(splits=args.train,
fraction=args.mixmatch_label_fraction)
unlabel_dset = NLVR2Dataset(splits=args.train,
fraction=args.mixmatch_label_fraction,
remove_labels=True,
keep_reverse=True)
print("Final Labelled / Unlabelled Split = %d / %d" %
(len(label_dset), len(unlabel_dset)))
self.img_data = []
label_tset = NLVR2TorchDataset(label_dset, img_data=self.img_data)
unlabel_tset = NLVR2TorchDataset(unlabel_dset, img_data=self.img_data)
label_data_loader = DataLoader(label_tset,
batch_size=args.batch_size // 2,
shuffle=True,
num_workers=args.num_workers,
drop_last=True,
pin_memory=True)
unlabel_data_loader = DataLoader(unlabel_tset,
batch_size=args.batch_size // 2,
shuffle=True,
num_workers=args.num_workers,
drop_last=True,
pin_memory=True)
evaluator = NLVR2Evaluator(label_dset)
self.train_tuple = MixMatchDataTuple(
dataset=label_dset,
loader=label_data_loader,
unlabel_dataset=unlabel_dset,
unlabel_loader=unlabel_data_loader,
evaluator=evaluator)
def get_tuple(splits: str, bs:int, shuffle=False, drop_last=False) -> DataTuple:
dset = NLVR2Dataset(splits)
tset = NLVR2TorchDataset(dset)
evaluator = NLVR2Evaluator(dset)
data_loader = DataLoader(
tset, batch_size=bs,
shuffle=shuffle, num_workers=args.num_workers,
drop_last=drop_last, pin_memory=True
)
return DataTuple(dataset=dset, loader=data_loader, evaluator=evaluator)
def reinitialize(self):
# Update train_tuple
dset = self.updated_dataset_for_self_train()
tset = NLVR2TorchDataset(dset, img_data=self.img_data)
evaluator = NLVR2Evaluator(dset)
data_loader = DataLoader(tset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
drop_last=True,
pin_memory=True)
self.train_tuple = DataTuple(dataset=dset,
loader=data_loader,
evaluator=evaluator)
del self.model
self.setup_model()
self.setup_optimizers()
def get_tuple(splits: str,
bs: int,
shuffle=False,
drop_last=False,
fraction=None,
img_data=None) -> DataTuple:
dset = NLVR2Dataset(splits, fraction=fraction)
tset = NLVR2TorchDataset(dset, img_data=img_data)
evaluator = NLVR2Evaluator(dset)
data_loader = DataLoader(tset,
batch_size=bs,
shuffle=shuffle,
num_workers=args.num_workers,
drop_last=drop_last,
pin_memory=True)
print("Final length of splits %s = %d" % (splits, len(tset)))
return DataTuple(dataset=dset, loader=data_loader, evaluator=evaluator)