def get_train_val_split(data: Union[Dataset, ConcatDataset],
**kwargs) -> Tuple[DataLoader, DataLoader]:
"""Creates validation and train dataloader from the Data_set object.
Args:
data (Data_Set): Object of the class Data_Set.
kwargs:
These arguments are passed as is to the pytorch DataLoader.
Returns:
Tuple[DataLoader, DataLoader]: train and validation data loader respectively.
"""
if isinstance(data, ConcatDataset):
val_datasets = []
train_weights, val_weights = [], []
for i in range(len(data.datasets)):
train_weights += [1.0 / len(data.datasets[i])] * len(
data.datasets[i])
val_datasets.append(copy.copy(data.datasets[i]))
val_datasets[-1].is_training(False)
val_weights += [1.0 / len(val_datasets[-1])] * len(
val_datasets[-1])
val_data = ConcatDataset(val_datasets)
train_weights = np.array(train_weights) / sum(train_weights)
val_weights = np.array(val_weights) / sum(val_weights)
return (
DataLoader(data,
sampler=WeightedRandomSampler(
weights=train_weights,
num_samples=len(train_weights),
replacement=True,
),
**kwargs),
DataLoader(val_data,
sampler=WeightedRandomSampler(
weights=val_weights,
num_samples=len(val_weights),
replacement=True),
**kwargs),
)
else:
data.is_training(True)
val_data = copy.copy(data)
val_data.is_training(False)
return (
DataLoader(data, **{
**kwargs, "shuffle": True
}),
DataLoader(val_data, **{
**kwargs, "shuffle": False
}),
)