def prepare_confounder_data(args, train, return_full_dataset=False):
full_dataset = confounder_settings[args.dataset]['constructor'](
root_dir=args.root_dir,
target_name=args.target_name,
confounder_names=args.confounder_names,
model_type=args.model,
augment_data=args.augment_data)
if return_full_dataset:
return DRODataset(full_dataset,
process_item_fn=None,
n_groups=full_dataset.n_groups,
n_classes=full_dataset.n_classes,
group_str_fn=full_dataset.group_str)
if train:
splits = ['train', 'val', 'test']
else:
splits = ['test']
subsets = full_dataset.get_splits(
splits,
train_frac=args.fraction,
subsample_to_minority=args.subsample_to_minority)
dro_subsets = [
DRODataset(
subsets[split],
process_item_fn=None,
n_groups=full_dataset.n_groups,
n_classes=full_dataset.n_classes,
group_str_fn=full_dataset.group_str) \
for split in splits]
return dro_subsets
def prepare_confounder_data(args, train, return_full_dataset=False):
full_dataset = confounder_settings[args.dataset]['constructor'](
root_dir=args.root_dir,
target_name=args.target_name,
confounder_names=args.confounder_names,
model_type=args.model,
augment_data=args.augment_data)
if return_full_dataset: # never used
return DRODataset(
full_dataset,
process_item_fn=None,
n_groups=full_dataset.n_groups,
n_classes=full_dataset.n_classes,
group_str_fn=full_dataset.group_str)
if train:
splits = ['train', 'val', 'test'] # used
else:
splits = ['test']
subsets = full_dataset.get_splits(splits, train_frac=args.fraction)
# if args.ood:
# dro_subsets = [ 1]
# dro_subsets.extend([DRODataset(subsets[split], process_item_fn=None, n_groups=full_dataset.n_groups,
# n_classes=full_dataset.n_classes, group_str_fn=full_dataset.group_str, args=args) \
# for split in ['val', 'test']] )
# else:
dro_subsets = [DRODataset(subsets[split], process_item_fn=None, n_groups=full_dataset.n_groups,
n_classes=full_dataset.n_classes, group_str_fn=full_dataset.group_str, args=args) \
for split in splits]
return dro_subsets
def prepare_label_shift_data(args, train):
settings = label_shift_settings[args.dataset]
data = settings["load_fn"](args, train)
n_classes = settings["n_classes"]
if train:
train_data, val_data = data
if args.fraction < 1:
train_data = subsample(train_data, args.fraction)
train_data = apply_label_shift(
train_data,
n_classes,
args.shift_type,
args.minority_fraction,
args.imbalance_ratio,
)
data = [train_data, val_data]
dro_data = [
DRODataset(
subset,
process_item_fn=settings["process_fn"],
n_groups=n_classes,
n_classes=n_classes,
group_str_fn=settings["group_str_fn"],
) for subset in data
]
return dro_data
def prepare_confounder_data(args, train, return_full_dataset=False):
if args.dataset != "jigsaw":
full_dataset = confounder_settings[args.dataset]["constructor"](
root_dir=args.root_dir,
target_name=args.target_name,
confounder_names=args.confounder_names,
model_type=args.model,
augment_data=args.augment_data,
metadata_csv_name=args.metadata_csv_name if
(args.metadata_csv_name is not None) else "metadata.csv")
else:
full_dataset = confounder_settings[args.dataset]["constructor"](
root_dir=args.root_dir,
target_name=args.target_name,
confounder_names=args.confounder_names,
model_type=args.model,
augment_data=args.augment_data,
metadata_csv_name=args.metadata_csv_name if
(args.metadata_csv_name is not None) else "metadata.csv",
batch_size=args.batch_size)
if return_full_dataset:
return DRODataset(
full_dataset,
process_item_fn=None,
n_groups=full_dataset.n_groups,
n_classes=full_dataset.n_classes,
group_str_fn=full_dataset.group_str,
)
if train:
splits = ["train", "val", "test"]
else:
splits = ["test"]
subsets = full_dataset.get_splits(splits, train_frac=args.fraction)
dro_subsets = [
DRODataset(
subsets[split],
process_item_fn=None,
n_groups=full_dataset.n_groups,
n_classes=full_dataset.n_classes,
group_str_fn=full_dataset.group_str,
) for split in splits
]
return dro_subsets
def prepare_confounder_data(args, train, return_full_dataset=False):
full_dataset = confounder_settings[args.dataset]['constructor'](args=args)
if return_full_dataset:
return DRODataset(full_dataset,
process_item_fn=None,
n_groups=full_dataset.n_groups,
n_classes=full_dataset.n_classes,
group_str_fn=full_dataset.group_str)
if train:
if args.dataset == 'CheXpert':
splits = ['train', 'val']
else:
splits = ['train', 'val', 'test']
else:
splits = ['test']
subsets = full_dataset.get_splits(splits)
dro_subsets = [DRODataset(subsets[split], process_item_fn=None, n_groups=full_dataset.n_groups,
n_classes=full_dataset.n_classes, group_str_fn=full_dataset.group_str) \
for split in splits]
return dro_subsets