def _make_datapipe(
self,
resource_dps: List[IterDataPipe],
*,
config: DatasetConfig,
decoder: Optional[Callable[[io.IOBase], torch.Tensor]],
) -> IterDataPipe[Dict[str, Any]]:
archive_dp = resource_dps[0]
splits_dp, joint_categories_dp, images_dp = Demultiplexer(
archive_dp, 3, self._classify_archive, drop_none=True, buffer_size=INFINITE_BUFFER_SIZE
)
splits_dp = Filter(splits_dp, path_comparator("name", f"{config.split}{config.fold}.txt"))
splits_dp = LineReader(splits_dp, decode=True, return_path=False)
splits_dp = Shuffler(splits_dp, buffer_size=INFINITE_BUFFER_SIZE)
splits_dp = hint_sharding(splits_dp)
joint_categories_dp = CSVParser(joint_categories_dp, delimiter=" ")
dp = IterKeyZipper(
splits_dp,
joint_categories_dp,
key_fn=getitem(),
ref_key_fn=getitem(0),
buffer_size=INFINITE_BUFFER_SIZE,
)
dp = IterKeyZipper(
dp,
images_dp,
key_fn=getitem(0),
ref_key_fn=self._image_key_fn,
buffer_size=INFINITE_BUFFER_SIZE,
)
return Mapper(dp, functools.partial(self._collate_and_decode_sample, decoder=decoder))
def _make_datapipe(
self,
resource_dps: List[IterDataPipe],
*,
config: DatasetConfig,
) -> IterDataPipe[Dict[str, Any]]:
archive_dp, extra_split_dp = resource_dps
archive_dp = resource_dps[0]
split_dp, images_dp, anns_dp = Demultiplexer(
archive_dp,
3,
self._classify_archive,
buffer_size=INFINITE_BUFFER_SIZE,
drop_none=True,
)
if config.split == "train_noval":
split_dp = extra_split_dp
split_dp = Filter(split_dp,
path_comparator("name", f"{config.split}.txt"))
split_dp = LineReader(split_dp, decode=True)
split_dp = hint_sharding(split_dp)
split_dp = hint_shuffling(split_dp)
dp = split_dp
for level, data_dp in enumerate((images_dp, anns_dp)):
dp = IterKeyZipper(
dp,
data_dp,
key_fn=getitem(*[0] * level, 1),
ref_key_fn=path_accessor("stem"),
buffer_size=INFINITE_BUFFER_SIZE,
)
return Mapper(dp, self._prepare_sample)
def _datapipe(self, resource_dps: List[IterDataPipe]) -> IterDataPipe[Dict[str, Any]]:
archive_dp = resource_dps[0]
split_dp, images_dp, anns_dp = Demultiplexer(
archive_dp,
3,
self._classify_archive,
drop_none=True,
buffer_size=INFINITE_BUFFER_SIZE,
)
split_dp = Filter(split_dp, functools.partial(self._is_in_folder, name=self._split_folder))
split_dp = Filter(split_dp, path_comparator("name", f"{self._split}.txt"))
split_dp = LineReader(split_dp, decode=True)
split_dp = hint_shuffling(split_dp)
split_dp = hint_sharding(split_dp)
dp = split_dp
for level, data_dp in enumerate((images_dp, anns_dp)):
dp = IterKeyZipper(
dp,
data_dp,
key_fn=getitem(*[0] * level, 1),
ref_key_fn=path_accessor("stem"),
buffer_size=INFINITE_BUFFER_SIZE,
)
return Mapper(dp, self._prepare_sample)
def _datapipe(
self,
resource_dps: List[IterDataPipe]) -> IterDataPipe[Dict[str, Any]]:
dp = Decompressor(resource_dps[0])
dp = LineReader(dp, decode=True, return_path=False)
dp = hint_shuffling(dp)
dp = hint_sharding(dp)
return Mapper(dp, self._prepare_sample)
def _make_datapipe(self, resource_dps: List[IterDataPipe], *,
config: DatasetConfig) -> IterDataPipe[Dict[str, Any]]:
if config.split in {"train", "test"}:
dp = resource_dps[0]
# the train archive is a tar of tars
if config.split == "train":
dp = TarArchiveReader(dp)
dp = hint_sharding(dp)
dp = hint_shuffling(dp)
dp = Mapper(
dp, self._prepare_train_data
if config.split == "train" else self._prepare_test_data)
else: # config.split == "val":
images_dp, devkit_dp = resource_dps
meta_dp, label_dp = Demultiplexer(devkit_dp,
2,
self._classifiy_devkit,
drop_none=True,
buffer_size=INFINITE_BUFFER_SIZE)
meta_dp = Mapper(meta_dp, self._extract_categories_and_wnids)
_, wnids = zip(*next(iter(meta_dp)))
label_dp = LineReader(label_dp, decode=True, return_path=False)
label_dp = Mapper(
label_dp,
functools.partial(self._imagenet_label_to_wnid, wnids=wnids))
label_dp: IterDataPipe[Tuple[int, str]] = Enumerator(label_dp, 1)
label_dp = hint_sharding(label_dp)
label_dp = hint_shuffling(label_dp)
dp = IterKeyZipper(
label_dp,
images_dp,
key_fn=getitem(0),
ref_key_fn=self._val_test_image_key,
buffer_size=INFINITE_BUFFER_SIZE,
)
dp = Mapper(dp, self._prepare_val_data)
return Mapper(dp, self._prepare_sample)
def _make_datapipe(
self,
resource_dps: List[IterDataPipe],
*,
config: DatasetConfig,
decoder: Optional[Callable[[io.IOBase], torch.Tensor]],
) -> IterDataPipe[Dict[str, Any]]:
images_dp, devkit_dp = resource_dps
if config.split == "train":
# the train archive is a tar of tars
dp = TarArchiveReader(images_dp)
dp = hint_sharding(dp)
dp = hint_shuffling(dp)
dp = Mapper(dp, self._collate_train_data)
elif config.split == "val":
devkit_dp = Filter(
devkit_dp,
path_comparator("name",
"ILSVRC2012_validation_ground_truth.txt"))
devkit_dp = LineReader(devkit_dp, return_path=False)
devkit_dp = Mapper(devkit_dp, int)
devkit_dp = Enumerator(devkit_dp, 1)
devkit_dp = hint_sharding(devkit_dp)
devkit_dp = hint_shuffling(devkit_dp)
dp = IterKeyZipper(
devkit_dp,
images_dp,
key_fn=getitem(0),
ref_key_fn=self._val_test_image_key,
buffer_size=INFINITE_BUFFER_SIZE,
)
dp = Mapper(dp, self._collate_val_data)
else: # config.split == "test"
dp = hint_sharding(images_dp)
dp = hint_shuffling(dp)
dp = Mapper(dp, self._collate_test_data)
return Mapper(
dp,
functools.partial(self._collate_and_decode_sample,
decoder=decoder))
def _make_datapipe(
self,
resource_dps: List[IterDataPipe],
*,
config: DatasetConfig,
) -> IterDataPipe[Dict[str, Any]]:
archive_dp = resource_dps[0]
split_dp, images_dp, anns_dp = Demultiplexer(
archive_dp,
3,
functools.partial(self._classify_archive, config=config),
drop_none=True,
buffer_size=INFINITE_BUFFER_SIZE,
)
split_dp = Filter(
split_dp,
functools.partial(self._is_in_folder,
name=self._SPLIT_FOLDER[config.task]))
split_dp = Filter(split_dp,
path_comparator("name", f"{config.split}.txt"))
split_dp = LineReader(split_dp, decode=True)
split_dp = hint_sharding(split_dp)
split_dp = hint_shuffling(split_dp)
dp = split_dp
for level, data_dp in enumerate((images_dp, anns_dp)):
dp = IterKeyZipper(
dp,
data_dp,
key_fn=getitem(*[0] * level, 1),
ref_key_fn=path_accessor("stem"),
buffer_size=INFINITE_BUFFER_SIZE,
)
return Mapper(
dp,
functools.partial(
self._prepare_sample,
prepare_ann_fn=self._prepare_detection_ann if config.task
== "detection" else self._prepare_segmentation_ann,
),
)
def _make_datapipe(
self,
resource_dps: List[IterDataPipe],
*,
config: DatasetConfig,
decoder: Optional[Callable[[io.IOBase], torch.Tensor]],
) -> IterDataPipe[Dict[str, Any]]:
archive_dp, extra_split_dp = resource_dps
archive_dp = resource_dps[0]
split_dp, images_dp, anns_dp = Demultiplexer(
archive_dp,
3,
self._classify_archive,
buffer_size=INFINITE_BUFFER_SIZE,
drop_none=True,
)
if config.split == "train_noval":
split_dp = extra_split_dp
split_dp = Filter(split_dp, path_comparator("stem", config.split))
split_dp = LineReader(split_dp, decode=True)
split_dp = hint_sharding(split_dp)
split_dp = hint_shuffling(split_dp)
dp = split_dp
for level, data_dp in enumerate((images_dp, anns_dp)):
dp = IterKeyZipper(
dp,
data_dp,
key_fn=getitem(*[0] * level, 1),
ref_key_fn=path_accessor("stem"),
buffer_size=INFINITE_BUFFER_SIZE,
)
return Mapper(
dp,
functools.partial(self._collate_and_decode_sample,
config=config,
decoder=decoder))
def _generate_categories(self) -> Tuple[str, ...]:
resources = self._resources()
dp = resources[0].load(self._root)
dp = Filter(dp, path_comparator("name", "category_names.m"))
dp = LineReader(dp)
dp = Mapper(dp, bytes.decode, input_col=1)
lines = tuple(zip(*iter(dp)))[1]
pattern = re.compile(r"\s*'(?P<category>\w+)';\s*%(?P<label>\d+)")
categories_and_labels = cast(
List[Tuple[str, ...]],
[
pattern.match(line).groups() # type: ignore[union-attr]
# the first and last line contain no information
for line in lines[1:-1]
],
)
categories_and_labels.sort(key=lambda category_and_label: int(category_and_label[1]))
categories, _ = zip(*categories_and_labels)
return categories
def _datapipe(
self,
resource_dps: List[IterDataPipe]) -> IterDataPipe[Dict[str, Any]]:
archive_dp = resource_dps[0]
images_dp, split_dp = Demultiplexer(archive_dp,
2,
self._classify_archive,
drop_none=True,
buffer_size=INFINITE_BUFFER_SIZE)
split_dp = Filter(split_dp,
path_comparator("name", f"{self._split}.txt"))
split_dp = LineReader(split_dp, decode=True, return_path=False)
split_dp = hint_sharding(split_dp)
split_dp = hint_shuffling(split_dp)
dp = IterKeyZipper(
split_dp,
images_dp,
key_fn=getitem(),
ref_key_fn=self._image_key,
buffer_size=INFINITE_BUFFER_SIZE,
)
return Mapper(dp, self._prepare_sample)
def _generate_categories(self) -> List[str]:
resources = self._resources()
dp = resources[0].load(self._root)
dp = Filter(dp, path_comparator("name", "classes.txt"))
dp = LineReader(dp, decode=True, return_path=False)
return list(dp)
def _datapipe(
self,
resource_dps: List[IterDataPipe]) -> IterDataPipe[Dict[str, Any]]:
prepare_ann_fn: Callable
if self._year == "2011":
archive_dp, segmentations_dp = resource_dps
images_dp, split_dp, image_files_dp, bounding_boxes_dp = Demultiplexer(
archive_dp,
4,
self._2011_classify_archive,
drop_none=True,
buffer_size=INFINITE_BUFFER_SIZE)
image_files_dp = CSVParser(image_files_dp, dialect="cub200")
image_files_map = dict(
(image_id, rel_posix_path.rsplit("/", maxsplit=1)[1])
for image_id, rel_posix_path in image_files_dp)
split_dp = CSVParser(split_dp, dialect="cub200")
split_dp = Filter(split_dp, self._2011_filter_split)
split_dp = Mapper(split_dp, getitem(0))
split_dp = Mapper(split_dp, image_files_map.get)
bounding_boxes_dp = CSVParser(bounding_boxes_dp, dialect="cub200")
bounding_boxes_dp = Mapper(bounding_boxes_dp,
image_files_map.get,
input_col=0)
anns_dp = IterKeyZipper(
bounding_boxes_dp,
segmentations_dp,
key_fn=getitem(0),
ref_key_fn=self._2011_segmentation_key,
keep_key=True,
buffer_size=INFINITE_BUFFER_SIZE,
)
prepare_ann_fn = self._2011_prepare_ann
else: # self._year == "2010"
split_dp, images_dp, anns_dp = resource_dps
split_dp = Filter(split_dp,
path_comparator("name", f"{self._split}.txt"))
split_dp = LineReader(split_dp, decode=True, return_path=False)
split_dp = Mapper(split_dp, self._2010_split_key)
anns_dp = Mapper(anns_dp, self._2010_anns_key)
prepare_ann_fn = self._2010_prepare_ann
split_dp = hint_shuffling(split_dp)
split_dp = hint_sharding(split_dp)
dp = IterKeyZipper(
split_dp,
images_dp,
getitem(),
path_accessor("name"),
buffer_size=INFINITE_BUFFER_SIZE,
)
dp = IterKeyZipper(
dp,
anns_dp,
getitem(0),
buffer_size=INFINITE_BUFFER_SIZE,
)
return Mapper(
dp,
functools.partial(self._prepare_sample,
prepare_ann_fn=prepare_ann_fn))
def _make_datapipe(
self,
resource_dps: List[IterDataPipe],
*,
config: DatasetConfig,
decoder: Optional[Callable[[io.IOBase], torch.Tensor]],
) -> IterDataPipe[Dict[str, Any]]:
if config.year == "2011":
archive_dp, segmentations_dp = resource_dps
images_dp, split_dp, image_files_dp, bounding_boxes_dp = Demultiplexer(
archive_dp,
4,
self._2011_classify_archive,
drop_none=True,
buffer_size=INFINITE_BUFFER_SIZE)
image_files_dp = CSVParser(image_files_dp, dialect="cub200")
image_files_map = dict(
(image_id, rel_posix_path.rsplit("/", maxsplit=1)[1])
for image_id, rel_posix_path in image_files_dp)
split_dp = CSVParser(split_dp, dialect="cub200")
split_dp = Filter(
split_dp,
functools.partial(self._2011_filter_split, split=config.split))
split_dp = Mapper(split_dp, getitem(0))
split_dp = Mapper(split_dp, image_files_map.get)
bounding_boxes_dp = CSVParser(bounding_boxes_dp, dialect="cub200")
bounding_boxes_dp = Mapper(bounding_boxes_dp,
image_files_map.get,
input_col=0)
anns_dp = IterKeyZipper(
bounding_boxes_dp,
segmentations_dp,
key_fn=getitem(0),
ref_key_fn=self._2011_segmentation_key,
keep_key=True,
buffer_size=INFINITE_BUFFER_SIZE,
)
else: # config.year == "2010"
split_dp, images_dp, anns_dp = resource_dps
split_dp = Filter(split_dp,
path_comparator("name", f"{config.split}.txt"))
split_dp = LineReader(split_dp, decode=True, return_path=False)
split_dp = Mapper(split_dp, self._2010_split_key)
anns_dp = Mapper(anns_dp, self._2010_anns_key)
split_dp = hint_sharding(split_dp)
split_dp = hint_shuffling(split_dp)
dp = IterKeyZipper(
split_dp,
images_dp,
getitem(),
path_accessor("name"),
buffer_size=INFINITE_BUFFER_SIZE,
)
dp = IterKeyZipper(
dp,
anns_dp,
getitem(0),
buffer_size=INFINITE_BUFFER_SIZE,
)
return Mapper(
dp,
functools.partial(self._collate_and_decode_sample,
year=config.year,
decoder=decoder))