def dataset_convert(dataset_identifier: str,
format: str,
output_dir: Optional[PathLike] = None) -> None:
"""
Converts the annotations from the given dataset to the given format.
Exits the application if no dataset with the given slug exists or no releases for the dataset
were previously pulled.
Parameters
----------
dataset_identifier: str
The dataset identifier, normally in the "<team-slug>/<dataset-slug>:<version>" form.
format: str
The format we want to convert to.
output_dir: Optional[PathLike]
The folder where the exported annotation files will be. If None it will be the inside the
annotations folder of the dataset under 'other_formats/{format}'. The Defaults to None.
"""
identifier: DatasetIdentifier = DatasetIdentifier.parse(dataset_identifier)
client: Client = _load_client(team_slug=identifier.team_slug)
try:
parser: ExportParser = get_exporter(format)
dataset: RemoteDataset = client.get_remote_dataset(
dataset_identifier=identifier)
if not dataset.local_path.exists():
_error(
f"No annotations downloaded for dataset f{dataset}, first pull a release using "
f"'darwin dataset pull {identifier}'")
release_path: Path = get_release_path(dataset.local_path,
identifier.version)
annotations_path: Path = release_path / "annotations"
if output_dir is None:
output_dir = release_path / "other_formats" / format
else:
output_dir = Path(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
export_annotations(parser, [annotations_path], output_dir)
except ExporterNotFoundError:
_error(
f"Unsupported export format: {format}, currently supported: {export_formats}"
)
except AttributeError:
_error(
f"Unsupported export format: {format}, currently supported: {export_formats}"
)
except NotFound as e:
_error(f"No dataset with name '{e.name}'")
def dataset_convert(dataset_slug: str, format: str, output_dir: Optional[Union[str, Path]]):
client = _load_client()
parser = find_supported_format(format, darwin.exporter.formats.supported_formats)
try:
dataset = client.get_remote_dataset(dataset_identifier=dataset_slug)
if not dataset.local_path.exists():
_error(
f"No annotations downloaded for dataset f{dataset}, first pull a release using "
f"'darwin dataset pull {dataset_slug}'"
)
release_path = get_release_path(dataset.local_path)
annotations_path = release_path / "annotations"
if output_dir is None:
output_dir = release_path / "other_formats" / f"{format}"
else:
output_dir = Path(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
exporter.export_annotations(parser, [annotations_path], output_dir)
except NotFound as e:
_error(f"No dataset with name '{e.name}'")
def __init__(
self,
dataset_path: Path,
annotation_type: str,
partition: Optional[str] = None,
split: str = "default",
split_type: str = "random",
release_name: Optional[str] = None,
):
""" Creates a dataset
Parameters
----------
dataset_path: Path, str
Path to the location of the dataset on the file system
annotation_type: str
The type of annotation classes [tag, bounding_box, polygon]
partition: str
Selects one of the partitions [train, val, test]
split: str
Selects the split that defines the percentages used (use 'default' to select the default split)
split_type: str
Heuristic used to do the split [random, stratified]
release_name: str
Version of the dataset
"""
assert dataset_path is not None
release_path = get_release_path(dataset_path, release_name)
annotations_dir = release_path / "annotations"
assert annotations_dir.exists()
images_dir = dataset_path / "images"
assert images_dir.exists()
if partition not in ["train", "val", "test", None]:
raise ValueError("partition should be either 'train', 'val', or 'test'")
if split_type not in ["random", "stratified"]:
raise ValueError("split_type should be either 'random', 'stratified'")
if annotation_type not in ["tag", "polygon", "bounding_box"]:
raise ValueError("annotation_type should be either 'tag', 'bounding_box', or 'polygon'")
self.dataset_path = dataset_path
self.annotation_type = annotation_type
self.images_path: List[Path] = []
self.annotations_path: List[Path] = []
self.original_classes = None
self.original_images_path: Optional[List[Path]] = None
self.original_annotations_path: Optional[List[Path]] = None
# Get the list of classes
self.classes = get_classes(
self.dataset_path, release_name, annotation_type=self.annotation_type, remove_background=True
)
self.num_classes = len(self.classes)
# Get the list of stems
if partition:
# Get the split
if split_type == "random":
split_file = f"{split_type}_{partition}.txt"
elif split_type == "stratified":
split_file = f"{split_type}_{annotation_type}_{partition}.txt"
split_path = release_path / "lists" / split / split_file
if split_path.is_file():
stems = (e.strip() for e in split_path.open())
else:
raise FileNotFoundError(
f"could not find a dataset partition. "
f"Split the dataset using `split_dataset()` from `darwin.dataset.utils`"
) from None
else:
# If the partition is not specified, get all the annotations
stems = [e.stem for e in annotations_dir.glob("*.json")]
# Find all the annotations and their corresponding images
for stem in stems:
annotation_path = annotations_dir / f"{stem}.json"
images = []
for ext in SUPPORTED_IMAGE_EXTENSIONS:
image_path = images_dir / f"{stem}{ext}"
if image_path.exists():
images.append(image_path)
if len(images) < 1:
raise ValueError(f"Annotation ({annotation_path}) does not have a corresponding image")
if len(images) > 1:
raise ValueError(f"Image ({stem}) is present with multiple extensions. This is forbidden.")
assert len(images) == 1
self.images_path.append(images[0])
self.annotations_path.append(annotation_path)
if len(self.images_path) == 0:
raise ValueError(f"Could not find any {SUPPORTED_IMAGE_EXTENSIONS} file", f" in {images_dir}")
assert len(self.images_path) == len(self.annotations_path)
def it_uses_provided_version_name_otherwise(team_dataset_path: Path):
test_release_path = team_dataset_path / "releases" / "test"
test_release_path.mkdir(parents=True)
assert get_release_path(team_dataset_path, "test") == test_release_path
def it_defaults_to_latest_version_if_no_version_provided(
team_dataset_path: Path):
latest_release_path = team_dataset_path / "releases" / "latest"
latest_release_path.mkdir(parents=True)
assert get_release_path(team_dataset_path) == latest_release_path
def split_dataset(
dataset_path: PathLike,
release_name: Optional[str] = None,
val_percentage: float = 0.1,
test_percentage: float = 0.2,
split_seed: int = 0,
make_default_split: bool = True,
stratified_types: List[str] = ["bounding_box", "polygon", "tag"],
) -> Path:
"""
Given a local a dataset (pulled from Darwin), split it by creating lists of filenames.
The partitions to split the dataset into are called train, val and test.
The dataset is always split randomly, and can be additionally split according to the
stratified strategy by providing a list of stratified types.
Parameters
----------
dataset_path : Path
Local path to the dataset
release_name : str
Version of the dataset
val_percentage : float
Percentage of images used in the validation set
test_percentage : float
Percentage of images used in the test set
split_seed : int
Fix seed for random split creation
make_default_split : bool
Makes this split the default split
stratified_types : List[str]
List of annotation types to split with the stratified strategy
Returns
-------
splits : Path
Keys are the different splits (random, tags, ...) and values are the relative file names
"""
# Requirements: scikit-learn
try:
import sklearn # noqa
except ImportError:
raise ImportError(
"Darwin requires scikit-learn to split a dataset. Install it using: pip install scikit-learn"
) from None
_validate_split(val_percentage, test_percentage)
# Infer release path
if isinstance(dataset_path, str):
dataset_path = Path(dataset_path)
release_path = get_release_path(dataset_path, release_name)
# List all annotation files in release
annotation_path = release_path / "annotations"
assert annotation_path.exists()
annotation_files = list(annotation_path.glob("**/*.json"))
# Prepare the "lists" folder, which is where we are going to save the split files
lists_path = release_path / "lists"
lists_path.mkdir(parents=True, exist_ok=True)
# Compute sizes of each dataset partition
dataset_size: int = len(annotation_files)
val_size: int = int(val_percentage * dataset_size)
test_size: int = int(test_percentage * dataset_size)
train_size: int = dataset_size - val_size - test_size
split_id = f"{train_size}_{val_size}_{test_size}"
# Compute split id, a combination of val precentage, test percentage and split seed
# The split id is used to create a folder with the same name in the "lists" folder
if split_seed != 0:
split_id += f"_s{split_seed}"
split_path = lists_path / split_id
# Build a split paths dictionary. The split paths are indexed by strategy (e.g. random
# or stratified), and by partition (train/val/test)
split = _build_split(split_path, stratified_types)
assert split.is_valid()
# Do the actual splitting
split_path.mkdir(exist_ok=True)
if split.random:
_random_split(
annotation_path=annotation_path,
annotation_files=annotation_files,
split=split.random,
train_size=train_size,
val_size=val_size,
test_size=test_size,
split_seed=split_seed,
)
if split.stratified:
_stratified_split(
annotation_path=annotation_path,
split=split.stratified,
annotation_files=annotation_files,
train_size=train_size,
val_size=val_size,
test_size=test_size,
stratified_types=stratified_types,
split_seed=split_seed,
)
# Create symlink for default split
default_split_path = lists_path / "default"
if make_default_split or not default_split_path.exists():
if default_split_path.exists():
default_split_path.unlink()
default_split_path.symlink_to(f"./{split_id}")
return split_path
