hafnia 0.4.2__py3-none-any.whl → 0.5.0__py3-none-any.whl

hafnia/dataset/format_conversions/format_coco.py

@@ -0,0 +1,490 @@
+import json
+import shutil
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
+
+import polars as pl
+from pycocotools import mask as coco_utils
+
+from hafnia.dataset import license_types
+from hafnia.dataset.dataset_names import SampleField, SplitName
+from hafnia.dataset.format_conversions import format_coco, format_helpers
+from hafnia.utils import progress_bar
+
+if TYPE_CHECKING:  # Using 'TYPE_CHECKING' to avoid circular imports during type checking
+    from hafnia.dataset.hafnia_dataset import HafniaDataset
+
+from hafnia.dataset.hafnia_dataset_types import Attribution, DatasetInfo, License, Sample, TaskInfo
+from hafnia.dataset.primitives import Bbox, Bitmask
+from hafnia.log import user_logger
+
+COCO_KEY_FILE_NAME = "file_name"
+
+HAFNIA_TO_ROBOFLOW_SPLIT_NAME = {
+    SplitName.TRAIN: "train",
+    SplitName.VAL: "valid",
+    SplitName.TEST: "test",
+}
+ROBOFLOW_ANNOTATION_FILE_NAME = "_annotations.coco.json"
+
+
+@dataclass
+class CocoSplitPaths:
+    split: str
+    path_images: Path
+    path_instances_json: Path
+
+
+def from_coco_format(
+    path_dataset: Path,
+    coco_format_type: str = "roboflow",
+    max_samples: Optional[int] = None,
+    dataset_name: str = "coco-2017",
+):
+    split_definitions = get_split_paths_for_coco_dataset_formats(
+        path_dataset=path_dataset, coco_format_type=coco_format_type
+    )
+
+    hafnia_dataset = from_coco_dataset_by_split_definitions(
+        split_definitions=split_definitions,
+        max_samples=max_samples,
+        dataset_name=dataset_name,
+    )
+
+    return hafnia_dataset
+
+
+def get_split_paths_for_coco_dataset_formats(
+    path_dataset: Path,
+    coco_format_type: str,
+) -> List[CocoSplitPaths]:
+    splits = []
+    if coco_format_type == "roboflow":
+        for split_def in format_helpers.get_splits_from_folder(path_dataset):
+            splits.append(
+                CocoSplitPaths(
+                    split=split_def.name,
+                    path_images=split_def.path,
+                    path_instances_json=split_def.path / ROBOFLOW_ANNOTATION_FILE_NAME,
+                )
+            )
+        return splits
+
+    raise ValueError(f"The specified '{coco_format_type=}' is not supported.")
+
+
+def from_coco_dataset_by_split_definitions(
+    split_definitions: List[CocoSplitPaths],
+    max_samples: Optional[int],
+    dataset_name: str,
+) -> "HafniaDataset":
+    from hafnia.dataset.hafnia_dataset import HafniaDataset
+
+    if max_samples is None:
+        max_samples_per_split = None
+    else:
+        max_samples_per_split = max_samples // len(split_definitions)
+    samples = []
+    tasks: List[TaskInfo] = []
+    for split_definition in split_definitions:
+        if split_definition.path_instances_json is None or not split_definition.path_instances_json.exists():
+            raise FileNotFoundError(
+                f"Expected COCO dataset files not found for split '{split_definition.split}'. "
+                f"Label file doesn't exist: {split_definition.path_instances_json}"
+            )
+        if not split_definition.path_images.exists():
+            raise FileNotFoundError(
+                f"Expected COCO dataset files not found for split '{split_definition.split}'. "
+                f"Images folder doesn't exist: {split_definition.path_images}"
+            )
+
+        samples_in_split, tasks_in_split = coco_format_folder_with_split_to_hafnia_samples(
+            path_label_file=split_definition.path_instances_json,
+            max_samples_per_split=max_samples_per_split,
+            path_images=split_definition.path_images,
+            split_name=split_definition.split,
+        )
+
+        for task_in_split in tasks_in_split:
+            matching_tasks = [task for task in tasks if task.name == task_in_split.name]
+
+            add_missing_task = len(matching_tasks) == 0
+            if add_missing_task:
+                tasks.append(task_in_split)
+                continue
+
+            if len(matching_tasks) != 1:
+                raise ValueError("Duplicate task names found across splits in the COCO dataset.")
+            match_task = matching_tasks[0]
+            if task_in_split != match_task:
+                raise ValueError(
+                    f"Inconsistent task found across splits in the COCO dataset for task name '{task_in_split.name}'. "
+                )
+
+        samples.extend(samples_in_split)
+
+    dataset_info = DatasetInfo(
+        dataset_name=dataset_name,
+        tasks=tasks,
+    )
+
+    hafnia_dataset = HafniaDataset.from_samples_list(samples, info=dataset_info)
+    return hafnia_dataset
+
+
+def coco_format_folder_with_split_to_hafnia_samples(
+    path_label_file: Path,
+    path_images: Path,
+    split_name: str,
+    max_samples_per_split: Optional[int],
+) -> Tuple[List[Sample], List[TaskInfo]]:
+    if not path_label_file.exists():
+        raise FileNotFoundError(f"Expected label file not found: {path_label_file}")
+    user_logger.info("Loading coco label file as json")
+    image_and_annotation_dict = json.loads(path_label_file.read_text())
+    user_logger.info("Converting coco dataset to HafniaDataset samples")
+
+    id_to_category, class_names = get_coco_id_category_mapping(image_and_annotation_dict.get("categories", []))
+    tasks = [
+        TaskInfo(primitive=Bbox, class_names=class_names),
+        TaskInfo(primitive=Bitmask, class_names=class_names),
+    ]
+
+    coco_licenses = image_and_annotation_dict.get("licenses", [])
+    id_to_license_mapping = {lic["id"]: license_types.get_license_by_url(lic["url"]) for lic in coco_licenses}
+
+    coco_images = image_and_annotation_dict.get("images", [])
+    if max_samples_per_split is not None:
+        coco_images = coco_images[:max_samples_per_split]
+    id_to_image = {img["id"]: img for img in coco_images}
+
+    img_id_to_annotations: Dict[int, List[dict]] = {}
+    coco_annotations = image_and_annotation_dict.get("annotations", [])
+    for annotation in coco_annotations:
+        img_id = annotation["image_id"]
+        if img_id not in img_id_to_annotations:
+            img_id_to_annotations[img_id] = []
+        img_id_to_annotations[img_id].append(annotation)
+
+    samples = []
+    for img_id, image_dict in progress_bar(
+        id_to_image.items(), description=f"Convert coco to hafnia sample '{split_name}'"
+    ):
+        image_annotations = img_id_to_annotations.get(img_id, [])
+
+        sample = fiftyone_coco_to_hafnia_sample(
+            path_images=path_images,
+            image_dict=image_dict,
+            image_annotations=image_annotations,
+            id_to_category=id_to_category,
+            class_names=class_names,
+            id_to_license_mapping=id_to_license_mapping,
+            split_name=split_name,
+        )
+        samples.append(sample)
+
+    return samples, tasks
+
+
+def get_coco_id_category_mapping(
+    coco_categories: List[dict],
+) -> Tuple[Dict[int, dict], List[str]]:
+    category_mapping = {}
+    for i_cat, category in enumerate(coco_categories):
+        category = category.copy()  # Create a copy to avoid modifying the original dictionary.
+        category["class_idx"] = i_cat  # Add an index to the category for easier access.
+        category_mapping[category["id"]] = category  # Map the category ID to the category dictionary.
+    sorted_category_mapping = dict(sorted(category_mapping.items(), key=lambda item: item[1]["class_idx"]))
+    class_names = [cat_data["name"] for cat_data in sorted_category_mapping.values()]
+    return sorted_category_mapping, class_names
+
+
+def convert_segmentation_to_rle_list(segmentation: Union[Dict, List], height: int, width: int) -> List[Dict]:
+    is_polygon_format = isinstance(segmentation, list)
+    if is_polygon_format:  # Multiple polygons format
+        rles = coco_utils.frPyObjects(segmentation, height, width)
+        return rles
+
+    is_rle_format = isinstance(segmentation, dict) and "counts" in segmentation
+    if is_rle_format:  # RLE format
+        counts = segmentation["counts"]  # type: ignore
+        uncompressed_list_of_ints = isinstance(counts, list)
+        if uncompressed_list_of_ints:  # Uncompressed RLE. Counts is List[int]
+            rles = coco_utils.frPyObjects([segmentation], height, width)
+            return rles
+
+        is_compressed_str_or_bytes = isinstance(counts, str | bytes)
+        if is_compressed_str_or_bytes:  # Compressed RLE. Counts is str
+            rles = [segmentation]
+            return rles
+
+    raise ValueError("Segmentation format not recognized for conversion to RLE.")
+
+
+def fiftyone_coco_to_hafnia_sample(
+    path_images: Path,
+    image_dict: Dict,
+    image_annotations: List[Dict],
+    id_to_category: Dict,
+    class_names: List[str],
+    id_to_license_mapping: Dict[int, License],
+    split_name: str,
+) -> Sample:
+    image_dict = image_dict.copy()  # Create a copy to avoid modifying the original dictionary.
+    file_name_relative = image_dict.pop(COCO_KEY_FILE_NAME)
+    file_name = path_images / file_name_relative
+    if not file_name.exists():
+        raise FileNotFoundError(f"Expected image file not found: {file_name}. Please check the dataset structure.")
+
+    img_width = image_dict.pop("width")
+    img_height = image_dict.pop("height")
+    bitmasks: List[Bitmask] = []
+    bboxes: List[Bbox] = []
+    for obj_instance in image_annotations:
+        category_data = id_to_category[obj_instance["category_id"]]
+        class_name = category_data["name"]  # Get the name of the category.
+        class_idx = class_names.index(class_name)
+        bbox_list = obj_instance["bbox"]
+        if isinstance(bbox_list[0], float):  # Polygon coordinates are often floats.
+            bbox_ints = [int(coord) for coord in bbox_list]
+        else:
+            bbox_ints = bbox_list
+        rle_list = convert_segmentation_to_rle_list(obj_instance["segmentation"], height=img_height, width=img_width)
+        rle = coco_utils.merge(rle_list)
+        rle_string = rle["counts"]
+        if isinstance(rle_string, bytes):
+            rle_string = rle_string.decode("utf-8")
+
+        if "area" in obj_instance and obj_instance["area"] is not None:
+            area_px = obj_instance["area"]
+        else:
+            area_px = coco_utils.area(rle).item()
+        area = float(area_px) / (img_height * img_width)
+        bitmask = Bitmask(
+            top=bbox_ints[1],
+            left=bbox_ints[0],
+            height=bbox_ints[3],
+            width=bbox_ints[2],
+            area=area,
+            rle_string=rle_string,
+            class_name=class_name,
+            class_idx=class_idx,
+            object_id=str(obj_instance["id"]),
+            meta={"iscrowd": obj_instance["iscrowd"]},
+        )
+        bitmasks.append(bitmask)
+
+        bbox = Bbox.from_coco(bbox=bbox_list, height=img_height, width=img_width)
+        bbox.class_name = class_name
+        bbox.class_idx = class_idx
+        bbox.object_id = str(obj_instance["id"])  # Use the ID from the instance if available.
+        bbox.meta = {"iscrowd": obj_instance["iscrowd"]}
+        bbox.area = bbox.calculate_area(image_height=img_height, image_width=img_width)
+        bboxes.append(bbox)
+
+    if "license" in image_dict:
+        license_data: License = id_to_license_mapping[image_dict["license"]]
+
+        capture_date = datetime.fromisoformat(image_dict["date_captured"])
+        source_url = image_dict["flickr_url"] if "flickr_url" in image_dict else image_dict.get("coco_url")
+        attribution = Attribution(
+            date_captured=capture_date,
+            licenses=[license_data],
+            source_url=source_url,
+        )
+    else:
+        attribution = None
+
+    return Sample(
+        file_path=str(file_name),
+        width=img_width,
+        height=img_height,
+        split=split_name,
+        bboxes=bboxes,  # Bboxes will be added later if needed.
+        bitmasks=bitmasks,  # Add the bitmask to the sample.
+        attribution=attribution,
+        meta=image_dict,
+    )
+
+
+def to_coco_format(
+    dataset: "HafniaDataset",
+    path_output: Path,
+    task_name: Optional[str] = None,
+    coco_format_type: str = "roboflow",
+) -> List[CocoSplitPaths]:
+    samples_modified_all = dataset.samples.with_row_index("id")
+
+    if SampleField.ATTRIBUTION in samples_modified_all.columns:
+        samples_modified_all = samples_modified_all.unnest(SampleField.ATTRIBUTION)
+        license_table = (
+            samples_modified_all["licenses"]
+            .explode()
+            .struct.unnest()
+            .unique()
+            .with_row_index("id")
+            .select(["id", "name", "url"])
+        )
+        license_mapping = {lic["name"]: lic["id"] for lic in license_table.iter_rows(named=True)}
+    else:
+        license_mapping = None
+        license_table = None
+
+    if task_name is not None:
+        task_info = dataset.info.get_task_by_name(task_name)
+    else:
+        # Auto derive the task to be used for COCO conversion as only one Bitmask/Bbox task can be present
+        # in the coco format. Will first search for Bitmask (because COCO supports segmentation), then Bbox afterwards.
+        tasks_info = dataset.info.get_tasks_by_primitive(Bitmask)
+        if len(tasks_info) == 0:
+            tasks_info = dataset.info.get_tasks_by_primitive(Bbox)
+        if len(tasks_info) == 0:
+            raise ValueError("No 'Bitmask' or 'Bbox' primitive found in dataset tasks for COCO conversion")
+        if len(tasks_info) > 1:
+            task_names = [task.name for task in tasks_info]
+            raise ValueError(
+                f"Found multiple tasks {task_names} for 'Bitmask'/'Bbox' primitive in dataset."
+                " Please specify 'task_name'."
+            )
+        task_info = tasks_info[0]
+
+    categories_list_dict = [
+        {"id": i, "name": c, "supercategory": "NotDefined"} for i, c in enumerate(task_info.class_names or [])
+    ]
+    category_mapping = {cat["name"]: cat["id"] for cat in categories_list_dict}
+
+    split_names = samples_modified_all[SampleField.SPLIT].unique().to_list()
+
+    list_split_paths = []
+    for split_name in split_names:
+        if coco_format_type == "roboflow":
+            path_split = path_output / HAFNIA_TO_ROBOFLOW_SPLIT_NAME[split_name]
+            split_paths = format_coco.CocoSplitPaths(
+                split=split_name,
+                path_images=path_split,
+                path_instances_json=path_split / ROBOFLOW_ANNOTATION_FILE_NAME,
+            )
+        else:
+            raise ValueError(f"The specified '{coco_format_type=}' is not supported.")
+        samples_in_split = samples_modified_all.filter(pl.col(SampleField.SPLIT) == split_name)
+        images_table, annotation_table = _convert_bbox_bitmask_to_coco_format(
+            samples_modified=samples_in_split,
+            license_mapping=license_mapping,
+            task_info=task_info,
+            category_mapping=category_mapping,  # type: ignore[arg-type]
+        )
+
+        split_paths.path_images.mkdir(parents=True, exist_ok=True)
+        src_paths = images_table[COCO_KEY_FILE_NAME].to_list()
+        new_relative_image_path = []
+        for src_path in src_paths:
+            dst_path = split_paths.path_images / Path(src_path).name
+            new_relative_image_path.append(dst_path.relative_to(split_paths.path_images).as_posix())
+            if dst_path.exists():
+                continue
+
+            shutil.copy2(src_path, dst_path)
+
+        images_table_files_moved = images_table.with_columns(
+            pl.Series(new_relative_image_path).alias(COCO_KEY_FILE_NAME)
+        )
+        split_labels = {
+            "info": dataset.info.model_dump(mode="json"),
+            "images": list(images_table_files_moved.iter_rows(named=True)),
+            "categories": categories_list_dict,
+            "annotations": list(annotation_table.iter_rows(named=True)),
+        }
+        if license_table is not None:
+            split_labels["licenses"] = list(license_table.iter_rows(named=True))
+        split_paths.path_instances_json.parent.mkdir(parents=True, exist_ok=True)
+        split_paths.path_instances_json.write_text(json.dumps(split_labels))
+
+        list_split_paths.append(split_paths)
+
+    return list_split_paths
+
+
+def _convert_bbox_bitmask_to_coco_format(
+    samples_modified: pl.DataFrame,
+    license_mapping: Optional[Dict[str, int]],
+    task_info: TaskInfo,
+    category_mapping: Dict[str, int],
+) -> Tuple[pl.DataFrame, pl.DataFrame]:
+    if task_info.primitive not in [Bbox, Bitmask]:
+        raise ValueError(f"Unsupported primitive '{task_info.primitive}' for COCO conversion")
+
+    task_sample_field = task_info.primitive.column_name()
+    select_image_table_columns = [
+        pl.col("id"),
+        pl.col(SampleField.WIDTH).alias("width"),
+        pl.col(SampleField.HEIGHT).alias("height"),
+        pl.col(SampleField.FILE_PATH).alias(COCO_KEY_FILE_NAME),
+    ]
+
+    if license_mapping is not None:
+        samples_modified = samples_modified.with_columns(pl.col("licenses").list.first().struct.unnest())
+        select_image_table_columns = select_image_table_columns + [
+            pl.col("name").replace_strict(license_mapping, return_dtype=pl.Int64).alias("license"),
+            pl.col("source_url").alias("flickr_url"),
+            pl.col("source_url").alias("coco_url"),
+            pl.col("date_captured"),
+        ]
+
+    images_table = samples_modified.select(select_image_table_columns)
+
+    annotation_table_full = (
+        samples_modified.select(
+            pl.col("id").alias("image_id"),
+            pl.col(SampleField.HEIGHT).alias("image_height"),
+            pl.col(SampleField.WIDTH).alias("image_width"),
+            pl.col(task_sample_field),
+        )
+        .explode(task_sample_field)
+        .with_row_index("id")
+        .unnest(task_sample_field)
+    )
+
+    iscrowd_list = [0 if row is None else row.get("iscrowd", 0) for row in annotation_table_full["meta"]]
+    annotation_table_full = annotation_table_full.with_columns(pl.Series(iscrowd_list).alias("iscrowd"))
+
+    if task_info.primitive == Bitmask:
+        annotation_table = annotation_table_full.select(
+            pl.col("id"),
+            pl.col("image_id"),
+            category_id=pl.col("class_name").replace_strict(category_mapping, return_dtype=pl.Int64),
+            segmentation=pl.struct(
+                counts=pl.col("rle_string"),
+                size=pl.concat_arr(
+                    pl.col("image_height"),
+                    pl.col("image_width"),
+                ),
+            ),
+            area=pl.col("area") * pl.col("image_height") * pl.col("image_width"),
+            bbox=pl.concat_arr(
+                pl.col("left"),  # bbox x coordinate
+                pl.col("top"),  # bbox y coordinate
+                pl.col("width"),  # bbox width
+                pl.col("height"),  # bbox height
+            ),
+            iscrowd=pl.col("iscrowd"),
+        )
+
+    elif task_info.primitive == Bbox:
+        annotation_table = annotation_table_full.select(
+            pl.col("id"),
+            pl.col("image_id"),
+            category_id=pl.col("class_name").replace_strict(category_mapping, return_dtype=pl.Int64),
+            segmentation=pl.lit([]),
+            area=pl.col("height") * pl.col("width") * pl.col("image_height") * pl.col("image_width"),
+            bbox=pl.concat_arr(
+                pl.col("top_left_x") * pl.col("image_width"),  # x coordinate
+                pl.col("top_left_y") * pl.col("image_height"),  # y coordinate
+                pl.col("width") * pl.col("image_width"),  # width
+                pl.col("height") * pl.col("image_height"),  # height
+            ),
+            iscrowd=pl.col("iscrowd"),
+        )
+
+    return images_table, annotation_table

hafnia/dataset/format_conversions/format_helpers.py

@@ -0,0 +1,33 @@
+from dataclasses import dataclass
+from pathlib import Path
+from typing import List
+
+from hafnia.dataset.dataset_names import SplitName
+from hafnia.log import user_logger
+
+
+@dataclass
+class SplitNameAndPath:
+    name: str
+    path: Path
+
+    def check(self) -> None:
+        if not self.path.is_dir():
+            raise ValueError(f"Path '{self.path}' is not a valid directory.")
+
+        if self.name not in SplitName.valid_splits():
+            raise ValueError(f"Split name '{self.name}' is not a valid split name.")
+
+
+def get_splits_from_folder(path_folder: Path) -> List[SplitNameAndPath]:
+    split_name_and_paths = []
+    for path_sub_folder in path_folder.iterdir():
+        if not path_sub_folder.is_dir():
+            continue
+        folder_split_name = path_sub_folder.name
+        split_name = SplitName.map_split_name(folder_split_name, strict=False)
+        if split_name == SplitName.UNDEFINED:
+            user_logger.warning(f"Skipping sub-folder with name '{folder_split_name}'")
+            continue
+        split_name_and_paths.append(SplitNameAndPath(name=split_name, path=path_sub_folder))
+    return split_name_and_paths

hafnia/dataset/format_conversions/format_image_classification_folder.py

@@ -5,29 +5,79 @@ from typing import TYPE_CHECKING, List, Optional
 import more_itertools
 import polars as pl
 from PIL import Image
-from rich.progress import track
 
 from hafnia.dataset.dataset_names import PrimitiveField, SampleField
+from hafnia.dataset.format_conversions.format_helpers import SplitNameAndPath, get_splits_from_folder
+from hafnia.dataset.hafnia_dataset_types import DatasetInfo, Sample, TaskInfo
 from hafnia.dataset.primitives import Classification
-from hafnia.utils import is_image_file
+from hafnia.utils import is_image_file, progress_bar
 
-if TYPE_CHECKING:
+if TYPE_CHECKING:  # Using 'TYPE_CHECKING' to avoid circular imports during type checking
     from hafnia.dataset.hafnia_dataset import HafniaDataset
 
 
+DEFAULT_DATASET_NAME = "ImageClassificationDataset"
+
+
 def from_image_classification_folder(
+    path_folder: Path,
+    n_samples: Optional[int] = None,
+    dataset_name: str = DEFAULT_DATASET_NAME,
+) -> "HafniaDataset":
+    list_split_name_and_path = get_splits_from_folder(path_folder)
+
+    dataset = from_image_classification_folder_by_split_paths(
+        n_samples=n_samples,
+        dataset_name=dataset_name,
+        list_split_paths=list_split_name_and_path,
+    )
+    return dataset
+
+
+def from_image_classification_folder_by_split_paths(
+    list_split_paths: List[SplitNameAndPath],
+    dataset_name: str = DEFAULT_DATASET_NAME,
+    n_samples: Optional[int] = None,
+) -> "HafniaDataset":
+    from hafnia.dataset.hafnia_dataset import HafniaDataset
+
+    class_names = sorted(more_itertools.collapse([class_names_from_folder(split.path) for split in list_split_paths]))
+
+    if n_samples is not None:
+        n_samples = n_samples // len(list_split_paths)  # Divide samples evenly across splits
+    datasets_per_split = []
+    for split in list_split_paths:
+        dataset_split = from_image_classification_split_folder(
+            path_folder=split.path,
+            split=split.name,
+            n_samples=n_samples,
+            class_names=class_names,
+            dataset_name=dataset_name,
+        )
+
+        datasets_per_split.append(dataset_split)
+
+    dataset = HafniaDataset.from_merger(datasets=datasets_per_split)
+    dataset.info.dataset_name = dataset_name
+    return dataset
+
+
+def from_image_classification_split_folder(
     path_folder: Path,
     split: str,
+    dataset_name: str,
     n_samples: Optional[int] = None,
+    class_names: Optional[List[str]] = None,
 ) -> "HafniaDataset":
-    from hafnia.dataset.hafnia_dataset import DatasetInfo, HafniaDataset, Sample, TaskInfo
+    from hafnia.dataset.hafnia_dataset import HafniaDataset
 
-    class_folder_paths = [path for path in path_folder.iterdir() if path.is_dir()]
-    class_names = sorted([folder.name for folder in class_folder_paths])  # Sort for determinism
+    if class_names is None:
+        class_names = class_names_from_folder(path_folder)
 
     # Gather all image paths per class
     path_images_per_class: List[List[Path]] = []
-    for path_class_folder in class_folder_paths:
+    for class_name in class_names:
+        path_class_folder = path_folder / class_name
         per_class_images = []
         for path_image in list(path_class_folder.rglob("*.*")):
             if is_image_file(path_image):
@@ -41,7 +91,9 @@ def from_image_classification_folder(
         path_images = path_images[:n_samples]
 
     samples = []
-    for path_image_org in track(path_images, description="Convert 'image classification' dataset to Hafnia Dataset"):
+    for path_image_org in progress_bar(
+        path_images, description="Convert 'image classification' dataset to Hafnia Dataset"
+    ):
         class_name = path_image_org.parent.name
 
         read_image = Image.open(path_image_org)
@@ -58,7 +110,7 @@ def from_image_classification_folder(
         samples.append(sample)
 
     dataset_info = DatasetInfo(
-        dataset_name="ImageClassificationFromDirectoryTree",
+        dataset_name=dataset_name,
         tasks=[TaskInfo(primitive=Classification, class_names=class_names)],
     )
 
@@ -71,6 +123,29 @@ def to_image_classification_folder(
     path_output: Path,
     task_name: Optional[str] = None,
     clean_folder: bool = False,
+) -> List[Path]:
+    task = dataset.info.get_task_by_task_name_and_primitive(task_name=task_name, primitive=Classification)
+
+    split_names = dataset.samples[SampleField.SPLIT].unique().to_list()
+    split_paths = []
+    for split_name in split_names:
+        dataset_split = dataset.create_split_dataset(split_name)
+        split_path = to_image_classification_split_folder(
+            dataset=dataset_split,
+            path_output_split=path_output / split_name,
+            task_name=task.name,
+            clean_folder=clean_folder,
+        )
+        split_paths.append(split_path)
+
+    return split_paths
+
+
+def to_image_classification_split_folder(
+    dataset: "HafniaDataset",
+    path_output_split: Path,
+    task_name: Optional[str] = None,
+    clean_folder: bool = False,
 ) -> Path:
     task = dataset.info.get_task_by_task_name_and_primitive(task_name=task_name, primitive=Classification)
 
@@ -90,21 +165,27 @@ def to_image_classification_folder(
         raise ValueError(f"Some samples have multiple classifications for task '{task.name}'.")
 
     if clean_folder:
-        shutil.rmtree(path_output, ignore_errors=True)
-    path_output.mkdir(parents=True, exist_ok=True)
+        shutil.rmtree(path_output_split, ignore_errors=True)
+    path_output_split.mkdir(parents=True, exist_ok=True)
 
     description = "Export Hafnia Dataset to directory tree"
-    for sample_dict in track(samples.iter_rows(named=True), total=len(samples), description=description):
+    for sample_dict in progress_bar(samples.iter_rows(named=True), total=len(samples), description=description):
         classifications = sample_dict[task.primitive.column_name()]
         if len(classifications) != 1:
             raise ValueError("Each sample should have exactly one classification.")
         classification = classifications[0]
         class_name = classification[PrimitiveField.CLASS_NAME].replace("/", "_")  # Avoid issues with subfolders
-        path_class_folder = path_output / class_name
+        path_class_folder = path_output_split / class_name
         path_class_folder.mkdir(parents=True, exist_ok=True)
 
         path_image_org = Path(sample_dict[SampleField.FILE_PATH])
         path_image_new = path_class_folder / path_image_org.name
         shutil.copy2(path_image_org, path_image_new)
 
-    return path_output
+    return path_output_split
+
+
+def class_names_from_folder(path_folder: Path) -> List[str]:
+    class_folder_paths = [path for path in path_folder.iterdir() if path.is_dir()]
+    class_names = sorted([folder.name for folder in class_folder_paths])  # Sort for determinism
+    return class_names