Reference for ultralytics/data/utils.py#

def img2label_paths(img_paths: list[str], label_dir: str = "labels", suffix: str = ".txt") -> list[str]:
    """Convert image paths to label paths by replacing 'images' with 'labels' and extension with '.txt'."""
    sa, sb = f"{os.sep}images{os.sep}", f"{os.sep}{label_dir}{os.sep}"  # /images/, /labels/ substrings
    return [sb.join(x.rsplit(sa, 1)).rsplit(".", 1)[0] + f"{suffix}" for x in img_paths]

def check_file_speeds(
    files: list[str], threshold_ms: float = 10, threshold_mb: float = 50, max_files: int = 5, prefix: str = ""
):
    """Check dataset file access speed and provide performance feedback.

    This function tests the access speed of dataset files by measuring ping (stat call) time and read speed. It samples
    up to `max_files` files from the provided list and warns if access times exceed the threshold.

    Args:
        files (list[str]): List of file paths to check for access speed.
        threshold_ms (float, optional): Threshold in milliseconds for ping time warnings.
        threshold_mb (float, optional): Threshold in megabytes per second for read speed warnings.
        max_files (int, optional): The maximum number of files to check.
        prefix (str, optional): Prefix string to add to log messages.

    Examples:
        >>> from pathlib import Path
        >>> image_files = list(Path("dataset/images").glob("*.jpg"))
        >>> check_file_speeds(image_files, threshold_ms=15)
    """
    if not files:
        LOGGER.warning(f"{prefix}Image speed checks: No files to check")
        return

    # Sample files (max 5)
    files = random.sample(files, min(max_files, len(files)))

    # Test ping (stat time)
    ping_times = []
    file_sizes = []
    read_speeds = []

    for f in files:
        try:
            # Measure ping (stat call)
            start = time.perf_counter()
            file_size = os.stat(f).st_size
            ping_times.append((time.perf_counter() - start) * 1000)  # ms
            file_sizes.append(file_size)

            # Measure read speed
            start = time.perf_counter()
            with open(f, "rb") as file_obj:
                _ = file_obj.read()
            read_time = time.perf_counter() - start
            if read_time > 0:  # Avoid division by zero
                read_speeds.append(file_size / (1 << 20) / read_time)  # MB/s
        except Exception:
            pass

    if not ping_times:
        LOGGER.warning(f"{prefix}Image speed checks: failed to access files")
        return

    # Calculate stats with uncertainties
    avg_ping = np.mean(ping_times)
    std_ping = np.std(ping_times, ddof=1) if len(ping_times) > 1 else 0
    size_msg = f", size: {np.mean(file_sizes) / (1 << 10):.1f} KB"
    ping_msg = f"ping: {avg_ping:.1f}±{std_ping:.1f} ms"

    if read_speeds:
        avg_speed = np.mean(read_speeds)
        std_speed = np.std(read_speeds, ddof=1) if len(read_speeds) > 1 else 0
        speed_msg = f", read: {avg_speed:.1f}±{std_speed:.1f} MB/s"
    else:
        avg_speed = float("inf")
        speed_msg = ""

    if avg_ping < threshold_ms and avg_speed > threshold_mb:
        LOGGER.info(f"{prefix}Fast image access ✅ ({ping_msg}{speed_msg}{size_msg})")
    else:
        LOGGER.warning(
            f"{prefix}Slow image access detected ({ping_msg}{speed_msg}{size_msg}). "
            f"Use local storage instead of remote/mounted storage for better performance. "
            f"See https://docs.ultralytics.com/guides/model-training-tips/"
        )

def get_hash(paths: list[str]) -> str:
    """Return a single hash value of a list of paths (files or dirs)."""
    size = 0
    for p in paths:
        try:
            size += os.stat(p).st_size
        except OSError:
            continue
    h = __import__("hashlib").sha256(str(size).encode())  # hash sizes
    h.update("".join(paths).encode())  # hash paths
    return h.hexdigest()  # return hash

def exif_size(img: Image.Image) -> tuple[int, int]:
    """Return exif-corrected PIL size."""
    s = img.size  # (width, height)
    if img.format == "JPEG":  # only support JPEG images
        try:
            if exif := img.getexif():
                rotation = exif.get(274, None)  # the EXIF key for the orientation tag is 274
                if rotation in {6, 8}:  # rotation 270 or 90
                    s = s[1], s[0]
        except Exception:
            pass
    return s

def check_image(im_file: str) -> tuple[str, tuple[int, int]]:
    """Verify an image file for integrity and correct corrupt JPEGs if found.

    Args:
        im_file (str): Path to the image file to check.

    Returns:
        (str): A message describing any corrective action taken, or an empty string if the image is valid.
        (tuple[int, int]): Image shape as (height, width) in pixels.

    Raises:
        AssertionError: If the image size is less than 10 pixels in any dimension or the format is invalid.
    """
    msg = ""
    im = Image.open(im_file)
    im.verify()  # PIL verify
    shape = exif_size(im)  # image size
    shape = (shape[1], shape[0])  # hw
    assert (shape[0] > 9) & (shape[1] > 9), f"image size {shape} <10 pixels"
    assert im.format.lower() in IMG_FORMATS, f"Invalid image format {im.format}. {FORMATS_HELP_MSG}"
    if im.format.lower() in {"jpg", "jpeg"}:
        with open(im_file, "rb") as f:
            f.seek(-2, 2)
            if f.read() != b"\xff\xd9":  # corrupt JPEG
                ImageOps.exif_transpose(Image.open(im_file)).save(im_file, "JPEG", subsampling=0, quality=100)
                msg = f"{im_file}: corrupt JPEG restored and saved"
    return msg, shape

def verify_image(args: tuple) -> tuple:
    """Verify one image."""
    (im_file, cls), prefix = args
    # Number (found, corrupt), message
    nf, nc, msg = 0, 0, ""
    try:
        msg = check_image(im_file)[0]
        msg = f"{prefix}{msg}" if msg else ""
        nf = 1
    except Exception as e:
        nc = 1
        msg = f"{prefix}{im_file}: ignoring corrupt image/label: {e}"
    return (im_file, cls), nf, nc, msg

def verify_image_depth(args: tuple) -> tuple:
    """Verify that an image and its paired depth .npy map exist and are readable."""
    im_file, depth_file, prefix = args
    # Number (found, missing, corrupt), message
    nf, nm, nc, msg = 0, 0, 0, ""
    try:
        msg, shape = check_image(im_file)
        msg = f"{prefix}{msg}" if msg else ""
        if not os.path.isfile(depth_file):
            nm = 1
            msg = f"{prefix}{im_file}: ignoring image with missing depth map {depth_file}"
            return None, None, nf, nm, nc, msg
        depth = np.load(depth_file, mmap_mode="r", allow_pickle=False)
        assert depth.ndim == 2, f"depth map {depth_file} expected a 2D array, got shape {depth.shape}"
        nf = 1
        return im_file, shape, nf, nm, nc, msg
    except Exception as e:
        nc = 1
        msg = f"{prefix}{im_file}: ignoring corrupt image/depth: {e}"
    return None, None, nf, nm, nc, msg

def verify_image_mask(args: tuple) -> tuple:
    """Verify that an image and its semantic mask exist, are readable, and have matching shapes."""
    im_file, mask_file, prefix, check_bit_depth = args
    # Number (found, missing, corrupt), message
    nf, nm, nc, msg = 0, 0, 0, ""
    try:
        msg, shape = check_image(im_file)
        msg = f"{prefix}{msg}" if msg else ""
        if not os.path.isfile(mask_file):
            for ext in IMG_FORMATS:  # check other suffixes
                alt_mask_file = mask_file.rsplit(".", 1)[0] + f".{ext}"
                if os.path.isfile(alt_mask_file):
                    mask_file = alt_mask_file
                    break
        if os.path.isfile(mask_file):
            mask = cv2.imread(mask_file, cv2.IMREAD_GRAYSCALE)
            assert mask is not None, f"mask file {mask_file} is unreadable"
            assert mask.shape[:2] == shape, f"mask size {mask.shape[:2]} does not match image size {shape}"
            is_1bit = False
            if check_bit_depth:
                with Image.open(mask_file) as im:
                    is_1bit = im.mode == "1"
            nf = 1
        else:
            nm = 1
            msg = f"{prefix}{im_file}: ignoring image with missing mask {mask_file}"
            return None, None, None, None, nm, nf, nc, msg
        return im_file, mask_file, shape, is_1bit, nm, nf, nc, msg
    except Exception as e:
        nc = 1
        msg = f"{prefix}{im_file}: ignoring corrupt image/mask: {e}"
    return None, None, None, None, nm, nf, nc, msg

def verify_image_label(args: tuple) -> list:
    """Verify one image-label pair."""
    im_file, lb_file, prefix, keypoint, num_cls, nkpt, ndim, single_cls = args
    # Number (missing, found, empty, corrupt), message, segments, keypoints
    nm, nf, ne, nc, msg, segments, keypoints = 0, 0, 0, 0, "", [], None
    try:
        # Verify images
        msg, shape = check_image(im_file)
        msg = f"{prefix}{msg}" if msg else ""

        # Verify labels
        if os.path.isfile(lb_file):
            nf = 1  # label found
            with open(lb_file, encoding="utf-8") as f:
                lb = [x.split() for x in f.read().strip().splitlines() if len(x)]
                if any(len(x) > 6 for x in lb) and (not keypoint):  # is segment
                    assert not any(len(x) == 5 for x in lb), "labels mix segment and detection rows"
                    classes = np.array([x[0] for x in lb], dtype=np.float32)
                    segments = [np.array(x[1:], dtype=np.float32).reshape(-1, 2) for x in lb]  # (cls, xy1...)
                    lb = np.concatenate((classes.reshape(-1, 1), segments2boxes(segments)), 1)  # (cls, xywh)
                lb = np.array(lb, dtype=np.float32)
            if nl := len(lb):
                if keypoint:
                    assert lb.shape[1] == (5 + nkpt * ndim), f"labels require {(5 + nkpt * ndim)} columns each"
                    points = lb[:, 5:].reshape(-1, ndim)[:, :2]
                else:
                    assert lb.shape[1] == 5, f"labels require 5 columns, {lb.shape[1]} columns detected"
                    points = lb[:, 1:]
                # Coordinate points check with 1% tolerance
                assert points.max() <= 1.01, f"non-normalized or out of bounds coordinates {points[points > 1.01]}"
                assert lb.min() >= -0.01, f"negative class labels or coordinate {lb[lb < -0.01]}"

                # All labels
                max_cls = 0 if single_cls else lb[:, 0].max()  # max label count
                assert max_cls < num_cls, (
                    f"Label class {int(max_cls)} exceeds dataset class count {num_cls}. "
                    f"Possible class labels are 0-{num_cls - 1}"
                )
                _, i = np.unique(lb, axis=0, return_index=True)
                if len(i) < nl:  # duplicate row check
                    lb = lb[i]  # remove duplicates
                    if segments:
                        segments = [segments[x] for x in i]
                    msg = f"{prefix}{im_file}: {nl - len(i)} duplicate labels removed"
            else:
                ne = 1  # label empty
                lb = np.zeros((0, (5 + nkpt * ndim) if keypoint else 5), dtype=np.float32)
        else:
            nm = 1  # label missing
            lb = np.zeros((0, (5 + nkpt * ndim) if keypoint else 5), dtype=np.float32)
        if keypoint:
            keypoints = lb[:, 5:].reshape(-1, nkpt, ndim)
            if ndim == 2:
                kpt_mask = np.where((keypoints[..., 0] < 0) | (keypoints[..., 1] < 0), 0.0, 1.0).astype(np.float32)
                keypoints = np.concatenate([keypoints, kpt_mask[..., None]], axis=-1)  # (nl, nkpt, 3)
        lb = lb[:, :5]
        return im_file, lb, shape, segments, keypoints, nm, nf, ne, nc, msg
    except Exception as e:
        nc = 1
        msg = f"{prefix}{im_file}: ignoring corrupt image/label: {e}"
        return [None, None, None, None, None, nm, nf, ne, nc, msg]

def visualize_image_annotations(image_path: str, txt_path: str, label_map: dict[int, str]):
    """Visualize YOLO annotations (bounding boxes and class labels) on an image.

    This function reads an image and its corresponding annotation file in YOLO format, then draws bounding boxes around
    detected objects and labels them with their respective class names. The bounding box colors are assigned based on
    the class ID, and the text color is dynamically adjusted for readability, depending on the background color's
    luminance.

    Args:
        image_path (str): Path to the image file to annotate. The file must be readable by PIL.
        txt_path (str): Path to the annotation file in YOLO format, which should contain one line per object.
        label_map (dict[int, str]): A dictionary that maps class IDs (integers) to class labels (strings).

    Examples:
        >>> label_map = {0: "cat", 1: "dog", 2: "bird"}  # Should include all annotated classes
        >>> visualize_image_annotations("path/to/image.jpg", "path/to/annotations.txt", label_map)
    """
    import matplotlib.pyplot as plt

    from ultralytics.utils.plotting import colors

    img = np.array(Image.open(image_path))
    img_height, img_width = img.shape[:2]
    annotations = []
    with open(txt_path, encoding="utf-8") as file:
        for line in file:
            class_id, x_center, y_center, width, height = map(float, line.split())
            x = (x_center - width / 2) * img_width
            y = (y_center - height / 2) * img_height
            w = width * img_width
            h = height * img_height
            annotations.append((x, y, w, h, int(class_id)))
    _, ax = plt.subplots(1)  # Plot the image and annotations
    for x, y, w, h, label in annotations:
        color = tuple(c / 255 for c in colors(label, False))  # Get and normalize an RGB color for Matplotlib
        rect = plt.Rectangle((x, y), w, h, linewidth=2, edgecolor=color, facecolor="none")  # Create a rectangle
        ax.add_patch(rect)
        luminance = 0.2126 * color[0] + 0.7152 * color[1] + 0.0722 * color[2]  # Formula for luminance
        ax.text(x, y - 5, label_map[label], color="white" if luminance < 0.5 else "black", backgroundcolor=color)
    ax.imshow(img)
    plt.show()

def polygon2mask(
    imgsz: tuple[int, int], polygons: list[np.ndarray], color: int = 1, downsample_ratio: int = 1
) -> np.ndarray:
    """Convert a list of polygons to a binary mask of the specified image size.

    Args:
        imgsz (tuple[int, int]): The size of the image as (height, width).
        polygons (list[np.ndarray]): A list of polygons. Each polygon is a 1D array of coordinates with length M, where
            M % 2 = 0 (alternating x, y values).
        color (int, optional): The color value to fill in the polygons on the mask.
        downsample_ratio (int, optional): Factor by which to downsample the mask.

    Returns:
        (np.ndarray): A binary mask of the specified image size with the polygons filled in.
    """
    mask = np.zeros(imgsz, dtype=np.uint8)
    polygons = np.asarray(polygons, dtype=np.int32)
    polygons = polygons.reshape((polygons.shape[0], -1, 2))
    cv2.fillPoly(mask, polygons, color=color)
    nh, nw = (imgsz[0] // downsample_ratio, imgsz[1] // downsample_ratio)
    # Note: fillPoly first then resize is trying to keep the same loss calculation method when mask-ratio=1
    return cv2.resize(mask, (nw, nh))

def polygons2masks(
    imgsz: tuple[int, int], polygons: list[np.ndarray], color: int, downsample_ratio: int = 1
) -> np.ndarray:
    """Convert a list of polygons to a set of binary masks of the specified image size.

    Args:
        imgsz (tuple[int, int]): The size of the image as (height, width).
        polygons (list[np.ndarray]): A list of polygons. Each polygon is an array of coordinates that can be reshaped to
            (-1, 2) as (x, y) point pairs.
        color (int): The color value to fill in the polygons on the masks.
        downsample_ratio (int, optional): Factor by which to downsample each mask.

    Returns:
        (np.ndarray): A set of binary masks of the specified image size with the polygons filled in.
    """
    return np.array([polygon2mask(imgsz, [x.reshape(-1)], color, downsample_ratio) for x in polygons])

def polygons2masks_overlap(
    imgsz: tuple[int, int], segments: list[np.ndarray], downsample_ratio: int = 1
) -> tuple[np.ndarray, np.ndarray]:
    """Return a downsampled overlap mask and sorted area indices."""
    masks = np.zeros(
        (imgsz[0] // downsample_ratio, imgsz[1] // downsample_ratio),
        dtype=np.int32 if len(segments) > 255 else np.uint8,
    )
    areas = []
    ms = []
    for segment in segments:
        mask = polygon2mask(
            imgsz,
            [segment.reshape(-1)],
            downsample_ratio=downsample_ratio,
            color=1,
        )
        ms.append(mask.astype(masks.dtype))
        areas.append(mask.sum())
    areas = np.asarray(areas)
    index = np.argsort(-areas)
    ms = np.array(ms)[index]
    # Running max: the old `masks + mask` sum hit 2 * i + 1 and overflowed uint8 past 128 overlapping instances
    for i in range(len(segments)):
        np.maximum(masks, ms[i] * (i + 1), out=masks)
    return masks, index

def find_dataset_yaml(path: Path) -> Path:
    """Find and return the YAML file associated with a Detect, Segment or Pose dataset.

    This function searches for a YAML file at the root level of the provided directory first, and if not found, it
    performs a recursive search. It prefers YAML files that have the same stem as the provided path.

    Args:
        path (Path): The directory path to search for the YAML file.

    Returns:
        (Path): The path of the found YAML file.
    """
    files = list(path.glob("*.yaml")) or list(path.rglob("*.yaml"))  # try root level first and then recursive
    assert files, f"No YAML file found in '{path.resolve()}'"
    if len(files) > 1:
        files = [f for f in files if f.stem == path.stem]  # prefer *.yaml files that match
    assert len(files) == 1, f"Expected 1 YAML file in '{path.resolve()}', but found {len(files)}.\n{files}"
    return files[0]

def convert_ndjson_to_yolo_if_needed(data: str | Path) -> str | Path:
    """Convert an NDJSON dataset or Platform dataset URI to YOLO format."""
    data = normalize_platform_uri(data)  # accept Platform web URLs (https://platform.ultralytics.com/.../datasets/...)
    data_str = str(data)
    if clean_url(data_str).endswith(".ndjson") or (data_str.startswith("ul://") and "/datasets/" in data_str):
        import asyncio

        from ultralytics.data.converter import convert_ndjson_to_yolo

        return asyncio.run(convert_ndjson_to_yolo(data))
    return data

def check_det_dataset(dataset: str, autodownload: bool = True, split: str = "") -> dict[str, Any]:
    """Download, verify, and/or unzip a dataset if not found locally.

    This function checks the availability of a specified dataset, and if not found, it has the option to download and
    unzip the dataset. It then reads and parses the accompanying YAML data, ensuring key requirements are met and also
    resolves paths related to the dataset.

    Args:
        dataset (str): Path to the dataset or dataset descriptor (like a YAML file).
        autodownload (bool, optional): Whether to automatically download the dataset if not found.
        split (str, optional): Dataset split required by the caller.

    Returns:
        (dict[str, Any]): Parsed dataset information and paths.
    """
    dataset = str(dataset)
    if "://" not in dataset and not Path(dataset).exists() and Path(dataset).suffix not in {".yaml", ".yml"}:
        # allow bare dataset names, e.g. 'coco8' -> 'coco8.yaml', 'DOTAv1.5' -> 'DOTAv1.5.yaml'
        dataset = next((f"{dataset}{x}" for x in (".yaml", ".yml") if check_file(f"{dataset}{x}", hard=False)), dataset)
    file = Path(check_file(dataset))
    if file.is_dir():
        file = find_dataset_yaml(file)

    # Download (optional)
    extract_dir = ""
    if zipfile.is_zipfile(file) or is_tarfile(file):
        new_dir = safe_download(file, dir=DATASETS_DIR, unzip=True, delete=False)
        file = find_dataset_yaml(DATASETS_DIR / new_dir)
        extract_dir, autodownload = file.parent, False

    # Read YAML
    data = YAML.load(file, append_filename=True)  # dictionary

    # Checks
    for k in "train", "val":
        if k not in data:
            if k != "val" or "validation" not in data:
                raise SyntaxError(
                    emojis(f"{dataset} '{k}:' key missing ❌.\n'train' and 'val' are required in all data YAMLs.")
                )
            LOGGER.warning("renaming data YAML 'validation' key to 'val' to match YOLO format.")
            data["val"] = data.pop("validation")  # replace 'validation' key with 'val' key
    if split and not data.get(split):
        raise FileNotFoundError(f"{dataset} '{split}:' images not found ❌")
    # `names` compared to None, not membership: a bare `names:` parses to None and len(None) below
    # raises. `nc` stays membership so a valueless `nc:` still reaches its "must be an integer" error.
    if data.get("names") is None and "nc" not in data:
        raise SyntaxError(emojis(f"{dataset} key missing ❌.\n either 'names' or 'nc' are required in all data YAMLs."))
    if "nc" in data and not isinstance(data["nc"], int):
        try:
            nc = float(data["nc"])  # accept integer-like values, e.g. '10' or 10.0, but not 1.9 or placeholders
            if nc != int(nc):
                raise ValueError
            data["nc"] = int(nc)
        except (TypeError, ValueError):
            raise SyntaxError(emojis(f"{dataset} 'nc: {data['nc']}' must be an integer ❌."))
    if data.get("names") is not None and data.get("nc") is not None and len(data["names"]) != data["nc"]:
        raise SyntaxError(emojis(f"{dataset} 'names' length {len(data['names'])} and 'nc: {data['nc']}' must match."))
    if data.get("names") is None:
        data["names"] = [f"class_{i}" for i in range(data["nc"])]
    else:
        data["nc"] = len(data["names"])

    data["names"] = check_class_names(data["names"])
    data["channels"] = data.get("channels", 3)  # get image channels, default to 3

    # Resolve paths
    path = Path(extract_dir or data.get("path") or Path(data.get("yaml_file", "")).parent)  # dataset root
    if not path.exists() and not path.is_absolute():
        path = (DATASETS_DIR / path).resolve()  # path relative to DATASETS_DIR

    # Set paths
    data["path"] = path  # download scripts
    for k in "train", "val", "test", "minival":
        if data.get(k):  # prepend path
            if isinstance(data[k], str):
                x = (path / data[k]).resolve()
                if not x.exists() and data[k].startswith("../"):
                    x = (path / data[k][3:]).resolve()
                data[k] = str(x)
            else:
                data[k] = [str((path / x).resolve()) for x in data[k]]

    # Parse YAML
    val, s = (data.get(x) for x in (split or "val", "download"))
    if val:
        val = [Path(x).resolve() for x in (val if isinstance(val, list) else [val])]  # val path
        if not all(x.exists() for x in val):
            name = clean_url(dataset)  # dataset name with URL auth stripped
            LOGGER.info("")
            m = f"Dataset '{name}' images not found, missing path '{next(x for x in val if not x.exists())}'"
            if s and autodownload:
                LOGGER.warning(m)
            else:
                m += f"\nNote dataset download directory is '{DATASETS_DIR}'. You can update this in '{SETTINGS_FILE}'"
                raise FileNotFoundError(m)
            t = time.time()
            r = None  # success
            if s.startswith("http") and s.endswith(".zip"):  # URL
                safe_download(url=s, dir=DATASETS_DIR, delete=True)
            elif s.startswith("bash "):  # bash script
                LOGGER.info(f"Running {s} ...")
                subprocess.run(s.split(), check=True)
            else:  # python script
                exec(s, {"yaml": data})  # noqa: S102
            dt = f"({round(time.time() - t, 1)}s)"
            s = f"success ✅ {dt}, saved to {colorstr('bold', DATASETS_DIR)}" if r in {0, None} else f"failure {dt} ❌"
            LOGGER.info(f"Dataset download {s}\n")
    check_font("Arial.ttf" if is_ascii(data["names"]) else "Arial.Unicode.ttf")  # download fonts

    return data  # dictionary

def check_cls_dataset(dataset: str | Path, split: str = "") -> dict[str, Any]:
    """Check a classification dataset such as Imagenet.

    This function accepts a `dataset` name and attempts to retrieve the corresponding dataset information. If the
    dataset is not found locally, it attempts to download the dataset from the internet and save it locally.

    Args:
        dataset (str | Path): The name of the dataset.
        split (str, optional): The split of the dataset. Either 'val', 'test', or ''.

    Returns:
        (dict[str, Any]): A dictionary containing the following keys:

            - 'train' (Path): The directory path containing the training set of the dataset.
            - 'val' (Path): The directory path containing the validation set of the dataset.
            - 'test' (Path): The directory path containing the test set of the dataset.
            - 'nc' (int): The number of classes in the dataset.
            - 'names' (dict[int, str]): A dictionary of class names in the dataset.
    """
    if split and split not in {"train", "val", "test"}:
        raise ValueError(f"Invalid classification dataset split '{split}'. Use 'train', 'val', or 'test'.")

    # Download (optional if dataset=https://file.zip is passed directly)
    if str(dataset).startswith(("http:/", "https:/")):
        dataset = safe_download(dataset, dir=DATASETS_DIR, unzip=True, delete=False)
    elif str(dataset).endswith((".zip", ".tar", ".gz")):
        file = check_file(dataset)
        dataset = safe_download(file, dir=DATASETS_DIR, unzip=True, delete=False)

    dataset = Path(dataset)
    data_dir = (dataset if dataset.is_dir() else (DATASETS_DIR / dataset)).resolve()
    if not data_dir.is_dir():
        if data_dir.suffix != "":
            raise ValueError(
                f'Classification datasets must be a directory (data="path/to/dir") not a file (data="{dataset}"), '
                "See https://docs.ultralytics.com/datasets/classify/"
            )
        LOGGER.info("")
        LOGGER.warning(f"Dataset not found, missing path {data_dir}, attempting download...")
        t = time.time()
        if str(dataset) == "imagenet":
            subprocess.run(["bash", str(ROOT / "data/scripts/get_imagenet.sh")], check=True)
        else:
            download(f"{ASSETS_URL}/{dataset}.zip", dir=data_dir.parent)
        LOGGER.info(f"Dataset download success ✅ ({time.time() - t:.1f}s), saved to {colorstr('bold', data_dir)}\n")
    train_set = data_dir / "train"
    if not train_set.is_dir():
        LOGGER.warning(f"Dataset 'split=train' not found at {train_set}")
        if image_files := [f for f in data_dir.rglob("*.*") if f.suffix[1:].lower() in IMG_FORMATS]:
            from ultralytics.data.split import split_classify_dataset

            LOGGER.info(f"Found {len(image_files)} images in subdirectories. Attempting to split...")
            data_dir = split_classify_dataset(data_dir, train_ratio=0.8)
            train_set = data_dir / "train"
        else:
            raise FileNotFoundError(f"No images found in {data_dir} or its subdirectories.")
    val_set = (
        data_dir / "val"
        if (data_dir / "val").exists()
        else data_dir / "validation"
        if (data_dir / "validation").exists()
        else data_dir / "valid"
        if (data_dir / "valid").exists()
        else None
    )  # data/test or data/val
    test_set = data_dir / "test" if (data_dir / "test").exists() else None  # data/val or data/test
    if split == "val" and not val_set:
        LOGGER.warning("Dataset 'split=val' not found, using 'split=test' instead.")
        val_set = test_set
    elif split == "test" and not test_set:
        LOGGER.warning("Dataset 'split=test' not found, using 'split=val' instead.")
        test_set = val_set

    if (ndjson_names := data_dir / ".ndjson.yaml").is_file():
        names = YAML.load(ndjson_names)["names"]
    else:
        names = dict(enumerate(sorted(x.name for x in (data_dir / "train").iterdir() if x.is_dir())))
    nc = len(names)

    # Print to console
    for k, v in {"train": train_set, "val": val_set, "test": test_set}.items():
        prefix = f"{colorstr(f'{k}:')} {v}..."
        if v is None:
            LOGGER.info(prefix)
        else:
            files = [path for path in v.rglob("*.*") if path.suffix[1:].lower() in IMG_FORMATS]
            nf = len(files)  # number of files
            nd = len({file.parent for file in files})  # number of directories
            if nf == 0:
                if k == "train":
                    raise FileNotFoundError(f"{dataset} '{k}:' no training images found")
                else:
                    LOGGER.warning(f"{prefix} found {nf} images in {nd} classes (no images found)")
            elif nd != nc and not ndjson_names.is_file():
                LOGGER.error(f"{prefix} found {nf} images in {nd} classes (requires {nc} classes, not {nd})")
            else:
                class_count = f"{nd}/{nc}" if ndjson_names.is_file() else nd
                LOGGER.info(f"{prefix} found {nf} images in {class_count} classes ✅ ")

    return {"train": train_set, "val": val_set, "test": test_set, "nc": nc, "names": names, "channels": 3}

def compress_one_image(f: str, f_new: str | None = None, max_dim: int = 1920, quality: int = 50):
    """Compress a single image file to reduced size while preserving its aspect ratio and quality using either the
    Python Imaging Library (PIL) or OpenCV library. If the input image is smaller than the maximum dimension, it
    will not be resized.

    Args:
        f (str): The path to the input image file.
        f_new (str, optional): The path to the output image file. If not specified, the input file will be overwritten.
        max_dim (int, optional): The maximum dimension (width or height) of the output image.
        quality (int, optional): The image compression quality as a percentage.

    Examples:
        >>> from pathlib import Path
        >>> from ultralytics.data.utils import compress_one_image
        >>> for f in Path("path/to/dataset").rglob("*.jpg"):
        >>>    compress_one_image(f)
    """
    try:  # use PIL
        Image.MAX_IMAGE_PIXELS = None  # Fix DecompressionBombError, allow optimization of image > ~178.9 million pixels
        im = Image.open(f)
        if im.mode in {"RGBA", "LA"}:  # Convert to RGB if needed (for JPEG)
            im = im.convert("RGB")
        r = max_dim / max(im.height, im.width)  # ratio
        if r < 1.0:  # image too large
            im = im.resize((int(im.width * r), int(im.height * r)))
        im.save(f_new or f, "JPEG", quality=quality, optimize=True)  # save
    except Exception as e:  # use OpenCV
        LOGGER.warning(f"Image compression PIL failure {f}: {e}")
        im = cv2.imread(f)
        im_height, im_width = im.shape[:2]
        r = max_dim / max(im_height, im_width)  # ratio
        if r < 1.0:  # image too large
            im = cv2.resize(im, (int(im_width * r), int(im_height * r)), interpolation=cv2.INTER_AREA)
        cv2.imwrite(str(f_new or f), im)

def load_dataset_cache_file(path: Path) -> dict:
    """Load an Ultralytics *.cache dictionary from path."""
    import gc

    gc.disable()  # reduce pickle load time https://github.com/ultralytics/ultralytics/pull/1585
    cache = np.load(str(path), allow_pickle=True).item()  # load dict
    gc.enable()
    return cache

def save_dataset_cache_file(prefix: str, path: Path, x: dict, version: str):
    """Save an Ultralytics dataset *.cache dictionary x to path."""
    x["version"] = version  # add cache version
    if is_dir_writeable(path.parent):
        if path.exists():
            path.unlink()  # remove *.cache file if exists
        try:
            with open(str(path), "wb") as file:  # context manager here fixes windows async np.save bug
                np.save(file, x)
            LOGGER.info(f"{prefix}New cache created: {path}")
        except Exception as e:
            Path(path).unlink(missing_ok=True)  # remove partially written file
            LOGGER.warning(f"{prefix}WARNING ⚠️ Failed to save cache to {path}: {e}")
    else:
        LOGGER.warning(f"{prefix}Cache directory {path.parent} is not writable, cache not saved.")

def add_polygon_background(data: dict) -> dict:
    """Set up the background class for polygon-based semantic datasets without 'masks_dir'.

    - nc > 1: appends a 'background' class at id=nc and bumps data['nc'] to nc+1; polygon
    cls values are kept as foreground ids.
    - nc == 1: keeps nc=1 (binary segmentation). Polygon rasterization
    yields a {0=bg, 1=fg} mask regardless of the label cls value.
    """
    if data.get("masks_dir") or data.get("_polygon_bg_added"):
        return data
    nc = int(data.get("nc") or len(data.get("names") or {}))
    if nc == 1:  # binary: bg=0, fg=1 (implicit); model uses BCE on a single output channel
        data["bg_class_idx"] = 0
    else:
        names = dict(data.get("names") or {})
        names[nc] = "background"
        data["bg_class_idx"] = nc
        data["nc"] = nc + 1
        data["names"] = names
    data["_polygon_bg_added"] = True
    return data