跳至内容

参考资料 ultralytics/data/augment.py

备注

该文件可从https://github.com/ultralytics/ultralytics/blob/main/ ultralytics/data/augment .py 获取。如果您发现问题,请通过提交 Pull Request🛠️ 帮助修复。谢谢🙏!



ultralytics.data.augment.BaseTransform

图像变换基类

这是一个通用转换类,可根据特定的图像处理需求进行扩展。 该类旨在兼容分类和语义分割任务。

方法

名称 说明
__init__

初始化 BaseTransform 对象。

apply_image

对标签进行图像转换。

apply_instances

对标签中的对象实例进行转换。

apply_semantic

对图像进行语义分割。

__call__

将所有标签变换应用于图像、实例和语义掩码。

源代码 ultralytics/data/augment.py
class BaseTransform:
    """
    Base class for image transformations.

    This is a generic transformation class that can be extended for specific image processing needs.
    The class is designed to be compatible with both classification and semantic segmentation tasks.

    Methods:
        __init__: Initializes the BaseTransform object.
        apply_image: Applies image transformation to labels.
        apply_instances: Applies transformations to object instances in labels.
        apply_semantic: Applies semantic segmentation to an image.
        __call__: Applies all label transformations to an image, instances, and semantic masks.
    """

    def __init__(self) -> None:
        """Initializes the BaseTransform object."""
        pass

    def apply_image(self, labels):
        """Applies image transformations to labels."""
        pass

    def apply_instances(self, labels):
        """Applies transformations to object instances in labels."""
        pass

    def apply_semantic(self, labels):
        """Applies semantic segmentation to an image."""
        pass

    def __call__(self, labels):
        """Applies all label transformations to an image, instances, and semantic masks."""
        self.apply_image(labels)
        self.apply_instances(labels)
        self.apply_semantic(labels)

__call__(labels)

将所有标签变换应用于图像、实例和语义掩码。

源代码 ultralytics/data/augment.py
def __call__(self, labels):
    """Applies all label transformations to an image, instances, and semantic masks."""
    self.apply_image(labels)
    self.apply_instances(labels)
    self.apply_semantic(labels)

__init__()

初始化 BaseTransform 对象。

源代码 ultralytics/data/augment.py
def __init__(self) -> None:
    """Initializes the BaseTransform object."""
    pass

apply_image(labels)

对标签进行图像变换。

源代码 ultralytics/data/augment.py
def apply_image(self, labels):
    """Applies image transformations to labels."""
    pass

apply_instances(labels)

对标签中的对象实例进行转换。

源代码 ultralytics/data/augment.py
def apply_instances(self, labels):
    """Applies transformations to object instances in labels."""
    pass

apply_semantic(labels)

对图像进行语义分割。

源代码 ultralytics/data/augment.py
def apply_semantic(self, labels):
    """Applies semantic segmentation to an image."""
    pass



ultralytics.data.augment.Compose

用于合成多种图像变换的类。

源代码 ultralytics/data/augment.py
class Compose:
    """Class for composing multiple image transformations."""

    def __init__(self, transforms):
        """Initializes the Compose object with a list of transforms."""
        self.transforms = transforms if isinstance(transforms, list) else [transforms]

    def __call__(self, data):
        """Applies a series of transformations to input data."""
        for t in self.transforms:
            data = t(data)
        return data

    def append(self, transform):
        """Appends a new transform to the existing list of transforms."""
        self.transforms.append(transform)

    def insert(self, index, transform):
        """Inserts a new transform to the existing list of transforms."""
        self.transforms.insert(index, transform)

    def __getitem__(self, index: Union[list, int]) -> "Compose":
        """Retrieve a specific transform or a set of transforms using indexing."""
        assert isinstance(index, (int, list)), f"The indices should be either list or int type but got {type(index)}"
        index = [index] if isinstance(index, int) else index
        return Compose([self.transforms[i] for i in index])

    def __setitem__(self, index: Union[list, int], value: Union[list, int]) -> None:
        """Retrieve a specific transform or a set of transforms using indexing."""
        assert isinstance(index, (int, list)), f"The indices should be either list or int type but got {type(index)}"
        if isinstance(index, list):
            assert isinstance(
                value, list
            ), f"The indices should be the same type as values, but got {type(index)} and {type(value)}"
        if isinstance(index, int):
            index, value = [index], [value]
        for i, v in zip(index, value):
            assert i < len(self.transforms), f"list index {i} out of range {len(self.transforms)}."
            self.transforms[i] = v

    def tolist(self):
        """Converts the list of transforms to a standard Python list."""
        return self.transforms

    def __repr__(self):
        """Returns a string representation of the object."""
        return f"{self.__class__.__name__}({', '.join([f'{t}' for t in self.transforms])})"

__call__(data)

对输入数据进行一系列转换。

源代码 ultralytics/data/augment.py
def __call__(self, data):
    """Applies a series of transformations to input data."""
    for t in self.transforms:
        data = t(data)
    return data

__getitem__(index)

使用索引检索特定变换或一组变换。

源代码 ultralytics/data/augment.py
def __getitem__(self, index: Union[list, int]) -> "Compose":
    """Retrieve a specific transform or a set of transforms using indexing."""
    assert isinstance(index, (int, list)), f"The indices should be either list or int type but got {type(index)}"
    index = [index] if isinstance(index, int) else index
    return Compose([self.transforms[i] for i in index])

__init__(transforms)

使用变换列表初始化 Compose 对象。

源代码 ultralytics/data/augment.py
def __init__(self, transforms):
    """Initializes the Compose object with a list of transforms."""
    self.transforms = transforms if isinstance(transforms, list) else [transforms]

__repr__()

返回对象的字符串表示。

源代码 ultralytics/data/augment.py
def __repr__(self):
    """Returns a string representation of the object."""
    return f"{self.__class__.__name__}({', '.join([f'{t}' for t in self.transforms])})"

__setitem__(index, value)

使用索引检索特定变换或一组变换。

源代码 ultralytics/data/augment.py
def __setitem__(self, index: Union[list, int], value: Union[list, int]) -> None:
    """Retrieve a specific transform or a set of transforms using indexing."""
    assert isinstance(index, (int, list)), f"The indices should be either list or int type but got {type(index)}"
    if isinstance(index, list):
        assert isinstance(
            value, list
        ), f"The indices should be the same type as values, but got {type(index)} and {type(value)}"
    if isinstance(index, int):
        index, value = [index], [value]
    for i, v in zip(index, value):
        assert i < len(self.transforms), f"list index {i} out of range {len(self.transforms)}."
        self.transforms[i] = v

append(transform)

将新的变换添加到现有的变换列表中。

源代码 ultralytics/data/augment.py
def append(self, transform):
    """Appends a new transform to the existing list of transforms."""
    self.transforms.append(transform)

insert(index, transform)

在现有变换列表中插入一个新变换。

源代码 ultralytics/data/augment.py
def insert(self, index, transform):
    """Inserts a new transform to the existing list of transforms."""
    self.transforms.insert(index, transform)

tolist()

将变换列表转换为标准Python 列表。

源代码 ultralytics/data/augment.py
def tolist(self):
    """Converts the list of transforms to a standard Python list."""
    return self.transforms



ultralytics.data.augment.BaseMixTransform

用于基本混合(MixUp/Mosaic)转换的类。

该实现来自 mmyolo。

源代码 ultralytics/data/augment.py
class BaseMixTransform:
    """
    Class for base mix (MixUp/Mosaic) transformations.

    This implementation is from mmyolo.
    """

    def __init__(self, dataset, pre_transform=None, p=0.0) -> None:
        """Initializes the BaseMixTransform object with dataset, pre_transform, and probability."""
        self.dataset = dataset
        self.pre_transform = pre_transform
        self.p = p

    def __call__(self, labels):
        """Applies pre-processing transforms and mixup/mosaic transforms to labels data."""
        if random.uniform(0, 1) > self.p:
            return labels

        # Get index of one or three other images
        indexes = self.get_indexes()
        if isinstance(indexes, int):
            indexes = [indexes]

        # Get images information will be used for Mosaic or MixUp
        mix_labels = [self.dataset.get_image_and_label(i) for i in indexes]

        if self.pre_transform is not None:
            for i, data in enumerate(mix_labels):
                mix_labels[i] = self.pre_transform(data)
        labels["mix_labels"] = mix_labels

        # Update cls and texts
        labels = self._update_label_text(labels)
        # Mosaic or MixUp
        labels = self._mix_transform(labels)
        labels.pop("mix_labels", None)
        return labels

    def _mix_transform(self, labels):
        """Applies MixUp or Mosaic augmentation to the label dictionary."""
        raise NotImplementedError

    def get_indexes(self):
        """Gets a list of shuffled indexes for mosaic augmentation."""
        raise NotImplementedError

    def _update_label_text(self, labels):
        """Update label text."""
        if "texts" not in labels:
            return labels

        mix_texts = sum([labels["texts"]] + [x["texts"] for x in labels["mix_labels"]], [])
        mix_texts = list({tuple(x) for x in mix_texts})
        text2id = {text: i for i, text in enumerate(mix_texts)}

        for label in [labels] + labels["mix_labels"]:
            for i, cls in enumerate(label["cls"].squeeze(-1).tolist()):
                text = label["texts"][int(cls)]
                label["cls"][i] = text2id[tuple(text)]
            label["texts"] = mix_texts
        return labels

__call__(labels)

对标签数据应用预处理变换和混合/马赛克变换。

源代码 ultralytics/data/augment.py
def __call__(self, labels):
    """Applies pre-processing transforms and mixup/mosaic transforms to labels data."""
    if random.uniform(0, 1) > self.p:
        return labels

    # Get index of one or three other images
    indexes = self.get_indexes()
    if isinstance(indexes, int):
        indexes = [indexes]

    # Get images information will be used for Mosaic or MixUp
    mix_labels = [self.dataset.get_image_and_label(i) for i in indexes]

    if self.pre_transform is not None:
        for i, data in enumerate(mix_labels):
            mix_labels[i] = self.pre_transform(data)
    labels["mix_labels"] = mix_labels

    # Update cls and texts
    labels = self._update_label_text(labels)
    # Mosaic or MixUp
    labels = self._mix_transform(labels)
    labels.pop("mix_labels", None)
    return labels

__init__(dataset, pre_transform=None, p=0.0)

使用数据集、预变换和概率初始化 BaseMixTransform 对象。

源代码 ultralytics/data/augment.py
def __init__(self, dataset, pre_transform=None, p=0.0) -> None:
    """Initializes the BaseMixTransform object with dataset, pre_transform, and probability."""
    self.dataset = dataset
    self.pre_transform = pre_transform
    self.p = p

get_indexes()

获取用于马赛克增强的洗牌索引列表。

源代码 ultralytics/data/augment.py
def get_indexes(self):
    """Gets a list of shuffled indexes for mosaic augmentation."""
    raise NotImplementedError



ultralytics.data.augment.Mosaic

垒球 BaseMixTransform

马赛克增强

该类通过将多幅(4 幅或 9 幅)图像合并为一幅马赛克图像来执行马赛克增强。 扩增以给定的概率应用于数据集。

属性

名称 类型 说明
dataset

应用马赛克增强的数据集。

imgsz int

单张图像马赛克管道后的图像大小(高度和宽度)。默认为 640。

p float

应用镶嵌增强的概率。范围必须为 0-1。默认为 1.0。

n int

网格大小,4(2x2)或 9(3x3)。

源代码 ultralytics/data/augment.py
class Mosaic(BaseMixTransform):
    """
    Mosaic augmentation.

    This class performs mosaic augmentation by combining multiple (4 or 9) images into a single mosaic image.
    The augmentation is applied to a dataset with a given probability.

    Attributes:
        dataset: The dataset on which the mosaic augmentation is applied.
        imgsz (int, optional): Image size (height and width) after mosaic pipeline of a single image. Default to 640.
        p (float, optional): Probability of applying the mosaic augmentation. Must be in the range 0-1. Default to 1.0.
        n (int, optional): The grid size, either 4 (for 2x2) or 9 (for 3x3).
    """

    def __init__(self, dataset, imgsz=640, p=1.0, n=4):
        """Initializes the object with a dataset, image size, probability, and border."""
        assert 0 <= p <= 1.0, f"The probability should be in range [0, 1], but got {p}."
        assert n in {4, 9}, "grid must be equal to 4 or 9."
        super().__init__(dataset=dataset, p=p)
        self.dataset = dataset
        self.imgsz = imgsz
        self.border = (-imgsz // 2, -imgsz // 2)  # width, height
        self.n = n

    def get_indexes(self, buffer=True):
        """Return a list of random indexes from the dataset."""
        if buffer:  # select images from buffer
            return random.choices(list(self.dataset.buffer), k=self.n - 1)
        else:  # select any images
            return [random.randint(0, len(self.dataset) - 1) for _ in range(self.n - 1)]

    def _mix_transform(self, labels):
        """Apply mixup transformation to the input image and labels."""
        assert labels.get("rect_shape", None) is None, "rect and mosaic are mutually exclusive."
        assert len(labels.get("mix_labels", [])), "There are no other images for mosaic augment."
        return (
            self._mosaic3(labels) if self.n == 3 else self._mosaic4(labels) if self.n == 4 else self._mosaic9(labels)
        )  # This code is modified for mosaic3 method.

    def _mosaic3(self, labels):
        """Create a 1x3 image mosaic."""
        mosaic_labels = []
        s = self.imgsz
        for i in range(3):
            labels_patch = labels if i == 0 else labels["mix_labels"][i - 1]
            # Load image
            img = labels_patch["img"]
            h, w = labels_patch.pop("resized_shape")

            # Place img in img3
            if i == 0:  # center
                img3 = np.full((s * 3, s * 3, img.shape[2]), 114, dtype=np.uint8)  # base image with 3 tiles
                h0, w0 = h, w
                c = s, s, s + w, s + h  # xmin, ymin, xmax, ymax (base) coordinates
            elif i == 1:  # right
                c = s + w0, s, s + w0 + w, s + h
            elif i == 2:  # left
                c = s - w, s + h0 - h, s, s + h0

            padw, padh = c[:2]
            x1, y1, x2, y2 = (max(x, 0) for x in c)  # allocate coords

            img3[y1:y2, x1:x2] = img[y1 - padh :, x1 - padw :]  # img3[ymin:ymax, xmin:xmax]
            # hp, wp = h, w  # height, width previous for next iteration

            # Labels assuming imgsz*2 mosaic size
            labels_patch = self._update_labels(labels_patch, padw + self.border[0], padh + self.border[1])
            mosaic_labels.append(labels_patch)
        final_labels = self._cat_labels(mosaic_labels)

        final_labels["img"] = img3[-self.border[0] : self.border[0], -self.border[1] : self.border[1]]
        return final_labels

    def _mosaic4(self, labels):
        """Create a 2x2 image mosaic."""
        mosaic_labels = []
        s = self.imgsz
        yc, xc = (int(random.uniform(-x, 2 * s + x)) for x in self.border)  # mosaic center x, y
        for i in range(4):
            labels_patch = labels if i == 0 else labels["mix_labels"][i - 1]
            # Load image
            img = labels_patch["img"]
            h, w = labels_patch.pop("resized_shape")

            # Place img in img4
            if i == 0:  # top left
                img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
                x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc  # xmin, ymin, xmax, ymax (large image)
                x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h  # xmin, ymin, xmax, ymax (small image)
            elif i == 1:  # top right
                x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
                x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
            elif i == 2:  # bottom left
                x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
                x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h)
            elif i == 3:  # bottom right
                x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
                x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)

            img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b]  # img4[ymin:ymax, xmin:xmax]
            padw = x1a - x1b
            padh = y1a - y1b

            labels_patch = self._update_labels(labels_patch, padw, padh)
            mosaic_labels.append(labels_patch)
        final_labels = self._cat_labels(mosaic_labels)
        final_labels["img"] = img4
        return final_labels

    def _mosaic9(self, labels):
        """Create a 3x3 image mosaic."""
        mosaic_labels = []
        s = self.imgsz
        hp, wp = -1, -1  # height, width previous
        for i in range(9):
            labels_patch = labels if i == 0 else labels["mix_labels"][i - 1]
            # Load image
            img = labels_patch["img"]
            h, w = labels_patch.pop("resized_shape")

            # Place img in img9
            if i == 0:  # center
                img9 = np.full((s * 3, s * 3, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
                h0, w0 = h, w
                c = s, s, s + w, s + h  # xmin, ymin, xmax, ymax (base) coordinates
            elif i == 1:  # top
                c = s, s - h, s + w, s
            elif i == 2:  # top right
                c = s + wp, s - h, s + wp + w, s
            elif i == 3:  # right
                c = s + w0, s, s + w0 + w, s + h
            elif i == 4:  # bottom right
                c = s + w0, s + hp, s + w0 + w, s + hp + h
            elif i == 5:  # bottom
                c = s + w0 - w, s + h0, s + w0, s + h0 + h
            elif i == 6:  # bottom left
                c = s + w0 - wp - w, s + h0, s + w0 - wp, s + h0 + h
            elif i == 7:  # left
                c = s - w, s + h0 - h, s, s + h0
            elif i == 8:  # top left
                c = s - w, s + h0 - hp - h, s, s + h0 - hp

            padw, padh = c[:2]
            x1, y1, x2, y2 = (max(x, 0) for x in c)  # allocate coords

            # Image
            img9[y1:y2, x1:x2] = img[y1 - padh :, x1 - padw :]  # img9[ymin:ymax, xmin:xmax]
            hp, wp = h, w  # height, width previous for next iteration

            # Labels assuming imgsz*2 mosaic size
            labels_patch = self._update_labels(labels_patch, padw + self.border[0], padh + self.border[1])
            mosaic_labels.append(labels_patch)
        final_labels = self._cat_labels(mosaic_labels)

        final_labels["img"] = img9[-self.border[0] : self.border[0], -self.border[1] : self.border[1]]
        return final_labels

    @staticmethod
    def _update_labels(labels, padw, padh):
        """Update labels."""
        nh, nw = labels["img"].shape[:2]
        labels["instances"].convert_bbox(format="xyxy")
        labels["instances"].denormalize(nw, nh)
        labels["instances"].add_padding(padw, padh)
        return labels

    def _cat_labels(self, mosaic_labels):
        """Return labels with mosaic border instances clipped."""
        if len(mosaic_labels) == 0:
            return {}
        cls = []
        instances = []
        imgsz = self.imgsz * 2  # mosaic imgsz
        for labels in mosaic_labels:
            cls.append(labels["cls"])
            instances.append(labels["instances"])
        # Final labels
        final_labels = {
            "im_file": mosaic_labels[0]["im_file"],
            "ori_shape": mosaic_labels[0]["ori_shape"],
            "resized_shape": (imgsz, imgsz),
            "cls": np.concatenate(cls, 0),
            "instances": Instances.concatenate(instances, axis=0),
            "mosaic_border": self.border,
        }
        final_labels["instances"].clip(imgsz, imgsz)
        good = final_labels["instances"].remove_zero_area_boxes()
        final_labels["cls"] = final_labels["cls"][good]
        if "texts" in mosaic_labels[0]:
            final_labels["texts"] = mosaic_labels[0]["texts"]
        return final_labels

__init__(dataset, imgsz=640, p=1.0, n=4)

使用数据集、图像大小、概率和边框初始化对象。

源代码 ultralytics/data/augment.py
def __init__(self, dataset, imgsz=640, p=1.0, n=4):
    """Initializes the object with a dataset, image size, probability, and border."""
    assert 0 <= p <= 1.0, f"The probability should be in range [0, 1], but got {p}."
    assert n in {4, 9}, "grid must be equal to 4 or 9."
    super().__init__(dataset=dataset, p=p)
    self.dataset = dataset
    self.imgsz = imgsz
    self.border = (-imgsz // 2, -imgsz // 2)  # width, height
    self.n = n

get_indexes(buffer=True)

返回数据集中的随机索引列表。

源代码 ultralytics/data/augment.py
def get_indexes(self, buffer=True):
    """Return a list of random indexes from the dataset."""
    if buffer:  # select images from buffer
        return random.choices(list(self.dataset.buffer), k=self.n - 1)
    else:  # select any images
        return [random.randint(0, len(self.dataset) - 1) for _ in range(self.n - 1)]



ultralytics.data.augment.MixUp

垒球 BaseMixTransform

用于对数据集进行 MixUp 增强的类。

源代码 ultralytics/data/augment.py
class MixUp(BaseMixTransform):
    """Class for applying MixUp augmentation to the dataset."""

    def __init__(self, dataset, pre_transform=None, p=0.0) -> None:
        """Initializes MixUp object with dataset, pre_transform, and probability of applying MixUp."""
        super().__init__(dataset=dataset, pre_transform=pre_transform, p=p)

    def get_indexes(self):
        """Get a random index from the dataset."""
        return random.randint(0, len(self.dataset) - 1)

    def _mix_transform(self, labels):
        """Applies MixUp augmentation as per https://arxiv.org/pdf/1710.09412.pdf."""
        r = np.random.beta(32.0, 32.0)  # mixup ratio, alpha=beta=32.0
        labels2 = labels["mix_labels"][0]
        labels["img"] = (labels["img"] * r + labels2["img"] * (1 - r)).astype(np.uint8)
        labels["instances"] = Instances.concatenate([labels["instances"], labels2["instances"]], axis=0)
        labels["cls"] = np.concatenate([labels["cls"], labels2["cls"]], 0)
        return labels

__init__(dataset, pre_transform=None, p=0.0)

使用数据集、预变换和应用 MixUp 的概率初始化 MixUp 对象。

源代码 ultralytics/data/augment.py
def __init__(self, dataset, pre_transform=None, p=0.0) -> None:
    """Initializes MixUp object with dataset, pre_transform, and probability of applying MixUp."""
    super().__init__(dataset=dataset, pre_transform=pre_transform, p=p)

get_indexes()

从数据集中获取一个随机索引。

源代码 ultralytics/data/augment.py
def get_indexes(self):
    """Get a random index from the dataset."""
    return random.randint(0, len(self.dataset) - 1)



ultralytics.data.augment.RandomPerspective

对图像和相应的边界框、线段和关键点执行随机透视和仿射变换。 关键点。这些变换包括旋转、平移、缩放和剪切。该类还提供 选项,以指定的概率有条件地应用这些变换。

属性

名称 类型 说明
degrees float

随机旋转的度数范围。

translate float

随机平移时总宽度和高度的分数。

scale float

缩放因子区间,例如,缩放因子为 0.1 时,可在 90%-110% 之间调整大小。

shear float

剪切强度(角度,单位为度)。

perspective float

透视失真系数

border tuple

指定镶嵌边框的元组。

pre_transform callable

在开始随机变换之前应用于图像的函数/变换。

方法

名称 说明
affine_transform

对图像进行一系列仿射变换。

apply_bboxes

使用计算出的仿射矩阵变换包围盒。

apply_segments

转换线段并生成新的边界框。

apply_keypoints

转换关键点。

__call__

对图像及其相应注释进行转换的主要方法。

box_candidates

在变换后过滤掉不符合特定条件的边框。

源代码 ultralytics/data/augment.py
class RandomPerspective:
    """
    Implements random perspective and affine transformations on images and corresponding bounding boxes, segments, and
    keypoints. These transformations include rotation, translation, scaling, and shearing. The class also offers the
    option to apply these transformations conditionally with a specified probability.

    Attributes:
        degrees (float): Degree range for random rotations.
        translate (float): Fraction of total width and height for random translation.
        scale (float): Scaling factor interval, e.g., a scale factor of 0.1 allows a resize between 90%-110%.
        shear (float): Shear intensity (angle in degrees).
        perspective (float): Perspective distortion factor.
        border (tuple): Tuple specifying mosaic border.
        pre_transform (callable): A function/transform to apply to the image before starting the random transformation.

    Methods:
        affine_transform(img, border): Applies a series of affine transformations to the image.
        apply_bboxes(bboxes, M): Transforms bounding boxes using the calculated affine matrix.
        apply_segments(segments, M): Transforms segments and generates new bounding boxes.
        apply_keypoints(keypoints, M): Transforms keypoints.
        __call__(labels): Main method to apply transformations to both images and their corresponding annotations.
        box_candidates(box1, box2): Filters out bounding boxes that don't meet certain criteria post-transformation.
    """

    def __init__(
        self, degrees=0.0, translate=0.1, scale=0.5, shear=0.0, perspective=0.0, border=(0, 0), pre_transform=None
    ):
        """Initializes RandomPerspective object with transformation parameters."""

        self.degrees = degrees
        self.translate = translate
        self.scale = scale
        self.shear = shear
        self.perspective = perspective
        self.border = border  # mosaic border
        self.pre_transform = pre_transform

    def affine_transform(self, img, border):
        """
        Applies a sequence of affine transformations centered around the image center.

        Args:
            img (ndarray): Input image.
            border (tuple): Border dimensions.

        Returns:
            img (ndarray): Transformed image.
            M (ndarray): Transformation matrix.
            s (float): Scale factor.
        """

        # Center
        C = np.eye(3, dtype=np.float32)

        C[0, 2] = -img.shape[1] / 2  # x translation (pixels)
        C[1, 2] = -img.shape[0] / 2  # y translation (pixels)

        # Perspective
        P = np.eye(3, dtype=np.float32)
        P[2, 0] = random.uniform(-self.perspective, self.perspective)  # x perspective (about y)
        P[2, 1] = random.uniform(-self.perspective, self.perspective)  # y perspective (about x)

        # Rotation and Scale
        R = np.eye(3, dtype=np.float32)
        a = random.uniform(-self.degrees, self.degrees)
        # a += random.choice([-180, -90, 0, 90])  # add 90deg rotations to small rotations
        s = random.uniform(1 - self.scale, 1 + self.scale)
        # s = 2 ** random.uniform(-scale, scale)
        R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)

        # Shear
        S = np.eye(3, dtype=np.float32)
        S[0, 1] = math.tan(random.uniform(-self.shear, self.shear) * math.pi / 180)  # x shear (deg)
        S[1, 0] = math.tan(random.uniform(-self.shear, self.shear) * math.pi / 180)  # y shear (deg)

        # Translation
        T = np.eye(3, dtype=np.float32)
        T[0, 2] = random.uniform(0.5 - self.translate, 0.5 + self.translate) * self.size[0]  # x translation (pixels)
        T[1, 2] = random.uniform(0.5 - self.translate, 0.5 + self.translate) * self.size[1]  # y translation (pixels)

        # Combined rotation matrix
        M = T @ S @ R @ P @ C  # order of operations (right to left) is IMPORTANT
        # Affine image
        if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any():  # image changed
            if self.perspective:
                img = cv2.warpPerspective(img, M, dsize=self.size, borderValue=(114, 114, 114))
            else:  # affine
                img = cv2.warpAffine(img, M[:2], dsize=self.size, borderValue=(114, 114, 114))
        return img, M, s

    def apply_bboxes(self, bboxes, M):
        """
        Apply affine to bboxes only.

        Args:
            bboxes (ndarray): list of bboxes, xyxy format, with shape (num_bboxes, 4).
            M (ndarray): affine matrix.

        Returns:
            new_bboxes (ndarray): bboxes after affine, [num_bboxes, 4].
        """
        n = len(bboxes)
        if n == 0:
            return bboxes

        xy = np.ones((n * 4, 3), dtype=bboxes.dtype)
        xy[:, :2] = bboxes[:, [0, 1, 2, 3, 0, 3, 2, 1]].reshape(n * 4, 2)  # x1y1, x2y2, x1y2, x2y1
        xy = xy @ M.T  # transform
        xy = (xy[:, :2] / xy[:, 2:3] if self.perspective else xy[:, :2]).reshape(n, 8)  # perspective rescale or affine

        # Create new boxes
        x = xy[:, [0, 2, 4, 6]]
        y = xy[:, [1, 3, 5, 7]]
        return np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1)), dtype=bboxes.dtype).reshape(4, n).T

    def apply_segments(self, segments, M):
        """
        Apply affine to segments and generate new bboxes from segments.

        Args:
            segments (ndarray): list of segments, [num_samples, 500, 2].
            M (ndarray): affine matrix.

        Returns:
            new_segments (ndarray): list of segments after affine, [num_samples, 500, 2].
            new_bboxes (ndarray): bboxes after affine, [N, 4].
        """
        n, num = segments.shape[:2]
        if n == 0:
            return [], segments

        xy = np.ones((n * num, 3), dtype=segments.dtype)
        segments = segments.reshape(-1, 2)
        xy[:, :2] = segments
        xy = xy @ M.T  # transform
        xy = xy[:, :2] / xy[:, 2:3]
        segments = xy.reshape(n, -1, 2)
        bboxes = np.stack([segment2box(xy, self.size[0], self.size[1]) for xy in segments], 0)
        segments[..., 0] = segments[..., 0].clip(bboxes[:, 0:1], bboxes[:, 2:3])
        segments[..., 1] = segments[..., 1].clip(bboxes[:, 1:2], bboxes[:, 3:4])
        return bboxes, segments

    def apply_keypoints(self, keypoints, M):
        """
        Apply affine to keypoints.

        Args:
            keypoints (ndarray): keypoints, [N, 17, 3].
            M (ndarray): affine matrix.

        Returns:
            new_keypoints (ndarray): keypoints after affine, [N, 17, 3].
        """
        n, nkpt = keypoints.shape[:2]
        if n == 0:
            return keypoints
        xy = np.ones((n * nkpt, 3), dtype=keypoints.dtype)
        visible = keypoints[..., 2].reshape(n * nkpt, 1)
        xy[:, :2] = keypoints[..., :2].reshape(n * nkpt, 2)
        xy = xy @ M.T  # transform
        xy = xy[:, :2] / xy[:, 2:3]  # perspective rescale or affine
        out_mask = (xy[:, 0] < 0) | (xy[:, 1] < 0) | (xy[:, 0] > self.size[0]) | (xy[:, 1] > self.size[1])
        visible[out_mask] = 0
        return np.concatenate([xy, visible], axis=-1).reshape(n, nkpt, 3)

    def __call__(self, labels):
        """
        Affine images and targets.

        Args:
            labels (dict): a dict of `bboxes`, `segments`, `keypoints`.
        """
        if self.pre_transform and "mosaic_border" not in labels:
            labels = self.pre_transform(labels)
        labels.pop("ratio_pad", None)  # do not need ratio pad

        img = labels["img"]
        cls = labels["cls"]
        instances = labels.pop("instances")
        # Make sure the coord formats are right
        instances.convert_bbox(format="xyxy")
        instances.denormalize(*img.shape[:2][::-1])

        border = labels.pop("mosaic_border", self.border)
        self.size = img.shape[1] + border[1] * 2, img.shape[0] + border[0] * 2  # w, h
        # M is affine matrix
        # Scale for func:`box_candidates`
        img, M, scale = self.affine_transform(img, border)

        bboxes = self.apply_bboxes(instances.bboxes, M)

        segments = instances.segments
        keypoints = instances.keypoints
        # Update bboxes if there are segments.
        if len(segments):
            bboxes, segments = self.apply_segments(segments, M)

        if keypoints is not None:
            keypoints = self.apply_keypoints(keypoints, M)
        new_instances = Instances(bboxes, segments, keypoints, bbox_format="xyxy", normalized=False)
        # Clip
        new_instances.clip(*self.size)

        # Filter instances
        instances.scale(scale_w=scale, scale_h=scale, bbox_only=True)
        # Make the bboxes have the same scale with new_bboxes
        i = self.box_candidates(
            box1=instances.bboxes.T, box2=new_instances.bboxes.T, area_thr=0.01 if len(segments) else 0.10
        )
        labels["instances"] = new_instances[i]
        labels["cls"] = cls[i]
        labels["img"] = img
        labels["resized_shape"] = img.shape[:2]
        return labels

    def box_candidates(self, box1, box2, wh_thr=2, ar_thr=100, area_thr=0.1, eps=1e-16):
        """
        Compute box candidates based on a set of thresholds. This method compares the characteristics of the boxes
        before and after augmentation to decide whether a box is a candidate for further processing.

        Args:
            box1 (numpy.ndarray): The 4,n bounding box before augmentation, represented as [x1, y1, x2, y2].
            box2 (numpy.ndarray): The 4,n bounding box after augmentation, represented as [x1, y1, x2, y2].
            wh_thr (float, optional): The width and height threshold in pixels. Default is 2.
            ar_thr (float, optional): The aspect ratio threshold. Default is 100.
            area_thr (float, optional): The area ratio threshold. Default is 0.1.
            eps (float, optional): A small epsilon value to prevent division by zero. Default is 1e-16.

        Returns:
            (numpy.ndarray): A boolean array indicating which boxes are candidates based on the given thresholds.
        """
        w1, h1 = box1[2] - box1[0], box1[3] - box1[1]
        w2, h2 = box2[2] - box2[0], box2[3] - box2[1]
        ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps))  # aspect ratio
        return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr)  # candidates

__call__(labels)

仿射图像和目标

参数

名称 类型 说明 默认值
labels dict

一个 bboxes, segments, keypoints.

所需
源代码 ultralytics/data/augment.py
def __call__(self, labels):
    """
    Affine images and targets.

    Args:
        labels (dict): a dict of `bboxes`, `segments`, `keypoints`.
    """
    if self.pre_transform and "mosaic_border" not in labels:
        labels = self.pre_transform(labels)
    labels.pop("ratio_pad", None)  # do not need ratio pad

    img = labels["img"]
    cls = labels["cls"]
    instances = labels.pop("instances")
    # Make sure the coord formats are right
    instances.convert_bbox(format="xyxy")
    instances.denormalize(*img.shape[:2][::-1])

    border = labels.pop("mosaic_border", self.border)
    self.size = img.shape[1] + border[1] * 2, img.shape[0] + border[0] * 2  # w, h
    # M is affine matrix
    # Scale for func:`box_candidates`
    img, M, scale = self.affine_transform(img, border)

    bboxes = self.apply_bboxes(instances.bboxes, M)

    segments = instances.segments
    keypoints = instances.keypoints
    # Update bboxes if there are segments.
    if len(segments):
        bboxes, segments = self.apply_segments(segments, M)

    if keypoints is not None:
        keypoints = self.apply_keypoints(keypoints, M)
    new_instances = Instances(bboxes, segments, keypoints, bbox_format="xyxy", normalized=False)
    # Clip
    new_instances.clip(*self.size)

    # Filter instances
    instances.scale(scale_w=scale, scale_h=scale, bbox_only=True)
    # Make the bboxes have the same scale with new_bboxes
    i = self.box_candidates(
        box1=instances.bboxes.T, box2=new_instances.bboxes.T, area_thr=0.01 if len(segments) else 0.10
    )
    labels["instances"] = new_instances[i]
    labels["cls"] = cls[i]
    labels["img"] = img
    labels["resized_shape"] = img.shape[:2]
    return labels

__init__(degrees=0.0, translate=0.1, scale=0.5, shear=0.0, perspective=0.0, border=(0, 0), pre_transform=None)

使用变换参数初始化 RandomPerspective 对象。

源代码 ultralytics/data/augment.py
def __init__(
    self, degrees=0.0, translate=0.1, scale=0.5, shear=0.0, perspective=0.0, border=(0, 0), pre_transform=None
):
    """Initializes RandomPerspective object with transformation parameters."""

    self.degrees = degrees
    self.translate = translate
    self.scale = scale
    self.shear = shear
    self.perspective = perspective
    self.border = border  # mosaic border
    self.pre_transform = pre_transform

affine_transform(img, border)

应用以图像中心为中心的仿射变换序列。

参数

名称 类型 说明 默认值
img ndarray

输入图像。

所需
border tuple

边界尺寸

所需

返回:

名称 类型 说明
img ndarray

转换后的图像。

M ndarray

转换矩阵。

s float

比例系数。

源代码 ultralytics/data/augment.py
def affine_transform(self, img, border):
    """
    Applies a sequence of affine transformations centered around the image center.

    Args:
        img (ndarray): Input image.
        border (tuple): Border dimensions.

    Returns:
        img (ndarray): Transformed image.
        M (ndarray): Transformation matrix.
        s (float): Scale factor.
    """

    # Center
    C = np.eye(3, dtype=np.float32)

    C[0, 2] = -img.shape[1] / 2  # x translation (pixels)
    C[1, 2] = -img.shape[0] / 2  # y translation (pixels)

    # Perspective
    P = np.eye(3, dtype=np.float32)
    P[2, 0] = random.uniform(-self.perspective, self.perspective)  # x perspective (about y)
    P[2, 1] = random.uniform(-self.perspective, self.perspective)  # y perspective (about x)

    # Rotation and Scale
    R = np.eye(3, dtype=np.float32)
    a = random.uniform(-self.degrees, self.degrees)
    # a += random.choice([-180, -90, 0, 90])  # add 90deg rotations to small rotations
    s = random.uniform(1 - self.scale, 1 + self.scale)
    # s = 2 ** random.uniform(-scale, scale)
    R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)

    # Shear
    S = np.eye(3, dtype=np.float32)
    S[0, 1] = math.tan(random.uniform(-self.shear, self.shear) * math.pi / 180)  # x shear (deg)
    S[1, 0] = math.tan(random.uniform(-self.shear, self.shear) * math.pi / 180)  # y shear (deg)

    # Translation
    T = np.eye(3, dtype=np.float32)
    T[0, 2] = random.uniform(0.5 - self.translate, 0.5 + self.translate) * self.size[0]  # x translation (pixels)
    T[1, 2] = random.uniform(0.5 - self.translate, 0.5 + self.translate) * self.size[1]  # y translation (pixels)

    # Combined rotation matrix
    M = T @ S @ R @ P @ C  # order of operations (right to left) is IMPORTANT
    # Affine image
    if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any():  # image changed
        if self.perspective:
            img = cv2.warpPerspective(img, M, dsize=self.size, borderValue=(114, 114, 114))
        else:  # affine
            img = cv2.warpAffine(img, M[:2], dsize=self.size, borderValue=(114, 114, 114))
    return img, M, s

apply_bboxes(bboxes, M)

仅对 bboxes 应用仿射。

参数

名称 类型 说明 默认值
bboxes ndarray

bboxes 列表,格式为 xyxy,形状为(num_bboxes, 4)。

所需
M ndarray

仿射矩阵

所需

返回:

名称 类型 说明
new_bboxes ndarray

仿射后的 bboxes,[num_bboxes, 4]。

源代码 ultralytics/data/augment.py
def apply_bboxes(self, bboxes, M):
    """
    Apply affine to bboxes only.

    Args:
        bboxes (ndarray): list of bboxes, xyxy format, with shape (num_bboxes, 4).
        M (ndarray): affine matrix.

    Returns:
        new_bboxes (ndarray): bboxes after affine, [num_bboxes, 4].
    """
    n = len(bboxes)
    if n == 0:
        return bboxes

    xy = np.ones((n * 4, 3), dtype=bboxes.dtype)
    xy[:, :2] = bboxes[:, [0, 1, 2, 3, 0, 3, 2, 1]].reshape(n * 4, 2)  # x1y1, x2y2, x1y2, x2y1
    xy = xy @ M.T  # transform
    xy = (xy[:, :2] / xy[:, 2:3] if self.perspective else xy[:, :2]).reshape(n, 8)  # perspective rescale or affine

    # Create new boxes
    x = xy[:, [0, 2, 4, 6]]
    y = xy[:, [1, 3, 5, 7]]
    return np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1)), dtype=bboxes.dtype).reshape(4, n).T

apply_keypoints(keypoints, M)

对关键点应用仿射。

参数

名称 类型 说明 默认值
keypoints ndarray

关键点,[N,17,3]。

所需
M ndarray

仿射矩阵

所需

返回:

名称 类型 说明
new_keypoints ndarray

仿射后的关键点,[N,17,3]。

源代码 ultralytics/data/augment.py
def apply_keypoints(self, keypoints, M):
    """
    Apply affine to keypoints.

    Args:
        keypoints (ndarray): keypoints, [N, 17, 3].
        M (ndarray): affine matrix.

    Returns:
        new_keypoints (ndarray): keypoints after affine, [N, 17, 3].
    """
    n, nkpt = keypoints.shape[:2]
    if n == 0:
        return keypoints
    xy = np.ones((n * nkpt, 3), dtype=keypoints.dtype)
    visible = keypoints[..., 2].reshape(n * nkpt, 1)
    xy[:, :2] = keypoints[..., :2].reshape(n * nkpt, 2)
    xy = xy @ M.T  # transform
    xy = xy[:, :2] / xy[:, 2:3]  # perspective rescale or affine
    out_mask = (xy[:, 0] < 0) | (xy[:, 1] < 0) | (xy[:, 0] > self.size[0]) | (xy[:, 1] > self.size[1])
    visible[out_mask] = 0
    return np.concatenate([xy, visible], axis=-1).reshape(n, nkpt, 3)

apply_segments(segments, M)

对线段应用仿射,并从线段生成新的 bboxes。

参数

名称 类型 说明 默认值
segments ndarray

分段列表,[num_samples, 500, 2]。

所需
M ndarray

仿射矩阵

所需

返回:

名称 类型 说明
new_segments ndarray

仿射后的线段列表,[num_samples, 500, 2]。

new_bboxes ndarray

仿射后的 bboxes,[N,4]。

源代码 ultralytics/data/augment.py
def apply_segments(self, segments, M):
    """
    Apply affine to segments and generate new bboxes from segments.

    Args:
        segments (ndarray): list of segments, [num_samples, 500, 2].
        M (ndarray): affine matrix.

    Returns:
        new_segments (ndarray): list of segments after affine, [num_samples, 500, 2].
        new_bboxes (ndarray): bboxes after affine, [N, 4].
    """
    n, num = segments.shape[:2]
    if n == 0:
        return [], segments

    xy = np.ones((n * num, 3), dtype=segments.dtype)
    segments = segments.reshape(-1, 2)
    xy[:, :2] = segments
    xy = xy @ M.T  # transform
    xy = xy[:, :2] / xy[:, 2:3]
    segments = xy.reshape(n, -1, 2)
    bboxes = np.stack([segment2box(xy, self.size[0], self.size[1]) for xy in segments], 0)
    segments[..., 0] = segments[..., 0].clip(bboxes[:, 0:1], bboxes[:, 2:3])
    segments[..., 1] = segments[..., 1].clip(bboxes[:, 1:2], bboxes[:, 3:4])
    return bboxes, segments

box_candidates(box1, box2, wh_thr=2, ar_thr=100, area_thr=0.1, eps=1e-16)

根据一组阈值计算候选方框。这种方法会比较方框 来决定一个方框是否可以进一步处理。

参数

名称 类型 说明 默认值
box1 ndarray

增强前的 4,n 边界框,表示为 [x1, y1, x2, y2]。

所需
box2 ndarray

增强后的 4,n 边界框,表示为 [x1, y1, x2, y2]。

所需
wh_thr float

以像素为单位的宽度和高度阈值。默认为 2。

2
ar_thr float

长宽比阈值。默认为 100。

100
area_thr float

面积比率阈值。默认为 0.1。

0.1
eps float

一个小的ε值,用于防止除以零。默认值为 1e-16。

1e-16

返回:

类型 说明
ndarray

布尔数组,表示根据给定的阈值哪些方框是候选方框。

源代码 ultralytics/data/augment.py
def box_candidates(self, box1, box2, wh_thr=2, ar_thr=100, area_thr=0.1, eps=1e-16):
    """
    Compute box candidates based on a set of thresholds. This method compares the characteristics of the boxes
    before and after augmentation to decide whether a box is a candidate for further processing.

    Args:
        box1 (numpy.ndarray): The 4,n bounding box before augmentation, represented as [x1, y1, x2, y2].
        box2 (numpy.ndarray): The 4,n bounding box after augmentation, represented as [x1, y1, x2, y2].
        wh_thr (float, optional): The width and height threshold in pixels. Default is 2.
        ar_thr (float, optional): The aspect ratio threshold. Default is 100.
        area_thr (float, optional): The area ratio threshold. Default is 0.1.
        eps (float, optional): A small epsilon value to prevent division by zero. Default is 1e-16.

    Returns:
        (numpy.ndarray): A boolean array indicating which boxes are candidates based on the given thresholds.
    """
    w1, h1 = box1[2] - box1[0], box1[3] - box1[1]
    w2, h2 = box2[2] - box2[0], box2[3] - box2[1]
    ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps))  # aspect ratio
    return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr)  # candidates



ultralytics.data.augment.RandomHSV

该类负责对图像的色调、饱和度和值 (HSV) 通道执行随机调整。 通道进行随机调整。

调整是随机的,但在 hgain、sgain 和 vgain 设置的范围内。

源代码 ultralytics/data/augment.py
class RandomHSV:
    """
    This class is responsible for performing random adjustments to the Hue, Saturation, and Value (HSV) channels of an
    image.

    The adjustments are random but within limits set by hgain, sgain, and vgain.
    """

    def __init__(self, hgain=0.5, sgain=0.5, vgain=0.5) -> None:
        """
        Initialize RandomHSV class with gains for each HSV channel.

        Args:
            hgain (float, optional): Maximum variation for hue. Default is 0.5.
            sgain (float, optional): Maximum variation for saturation. Default is 0.5.
            vgain (float, optional): Maximum variation for value. Default is 0.5.
        """
        self.hgain = hgain
        self.sgain = sgain
        self.vgain = vgain

    def __call__(self, labels):
        """
        Applies random HSV augmentation to an image within the predefined limits.

        The modified image replaces the original image in the input 'labels' dict.
        """
        img = labels["img"]
        if self.hgain or self.sgain or self.vgain:
            r = np.random.uniform(-1, 1, 3) * [self.hgain, self.sgain, self.vgain] + 1  # random gains
            hue, sat, val = cv2.split(cv2.cvtColor(img, cv2.COLOR_BGR2HSV))
            dtype = img.dtype  # uint8

            x = np.arange(0, 256, dtype=r.dtype)
            lut_hue = ((x * r[0]) % 180).astype(dtype)
            lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)
            lut_val = np.clip(x * r[2], 0, 255).astype(dtype)

            im_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val)))
            cv2.cvtColor(im_hsv, cv2.COLOR_HSV2BGR, dst=img)  # no return needed
        return labels

__call__(labels)

在预定义范围内对图像进行随机 HSV 增强。

修改后的图像会取代输入 "标签 "字段中的原始图像。

源代码 ultralytics/data/augment.py
def __call__(self, labels):
    """
    Applies random HSV augmentation to an image within the predefined limits.

    The modified image replaces the original image in the input 'labels' dict.
    """
    img = labels["img"]
    if self.hgain or self.sgain or self.vgain:
        r = np.random.uniform(-1, 1, 3) * [self.hgain, self.sgain, self.vgain] + 1  # random gains
        hue, sat, val = cv2.split(cv2.cvtColor(img, cv2.COLOR_BGR2HSV))
        dtype = img.dtype  # uint8

        x = np.arange(0, 256, dtype=r.dtype)
        lut_hue = ((x * r[0]) % 180).astype(dtype)
        lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)
        lut_val = np.clip(x * r[2], 0, 255).astype(dtype)

        im_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val)))
        cv2.cvtColor(im_hsv, cv2.COLOR_HSV2BGR, dst=img)  # no return needed
    return labels

__init__(hgain=0.5, sgain=0.5, vgain=0.5)

使用每个 HSV 通道的增益初始化 RandomHSV 类。

参数

名称 类型 说明 默认值
hgain float

色调的最大变化。默认值为 0.5。

0.5
sgain float

饱和度的最大变化。默认为 0.5。

0.5
vgain float

值的最大变化。默认为 0.5。

0.5
源代码 ultralytics/data/augment.py
def __init__(self, hgain=0.5, sgain=0.5, vgain=0.5) -> None:
    """
    Initialize RandomHSV class with gains for each HSV channel.

    Args:
        hgain (float, optional): Maximum variation for hue. Default is 0.5.
        sgain (float, optional): Maximum variation for saturation. Default is 0.5.
        vgain (float, optional): Maximum variation for value. Default is 0.5.
    """
    self.hgain = hgain
    self.sgain = sgain
    self.vgain = vgain



ultralytics.data.augment.RandomFlip

以给定的概率对图像进行随机水平或垂直翻转。

还会相应更新任何实例(边框、关键点等)。

源代码 ultralytics/data/augment.py
class RandomFlip:
    """
    Applies a random horizontal or vertical flip to an image with a given probability.

    Also updates any instances (bounding boxes, keypoints, etc.) accordingly.
    """

    def __init__(self, p=0.5, direction="horizontal", flip_idx=None) -> None:
        """
        Initializes the RandomFlip class with probability and direction.

        Args:
            p (float, optional): The probability of applying the flip. Must be between 0 and 1. Default is 0.5.
            direction (str, optional): The direction to apply the flip. Must be 'horizontal' or 'vertical'.
                Default is 'horizontal'.
            flip_idx (array-like, optional): Index mapping for flipping keypoints, if any.
        """
        assert direction in {"horizontal", "vertical"}, f"Support direction `horizontal` or `vertical`, got {direction}"
        assert 0 <= p <= 1.0

        self.p = p
        self.direction = direction
        self.flip_idx = flip_idx

    def __call__(self, labels):
        """
        Applies random flip to an image and updates any instances like bounding boxes or keypoints accordingly.

        Args:
            labels (dict): A dictionary containing the keys 'img' and 'instances'. 'img' is the image to be flipped.
                           'instances' is an object containing bounding boxes and optionally keypoints.

        Returns:
            (dict): The same dict with the flipped image and updated instances under the 'img' and 'instances' keys.
        """
        img = labels["img"]
        instances = labels.pop("instances")
        instances.convert_bbox(format="xywh")
        h, w = img.shape[:2]
        h = 1 if instances.normalized else h
        w = 1 if instances.normalized else w

        # Flip up-down
        if self.direction == "vertical" and random.random() < self.p:
            img = np.flipud(img)
            instances.flipud(h)
        if self.direction == "horizontal" and random.random() < self.p:
            img = np.fliplr(img)
            instances.fliplr(w)
            # For keypoints
            if self.flip_idx is not None and instances.keypoints is not None:
                instances.keypoints = np.ascontiguousarray(instances.keypoints[:, self.flip_idx, :])
        labels["img"] = np.ascontiguousarray(img)
        labels["instances"] = instances
        return labels

__call__(labels)

对图像进行随机翻转,并相应更新任何实例(如边界框或关键点)。

参数

名称 类型 说明 默认值
labels dict

包含键 "img "和 "实例 "的字典。img "是要翻转的图像。 实例 "是一个包含边界框和可选关键点的对象。

所需

返回:

类型 说明
dict

在 "img "和 "instances "键下,对翻转图像和更新实例进行同样的操作。

源代码 ultralytics/data/augment.py
def __call__(self, labels):
    """
    Applies random flip to an image and updates any instances like bounding boxes or keypoints accordingly.

    Args:
        labels (dict): A dictionary containing the keys 'img' and 'instances'. 'img' is the image to be flipped.
                       'instances' is an object containing bounding boxes and optionally keypoints.

    Returns:
        (dict): The same dict with the flipped image and updated instances under the 'img' and 'instances' keys.
    """
    img = labels["img"]
    instances = labels.pop("instances")
    instances.convert_bbox(format="xywh")
    h, w = img.shape[:2]
    h = 1 if instances.normalized else h
    w = 1 if instances.normalized else w

    # Flip up-down
    if self.direction == "vertical" and random.random() < self.p:
        img = np.flipud(img)
        instances.flipud(h)
    if self.direction == "horizontal" and random.random() < self.p:
        img = np.fliplr(img)
        instances.fliplr(w)
        # For keypoints
        if self.flip_idx is not None and instances.keypoints is not None:
            instances.keypoints = np.ascontiguousarray(instances.keypoints[:, self.flip_idx, :])
    labels["img"] = np.ascontiguousarray(img)
    labels["instances"] = instances
    return labels

__init__(p=0.5, direction='horizontal', flip_idx=None)

用概率和方向初始化 RandomFlip 类。

参数

名称 类型 说明 默认值
p float

应用翻转的概率。默认值为 0.5。

0.5
direction str

应用翻转的方向。必须是 "水平 "或 "垂直"。 默认为 "水平"。

'horizontal'
flip_idx array - like

用于翻转关键点的索引映射(如果有)。

None
源代码 ultralytics/data/augment.py
def __init__(self, p=0.5, direction="horizontal", flip_idx=None) -> None:
    """
    Initializes the RandomFlip class with probability and direction.

    Args:
        p (float, optional): The probability of applying the flip. Must be between 0 and 1. Default is 0.5.
        direction (str, optional): The direction to apply the flip. Must be 'horizontal' or 'vertical'.
            Default is 'horizontal'.
        flip_idx (array-like, optional): Index mapping for flipping keypoints, if any.
    """
    assert direction in {"horizontal", "vertical"}, f"Support direction `horizontal` or `vertical`, got {direction}"
    assert 0 <= p <= 1.0

    self.p = p
    self.direction = direction
    self.flip_idx = flip_idx



ultralytics.data.augment.LetterBox

为检测、实例分割和姿势调整图像大小和填充。

源代码 ultralytics/data/augment.py
class LetterBox:
    """Resize image and padding for detection, instance segmentation, pose."""

    def __init__(self, new_shape=(640, 640), auto=False, scaleFill=False, scaleup=True, center=True, stride=32):
        """Initialize LetterBox object with specific parameters."""
        self.new_shape = new_shape
        self.auto = auto
        self.scaleFill = scaleFill
        self.scaleup = scaleup
        self.stride = stride
        self.center = center  # Put the image in the middle or top-left

    def __call__(self, labels=None, image=None):
        """Return updated labels and image with added border."""
        if labels is None:
            labels = {}
        img = labels.get("img") if image is None else image
        shape = img.shape[:2]  # current shape [height, width]
        new_shape = labels.pop("rect_shape", self.new_shape)
        if isinstance(new_shape, int):
            new_shape = (new_shape, new_shape)

        # Scale ratio (new / old)
        r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
        if not self.scaleup:  # only scale down, do not scale up (for better val mAP)
            r = min(r, 1.0)

        # Compute padding
        ratio = r, r  # width, height ratios
        new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
        dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
        if self.auto:  # minimum rectangle
            dw, dh = np.mod(dw, self.stride), np.mod(dh, self.stride)  # wh padding
        elif self.scaleFill:  # stretch
            dw, dh = 0.0, 0.0
            new_unpad = (new_shape[1], new_shape[0])
            ratio = new_shape[1] / shape[1], new_shape[0] / shape[0]  # width, height ratios

        if self.center:
            dw /= 2  # divide padding into 2 sides
            dh /= 2

        if shape[::-1] != new_unpad:  # resize
            img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
        top, bottom = int(round(dh - 0.1)) if self.center else 0, int(round(dh + 0.1))
        left, right = int(round(dw - 0.1)) if self.center else 0, int(round(dw + 0.1))
        img = cv2.copyMakeBorder(
            img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=(114, 114, 114)
        )  # add border
        if labels.get("ratio_pad"):
            labels["ratio_pad"] = (labels["ratio_pad"], (left, top))  # for evaluation

        if len(labels):
            labels = self._update_labels(labels, ratio, dw, dh)
            labels["img"] = img
            labels["resized_shape"] = new_shape
            return labels
        else:
            return img

    def _update_labels(self, labels, ratio, padw, padh):
        """Update labels."""
        labels["instances"].convert_bbox(format="xyxy")
        labels["instances"].denormalize(*labels["img"].shape[:2][::-1])
        labels["instances"].scale(*ratio)
        labels["instances"].add_padding(padw, padh)
        return labels

__call__(labels=None, image=None)

返回更新后的标签和图像,并添加边框。

源代码 ultralytics/data/augment.py
def __call__(self, labels=None, image=None):
    """Return updated labels and image with added border."""
    if labels is None:
        labels = {}
    img = labels.get("img") if image is None else image
    shape = img.shape[:2]  # current shape [height, width]
    new_shape = labels.pop("rect_shape", self.new_shape)
    if isinstance(new_shape, int):
        new_shape = (new_shape, new_shape)

    # Scale ratio (new / old)
    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
    if not self.scaleup:  # only scale down, do not scale up (for better val mAP)
        r = min(r, 1.0)

    # Compute padding
    ratio = r, r  # width, height ratios
    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
    if self.auto:  # minimum rectangle
        dw, dh = np.mod(dw, self.stride), np.mod(dh, self.stride)  # wh padding
    elif self.scaleFill:  # stretch
        dw, dh = 0.0, 0.0
        new_unpad = (new_shape[1], new_shape[0])
        ratio = new_shape[1] / shape[1], new_shape[0] / shape[0]  # width, height ratios

    if self.center:
        dw /= 2  # divide padding into 2 sides
        dh /= 2

    if shape[::-1] != new_unpad:  # resize
        img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
    top, bottom = int(round(dh - 0.1)) if self.center else 0, int(round(dh + 0.1))
    left, right = int(round(dw - 0.1)) if self.center else 0, int(round(dw + 0.1))
    img = cv2.copyMakeBorder(
        img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=(114, 114, 114)
    )  # add border
    if labels.get("ratio_pad"):
        labels["ratio_pad"] = (labels["ratio_pad"], (left, top))  # for evaluation

    if len(labels):
        labels = self._update_labels(labels, ratio, dw, dh)
        labels["img"] = img
        labels["resized_shape"] = new_shape
        return labels
    else:
        return img

__init__(new_shape=(640, 640), auto=False, scaleFill=False, scaleup=True, center=True, stride=32)

使用特定参数初始化 LetterBox 对象。

源代码 ultralytics/data/augment.py
def __init__(self, new_shape=(640, 640), auto=False, scaleFill=False, scaleup=True, center=True, stride=32):
    """Initialize LetterBox object with specific parameters."""
    self.new_shape = new_shape
    self.auto = auto
    self.scaleFill = scaleFill
    self.scaleup = scaleup
    self.stride = stride
    self.center = center  # Put the image in the middle or top-left



ultralytics.data.augment.CopyPaste

实现论文 https://arxiv.org/abs/2012.07177 中描述的复制粘贴增强功能。该类 负责在图像及其相应实例上应用复制粘贴增强功能。

源代码 ultralytics/data/augment.py
class CopyPaste:
    """
    Implements the Copy-Paste augmentation as described in the paper https://arxiv.org/abs/2012.07177. This class is
    responsible for applying the Copy-Paste augmentation on images and their corresponding instances.
    """

    def __init__(self, p=0.5) -> None:
        """
        Initializes the CopyPaste class with a given probability.

        Args:
            p (float, optional): The probability of applying the Copy-Paste augmentation. Must be between 0 and 1.
                                 Default is 0.5.
        """
        self.p = p

    def __call__(self, labels):
        """
        Applies the Copy-Paste augmentation to the given image and instances.

        Args:
            labels (dict): A dictionary containing:
                           - 'img': The image to augment.
                           - 'cls': Class labels associated with the instances.
                           - 'instances': Object containing bounding boxes, and optionally, keypoints and segments.

        Returns:
            (dict): Dict with augmented image and updated instances under the 'img', 'cls', and 'instances' keys.

        Notes:
            1. Instances are expected to have 'segments' as one of their attributes for this augmentation to work.
            2. This method modifies the input dictionary 'labels' in place.
        """
        im = labels["img"]
        cls = labels["cls"]
        h, w = im.shape[:2]
        instances = labels.pop("instances")
        instances.convert_bbox(format="xyxy")
        instances.denormalize(w, h)
        if self.p and len(instances.segments):
            n = len(instances)
            _, w, _ = im.shape  # height, width, channels
            im_new = np.zeros(im.shape, np.uint8)

            # Calculate ioa first then select indexes randomly
            ins_flip = deepcopy(instances)
            ins_flip.fliplr(w)

            ioa = bbox_ioa(ins_flip.bboxes, instances.bboxes)  # intersection over area, (N, M)
            indexes = np.nonzero((ioa < 0.30).all(1))[0]  # (N, )
            n = len(indexes)
            for j in random.sample(list(indexes), k=round(self.p * n)):
                cls = np.concatenate((cls, cls[[j]]), axis=0)
                instances = Instances.concatenate((instances, ins_flip[[j]]), axis=0)
                cv2.drawContours(im_new, instances.segments[[j]].astype(np.int32), -1, (1, 1, 1), cv2.FILLED)

            result = cv2.flip(im, 1)  # augment segments (flip left-right)
            i = cv2.flip(im_new, 1).astype(bool)
            im[i] = result[i]

        labels["img"] = im
        labels["cls"] = cls
        labels["instances"] = instances
        return labels

__call__(labels)

将复制粘贴增强功能应用于给定的图像和实例。

参数

名称 类型 说明 默认值
labels dict

包含以下内容的词典 - img":要增强的图像。 - 'cls':与实例相关的类标签。 - 实例包含边界框以及关键点和线段的对象。

所需

返回:

类型 说明
dict

Dict 在 "img"、"cls "和 "instances "键下包含增强的图像和更新的实例。

说明
  1. 要使该增强功能生效,实例必须将 "段 "作为其属性之一。
  2. 该方法会对输入字典 "标签 "进行适当修改。
源代码 ultralytics/data/augment.py
def __call__(self, labels):
    """
    Applies the Copy-Paste augmentation to the given image and instances.

    Args:
        labels (dict): A dictionary containing:
                       - 'img': The image to augment.
                       - 'cls': Class labels associated with the instances.
                       - 'instances': Object containing bounding boxes, and optionally, keypoints and segments.

    Returns:
        (dict): Dict with augmented image and updated instances under the 'img', 'cls', and 'instances' keys.

    Notes:
        1. Instances are expected to have 'segments' as one of their attributes for this augmentation to work.
        2. This method modifies the input dictionary 'labels' in place.
    """
    im = labels["img"]
    cls = labels["cls"]
    h, w = im.shape[:2]
    instances = labels.pop("instances")
    instances.convert_bbox(format="xyxy")
    instances.denormalize(w, h)
    if self.p and len(instances.segments):
        n = len(instances)
        _, w, _ = im.shape  # height, width, channels
        im_new = np.zeros(im.shape, np.uint8)

        # Calculate ioa first then select indexes randomly
        ins_flip = deepcopy(instances)
        ins_flip.fliplr(w)

        ioa = bbox_ioa(ins_flip.bboxes, instances.bboxes)  # intersection over area, (N, M)
        indexes = np.nonzero((ioa < 0.30).all(1))[0]  # (N, )
        n = len(indexes)
        for j in random.sample(list(indexes), k=round(self.p * n)):
            cls = np.concatenate((cls, cls[[j]]), axis=0)
            instances = Instances.concatenate((instances, ins_flip[[j]]), axis=0)
            cv2.drawContours(im_new, instances.segments[[j]].astype(np.int32), -1, (1, 1, 1), cv2.FILLED)

        result = cv2.flip(im, 1)  # augment segments (flip left-right)
        i = cv2.flip(im_new, 1).astype(bool)
        im[i] = result[i]

    labels["img"] = im
    labels["cls"] = cls
    labels["instances"] = instances
    return labels

__init__(p=0.5)

以给定的概率初始化 CopyPaste 类。

参数

名称 类型 说明 默认值
p float

应用复制粘贴增强功能的概率。必须介于 0 和 1 之间。 默认值为 0.5。

0.5
源代码 ultralytics/data/augment.py
def __init__(self, p=0.5) -> None:
    """
    Initializes the CopyPaste class with a given probability.

    Args:
        p (float, optional): The probability of applying the Copy-Paste augmentation. Must be between 0 and 1.
                             Default is 0.5.
    """
    self.p = p



ultralytics.data.augment.Albumentations

Albumentations transformations.

可选,卸载软件包即可禁用。应用模糊、中值模糊、灰度转换、对比度限制自适应 直方图均衡化、亮度和对比度随机变化、RandomGamma 以及通过压缩降低图像质量。 压缩。

源代码 ultralytics/data/augment.py
class Albumentations:
    """
    Albumentations transformations.

    Optional, uninstall package to disable. Applies Blur, Median Blur, convert to grayscale, Contrast Limited Adaptive
    Histogram Equalization, random change of brightness and contrast, RandomGamma and lowering of image quality by
    compression.
    """

    def __init__(self, p=1.0):
        """Initialize the transform object for YOLO bbox formatted params."""
        self.p = p
        self.transform = None
        prefix = colorstr("albumentations: ")

        try:
            import albumentations as A

            check_version(A.__version__, "1.0.3", hard=True)  # version requirement

            # List of possible spatial transforms
            spatial_transforms = {
                "Affine",
                "BBoxSafeRandomCrop",
                "CenterCrop",
                "CoarseDropout",
                "Crop",
                "CropAndPad",
                "CropNonEmptyMaskIfExists",
                "D4",
                "ElasticTransform",
                "Flip",
                "GridDistortion",
                "GridDropout",
                "HorizontalFlip",
                "Lambda",
                "LongestMaxSize",
                "MaskDropout",
                "MixUp",
                "Morphological",
                "NoOp",
                "OpticalDistortion",
                "PadIfNeeded",
                "Perspective",
                "PiecewiseAffine",
                "PixelDropout",
                "RandomCrop",
                "RandomCropFromBorders",
                "RandomGridShuffle",
                "RandomResizedCrop",
                "RandomRotate90",
                "RandomScale",
                "RandomSizedBBoxSafeCrop",
                "RandomSizedCrop",
                "Resize",
                "Rotate",
                "SafeRotate",
                "ShiftScaleRotate",
                "SmallestMaxSize",
                "Transpose",
                "VerticalFlip",
                "XYMasking",
            }  # from https://albumentations.ai/docs/getting_started/transforms_and_targets/#spatial-level-transforms

            # Transforms
            T = [
                A.Blur(p=0.01),
                A.MedianBlur(p=0.01),
                A.ToGray(p=0.01),
                A.CLAHE(p=0.01),
                A.RandomBrightnessContrast(p=0.0),
                A.RandomGamma(p=0.0),
                A.ImageCompression(quality_lower=75, p=0.0),
            ]

            # Compose transforms
            self.contains_spatial = any(transform.__class__.__name__ in spatial_transforms for transform in T)
            self.transform = (
                A.Compose(T, bbox_params=A.BboxParams(format="yolo", label_fields=["class_labels"]))
                if self.contains_spatial
                else A.Compose(T)
            )
            LOGGER.info(prefix + ", ".join(f"{x}".replace("always_apply=False, ", "") for x in T if x.p))
        except ImportError:  # package not installed, skip
            pass
        except Exception as e:
            LOGGER.info(f"{prefix}{e}")

    def __call__(self, labels):
        """Generates object detections and returns a dictionary with detection results."""
        if self.transform is None or random.random() > self.p:
            return labels

        if self.contains_spatial:
            cls = labels["cls"]
            if len(cls):
                im = labels["img"]
                labels["instances"].convert_bbox("xywh")
                labels["instances"].normalize(*im.shape[:2][::-1])
                bboxes = labels["instances"].bboxes
                # TODO: add supports of segments and keypoints
                new = self.transform(image=im, bboxes=bboxes, class_labels=cls)  # transformed
                if len(new["class_labels"]) > 0:  # skip update if no bbox in new im
                    labels["img"] = new["image"]
                    labels["cls"] = np.array(new["class_labels"])
                    bboxes = np.array(new["bboxes"], dtype=np.float32)
                labels["instances"].update(bboxes=bboxes)
        else:
            labels["img"] = self.transform(image=labels["img"])["image"]  # transformed

        return labels

__call__(labels)

生成对象检测并返回包含检测结果的字典。

源代码 ultralytics/data/augment.py
def __call__(self, labels):
    """Generates object detections and returns a dictionary with detection results."""
    if self.transform is None or random.random() > self.p:
        return labels

    if self.contains_spatial:
        cls = labels["cls"]
        if len(cls):
            im = labels["img"]
            labels["instances"].convert_bbox("xywh")
            labels["instances"].normalize(*im.shape[:2][::-1])
            bboxes = labels["instances"].bboxes
            # TODO: add supports of segments and keypoints
            new = self.transform(image=im, bboxes=bboxes, class_labels=cls)  # transformed
            if len(new["class_labels"]) > 0:  # skip update if no bbox in new im
                labels["img"] = new["image"]
                labels["cls"] = np.array(new["class_labels"])
                bboxes = np.array(new["bboxes"], dtype=np.float32)
            labels["instances"].update(bboxes=bboxes)
    else:
        labels["img"] = self.transform(image=labels["img"])["image"]  # transformed

    return labels

__init__(p=1.0)

初始化YOLO bbox 格式参数的变换对象。

源代码 ultralytics/data/augment.py
def __init__(self, p=1.0):
    """Initialize the transform object for YOLO bbox formatted params."""
    self.p = p
    self.transform = None
    prefix = colorstr("albumentations: ")

    try:
        import albumentations as A

        check_version(A.__version__, "1.0.3", hard=True)  # version requirement

        # List of possible spatial transforms
        spatial_transforms = {
            "Affine",
            "BBoxSafeRandomCrop",
            "CenterCrop",
            "CoarseDropout",
            "Crop",
            "CropAndPad",
            "CropNonEmptyMaskIfExists",
            "D4",
            "ElasticTransform",
            "Flip",
            "GridDistortion",
            "GridDropout",
            "HorizontalFlip",
            "Lambda",
            "LongestMaxSize",
            "MaskDropout",
            "MixUp",
            "Morphological",
            "NoOp",
            "OpticalDistortion",
            "PadIfNeeded",
            "Perspective",
            "PiecewiseAffine",
            "PixelDropout",
            "RandomCrop",
            "RandomCropFromBorders",
            "RandomGridShuffle",
            "RandomResizedCrop",
            "RandomRotate90",
            "RandomScale",
            "RandomSizedBBoxSafeCrop",
            "RandomSizedCrop",
            "Resize",
            "Rotate",
            "SafeRotate",
            "ShiftScaleRotate",
            "SmallestMaxSize",
            "Transpose",
            "VerticalFlip",
            "XYMasking",
        }  # from https://albumentations.ai/docs/getting_started/transforms_and_targets/#spatial-level-transforms

        # Transforms
        T = [
            A.Blur(p=0.01),
            A.MedianBlur(p=0.01),
            A.ToGray(p=0.01),
            A.CLAHE(p=0.01),
            A.RandomBrightnessContrast(p=0.0),
            A.RandomGamma(p=0.0),
            A.ImageCompression(quality_lower=75, p=0.0),
        ]

        # Compose transforms
        self.contains_spatial = any(transform.__class__.__name__ in spatial_transforms for transform in T)
        self.transform = (
            A.Compose(T, bbox_params=A.BboxParams(format="yolo", label_fields=["class_labels"]))
            if self.contains_spatial
            else A.Compose(T)
        )
        LOGGER.info(prefix + ", ".join(f"{x}".replace("always_apply=False, ", "") for x in T if x.p))
    except ImportError:  # package not installed, skip
        pass
    except Exception as e:
        LOGGER.info(f"{prefix}{e}")



ultralytics.data.augment.Format

为物体检测、实例分割和姿态估计任务格式化图像注释。该类 类将图像和实例注释标准化,供 collate_fn PyTorch DataLoader 中。

属性

名称 类型 说明
bbox_format str

边界框的格式。默认为 "xywh"。

normalize bool

是否规范化边界框。默认为 True。

return_mask bool

返回用于分割的实例掩码。默认为假。

return_keypoint bool

返回姿态估计的关键点。默认为假。

mask_ratio int

掩码的下采样率。默认值为 4。

mask_overlap bool

是否重叠遮罩。默认为 True。

batch_idx bool

保留批处理索引。默认为 True。

bgr float

返回 BGR 图像的概率。默认值为 0.0。

源代码 ultralytics/data/augment.py
class Format:
    """
    Formats image annotations for object detection, instance segmentation, and pose estimation tasks. The class
    standardizes the image and instance annotations to be used by the `collate_fn` in PyTorch DataLoader.

    Attributes:
        bbox_format (str): Format for bounding boxes. Default is 'xywh'.
        normalize (bool): Whether to normalize bounding boxes. Default is True.
        return_mask (bool): Return instance masks for segmentation. Default is False.
        return_keypoint (bool): Return keypoints for pose estimation. Default is False.
        mask_ratio (int): Downsample ratio for masks. Default is 4.
        mask_overlap (bool): Whether to overlap masks. Default is True.
        batch_idx (bool): Keep batch indexes. Default is True.
        bgr (float): The probability to return BGR images. Default is 0.0.
    """

    def __init__(
        self,
        bbox_format="xywh",
        normalize=True,
        return_mask=False,
        return_keypoint=False,
        return_obb=False,
        mask_ratio=4,
        mask_overlap=True,
        batch_idx=True,
        bgr=0.0,
    ):
        """Initializes the Format class with given parameters."""
        self.bbox_format = bbox_format
        self.normalize = normalize
        self.return_mask = return_mask  # set False when training detection only
        self.return_keypoint = return_keypoint
        self.return_obb = return_obb
        self.mask_ratio = mask_ratio
        self.mask_overlap = mask_overlap
        self.batch_idx = batch_idx  # keep the batch indexes
        self.bgr = bgr

    def __call__(self, labels):
        """Return formatted image, classes, bounding boxes & keypoints to be used by 'collate_fn'."""
        img = labels.pop("img")
        h, w = img.shape[:2]
        cls = labels.pop("cls")
        instances = labels.pop("instances")
        instances.convert_bbox(format=self.bbox_format)
        instances.denormalize(w, h)
        nl = len(instances)

        if self.return_mask:
            if nl:
                masks, instances, cls = self._format_segments(instances, cls, w, h)
                masks = torch.from_numpy(masks)
            else:
                masks = torch.zeros(
                    1 if self.mask_overlap else nl, img.shape[0] // self.mask_ratio, img.shape[1] // self.mask_ratio
                )
            labels["masks"] = masks
        labels["img"] = self._format_img(img)
        labels["cls"] = torch.from_numpy(cls) if nl else torch.zeros(nl)
        labels["bboxes"] = torch.from_numpy(instances.bboxes) if nl else torch.zeros((nl, 4))
        if self.return_keypoint:
            labels["keypoints"] = torch.from_numpy(instances.keypoints)
            if self.normalize:
                labels["keypoints"][..., 0] /= w
                labels["keypoints"][..., 1] /= h
        if self.return_obb:
            labels["bboxes"] = (
                xyxyxyxy2xywhr(torch.from_numpy(instances.segments)) if len(instances.segments) else torch.zeros((0, 5))
            )
        # NOTE: need to normalize obb in xywhr format for width-height consistency
        if self.normalize:
            labels["bboxes"][:, [0, 2]] /= w
            labels["bboxes"][:, [1, 3]] /= h
        # Then we can use collate_fn
        if self.batch_idx:
            labels["batch_idx"] = torch.zeros(nl)
        return labels

    def _format_img(self, img):
        """Format the image for YOLO from Numpy array to PyTorch tensor."""
        if len(img.shape) < 3:
            img = np.expand_dims(img, -1)
        img = img.transpose(2, 0, 1)
        img = np.ascontiguousarray(img[::-1] if random.uniform(0, 1) > self.bgr else img)
        img = torch.from_numpy(img)
        return img

    def _format_segments(self, instances, cls, w, h):
        """Convert polygon points to bitmap."""
        segments = instances.segments
        if self.mask_overlap:
            masks, sorted_idx = polygons2masks_overlap((h, w), segments, downsample_ratio=self.mask_ratio)
            masks = masks[None]  # (640, 640) -> (1, 640, 640)
            instances = instances[sorted_idx]
            cls = cls[sorted_idx]
        else:
            masks = polygons2masks((h, w), segments, color=1, downsample_ratio=self.mask_ratio)

        return masks, instances, cls

__call__(labels)

返回格式化的图像、类、边界框和关键点,供 "collate_fn "使用。

源代码 ultralytics/data/augment.py
def __call__(self, labels):
    """Return formatted image, classes, bounding boxes & keypoints to be used by 'collate_fn'."""
    img = labels.pop("img")
    h, w = img.shape[:2]
    cls = labels.pop("cls")
    instances = labels.pop("instances")
    instances.convert_bbox(format=self.bbox_format)
    instances.denormalize(w, h)
    nl = len(instances)

    if self.return_mask:
        if nl:
            masks, instances, cls = self._format_segments(instances, cls, w, h)
            masks = torch.from_numpy(masks)
        else:
            masks = torch.zeros(
                1 if self.mask_overlap else nl, img.shape[0] // self.mask_ratio, img.shape[1] // self.mask_ratio
            )
        labels["masks"] = masks
    labels["img"] = self._format_img(img)
    labels["cls"] = torch.from_numpy(cls) if nl else torch.zeros(nl)
    labels["bboxes"] = torch.from_numpy(instances.bboxes) if nl else torch.zeros((nl, 4))
    if self.return_keypoint:
        labels["keypoints"] = torch.from_numpy(instances.keypoints)
        if self.normalize:
            labels["keypoints"][..., 0] /= w
            labels["keypoints"][..., 1] /= h
    if self.return_obb:
        labels["bboxes"] = (
            xyxyxyxy2xywhr(torch.from_numpy(instances.segments)) if len(instances.segments) else torch.zeros((0, 5))
        )
    # NOTE: need to normalize obb in xywhr format for width-height consistency
    if self.normalize:
        labels["bboxes"][:, [0, 2]] /= w
        labels["bboxes"][:, [1, 3]] /= h
    # Then we can use collate_fn
    if self.batch_idx:
        labels["batch_idx"] = torch.zeros(nl)
    return labels

__init__(bbox_format='xywh', normalize=True, return_mask=False, return_keypoint=False, return_obb=False, mask_ratio=4, mask_overlap=True, batch_idx=True, bgr=0.0)

使用给定参数初始化格式类。

源代码 ultralytics/data/augment.py
def __init__(
    self,
    bbox_format="xywh",
    normalize=True,
    return_mask=False,
    return_keypoint=False,
    return_obb=False,
    mask_ratio=4,
    mask_overlap=True,
    batch_idx=True,
    bgr=0.0,
):
    """Initializes the Format class with given parameters."""
    self.bbox_format = bbox_format
    self.normalize = normalize
    self.return_mask = return_mask  # set False when training detection only
    self.return_keypoint = return_keypoint
    self.return_obb = return_obb
    self.mask_ratio = mask_ratio
    self.mask_overlap = mask_overlap
    self.batch_idx = batch_idx  # keep the batch indexes
    self.bgr = bgr



ultralytics.data.augment.RandomLoadText

随机抽取正面文本和负面文本,并根据样本数量更新类别指数。

属性

名称 类型 说明
prompt_format str

提示格式。默认为"{}"。

neg_samples tuple[int]

随机抽取负文本的游侠,默认值为 (80,80)。

max_samples int

一幅图像中不同文字样本的最大数量,默认为 80。

padding bool

是否将文本填充到 max_samples。默认为 "假"。

padding_value str

填充文本。默认为""。

源代码 ultralytics/data/augment.py
class RandomLoadText:
    """
    Randomly sample positive texts and negative texts and update the class indices accordingly to the number of samples.

    Attributes:
        prompt_format (str): Format for prompt. Default is '{}'.
        neg_samples (tuple[int]): A ranger to randomly sample negative texts, Default is (80, 80).
        max_samples (int): The max number of different text samples in one image, Default is 80.
        padding (bool): Whether to pad texts to max_samples. Default is False.
        padding_value (str): The padding text. Default is "".
    """

    def __init__(
        self,
        prompt_format: str = "{}",
        neg_samples: Tuple[int, int] = (80, 80),
        max_samples: int = 80,
        padding: bool = False,
        padding_value: str = "",
    ) -> None:
        """Initializes the RandomLoadText class with given parameters."""
        self.prompt_format = prompt_format
        self.neg_samples = neg_samples
        self.max_samples = max_samples
        self.padding = padding
        self.padding_value = padding_value

    def __call__(self, labels: dict) -> dict:
        """Return updated classes and texts."""
        assert "texts" in labels, "No texts found in labels."
        class_texts = labels["texts"]
        num_classes = len(class_texts)
        cls = np.asarray(labels.pop("cls"), dtype=int)
        pos_labels = np.unique(cls).tolist()

        if len(pos_labels) > self.max_samples:
            pos_labels = set(random.sample(pos_labels, k=self.max_samples))

        neg_samples = min(min(num_classes, self.max_samples) - len(pos_labels), random.randint(*self.neg_samples))
        neg_labels = [i for i in range(num_classes) if i not in pos_labels]
        neg_labels = random.sample(neg_labels, k=neg_samples)

        sampled_labels = pos_labels + neg_labels
        random.shuffle(sampled_labels)

        label2ids = {label: i for i, label in enumerate(sampled_labels)}
        valid_idx = np.zeros(len(labels["instances"]), dtype=bool)
        new_cls = []
        for i, label in enumerate(cls.squeeze(-1).tolist()):
            if label not in label2ids:
                continue
            valid_idx[i] = True
            new_cls.append([label2ids[label]])
        labels["instances"] = labels["instances"][valid_idx]
        labels["cls"] = np.array(new_cls)

        # Randomly select one prompt when there's more than one prompts
        texts = []
        for label in sampled_labels:
            prompts = class_texts[label]
            assert len(prompts) > 0
            prompt = self.prompt_format.format(prompts[random.randrange(len(prompts))])
            texts.append(prompt)

        if self.padding:
            valid_labels = len(pos_labels) + len(neg_labels)
            num_padding = self.max_samples - valid_labels
            if num_padding > 0:
                texts += [self.padding_value] * num_padding

        labels["texts"] = texts
        return labels

__call__(labels)

返回更新的课程和课文。

源代码 ultralytics/data/augment.py
def __call__(self, labels: dict) -> dict:
    """Return updated classes and texts."""
    assert "texts" in labels, "No texts found in labels."
    class_texts = labels["texts"]
    num_classes = len(class_texts)
    cls = np.asarray(labels.pop("cls"), dtype=int)
    pos_labels = np.unique(cls).tolist()

    if len(pos_labels) > self.max_samples:
        pos_labels = set(random.sample(pos_labels, k=self.max_samples))

    neg_samples = min(min(num_classes, self.max_samples) - len(pos_labels), random.randint(*self.neg_samples))
    neg_labels = [i for i in range(num_classes) if i not in pos_labels]
    neg_labels = random.sample(neg_labels, k=neg_samples)

    sampled_labels = pos_labels + neg_labels
    random.shuffle(sampled_labels)

    label2ids = {label: i for i, label in enumerate(sampled_labels)}
    valid_idx = np.zeros(len(labels["instances"]), dtype=bool)
    new_cls = []
    for i, label in enumerate(cls.squeeze(-1).tolist()):
        if label not in label2ids:
            continue
        valid_idx[i] = True
        new_cls.append([label2ids[label]])
    labels["instances"] = labels["instances"][valid_idx]
    labels["cls"] = np.array(new_cls)

    # Randomly select one prompt when there's more than one prompts
    texts = []
    for label in sampled_labels:
        prompts = class_texts[label]
        assert len(prompts) > 0
        prompt = self.prompt_format.format(prompts[random.randrange(len(prompts))])
        texts.append(prompt)

    if self.padding:
        valid_labels = len(pos_labels) + len(neg_labels)
        num_padding = self.max_samples - valid_labels
        if num_padding > 0:
            texts += [self.padding_value] * num_padding

    labels["texts"] = texts
    return labels

__init__(prompt_format='{}', neg_samples=(80, 80), max_samples=80, padding=False, padding_value='')

使用给定参数初始化 RandomLoadText 类。

源代码 ultralytics/data/augment.py
def __init__(
    self,
    prompt_format: str = "{}",
    neg_samples: Tuple[int, int] = (80, 80),
    max_samples: int = 80,
    padding: bool = False,
    padding_value: str = "",
) -> None:
    """Initializes the RandomLoadText class with given parameters."""
    self.prompt_format = prompt_format
    self.neg_samples = neg_samples
    self.max_samples = max_samples
    self.padding = padding
    self.padding_value = padding_value



ultralytics.data.augment.ClassifyLetterBox

YOLOv8 用于图像预处理的 LetterBox 类,旨在成为转换管道的一部分,例如:T.Compose([LetterBox(size, ToTensor()])、 T.Compose([LetterBox(size), ToTensor()]).

属性

名称 类型 说明
h int

图像的目标高度。

w int

图像的目标宽度。

auto bool

如果为 True,则使用步长自动求解短边。

stride int

跨距值,当 "auto "为 True 时使用。

源代码 ultralytics/data/augment.py
class ClassifyLetterBox:
    """
    YOLOv8 LetterBox class for image preprocessing, designed to be part of a transformation pipeline, e.g.,
    T.Compose([LetterBox(size), ToTensor()]).

    Attributes:
        h (int): Target height of the image.
        w (int): Target width of the image.
        auto (bool): If True, automatically solves for short side using stride.
        stride (int): The stride value, used when 'auto' is True.
    """

    def __init__(self, size=(640, 640), auto=False, stride=32):
        """
        Initializes the ClassifyLetterBox class with a target size, auto-flag, and stride.

        Args:
            size (Union[int, Tuple[int, int]]): The target dimensions (height, width) for the letterbox.
            auto (bool): If True, automatically calculates the short side based on stride.
            stride (int): The stride value, used when 'auto' is True.
        """
        super().__init__()
        self.h, self.w = (size, size) if isinstance(size, int) else size
        self.auto = auto  # pass max size integer, automatically solve for short side using stride
        self.stride = stride  # used with auto

    def __call__(self, im):
        """
        Resizes the image and pads it with a letterbox method.

        Args:
            im (numpy.ndarray): The input image as a numpy array of shape HWC.

        Returns:
            (numpy.ndarray): The letterboxed and resized image as a numpy array.
        """
        imh, imw = im.shape[:2]
        r = min(self.h / imh, self.w / imw)  # ratio of new/old dimensions
        h, w = round(imh * r), round(imw * r)  # resized image dimensions

        # Calculate padding dimensions
        hs, ws = (math.ceil(x / self.stride) * self.stride for x in (h, w)) if self.auto else (self.h, self.w)
        top, left = round((hs - h) / 2 - 0.1), round((ws - w) / 2 - 0.1)

        # Create padded image
        im_out = np.full((hs, ws, 3), 114, dtype=im.dtype)
        im_out[top : top + h, left : left + w] = cv2.resize(im, (w, h), interpolation=cv2.INTER_LINEAR)
        return im_out

__call__(im)

使用信箱方法调整图像大小并填充。

参数

名称 类型 说明 默认值
im ndarray

输入图像是一个 HWC 形状的 numpy 数组。

所需

返回:

类型 说明
ndarray

以 numpy 数组的形式显示经过信箱化和大小调整的图像。

源代码 ultralytics/data/augment.py
def __call__(self, im):
    """
    Resizes the image and pads it with a letterbox method.

    Args:
        im (numpy.ndarray): The input image as a numpy array of shape HWC.

    Returns:
        (numpy.ndarray): The letterboxed and resized image as a numpy array.
    """
    imh, imw = im.shape[:2]
    r = min(self.h / imh, self.w / imw)  # ratio of new/old dimensions
    h, w = round(imh * r), round(imw * r)  # resized image dimensions

    # Calculate padding dimensions
    hs, ws = (math.ceil(x / self.stride) * self.stride for x in (h, w)) if self.auto else (self.h, self.w)
    top, left = round((hs - h) / 2 - 0.1), round((ws - w) / 2 - 0.1)

    # Create padded image
    im_out = np.full((hs, ws, 3), 114, dtype=im.dtype)
    im_out[top : top + h, left : left + w] = cv2.resize(im, (w, h), interpolation=cv2.INTER_LINEAR)
    return im_out

__init__(size=(640, 640), auto=False, stride=32)

使用目标尺寸、自动标记和 stride 初始化 ClassifyLetterBox 类。

参数

名称 类型 说明 默认值
size Union[int, Tuple[int, int]]

信箱的目标尺寸(高、宽)。

(640, 640)
auto bool

如果为 True,则根据步幅自动计算短边。

False
stride int

跨距值,当 "auto "为 True 时使用。

32
源代码 ultralytics/data/augment.py
def __init__(self, size=(640, 640), auto=False, stride=32):
    """
    Initializes the ClassifyLetterBox class with a target size, auto-flag, and stride.

    Args:
        size (Union[int, Tuple[int, int]]): The target dimensions (height, width) for the letterbox.
        auto (bool): If True, automatically calculates the short side based on stride.
        stride (int): The stride value, used when 'auto' is True.
    """
    super().__init__()
    self.h, self.w = (size, size) if isinstance(size, int) else size
    self.auto = auto  # pass max size integer, automatically solve for short side using stride
    self.stride = stride  # used with auto



ultralytics.data.augment.CenterCrop

YOLOv8 用于图像预处理的 CenterCrop 类,旨在成为转换管道的一部分,例如:T.Compose([CenterCrop(size, ToTensor()])、 T.Compose([CenterCrop(size), ToTensor()]).

源代码 ultralytics/data/augment.py
class CenterCrop:
    """YOLOv8 CenterCrop class for image preprocessing, designed to be part of a transformation pipeline, e.g.,
    T.Compose([CenterCrop(size), ToTensor()]).
    """

    def __init__(self, size=640):
        """Converts an image from numpy array to PyTorch tensor."""
        super().__init__()
        self.h, self.w = (size, size) if isinstance(size, int) else size

    def __call__(self, im):
        """
        Resizes and crops the center of the image using a letterbox method.

        Args:
            im (numpy.ndarray): The input image as a numpy array of shape HWC.

        Returns:
            (numpy.ndarray): The center-cropped and resized image as a numpy array.
        """
        imh, imw = im.shape[:2]
        m = min(imh, imw)  # min dimension
        top, left = (imh - m) // 2, (imw - m) // 2
        return cv2.resize(im[top : top + m, left : left + m], (self.w, self.h), interpolation=cv2.INTER_LINEAR)

__call__(im)

使用信箱方法调整和裁剪图像中心。

参数

名称 类型 说明 默认值
im ndarray

输入图像是一个 HWC 形状的 numpy 数组。

所需

返回:

类型 说明
ndarray

经过居中裁剪和大小调整的图像,作为一个 numpy 数组。

源代码 ultralytics/data/augment.py
def __call__(self, im):
    """
    Resizes and crops the center of the image using a letterbox method.

    Args:
        im (numpy.ndarray): The input image as a numpy array of shape HWC.

    Returns:
        (numpy.ndarray): The center-cropped and resized image as a numpy array.
    """
    imh, imw = im.shape[:2]
    m = min(imh, imw)  # min dimension
    top, left = (imh - m) // 2, (imw - m) // 2
    return cv2.resize(im[top : top + m, left : left + m], (self.w, self.h), interpolation=cv2.INTER_LINEAR)

__init__(size=640)

将图像从 numpy 数组转换为PyTorch tensor 。

源代码 ultralytics/data/augment.py
def __init__(self, size=640):
    """Converts an image from numpy array to PyTorch tensor."""
    super().__init__()
    self.h, self.w = (size, size) if isinstance(size, int) else size



ultralytics.data.augment.ToTensor

YOLOv8 用于图像预处理的 ToTensor 类,即 T.Compose([LetterBox(size), ToTensor()])。

源代码 ultralytics/data/augment.py
class ToTensor:
    """YOLOv8 ToTensor class for image preprocessing, i.e., T.Compose([LetterBox(size), ToTensor()])."""

    def __init__(self, half=False):
        """Initialize YOLOv8 ToTensor object with optional half-precision support."""
        super().__init__()
        self.half = half

    def __call__(self, im):
        """
        Transforms an image from a numpy array to a PyTorch tensor, applying optional half-precision and normalization.

        Args:
            im (numpy.ndarray): Input image as a numpy array with shape (H, W, C) in BGR order.

        Returns:
            (torch.Tensor): The transformed image as a PyTorch tensor in float32 or float16, normalized to [0, 1].
        """
        im = np.ascontiguousarray(im.transpose((2, 0, 1))[::-1])  # HWC to CHW -> BGR to RGB -> contiguous
        im = torch.from_numpy(im)  # to torch
        im = im.half() if self.half else im.float()  # uint8 to fp16/32
        im /= 255.0  # 0-255 to 0.0-1.0
        return im

__call__(im)

将图像从 numpy 数组转换为PyTorch tensor ,应用可选的半精度和归一化。

参数

名称 类型 说明 默认值
im ndarray

输入图像为一个按 BGR 顺序排列的形状为(高、宽、高)的 numpy 数组。

所需

返回:

类型 说明
Tensor

变换后的图像PyTorch tensor ,单位为 float32 或 float16,归一化为 [0,1]。

源代码 ultralytics/data/augment.py
def __call__(self, im):
    """
    Transforms an image from a numpy array to a PyTorch tensor, applying optional half-precision and normalization.

    Args:
        im (numpy.ndarray): Input image as a numpy array with shape (H, W, C) in BGR order.

    Returns:
        (torch.Tensor): The transformed image as a PyTorch tensor in float32 or float16, normalized to [0, 1].
    """
    im = np.ascontiguousarray(im.transpose((2, 0, 1))[::-1])  # HWC to CHW -> BGR to RGB -> contiguous
    im = torch.from_numpy(im)  # to torch
    im = im.half() if self.half else im.float()  # uint8 to fp16/32
    im /= 255.0  # 0-255 to 0.0-1.0
    return im

__init__(half=False)

初始化YOLOv8 ToTensor 对象,可选择支持半精度。

源代码 ultralytics/data/augment.py
def __init__(self, half=False):
    """Initialize YOLOv8 ToTensor object with optional half-precision support."""
    super().__init__()
    self.half = half



ultralytics.data.augment.v8_transforms(dataset, imgsz, hyp, stretch=False)

将图像转换为适合YOLOv8 训练的大小。

源代码 ultralytics/data/augment.py
def v8_transforms(dataset, imgsz, hyp, stretch=False):
    """Convert images to a size suitable for YOLOv8 training."""
    pre_transform = Compose(
        [
            Mosaic(dataset, imgsz=imgsz, p=hyp.mosaic),
            CopyPaste(p=hyp.copy_paste),
            RandomPerspective(
                degrees=hyp.degrees,
                translate=hyp.translate,
                scale=hyp.scale,
                shear=hyp.shear,
                perspective=hyp.perspective,
                pre_transform=None if stretch else LetterBox(new_shape=(imgsz, imgsz)),
            ),
        ]
    )
    flip_idx = dataset.data.get("flip_idx", [])  # for keypoints augmentation
    if dataset.use_keypoints:
        kpt_shape = dataset.data.get("kpt_shape", None)
        if len(flip_idx) == 0 and hyp.fliplr > 0.0:
            hyp.fliplr = 0.0
            LOGGER.warning("WARNING ⚠️ No 'flip_idx' array defined in data.yaml, setting augmentation 'fliplr=0.0'")
        elif flip_idx and (len(flip_idx) != kpt_shape[0]):
            raise ValueError(f"data.yaml flip_idx={flip_idx} length must be equal to kpt_shape[0]={kpt_shape[0]}")

    return Compose(
        [
            pre_transform,
            MixUp(dataset, pre_transform=pre_transform, p=hyp.mixup),
            Albumentations(p=1.0),
            RandomHSV(hgain=hyp.hsv_h, sgain=hyp.hsv_s, vgain=hyp.hsv_v),
            RandomFlip(direction="vertical", p=hyp.flipud),
            RandomFlip(direction="horizontal", p=hyp.fliplr, flip_idx=flip_idx),
        ]
    )  # transforms



ultralytics.data.augment.classify_transforms(size=224, mean=DEFAULT_MEAN, std=DEFAULT_STD, interpolation=Image.BILINEAR, crop_fraction=DEFAULT_CROP_FRACTION)

用于评估/推断的分类变换。灵感来自 timm/data/transforms_factory.py。

参数

名称 类型 说明 默认值
size int

图像大小

224
mean tuple

RGB 通道的平均值

DEFAULT_MEAN
std tuple

RGB 通道的标准值

DEFAULT_STD
interpolation InterpolationMode

默认为 T.InterpolationMode.BILINEAR。

BILINEAR
crop_fraction float

要裁剪图像的分数,默认为 1.0。

DEFAULT_CROP_FRACTION

返回:

类型 说明
Compose

火炬视觉转换

源代码 ultralytics/data/augment.py
def classify_transforms(
    size=224,
    mean=DEFAULT_MEAN,
    std=DEFAULT_STD,
    interpolation=Image.BILINEAR,
    crop_fraction: float = DEFAULT_CROP_FRACTION,
):
    """
    Classification transforms for evaluation/inference. Inspired by timm/data/transforms_factory.py.

    Args:
        size (int): image size
        mean (tuple): mean values of RGB channels
        std (tuple): std values of RGB channels
        interpolation (T.InterpolationMode): interpolation mode. default is T.InterpolationMode.BILINEAR.
        crop_fraction (float): fraction of image to crop. default is 1.0.

    Returns:
        (T.Compose): torchvision transforms
    """
    import torchvision.transforms as T  # scope for faster 'import ultralytics'

    if isinstance(size, (tuple, list)):
        assert len(size) == 2
        scale_size = tuple(math.floor(x / crop_fraction) for x in size)
    else:
        scale_size = math.floor(size / crop_fraction)
        scale_size = (scale_size, scale_size)

    # Aspect ratio is preserved, crops center within image, no borders are added, image is lost
    if scale_size[0] == scale_size[1]:
        # Simple case, use torchvision built-in Resize with the shortest edge mode (scalar size arg)
        tfl = [T.Resize(scale_size[0], interpolation=interpolation)]
    else:
        # Resize the shortest edge to matching target dim for non-square target
        tfl = [T.Resize(scale_size)]
    tfl += [T.CenterCrop(size)]

    tfl += [
        T.ToTensor(),
        T.Normalize(
            mean=torch.tensor(mean),
            std=torch.tensor(std),
        ),
    ]

    return T.Compose(tfl)



ultralytics.data.augment.classify_augmentations(size=224, mean=DEFAULT_MEAN, std=DEFAULT_STD, scale=None, ratio=None, hflip=0.5, vflip=0.0, auto_augment=None, hsv_h=0.015, hsv_s=0.4, hsv_v=0.4, force_color_jitter=False, erasing=0.0, interpolation=Image.BILINEAR)

用于训练的增强分类变换。灵感来自 timm/data/transforms_factory.py。

参数

名称 类型 说明 默认值
size int

图像大小

224
scale tuple

图像的比例范围。默认为(0.08, 1.0)

None
ratio tuple

默认为(3./4.、4./3.)。

None
mean tuple

RGB 通道的平均值

DEFAULT_MEAN
std tuple

RGB 通道的标准值

DEFAULT_STD
hflip float

横向翻转概率

0.5
vflip float

垂直翻转概率

0.0
auto_augment str

自动增强策略。可以是 "randaugment"、"augmix"、"autoaugment "或无。

None
hsv_h float

图像 HSV 色调增强(分数)

0.015
hsv_s float

图像 HSV 饱和度增强(分数)

0.4
hsv_v float

图像 HSV 值增强(分数)

0.4
force_color_jitter bool

即使启用了自动增强功能,也会强制应用色彩抖动

False
erasing float

随机擦除概率

0.0
interpolation InterpolationMode

默认为 T.InterpolationMode.BILINEAR。

BILINEAR

返回:

类型 说明
Compose

火炬视觉转换

源代码 ultralytics/data/augment.py
def classify_augmentations(
    size=224,
    mean=DEFAULT_MEAN,
    std=DEFAULT_STD,
    scale=None,
    ratio=None,
    hflip=0.5,
    vflip=0.0,
    auto_augment=None,
    hsv_h=0.015,  # image HSV-Hue augmentation (fraction)
    hsv_s=0.4,  # image HSV-Saturation augmentation (fraction)
    hsv_v=0.4,  # image HSV-Value augmentation (fraction)
    force_color_jitter=False,
    erasing=0.0,
    interpolation=Image.BILINEAR,
):
    """
    Classification transforms with augmentation for training. Inspired by timm/data/transforms_factory.py.

    Args:
        size (int): image size
        scale (tuple): scale range of the image. default is (0.08, 1.0)
        ratio (tuple): aspect ratio range of the image. default is (3./4., 4./3.)
        mean (tuple): mean values of RGB channels
        std (tuple): std values of RGB channels
        hflip (float): probability of horizontal flip
        vflip (float): probability of vertical flip
        auto_augment (str): auto augmentation policy. can be 'randaugment', 'augmix', 'autoaugment' or None.
        hsv_h (float): image HSV-Hue augmentation (fraction)
        hsv_s (float): image HSV-Saturation augmentation (fraction)
        hsv_v (float): image HSV-Value augmentation (fraction)
        force_color_jitter (bool): force to apply color jitter even if auto augment is enabled
        erasing (float): probability of random erasing
        interpolation (T.InterpolationMode): interpolation mode. default is T.InterpolationMode.BILINEAR.

    Returns:
        (T.Compose): torchvision transforms
    """
    # Transforms to apply if Albumentations not installed
    import torchvision.transforms as T  # scope for faster 'import ultralytics'

    if not isinstance(size, int):
        raise TypeError(f"classify_transforms() size {size} must be integer, not (list, tuple)")
    scale = tuple(scale or (0.08, 1.0))  # default imagenet scale range
    ratio = tuple(ratio or (3.0 / 4.0, 4.0 / 3.0))  # default imagenet ratio range
    primary_tfl = [T.RandomResizedCrop(size, scale=scale, ratio=ratio, interpolation=interpolation)]
    if hflip > 0.0:
        primary_tfl += [T.RandomHorizontalFlip(p=hflip)]
    if vflip > 0.0:
        primary_tfl += [T.RandomVerticalFlip(p=vflip)]

    secondary_tfl = []
    disable_color_jitter = False
    if auto_augment:
        assert isinstance(auto_augment, str)
        # color jitter is typically disabled if AA/RA on,
        # this allows override without breaking old hparm cfgs
        disable_color_jitter = not force_color_jitter

        if auto_augment == "randaugment":
            if TORCHVISION_0_11:
                secondary_tfl += [T.RandAugment(interpolation=interpolation)]
            else:
                LOGGER.warning('"auto_augment=randaugment" requires torchvision >= 0.11.0. Disabling it.')

        elif auto_augment == "augmix":
            if TORCHVISION_0_13:
                secondary_tfl += [T.AugMix(interpolation=interpolation)]
            else:
                LOGGER.warning('"auto_augment=augmix" requires torchvision >= 0.13.0. Disabling it.')

        elif auto_augment == "autoaugment":
            if TORCHVISION_0_10:
                secondary_tfl += [T.AutoAugment(interpolation=interpolation)]
            else:
                LOGGER.warning('"auto_augment=autoaugment" requires torchvision >= 0.10.0. Disabling it.')

        else:
            raise ValueError(
                f'Invalid auto_augment policy: {auto_augment}. Should be one of "randaugment", '
                f'"augmix", "autoaugment" or None'
            )

    if not disable_color_jitter:
        secondary_tfl += [T.ColorJitter(brightness=hsv_v, contrast=hsv_v, saturation=hsv_s, hue=hsv_h)]

    final_tfl = [
        T.ToTensor(),
        T.Normalize(mean=torch.tensor(mean), std=torch.tensor(std)),
        T.RandomErasing(p=erasing, inplace=True),
    ]

    return T.Compose(primary_tfl + secondary_tfl + final_tfl)





Created 2023-11-12, Updated 2024-06-02
Authors: glenn-jocher (6), Burhan-Q (1), Laughing-q (1)