5. Yolov8 Mixup数据增强详解

Mixup简介

        mixup是一种运用在计算机视觉中的对图像进行混类增强的算法，它可以将不同的图像进行混合，从而扩充训练数据集。
$$\lambda =Beta(\alpha ,\beta )$$$$img\_mix=im{g}_1\ast \lambda +im{g}_2\ast (1-\lambda )$$其中$\lambda$是由参数为$\alpha$，$\beta$的贝塔分布计算出来的混合系数，$im{g}_1$和$im{g}_2$是进行混合的样本。无论如何设置$\alpha$，$\beta$的值，期望$\frac{\alpha }{\alpha +\beta }$ 始终近似为0.5。

代码详解

import random
import numpy as np

class BaseMixTransform:
    """This implementation is from mmyolo"""

    def __init__(self, dataset, pre_transform=None, p=0.0) -> None:
        # dataset中存储着读取的数据（图片、类别和真实框位置信息）
        self.dataset = dataset
        # 进行数据增强前进行的数据预处理
        self.pre_transform = pre_transform
        # 进行数据增强的概率
        self.p = p

    def __call__(self, labels):
        # 随机进行数据增强
        if random.uniform(0, 1) > self.p:
            return labels

        # 获取一个或三个其他图像的索引
        indexes = self.get_indexes()
        if isinstance(indexes, int):
            indexes = [indexes]

        # 获取图像信息用于 Mosaic 或者 MixUp 数据增强
        mix_labels = [self.dataset.get_label_info(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

        # 进行 Mosaic 或 MixUp
        labels = self._mix_transform(labels)
        labels.pop("mix_labels", None)
        return labels

    def _mix_transform(self, labels):
        raise NotImplementedError

    def get_indexes(self):
        raise NotImplementedError

class MixUp(BaseMixTransform):

    def __init__(self, dataset, pre_transform=None, p=0.0) -> None:
        super().__init__(dataset=dataset, pre_transform=pre_transform, p=p)

    # 获取另一个图像的索引
    def get_indexes(self):
        return random.randint(0, len(self.dataset) - 1)

    def _mix_transform(self, labels):
        # Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf
        # 混合系数
        r = np.random.beta(32.0, 32.0)  # mixup ratio, alpha=beta=32.0
        labels2 = labels["mix_labels"][0]
        # 根据混合系数将两张图片mixup
        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
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62