paper：CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features

前言

之前的数据增强方法存在的问题：

mixup：混合后的图像在局部是模糊和不自然的，因此会混淆模型，尤其是在定位方面。

cutout：被cutout的部分通常用0或者随机噪声填充，这就导致在训练过程中这部分的信息被浪费掉了。

cutmix在cutout的基础上进行改进，cutout的部分用另一张图像上cutout的部分进行填充，这样即保留了cutout的优点：让模型从目标的部分视图去学习目标的特征，让模型更关注那些less discriminative的部分。同时比cutout更高效，cutout的部分用另一张图像的部分进行填充，让模型同时学习两个目标的特征。

从下图可以看出，虽然Mixup和Cutout都提升了模型的分类精度，但在若监督定位和目标检测性能上都有不同程度的下降，而CutMix则在各个任务上都获得了显著的性能提升。

CutMix

cutmix的具体过程如下

其中\(M\in\left \{ 0,1 \right \}^{W\times H}\)是一个binary mask表明从两张图中裁剪的patch的位置，和mixup一样，\(\lambda\)也是通过\(\beta(\alpha, \alpha)\)分布得到的，在文章中作者设置\(\alpha=1\)，因此\(\lambda\)是从均匀分布\((0,1)\)中采样的。

为了得到mask，首先要确定cutmix的bounding box的坐标\(B=(r_{x},r_{y},r_{w},r_{h})\)，其值通过下式得到

即 \(\lambda\) 确定了patch与原图的面积比，即A图cutout的面积越大，标签融合时A图的比例越小。

代码实现

下面是torchvision的官方实现

class RandomCutmix(torch.nn.Module):
    """Randomly apply Cutmix to the provided batch and targets.
    The class implements the data augmentations as described in the paper
    `"CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features"
    <https://arxiv.org/abs/1905.04899>`_.

    Args:
        num_classes (int): number of classes used for one-hot encoding.
        p (float): probability of the batch being transformed. Default value is 0.5.
        alpha (float): hyperparameter of the Beta distribution used for cutmix.
            Default value is 1.0.
        inplace (bool): boolean to make this transform inplace. Default set to False.
    """

    def __init__(self, num_classes: int, p: float = 0.5, alpha: float = 1.0, inplace: bool = False) -> None:
        super().__init__()
        if num_classes < 1:
            raise ValueError("Please provide a valid positive value for the num_classes.")
        if alpha <= 0:
            raise ValueError("Alpha param can't be zero.")

        self.num_classes = num_classes
        self.p = p
        self.alpha = alpha
        self.inplace = inplace

    def forward(self, batch: Tensor, target: Tensor) -> Tuple[Tensor, Tensor]:
        """
        Args:
            batch (Tensor): Float tensor of size (B, C, H, W)
            target (Tensor): Integer tensor of size (B, )

        Returns:
            Tensor: Randomly transformed batch.
        """
        if batch.ndim != 4:
            raise ValueError(f"Batch ndim should be 4. Got {batch.ndim}")
        if target.ndim != 1:
            raise ValueError(f"Target ndim should be 1. Got {target.ndim}")
        if not batch.is_floating_point():
            raise TypeError(f"Batch dtype should be a float tensor. Got {batch.dtype}.")
        if target.dtype != torch.int64:
            raise TypeError(f"Target dtype should be torch.int64. Got {target.dtype}")

        if not self.inplace:
            batch = batch.clone()
            target = target.clone()

        if target.ndim == 1:
            target = torch.nn.functional.one_hot(target, num_classes=self.num_classes).to(dtype=batch.dtype)

        if torch.rand(1).item() >= self.p:
            return batch, target

        # It's faster to roll the batch by one instead of shuffling it to create image pairs
        batch_rolled = batch.roll(1, 0)
        target_rolled = target.roll(1, 0)

        # Implemented as on cutmix paper, page 12 (with minor corrections on typos).
        lambda_param = float(torch._sample_dirichlet(torch.tensor([self.alpha, self.alpha]))[0])
        _, H, W = F.get_dimensions(batch)

        r_x = torch.randint(W, (1,))
        r_y = torch.randint(H, (1,))

        r = 0.5 * math.sqrt(1.0 - lambda_param)
        r_w_half = int(r * W)
        r_h_half = int(r * H)

        x1 = int(torch.clamp(r_x - r_w_half, min=0))
        y1 = int(torch.clamp(r_y - r_h_half, min=0))
        x2 = int(torch.clamp(r_x + r_w_half, max=W))
        y2 = int(torch.clamp(r_y + r_h_half, max=H))

        batch[:, :, y1:y2, x1:x2] = batch_rolled[:, :, y1:y2, x1:x2]
        lambda_param = float(1.0 - (x2 - x1) * (y2 - y1) / (W * H))

        target_rolled.mul_(1.0 - lambda_param)
        target.mul_(lambda_param).add_(target_rolled)

        return batch, target

    def __repr__(self) -> str:
        s = (
            f"{self.__class__.__name__}("
            f"num_classes={self.num_classes}"
            f", p={self.p}"
            f", alpha={self.alpha}"
            f", inplace={self.inplace}"
            f")"
        )
        return s

实验结果

从下图可以看出，CutMix在ImageNet上的精度超过了Cutout和Mixup等数据增强方法

在若监督目标定位方面，CutMix也超过了Mixup和Cutout

当作为预训练模型迁移到其它下游任务比如目标检测和图像描述时，CutMix也取得了最好的效果