tfm.vision.augment.MixupAndCutmix

Applies Mixup and/or Cutmix to a batch of images.

tfm.vision.augment.MixupAndCutmix(
    num_classes: int,
    mixup_alpha: float = 0.8,
    cutmix_alpha: float = 1.0,
    prob: float = 1.0,
    switch_prob: float = 0.5,
    label_smoothing: float = 0.1
)

Implementaion is inspired by https://github.com/rwightman/pytorch-image-models

num_classes int

Number of classes.

mixup_alpha float, optional

For drawing a random lambda (lam) from a beta distribution (for each image). If zero Mixup is deactivated. Defaults to .8.

cutmix_alpha float, optional

For drawing a random lambda (lam) from a beta distribution (for each image). If zero Cutmix is deactivated. Defaults to 1..

prob float, optional

Of augmenting the batch. Defaults to 1.0.

switch_prob float, optional

Probability of applying Cutmix for the batch. Defaults to 0.5.

label_smoothing float, optional

Constant for label smoothing. Defaults to 0.1.

Methods

distort

View source

distort(
    images: tf.Tensor, labels: tf.Tensor
) -> Tuple[tf.Tensor, tf.Tensor]

Applies Mixup and/or Cutmix to batch of images and transforms labels.

Args
images tf.Tensor

Of shape [batch_size, height, width, 3] representing a batch of image, or [batch_size, time, height, width, 3] representing a batch of video.

labels tf.Tensor

Of shape [batch_size, ] representing the class id for each image of the batch.

Returns
Tuple[tf.Tensor, tf.Tensor]: The augmented version of image and labels.

__call__

View source

__call__(
    images: tf.Tensor, labels: tf.Tensor
) -> Tuple[tf.Tensor, tf.Tensor]

Call self as a function.