class

MixupCollator

MixupCollator(alpha: float = 0.2, num_classes: int, p: float = 1.0)

source edit

Mixup collator (Zhang et al., 2018 — arXiv:1710.09412).

For each batch:

Run default_collate to get (images, labels).
One-hot encode labels to (B, num_classes) soft targets.
With probability p, sample $\lambda \sim \mathrm{Beta}(\alpha, \alpha)$ , choose a random permutation of the batch, and produce

\begin{aligned} x_{\mathrm{mix}} &= \lambda x + (1 - \lambda) x_{\pi} \\ y_{\mathrm{mix}} &= \lambda y + (1 - \lambda) y_{\pi} \end{aligned}

Otherwise the batch passes through with one-hot soft targets.

Parameters

alphafloat= 0.2

Beta-distribution concentration. Reference-framework recipes use 0.2 for ImageNet, 1.0 for CIFAR.

num_classesint

Output class count — required so soft targets have the right width.

pfloat= 1.0

Probability of applying Mixup per batch.

Notes

The companion loss is lucid.nn.functional.cross_entropy with the soft target shape (B, num_classes) — Lucid's cross_entropy accepts that natively.

Used by 1

lucid.utils.data

Constructors

dunder

init

→None

__init__(alpha: float = 0.2, num_classes: int, p: float = 1.0)

source edit

dunder

call

→tuple[Tensor, Tensor]

__call__(batch: list[object])

source edit

Dunder methods

dunder

repr

→str

__repr__()

source edit

class

MixupCollator

MixupCollator(alpha: float = 0.2, num_classes: int, p: float = 1.0)

source edit

Mixup collator (Zhang et al., 2018 — arXiv:1710.09412).

For each batch:

Run default_collate to get (images, labels).
One-hot encode labels to (B, num_classes) soft targets.
With probability p, sample $\lambda \sim \mathrm{Beta}(\alpha, \alpha)$ , choose a random permutation of the batch, and produce

\begin{aligned} x_{\mathrm{mix}} &= \lambda x + (1 - \lambda) x_{\pi} \\ y_{\mathrm{mix}} &= \lambda y + (1 - \lambda) y_{\pi} \end{aligned}

Otherwise the batch passes through with one-hot soft targets.

Parameters

alphafloat= 0.2

Beta-distribution concentration. Reference-framework recipes use 0.2 for ImageNet, 1.0 for CIFAR.

num_classesint

Output class count — required so soft targets have the right width.

pfloat= 1.0

Probability of applying Mixup per batch.

Notes

The companion loss is lucid.nn.functional.cross_entropy with the soft target shape (B, num_classes) — Lucid's cross_entropy accepts that natively.

Used by 1

lucid.utils.data

Constructors

dunder

init

→None

__init__(alpha: float = 0.2, num_classes: int, p: float = 1.0)

source edit

dunder

call

→tuple[Tensor, Tensor]

__call__(batch: list[object])

source edit

Dunder methods

dunder

repr

→str

__repr__()

source edit