class

RandomErasing

extendsPhotometricTransform[_ErasingParams]

RandomErasing(p: float = 0.5, scale: tuple[float, float] = (0.02, 0.33), ratio: tuple[float, float] = (0.3, 3.3), value: float | tuple[float, ...] | str = 0.0)

source edit

Erase a random rectangular region of the image (Zhong et al., 2017).

At each call, with probability p one rectangle whose area is a fraction of the image (sampled uniformly from scale) and whose aspect ratio is sampled log-uniformly from ratio is replaced by value. If no rectangle fits after 10 attempts the call is a no-op (matches the torchvision contract).

Parameters

pfloat= 0.5

Probability of applying the erase.

scale(float, float)= (0.02, 0.33)

Range of the erased-region area as a fraction of the image area. Sampled uniformly per call.

ratio(float, float)= (0.3, 3.3)

Range of the erased-region aspect ratio (W / H). Sampled log-uniformly per call (matches the reference implementation).

valuefloat or tuple of float or {"random"}= 0.0

Fill value:

float → fill the rectangle with this scalar (broadcast across channels).
tuple of length C → per-channel constants (e.g. the ImageNet pixel mean).
"random" → fill with i.i.d. samples from $\mathcal{N}(0, 1)$ , drawn fresh on each call.

Notes

The erase is performed via a multiplicative keep-mask + additive fill (same pattern as CoarseDropout) so the operation composes cleanly with autograd and avoids any in-place writes.

The log-uniform aspect ratio means ratio=(0.3, 3.3) samples landscape and portrait rectangles symmetrically, exactly matching the spec in Zhong et al., 2017 §3.2.

Examples

Standard ImageNet recipe (p=0.25):
>>> import lucid
>>> from lucid.utils.transforms import RandomErasing
>>> tf = RandomErasing(p=0.25, scale=(0.02, 0.33), ratio=(0.3, 3.3))
>>> x = lucid.rand(3, 224, 224)
>>> y = tf(x)
>>> tuple(y.shape)
(3, 224, 224)
Per-channel ImageNet-mean fill:
>>> tf = RandomErasing(p=1.0, value=(0.485, 0.456, 0.406))

Used by 2

Constructors

dunder

init

→None

__init__(p: float = 0.5, scale: tuple[float, float] = (0.02, 0.33), ratio: tuple[float, float] = (0.3, 3.3), value: float | tuple[float, ...] | str = 0.0)

source edit

Instance methods

make_params

→_ErasingParams

make_params(img: Tensor)

source edit

Dunder methods

dunder

repr

→str

__repr__()

source edit

Standard ImageNet recipe (p=0.25): >>> import lucid >>> from lucid.utils.transforms import RandomErasing >>> tf = RandomErasing(p=0.25, scale=(0.02, 0.33), ratio=(0.3, 3.3)) >>> x = lucid.rand(3, 224, 224) >>> y = tf(x) >>> tuple(y.shape) (3, 224, 224) Per-channel ImageNet-mean fill: >>> tf = RandomErasing(p=1.0, value=(0.485, 0.456, 0.406))