class

ReflectionPad2d

extendsModule

ReflectionPad2d(padding: int | tuple[int, int, int, int])

source

Pad a 4-D tensor (N, C, H, W) by reflecting at all four spatial edges.

Reflection padding mirrors the pixel values at each boundary, producing a smooth, artefact-free extension of the image. This is particularly effective for convolutional image processing tasks where constant-value borders would introduce ringing or aliasing.

The reflected value at an out-of-bounds index $i$ (with respect to input size $S$ and left-pad $p$ ) is:

\text{output}[\dots, i] = x[\dots,\; |i - p|] \quad (i < p)

Parameters

paddingint or tuple[int, int, int, int]

(left, right, top, bottom) reflection padding sizes. Each value must be strictly less than the corresponding spatial input size. A single int pads equally on all four sides.

Attributes

paddingtuple[int, int, int, int]

Normalised (left, right, top, bottom) padding.

Notes

Input: $(N, C, H, W)$ .
Output: $(N, C, H + p_{\text{top}} + p_{\text{bottom}}, W + p_{\text{left}} + p_{\text{right}})$ .

Widely used in Neural Style Transfer and image-to-image translation architectures (e.g. Johnson et al., 2016) to avoid border colour bleeding caused by zero padding.

Examples

**Style-transfer encoder with reflection padding:**
>>> import lucid
>>> import lucid.nn as nn
>>>
>>> pad = nn.ReflectionPad2d(padding=3)
>>> conv = nn.Conv2d(3, 64, kernel_size=7)
>>> x = lucid.zeros(1, 3, 256, 256)
>>> conv(pad(x)).shape
(1, 64, 256, 256)    # 256 + 2*3 - 6 = 256 (same-size output)
**Asymmetric reflection:**
>>> pad = nn.ReflectionPad2d(padding=(1, 2, 0, 3))
>>> x = lucid.zeros(2, 8, 32, 32)
>>> pad(x).shape
(2, 8, 35, 35)

Methods (3)

dunder

init

→None

__init__(padding: int | tuple[int, int, int, int])

source

Initialise the ReflectionPad2d module. See the class docstring for parameter semantics.

forward

→Tensor

forward(x: Tensor)

source

Pad the input tensor according to the configured padding.

Parameters

inputTensor

Input tensor of shape

(N, C, *)

Returns

Tensor

Padded tensor with spatial dimensions expanded by the configured padding amounts.

extra_repr

→str

extra_repr()

source

Return a string representation of the layer's configuration.