class

ReplicationPad2d

extendsModule

ReplicationPad2d(padding: int | tuple[int, int, int, int])

source edit

Pad a 4-D tensor (N, C, H, W) by replicating the edge pixels.

Each out-of-bounds position is filled with the value of the nearest in-bounds pixel. Corner out-of-bounds positions replicate the nearest corner pixel.

\text{output}[n, c, i, j] = x\!\left[n, c,\; \text{clip}(i - p_t, 0, H-1),\; \text{clip}(j - p_l, 0, W-1) \right]

Parameters

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

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

Replication padding is preferred over zero padding in networks that process images with meaningful content close to the borders (e.g. panoramic or satellite imagery) because it avoids introducing a dark border artefact that the network would need to learn to ignore.

Examples

**Replicate borders before a 5×5 convolution:**
>>> import lucid
>>> import lucid.nn as nn
>>>
>>> pad = nn.ReplicationPad2d(padding=2)
>>> conv = nn.Conv2d(3, 32, kernel_size=5)
>>> x = lucid.zeros(1, 3, 64, 64)
>>> conv(pad(x)).shape    # same-size output
(1, 32, 64, 64)
**Asymmetric replication:**
>>> pad = nn.ReplicationPad2d(padding=(0, 0, 1, 1))   # top and bottom only
>>> x = lucid.zeros(4, 16, 30, 30)
>>> pad(x).shape
(4, 16, 32, 30)

Used by 1

lucid.nn.modules

Constructors

dunder

init

→None

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

source edit

Initialise the ReplicationPad2d module. See the class docstring for parameter semantics.

Instance methods

extra_repr

→str

extra_repr()

source edit

Return a string representation of the layer's configuration.

forward

→Tensor

forward(x: Tensor)

source edit

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.

**Replicate borders before a 5×5 convolution:** >>> import lucid >>> import lucid.nn as nn >>> >>> pad = nn.ReplicationPad2d(padding=2) >>> conv = nn.Conv2d(3, 32, kernel_size=5) >>> x = lucid.zeros(1, 3, 64, 64) >>> conv(pad(x)).shape # same-size output (1, 32, 64, 64) **Asymmetric replication:** >>> pad = nn.ReplicationPad2d(padding=(0, 0, 1, 1)) # top and bottom only >>> x = lucid.zeros(4, 16, 30, 30) >>> pad(x).shape (4, 16, 32, 30)