class

ReflectionPad1d

extendsModule

ReflectionPad1d(padding: _Size2d)

source edit

Pad a 3-D tensor (N, C, L) by reflecting at the sequence boundaries.

Unlike constant padding, reflection padding mirrors the input signal itself at the boundary rather than filling with a fixed value. For a left pad of size $p$ , the new values at positions $0, 1, \dots, p-1$ are set to $x[\dots, p], x[\dots, p-1], \dots, x[\dots, 1]$ respectively (the boundary pixel is excluded).

\text{output}[n, c, i] = x\!\left[n, c,\; |i - p_{\text{left}}|\right] \quad \text{for } i < p_{\text{left}}

Parameters

paddingint or tuple[int, int]

(left, right) reflection padding sizes. Each value must be strictly less than the corresponding input dimension. A single int pads equally on both sides.

Attributes

paddingtuple[int, int]

Normalised (left, right) padding.

Notes

Input: $(N, C, L)$ .
Output: $(N, C, L + p_{\text{left}} + p_{\text{right}})$ .

Reflection padding avoids the abrupt discontinuity introduced by zero or constant padding, making it preferable for tasks sensitive to border artefacts (image style transfer, super-resolution).
The padding size must be strictly less than the input size on the padded dimension.

Examples

**Mirror the ends of an audio waveform to avoid boundary clicks:**
>>> import lucid
>>> import lucid.nn as nn
>>>
>>> pad = nn.ReflectionPad1d(padding=4)
>>> x = lucid.zeros(2, 1, 1024)     # (N, channels, T)
>>> pad(x).shape
(2, 1, 1032)
**Asymmetric reflection for causal-ish temporal convolution:**
>>> pad = nn.ReflectionPad1d(padding=(8, 0))
>>> x = lucid.zeros(4, 16, 200)
>>> pad(x).shape
(4, 16, 208)

Used by 1

lucid.nn.modules

Constructors

dunder

init

→None

__init__(padding: _Size2d)

source edit

Initialise the ReflectionPad1d 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.

**Mirror the ends of an audio waveform to avoid boundary clicks:** >>> import lucid >>> import lucid.nn as nn >>> >>> pad = nn.ReflectionPad1d(padding=4) >>> x = lucid.zeros(2, 1, 1024) # (N, channels, T) >>> pad(x).shape (2, 1, 1032) **Asymmetric reflection for causal-ish temporal convolution:** >>> pad = nn.ReflectionPad1d(padding=(8, 0)) >>> x = lucid.zeros(4, 16, 200) >>> pad(x).shape (4, 16, 208)