class

Upsample

extendsModule

Upsample(size: int | tuple[int, ...] | None = None, scale_factor: float | tuple[float, ...] | None = None, mode: str = 'nearest', align_corners: bool | None = None)

source

Upsample an input tensor to a given spatial size or scale factor.

Upsample resizes the spatial dimensions of an input tensor using one of several interpolation algorithms. The batch and channel dimensions are never modified; only the trailing spatial axes are scaled.

Exactly one of size or scale_factor must be provided.

Nearest-neighbour (mode='nearest'):

y[i] = x\!\left[\left\lfloor \frac{i}{\text{scale}} \right\rfloor\right]

Each output element copies the value of the nearest input element.

Bilinear (mode='bilinear', 2-D only):

y[i, j] = (1 - t_h)(1 - t_w)\, x[i_0, j_0] + t_h(1 - t_w)\, x[i_1, j_0] + (1 - t_h) t_w\, x[i_0, j_1] + t_h t_w\, x[i_1, j_1]

where $i_0, i_1$ are the two nearest row indices in the input and $t_h$ is the fractional offset between them (analogously for $j_0, j_1, t_w$ ).

Parameters

sizeint or tuple[int, ...] or None= None

Output spatial size. Pass a single int to set all spatial dimensions to the same value, or a tuple matching the number of spatial dimensions. Mutually exclusive with scale_factor.

scale_factorfloat or tuple[float, ...] or None= None

Multiplier for each spatial dimension. Values > 1 upsample; values < 1 downsample. Mutually exclusive with size.

modestr= 'nearest'

Interpolation algorithm. One of:

'nearest' — no learnable parameters, fastest (default).
'linear' — 1-D linear interpolation (3-D input).
'bilinear' — 2-D bilinear interpolation (4-D input).
'bicubic' — 2-D bicubic interpolation (4-D input).
'trilinear' — 3-D trilinear interpolation (5-D input).

align_cornersbool or None= None

When True, the corner pixels of input and output are aligned, meaning the interpolation grid spans exactly from the first to the last pixel centre. When False (or None), the grid is scaled by the ratio of sizes, which may shift the grid by half a pixel relative to the corners. Ignored for 'nearest' mode.

Attributes

sizeint or tuple[int, ...] or None

Stored output size.

scale_factorfloat or tuple[float, ...] or None

Stored scale factor.

modestr

Stored interpolation mode.

align_cornersbool or None

Stored corner-alignment flag.

Notes

Input: $(N, C, d_1, d_2, \dots, d_k)$ .
Output: $(N, C, d_1', d_2', \dots, d_k')$ where each $d_i'$ is either given by size or $d_i' = \lfloor d_i \cdot \text{scale\_factor} \rfloor$ .

The align_corners flag has a significant effect on output values for bilinear / bicubic / trilinear modes. When True, corner pixels map exactly onto each other; when False the mapping preserves the overall scale but shifts by 0.5 input pixels at the boundary.
For 4-D inputs with mode='nearest' this module is equivalent to UpsamplingNearest2d; for mode='bilinear' with align_corners=True it is equivalent to UpsamplingBilinear2d.
Delegates to nn.functional.interpolate.

Examples

**2× nearest-neighbour upsampling of a feature map:**
>>> import lucid
>>> import lucid.nn as nn
>>>
>>> up = nn.Upsample(scale_factor=2, mode="nearest")
>>> x = lucid.zeros(1, 16, 14, 14)
>>> up(x).shape
(1, 16, 28, 28)
**Fixed output size with bilinear interpolation:**
>>> up = nn.Upsample(size=(256, 256), mode="bilinear", align_corners=False)
>>> x = lucid.zeros(4, 3, 64, 64)
>>> up(x).shape
(4, 3, 256, 256)
**Trilinear upsampling for 3-D volumetric data:**
>>> up_3d = nn.Upsample(scale_factor=2, mode="trilinear", align_corners=True)
>>> x = lucid.zeros(2, 8, 16, 16, 16)
>>> up_3d(x).shape
(2, 8, 32, 32, 32)

Methods (3)

dunder

init

→None

__init__(size: int | tuple[int, ...] | None = None, scale_factor: float | tuple[float, ...] | None = None, mode: str = 'nearest', align_corners: bool | None = None)

source

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

forward

→Tensor

forward(x: Tensor)

source

Upsample the input tensor.

Parameters

inputTensor

Input tensor of shape

(N, C, *)

Returns

Tensor

Upsampled output tensor.

extra_repr

→str

extra_repr()

source

Return a string representation of the layer's configuration.