interpolate

→Tensor

interpolate(x: Tensor, size: int | tuple[int, ...] | None = None, scale_factor: float | tuple[float, ...] | None = None, mode: str = 'nearest', align_corners: bool | None = None, recompute_scale_factor: bool | None = None, antialias: bool = False)

source edit

Resample an N-D tensor to a target spatial size or scale factor.

Supports the standard family of image / volume resampling kernels. The leading two axes (N batch and C channels) are preserved; only the trailing spatial axes are resampled.

Parameters

xTensor

Input of shape (N, C, *spatial). "nearest", "nearest-exact" and "area" take 1, 2 or 3 spatial axes; "linear" takes 1, "bilinear" / "bicubic" 2 and "trilinear" 3.

sizeint or tuple of int= None

Target spatial size. Mutually exclusive with scale_factor.

scale_factorfloat or tuple of float= None

Multiplicative factor applied to each spatial axis. Output size is floor(in_size * scale_factor).

modestr= 'nearest'

One of "nearest", "nearest-exact", "linear", "bilinear", "bicubic", "trilinear", "area". Default "nearest".

align_cornersbool= None

Only accepted by "linear" / "bilinear" / "bicubic" / "trilinear". If True, the corner pixels of the input and output are exactly aligned, and the values at those corners are preserved. If False (default), pixels are treated as 1×1 squares and the sampling grid is centred.

recompute_scale_factorbool= None

If True, the sampling step is in_size / out_size from the resolved output size rather than 1 / scale_factor. Not allowed together with size.

antialiasbool= False

For "bilinear" and "bicubic": when downsampling, widen the filter over every input sample an output sample covers, so detail finer than the output grid is averaged rather than aliased. Default False.

Returns

Tensor

Resampled tensor of shape (N, C, *out_spatial).

Raises

ValueError

For an unknown mode, a mode given the wrong number of spatial axes, both or neither of size and scale_factor, align_corners or antialias with a mode that has no use for it, or a non-positive size.

TypeError

For an integer input with any mode but the two nearest ones.

Notes

Output sample $d$ reads the input around

x_{\mathrm{src}} = \begin{cases} d \cdot \frac{S_{\mathrm{in}} - 1}{S_{\mathrm{out}} - 1} & \text{align\_corners} \\ (d + 0.5) \cdot r - 0.5 & \text{otherwise} \end{cases}

where $r = 1 / \text{scale\_factor}$ when a scale factor is given and recompute_scale_factor is not True, and $r = S_{\mathrm{in}} / S_{\mathrm{out}}$ otherwise. The two differ whenever $S_{\mathrm{in}} \cdot \text{scale\_factor}$ is not a whole number. "nearest" copies sample $\lfloor d \cdot r \rfloor$ and "nearest-exact" sample $\lfloor (d + 0.5) \cdot r \rfloor$ ; "area" is adaptive average pooling to the output size.

Examples

>>> import lucid
>>> from lucid.nn.functional import interpolate
>>> x = lucid.randn(1, 3, 32, 32)
>>> y = interpolate(x, scale_factor=2.0, mode="bilinear", align_corners=False)
>>> y.shape
(1, 3, 64, 64)
>>> interpolate(x, scale_factor=0.25, mode="bicubic", antialias=True).shape
(1, 3, 8, 8)

Used by 2

>>> import lucid >>> from lucid.nn.functional import interpolate >>> x = lucid.randn(1, 3, 32, 32) >>> y = interpolate(x, scale_factor=2.0, mode="bilinear", align_corners=False) >>> y.shape (1, 3, 64, 64) >>> interpolate(x, scale_factor=0.25, mode="bicubic", antialias=True).shape (1, 3, 8, 8)