adaptive_max_pool3d

→Tensor

adaptive_max_pool3d(x: Tensor, output_size: int | tuple[int, int, int], return_indices: bool = False)

source

3-D adaptive max pooling — produces a fixed (D, H, W).

Volumetric analogue of adaptive_max_pool2d. Computes per-axis kernel and stride so the output spatial shape exactly matches output_size regardless of input shape.

Parameters

xTensor

Input of shape (N, C, D, H, W).

output_sizeint or (int, int, int)

Desired output spatial shape.

return_indicesbool= False

Must currently be False.

Returns

Tensor

Output of shape (N, C, oD, oH, oW).

Notes

The window for output cell $(i, j, k)$ is the 3-way product of axis-wise floor / ceil intervals (see adaptive_avg_pool2d for the 2-D version). The cell takes the maximum over that volume:

y_{n,c,d,h,w} = \max_{(p, q, r) \in R(d, h, w)} x_{n,\,c,\,p,\,q,\,r}

Examples

>>> import lucid
>>> from lucid.nn.functional import adaptive_max_pool3d
>>> x = lucid.randn(1, 8, 5, 7, 9)
>>> y = adaptive_max_pool3d(x, output_size=(1, 1, 1))
>>> y.shape
(1, 8, 1, 1, 1)