adaptive_max_pool2d

→Tensor

adaptive_max_pool2d(x: Tensor, output_size: int | tuple[int, int], return_indices: bool = False)

source

2-D adaptive max pooling — fixed-shape (H, W) via per-cell max.

Like adaptive_avg_pool2d but takes the maximum over each dynamically computed window instead of the mean. Useful when salient peaks should drive the downstream representation (e.g. object detectors that rely on strong activations).

Parameters

xTensor

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

output_sizeint or (int, int)

Desired spatial output shape.

return_indicesbool= False

Must currently be False.

Returns

Tensor

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

Notes

For each output cell, the window definition is the same as in adaptive_avg_pool2d; the cell value is

y_{i,c,h,w} = \max_{(m, n) \in R(h, w)} x_{i,\,c,\,m,\,n}

Gradient flows only through the per-window argmax position.

Examples

>>> import lucid
>>> from lucid.nn.functional import adaptive_max_pool2d
>>> x = lucid.randn(1, 64, 11, 13)
>>> y = adaptive_max_pool2d(x, output_size=(3, 3))
>>> y.shape
(1, 64, 3, 3)