Pooling

Spatial reduction and gradient distribution.

2 min readv1.0.0C++20

From the implementationinclude/nexus_model/layers/pooling.hpp

01MaxPool2D#

MaxPool2D(kernel,stride=-1,padding=0). Negative default stride becomes the kernel size. Requires rank-4 NCHW. First maximum wins; forward stores argmax. Backward scatter-adds to winning cells.

1MaxPool2D pool(2);2auto x=Tensor::zeros({8,32,28,28});3auto y=pool.forward(x); // [8,32,14,14]

AvgPool2D(kernel,stride=-1,padding=0,count_include_pad=true). The default denominator includes padding. With false, only valid cells count. Backward distributes gradients with the same denominator and sums overlaps.

y=\frac1m\sum_{i\in\mathcal W}x_i,\qquad dx_i\mathrel{+}=\frac{dy}{m}

03GlobalAvgPool2D#

GlobalAvgPool2D() averages H×W per channel. Output is directly [N,C]; no extra Flatten is needed. Every cell receives gradient scaled by 1/(HW).

◆Read alongside the source. API behavior reflects the current implementation.

02AvgPool2D#

y=\frac1m\sum_{i\in\mathcal W}x_i,\qquad dx_i\mathrel{+}=\frac{dy}{m}