Average Pooling: Downsampling by Smoothing Instead of Selecting

The problem, from first principles

01-max-pooling keeps only a window's single loudest value, discarding everything else outright — a deliberate, sometimes aggressive choice. Average pooling makes the opposite choice: instead of picking a winner, it blends every value in the window into one smoothed number, so a single unusually large or small pixel doesn't dominate the result the way it would under max pooling.

Both are valid ways to shrink a feature map; which one helps depends on whether you want the network to notice "did this pattern fire strongly anywhere" (max) or "what's the overall level of activity here" (average).

From theory to code

Theory changes exactly one operation from 01-max-pooling: replace the per-window maximum with the per-window mean. Everything about how windows are laid out and stepped through stays identical.

Implement avg_pool2d(image, kernel_size, stride=None) against that reasoning.

Constraints

• image: shape (C, H, W).
• kernel_size: the pooling window's height and width.
• stride: step between windows; if None, defaults to kernel_size (non-overlapping pooling).
• Output: shape (C, (H - kernel_size)//stride + 1, (W - kernel_size)//stride + 1).
• Every value inside a window contributes equally to that window's output — no weighting, no learned parameters.
• image is never modified in place.

Hints

Open one at a time. Each gives away a little more than the last.