Global average pooling

Background

Modern CNN classifier heads often replace "flatten + big dense layer" with global average pooling: collapse each feature map to a single number by averaging over all its spatial positions. A (H, W, C) feature volume becomes a length-C vector — one summary per channel.

$$\text{out}_c = \frac{1}{HW}\sum_{i,j} \text{volume}_{i,j,c}$$

Problem statement

Implement global_avg_pool(volume) returning the per-channel spatial mean.

Input

• volume — array-like of shape (H, W, C).

Output

Returns an np.ndarray of shape (C,).

Examples

Example 1 — single channel

Input:  volume = [[[1], [3]], [[2], [6]]]
Output: [3.]

Explanation: mean of $\{1,3,2,6\}$ is $3$.

Example 2 — two channels

Input:  volume = [[[1, 10], [3, 30]], [[2, 20], [6, 60]]]
Output: [ 3. 30.]

Constraints

• Average over the spatial axes (0 and 1), keeping the channel axis.
• Return shape (C,).

Notes

• Global average pooling has no parameters and is translation-tolerant — a big reason it replaced giant flatten+dense heads in architectures like ResNet.