A Volume Knob for Every Map

Every feature map in a CNN answers one question about the image, such as "is there fur here?" or "is there a wheel-like shape here?". Not every question matters equally for every photo. A squeeze-and-excitation block lets the network turn some feature maps up and others down, depending on the image in front of it. It is built from pieces you already know:

Squeeze. Average each of the $C$ feature maps down to one number, exactly as global average pooling does. This gives a summary $s_1, \dots, s_C$ .
Excite. Pass the summary through two small dense layers, neither with a bias. The first is narrower than $C$ and uses ReLU. The second goes back to $C$ outputs and uses the sigmoid $\sigma(z) = \dfrac{1}{1 + e^{-z}}$ , so each output is a gate between 0 and 1:

h_k = \max\left(0,\ \sum_{i=1}^{C} w^{\text{down}}_{k,i}\, s_i\right), \qquad g_i = \sigma\left(\sum_{k} w^{\text{up}}_{i,k}\, h_k\right)

Scale. Multiply every value in feature map $i$ by its gate $g_i$ .

Task: write reweight_maps(maps, w_down, w_up).

maps is a list of $C$ feature maps. Each map is a grid (a list of rows), and all maps have the same height and width.
w_down has one row per hidden unit $k$ , and each row holds $C$ weights: w_down[k][i] is $w^{\text{down}}_{k,i}$ .
w_up has one row per feature map $i$ , and each row holds one weight per hidden unit: w_up[i][k] is $w^{\text{up}}_{i,k}$ .
Return the reweighted maps in the same nested shape as maps, with every value rounded to 4 decimal places.

A Volume Knob for Every Map

Examples

Hints