📉 Pooling Layers

Pooling layers are used to shrink the image down.

🗺️ The Map Zoom

When you open Google Maps, you don't need to see every single side street if you are zoomed out looking at the entire country. Pooling takes a 2x2 square of pixels and squishes them into a single pixel (usually by just keeping the highest number, known as Max Pooling).

This saves massive amounts of memory and helps the network look at the "big picture" rather than getting distracted by tiny details!

🐍 Python Implementation

import torch
import torch.nn as nn

# Our output from the last chapter: (1 batch, 16 channels, 64x64)
features = torch.randn(1, 16, 64, 64)

# Max Pooling (2x2 square)
pool_layer = nn.MaxPool2d(kernel_size=2)

# Shrunk!
shrunk_features = pool_layer(features)

# The width and height are cut exactly in half! (64 -> 32)
print("Shrunk shape:", shrunk_features.shape) # (1, 16, 32, 32)