Feature map

The Conv2d output is a learned feature map.

The feature map

Each filter's output is a feature map: it shows where in the image, and how strongly, the pattern the filter looks for is found. As the layers progress the maps shrink but the channel count grows:

# input      (N, 3, 32, 32)
# block 1 -> (N, 32, 16, 16)   edges, colours
# block 2 -> (N, 64, 8, 8)     textures, parts
# block 3 -> (N, 128, 4, 4)    object parts

Each neuron in a deep layer looks at a larger region of the original image; this is called the receptive field.

Block structure

A CNN block usually consists of convolution, normalisation, activation and pooling. Producing blocks with a function shortens the code:

def blok(giris, cikis):
    return nn.Sequential(
        nn.Conv2d(giris, cikis, 3, padding=1),
        nn.BatchNorm2d(cikis),
        nn.ReLU(),
        nn.MaxPool2d(2))

govde = nn.Sequential(blok(3, 32), blok(32, 64), blok(64, 128))
print(govde(torch.zeros(1, 3, 32, 32)).shape)   # torch.Size([1, 128, 4, 4])

BatchNorm2d normalises each channel's activations, which speeds up training and makes it stable.

Exercise Body shapes

Build a three-block body with the given blok function (3→16→32→64 channels). For a (2, 3, 64, 64) input, print the shape after every block on separate lines.

Show example solution

import torch
from torch import nn

def blok(giris, son):
    return nn.Sequential(nn.Conv2d(giris, son, 3, padding=1), nn.BatchNorm2d(son), nn.ReLU(), nn.MaxPool2d(2))

x = torch.zeros(2, 3, 64, 64)

govde = nn.Sequential(blok(3, 16), blok(16, 32), blok(32, 64))
for b in govde:
    x = b(x)
    print(x.shape)

Test yourself

How many pixels wide is the receptive field of two consecutive 3×3 convolutions (stride 1)?
- 3
- 9
- 5
- 6
Show answer

5 — Each 3×3 layer grows the receptive field by 2 pixels: 3 + 2 = 5.
Add per-channel normalisation.
```
nn.____(32)
```
Show answer
```
nn.BatchNorm2d(32)
```
Match the layers.

Conv2d

Extracts patterns
BatchNorm2d

Normalises
ReLU

Non-linearity
MaxPool2d

Shrinks

Next lesson: Flatten

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies