Summary - CNN

In this group you learned Conv2d, MaxPool2d, and feature maps.

Summary: CNN

In this group:

Convolution slides small filters; weight sharing gives few parameters and position-independent recognition.
nn.Conv2d: (N, C, H, W); output ⌊(H + 2p - k)/s⌋ + 1; parameters out·in·k·k + out.
Pooling: MaxPool2d, AvgPool2d, global AdaptiveAvgPool2d(1).
Blocks: Conv → BatchNorm → ReLU → Pool; maps shrink, channels grow, the receptive field widens.
A linear classifier after nn.Flatten or global pooling.

Exercise CNN with global pooling

Build a model with the body Conv2d(3, 16, 3, padding=1) → ReLU → Conv2d(16, 32, 3, padding=1) → ReLU, followed by AdaptiveAvgPool2d(1), Flatten and Linear(32, 5). Print the output shapes for both a (2, 3, 32, 32) and a (2, 3, 50, 70) input on separate lines.

Show example solution

import torch
from torch import nn

model = nn.Sequential(
    nn.Conv2d(3, 16, 3, padding=1), nn.ReLU(),
    nn.Conv2d(16, 32, 3, padding=1), nn.ReLU(),
    nn.AdaptiveAvgPool2d(1), nn.Flatten(), nn.Linear(32, 5))
print(model(torch.zeros(2, 3, 32, 32)).shape)
print(model(torch.zeros(2, 3, 50, 70)).shape)

Test yourself

What is the parameter count of a 3×3 convolution with 64 input and 128 output channels?
- 73856
- 73728
- 8192
- 1152
Show answer

73856 — 128 · 64 · 3 · 3 = 73728 weights + 128 biases = 73856.
Reduce each channel to a single value.
```
gap = nn.____(1)
```
Show answer
```
gap = nn.AdaptiveAvgPool2d(1)
```
Match the formulas.

output size

⌊(H + 2p - k)/s⌋ + 1
Conv2d parameters

out·in·k·k + out
MaxPool2d(2)

H/2, W/2
Flatten

C·H·W

Next lesson: What is RNN?

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies