Conv2d
- Machine Learning course
- Advanced
- Lesson 42/78
- 5 min
Conv2d uses a kernel to scan for patterns in images.
nn.Conv2d
nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=0) wants input of shape (N, C, H, W). Each output channel is a separate filter that looks at all input channels:
conv = nn.Conv2d(3, 16, kernel_size=3, padding=1)
x = torch.zeros(8, 3, 32, 32)
print(conv(x).shape) # torch.Size([8, 16, 32, 32])
print(conv.weight.shape) # torch.Size([16, 3, 3, 3])The output size is ⌊(H + 2·padding - kernel) / stride⌋ + 1. With padding=1 a 3×3 kernel keeps the size.
Parameter count and stride
A Conv2d layer's parameter count is out · in · k · k + out; it doesn't depend on the image size. stride=2 moves the filter two pixels at a time and halves the output:
conv = nn.Conv2d(3, 16, 3)
print(sum(p.numel() for p in conv.parameters())) # 16·3·3·3 + 16 = 448
s2 = nn.Conv2d(16, 32, 3, stride=2, padding=1)
print(s2(torch.zeros(1, 16, 32, 32)).shape) # torch.Size([1, 32, 16, 16])The same layer uses the same 448 parameters on a 32×32 or a 1024×1024 image.
Practice Convolution shapes
Apply the layers nn.Conv2d(1, 8, 5), nn.Conv2d(1, 8, 3, padding=1) and nn.Conv2d(1, 8, 3, stride=2, padding=1) to a (1, 1, 28, 28) input; print the three output shapes on separate lines. Then print the first layer's parameter count.
Starter code
import torch
from torch import nn
x = torch.zeros(1, 1, 28, 28)
# buraya yazSample solution
import torch
from torch import nn
x = torch.zeros(1, 1, 28, 28)
katmanlar = [nn.Conv2d(1, 8, 5), nn.Conv2d(1, 8, 3, padding=1), nn.Conv2d(1, 8, 3, stride=2, padding=1)]
for k in katmanlar:
print(k(x).shape)
print(sum(p.numel() for p in katmanlar[0].parameters()))Expected output
torch.Size([1, 8, 24, 24]) torch.Size([1, 8, 28, 28]) torch.Size([1, 8, 14, 14]) 208
Questions and answers
Review the lesson questions together with their answers. To test yourself, solve the lesson interactively.
1 With a 32×32 input, kernel 5, stride 1 and padding 0, how many pixels is the output?
- 28×28 Correct answer
- 32×32
- 27×27
- 30×30
Explanation: (32 + 0 - 5) / 1 + 1 = 28.
2 Build a 3×3 convolution that keeps the size.
conv = nn.Conv2d(3, 16, kernel_size=3, ____=1)conv = nn.Conv2d(3, 16, kernel_size=3, padding=1)Explanation: padding=1 adds one pixel of zeros at the edges.
3 Match the settings with their effects.
| Term | Meaning |
|---|---|
kernel_size |
Filter size |
stride=2 |
Halves the output |
padding=1 |
Keeps the size (3×3) |
out_channels |
Number of filters |