Stop 10 · Computer VisionTopics · 3/9

CNN Architectures (ResNet, VGG, EfficientNet)

In short

You will rarely design a CNN from scratch. You'll pick ResNet, VGG or EfficientNet off the shelf. Knowing what each one changed helps you choose the right one for speed, size and accuracy.

Landmark CNN designs — ResNet introduced skip connections to train very deep networks, VGG popularized deep stacks of small filters, EfficientNet optimized accuracy-per-parameter.

The blob compares three toy towers, one plain, one with an express lift, one neatly proportioned.

Think of it like this

Three ways to build a tall building

VGG is a plain building: the same floor repeated, very tall and heavy. ResNet adds an express lift so people can skip floors, which lets it go much taller safely. EfficientNet is the architect who balances height, width and window size together for the best value.

Watch it happen

What each family is known for

CNN families

VGG (2014)

ResNet (2015)

EfficientNet

Stacked 3x3s

Skip connections

50 to 152 layers

Compound scaling

B0 to B7 sizes

1/5All three are convolutional networks trained on ImageNet. They differ in how they stack layers.

The big ideas

What you need to know

Deeper is not free

Before ResNet, adding layers often made training worse, because gradients faded on the way back.

Skip connections

ResNet computes output = F(x) + x. The block only learns the change, and the shortcut keeps gradients alive.

Accuracy per parameter

VGG-16 has around 138 million parameters. EfficientNet-B0 has about 5 million and does better on ImageNet.

Pick for your device

On a phone, look at small models like EfficientNet-B0 or MobileNet. On a server GPU, ResNet-50 is a safe start.

See it in code

Walk through the code

python

1import torch.nn as nn2 3class ResidualBlock(nn.Module):4    def __init__(self, ch):5        super().__init__()6        self.conv1 = nn.Conv2d(ch, ch, 3, padding=1)7        self.bn1 = nn.BatchNorm2d(ch)8        self.conv2 = nn.Conv2d(ch, ch, 3, padding=1)9        self.bn2 = nn.BatchNorm2d(ch)10        self.relu = nn.ReLU()11 12    def forward(self, x):13        out = self.relu(self.bn1(self.conv1(x)))14        out = self.bn2(self.conv2(out))15        return self.relu(out + x)  # the skip connection

1/3Two 3x3 convolutions with batch norm. Padding 1 keeps height and width the same.

A simplified ResNet basic block. The whole trick is adding x back on the last line.

Save yourself the pain

Common mistakes

Oops

Picking the biggest model because it tops the leaderboard.

Do this instead

Check latency and memory on your target device. A smaller model that runs in 20 ms often beats a better one that takes 2 seconds.

Oops

Thinking skip connections just copy the input forward.

Do this instead

They add the input to what the block learned, so the block learns a correction. Shapes must match, or a 1x1 conv adjusts them.

Your mission

Compare three backbones

~30 min

Quick check

What problem did ResNet's skip connections mainly solve?

Go deeper

1Load a pretrained ResNet and use it for transfer learning on a custom dataset.
2Learn what skip connections in ResNet solve (vanishing gradients in deep networks).
3Compare model size and accuracy trade-offs across ResNet, VGG and EfficientNet.