6.3.1 CNN Roadmap: Turn Images Into Feature Maps

CNNs learn local visual patterns. Instead of reading an image as one flat row of numbers, they scan small regions and build feature maps.

Look at the Image Flow First

Concept	First meaning
channel	color or learned feature dimension
kernel	small sliding filter
feature map	output after filters scan the image
pooling / stride	shrink spatial size
transfer learning	reuse a pretrained vision backbone

Run One Convolution

Create cnn_first_loop.py and run it after installing torch.

import torch

image = torch.randn(1, 3, 32, 32)
conv = torch.nn.Conv2d(in_channels=3, out_channels=8, kernel_size=3, padding=1)
features = conv(image)

print("input_shape:", tuple(image.shape))
print("feature_shape:", tuple(features.shape))

Expected output:

input_shape: (1, 3, 32, 32)
feature_shape: (1, 8, 32, 32)

Read the shape as [batch, channels, height, width]. The convolution changed 3 input channels into 8 learned feature channels.

Learn in This Order

Order	Read	What to practice
1	6.3.2 Convolution Basics	kernel, stride, padding, channel
2	6.3.3 CNN Structure	conv block, pooling, classifier head
3	6.3.4 Classic Architectures	LeNet, AlexNet, VGG, ResNet intuition
4	6.3.5 Transfer Learning	frozen backbone, fine-tuning
5	6.3.6 Image Classification Practice	dataset, training, prediction examples

Evidence to Keep

Keep one CNN shape note:

Input

[batch, channels, height, width]

Conv Output

out_channels becomes new feature maps

Spatial Change

stride/padding/pooling change height and width

Classifier Bridge

conv features eventually become class logits

Transfer Choice

freeze first, fine-tune only if validation improves

Pass Check

You pass this roadmap when you can explain what changed between input image shape and feature map shape, and why pretrained CNN backbones are useful for small datasets.

Check reasoning and explanation

1. A passing answer connects tensors, model layers, loss, backward(), and optimizer updates into one training loop.
2. The evidence should include a runnable mini experiment, tensor-shape checks, and a loss or validation curve you can explain.
3. A good self-check names one failure mode such as shape mismatch, no loss decrease, overfitting, data leakage, or using Attention/Transformer words without explaining the data flow.