06_cnn

06 — Convolutional Neural Network

A Conv2d layer slides a small set of weights (a "kernel") across the image and produces a feature map. Stacking convolutions + pooling builds increasingly abstract features: edges → textures → parts → objects.

Why convolutions instead of a giant Linear layer?

A 224×224 RGB image flattened is 150,528 numbers. A single fully-connected layer with 1000 hidden units would have 150 million parameters — wasteful and prone to overfitting. A convolution re-uses the same handful of weights at every position, so it has thousands of parameters and naturally generalizes across translations.

Shapes you'll see

• Input: (batch, channels, height, width) — e.g. (64, 1, 16, 16).
• After Conv2d(1, 8, kernel_size=3, padding=1): (64, 8, 16, 16).
• After MaxPool2d(2): spatial dims halve.
• Eventually Flatten() collapses everything but the batch dim for the classifier head.

model.train() vs model.eval()

Some layers (Dropout, BatchNorm) behave differently between training and evaluation. Always toggle modes around training/eval steps. We also wrap eval in torch.no_grad() to skip building the autograd graph — faster and uses less memory.

Run

python 06_cnn/main.py