23. Convolution

A convolution applies a learnable filter across an image.

Image
 ↓
Kernel / Filter
 ↓
Feature Map

Important concepts:

• kernel size
• stride
• padding
• channels
• feature maps

A pytorchconvolution:

nn.Conv2d(
    in_channels=3,
    out_channels=16,
    kernel_size=3
)

learns 16 filters that operate on three-channel input images.