Introduction
A convolution has a few settings that control how big the output is and how much of the image each output value "sees". Getting the shapes right is half the battle when building a CNN. Most error messages in vision code are shape mismatches.
Think of mowing a lawn with a small mower. The stride is how far you move between passes (overlapping a lot, or leaving gaps). Padding is like adding a strip of path around the lawn so the mower can reach the very edges. The channels are like having several different mowers (filters), each leaving its own pattern.