stride and padding

Stride and padding are the two dials that decide how a sliding filter travels across an image — and what size the result comes out. Both are easiest to picture by going back to the stencil you slide over a photo. Stride is how far you jump between stops. A stride of 1 means you nudge the filter one pixel at a time, examining every position and producing a large, detailed output. A stride of 2 means you skip every other position, so you do half as much work in each direction and the output shrinks to roughly a quarter the size — a built-in way to downsize, much like pooling.

Padding deals with a nuisance at the edges. When the filter reaches the border of the image, part of it hangs off into nothing, so without help the output map ends up a little smaller than the input, and the very edges get visited fewer times than the middle. Padding fixes this by adding a frame of extra pixels around the image — usually zeros, hence "zero-padding." With the right padding, the output can stay exactly the same size as the input ("same" padding), and the corners get a fair shot. Without padding ("valid" padding), the map deliberately shrinks at each layer.

Together these two settings let a designer control the trade-off between detail and cost. Small stride and generous padding preserve resolution but pile on computation; large stride and no padding shrink things fast and cheaply but throw away fine spatial information. They are hyperparameters — choices made before training, not things the network learns — and getting them wrong is a common cause of feature maps that come out the wrong shape and refuse to connect to the next layer.