stride

Stride is how far the kernel hops between two consecutive positions. With stride 1 the kernel moves one pixel at a time and visits every location; with stride 2 it skips every other position, so it visits only half as many spots per axis. Larger strides therefore produce a smaller output grid — they downsample. Intuitively, stride is a coarseness dial: a big stride samples the input on a sparser grid, throwing away spatial resolution in exchange for cheaper computation and a faster-growing receptive field.

Precisely, for input size W, kernel size k, padding P, and stride S, the output size is floor((W + 2P - k) / S) + 1 (with dilation d, replace k by d(k-1)+1). A stride of 2 roughly halves each spatial dimension and quarters the number of output positions, cutting both the activation memory and the work of subsequent layers. Because earlier receptive fields end up spaced S pixels apart, every strided layer also multiplies how fast the receptive field grows downstream.

Strided convolution is one of two main ways to downsample inside a CNN, the other being pooling. Modern architectures increasingly prefer strided (or learnable) downsampling: the 'All Convolutional Net' (Springenberg et al., 2015) showed pooling can be replaced by stride-2 convolutions with no loss, and ResNet downsamples with stride-2 convs in its bottleneck blocks. The trade-off is that aggressive striding loses fine detail early, which hurts tasks needing precise localization (segmentation, keypoints) — there, designers keep stride low and use dilation instead to grow context without shrinking the maps.