Stride is how many pixels the filter jumps between positions.
A stride of 1 moves one pixel at a time and keeps the most detail.
A stride of 2 skips every other position and halves the output size.
Larger strides mean less computation but coarser output.
Strided convolution is one common way to downsample instead of pooling.
Stride is how far the kernel jumps between one position and the next.
A stride of one moves a single pixel at a time and looks at every position.
That keeps the most detail and costs the most computation.
A stride of two skips every other position, halving both output dimensions.
The output then has a quarter of the values, which is a large saving.
So stride is a downsampling control as well as a movement setting.
Strided convolution is one of the two standard ways to shrink a feature map.
Pooling is the other, and modern architectures increasingly prefer stride.
The difference is that a strided convolution learns how to downsample.
Pooling applies a fixed rule with no parameters at all.
Larger strides lose information that cannot be recovered later.
A stride of two is common and anything above three is unusual in practice.
Stride interacts with padding to determine the output size exactly.
Getting that arithmetic wrong is the classic shape mismatch error when building a network.
Most frameworks will print the expected and actual shapes, which usually names the culprit.
One question on this card. Every option explains itself once you pick, right or wrong.
22 concepts in this module, then a graded test with every answer explained.