Padding and Stride

Padding adds values around the input border.

Stride controls how far the kernel moves each step.

Without padding, a kernel cannot be centered at the border in the same way it can be centered inside the input. Padding lets the layer produce outputs near edges.

Stride changes output size. In one dimension, a simple valid convolution output length is:

output_length = floor((input_length - kernel_size) / stride) + 1

For input_length = 7, kernel_size = 3, and stride = 2:

floor((7 - 3) / 2) + 1 = 3