How to split tensors with overlap and then reconstruct the original tensor?

vision
10

This post is using padding so that the original input tensor can be restored.