Is your feature request related to a problem? Please describe.
CoordConv gives networks an idea of where they are in an input by adding additional channels to the input that simply contain the spatial coordinate system. It has been shown to improve performance in various computer vision tasks, paper here.

Describe the solution you'd like
This could be simply implemented as a transform. If you think this would be useful I can make a PR to add.

Describe alternatives you've considered