Re-parametrization of PyTorch Models

Say I want to train a model f_W indirectly by learning a displacement D to a (constant) intiallization W0. In other words I want to re-parametrize f_W = f_(W0+D) and then optimize D via SGD while W0 remains unchanged.

Is there a simple way to modify an existing PyTorch model to use a different parametrization (preferably without having to re-write the entire model class by hand)?