If you run your code with 0 multiplied with all the loss terms, you will still see the image gradually get smoother. in the Denormalize function because the tensor is not getting cloned, it is doing the multiplication by s and addition of mu in place.
If you run your code with 0 multiplied with all the loss terms, you will still see the image gradually get smoother. in the Denormalize function because the tensor is not getting cloned, it is doing the multiplication by s and addition of mu in place.