what is the difference between do output.backward(gradient=df) and loss.backward()?

Question

I am currently testing an open source repository(https://github.com/jzengust/RGBD2Normal) for surface normal estimation, the following function is used as the surface normal