In the flux.jl documentation, it says that I can get the gradients by declaring weights W as a param, calling back!(l), and then inspecting the grad field. How do I do this for a recurrent system? I want to get the sequence of gradients for a sequence of inputs.
Thanks!