Backpropagation by Hand: Two Layers, a Pen, and Then Autograd Agrees

in the last piece i did autograd, and the whole point of it was that you call loss.backward() once...