Simple implementation of reverse-mode, vectorized autodiff. Only dependencies are numpy, more_itertools and graphviz (for visualising the computation graph).
It is an eager implementation. As computations are completed, appropriate edges are added to the backwards graph (i.e. like PyTorch) rather than the user building a graph upfront.
For an example MLP see examples/mlp.py.