Be careful with that, numpy arrays can be slower than Python tuples for some operations. The creation is always slower and the overhead has to be worth it.

You can turn trees into numpy-style matrix operations because graphs and matrices are two sides of the same coin. I don't see the code for the binary-tree benchmark in the repo to see how it's written, but there are libraries like graphblas that use the equivalence for optimization.