Out of curiosity, I benchmarked our normalisation code(which currently uses numexpr) against vanilla numpy implementation and Numba CPU implementation. Here are the results:

Array size 2560:
numba CPU 7.586696507001761
numexpr 246.33255801402265
numpy 191.35872586001642

Array size 256:
numba CPU 5.624689556017984
numexpr 32.61893176403828
numpy 24.161521231988445

Given these results, I like to suggest a refactor where we replace numexpr code with numba code. Refactor PR will include a more complete benchmark result as well.