This is just a brief public service announcement reporting something that I’ve just found: np.percentile is a lot faster than scipy.stats.scoreatpercentile– almost an order of magnitude faster in some cases.
Someone recently asked me why on earth I was using scoreatpercentile anyway – and it turns out that np.percentile was only added in numpy 1.7, which was released part-way through my PhD in Feb 2013, hence why the scipy function is used in some of my code.
In my code I frequently calculate percentiles from satellite images represented as large 2D numpy arrays – and the speed differences can be quite astounding:
| Image size | scoreatpercentile | percentile | speedup |
| 100 | 595us | 169us | 3.5x |
| 1000 | 84ms | 13ms | 6.5x |
| 3000 | 927ms | 104ms | 9x |
| 8000 | 8s | 1s | 8x |
As you can see, we get 3-4 times speedup for even small arrays (100 x 100, so 10,000 elements), and up to 8-9 times speedup for large arrays (tens of millions of elements).
Anyway, the two functions have very similar signatures and options – the only thing missing from np.percentile is the ability to set hard upper or lower limits – so it should be fairly easy to switch over, and it’s worth it for the speed boost!