My effort was rather in trying to speed up all the python looping, etc. around the np calls. But I never went far trying to actually benchmark the entire pipeline in order to find out what was the actual bottleneck.