decouple pre-processing prediction and post-processing?

[Merlin2013]

Just like the common object detection, three phases are included for a whole cycle.
pre-processing: read image, resize or crop (run on cpu)
prediction: load inputs, run on GPU, and get the last conv (run on gpu)
post-processing: do NMS, generate the coordinates, and callback (run on cpu)

how can i decouple the different tasks and imporve throughout ( i think it's more complex here),
is there any reference for this?