Delaunay Interpolation about pyautolens HOT 7 CLOSED

Just test how long it will take to use scipy's in-built functionality to find pairs. For 120000 sub image pixels and 5000 source pixels, the time cost for pairing (building the Delaunay and finding pairs for all sub image pixels) is only ~ 0.05s for one cpu. The major cost is next step which is to compute the weight of each point. It takes ~ 4s. So probably we can just use the scipy built-in function to find pairs but for the weight computing part we need to use numba to speed it up. If we choose to use scipy built-in functionality, do we still need to compute pixelization_index_for_voronoi_sub_slim_index_from?

from pyautolens.

That is extremely fast for pairing! I would expect the function pixelization_index_for_voronoi_sub_slim_index_from to take > 5 seconds for that number of image / source pixels. There must be a geometric exploit for a Delaunay that speeds this up a lot compared to the Voronoi (for a Rectangular grid this process is extremely fast).

I would imagine that the SciPy computation of interpolation weights is as good as we can hope for in terms of run time. Its a very well optimized library. What are the run times for 30000 / 60000 / 120000 / 240000 sub-pixels and 1500 / 2500 source pixels?

from pyautolens.

sub_pixels | source_pixels | pairing time | weight time
65000 2600 0.03 1.8
125000 2600 0.03 4.6
215000 2600 0.03 7.1

65000 1500 0.02 1.8
125000 1500 0.02 3.8
215000 1500 0.03 7.3

The part computing the weights is written by myself and they are pure python scripts, so it's slow.

from pyautolens.

Ah, that is good news! Sounds like this can be lightning fast, once we speed up the weights.

Can you use this to get the weights efficiently?

https://docs.scipy.org/doc/scipy-0.15.1/reference/generated/scipy.interpolate.LinearNDInterpolator.html

from pyautolens.

Ah! Thanks! Didn't expect scipy actually implemented Delaunay interpolation. Will try to build it in.

from pyautolens.

Complied the weighting part with Numba, now with 125000 sub pixels and 5000 source pixels, the time cost for weighting computation is ~ 0.3 s (using one core) on my computer. Is this speed acceptable?

from pyautolens.

Amazing, signifcantly faster overall than the non interpolated Voronoi 🥳

from pyautolens.