Optimized binning into hex mesh

[danilexn]

This PR adapts code from matplotlib to mesh into hexagonal bins more efficiently.

With this, it is no longer necessary to compute the pairwise Euclidean distance between spots and bins, since spot coordinates are directly binned into hex coordinates.

E.g., when binning a Seq-scope tile into 7 um hex bins, time is reduced from ~15 min to ~3 min; memory usage is ~10x lower.

[mschilli87]

LGTM license-wise. 😃
Did not perform any functional test though. 😉

[nukappa]

PR looks good, tested it on 10 tiles and memory usage was ~100x lower, speed reduced from ~35min to ~10.

@danilexn before we merge can you pls check the following:

• Make the optimize_binning default by setting True in the create_mesh function
• I guess this extends beyond hexagons to circles? Can you pls adapt that part of the function, too?

[danilexn]

PR looks good, tested it on 10 tiles and memory usage was ~100x lower, speed reduced from ~35min to ~10.

@danilexn before we merge can you pls check the following:

• Make the optimize_binning default by setting True in the create_mesh function

Done, see 329345a

• I guess this extends beyond hexagons to circles? Can you pls adapt that part of the function, too?

Yes indeed, I am extending it to circle meshing – now testing if outputs are similar to the current approach.

Thanks for the review!

[danilexn]

PR looks good, tested it on 10 tiles and memory usage was ~100x lower, speed reduced from ~35min to ~10.

@danilexn before we merge can you pls check the following:

• Make the optimize_binning default by setting True in the create_mesh function
• I guess this extends beyond hexagons to circles? Can you pls adapt that part of the function, too?

962ae29 implements optimized binning to circle mesh. As well, meshes (for circle and hex cases) should be identical to non-optimized cases, now. I measured similar speedups and memory usage, for circular and hex cases.