Not reproducible

[skygering]

If I run the example in the documentation twice, even using the exact same points, I get very very slightly different answers on some of the points.

For example, if I run:

using VoronoiCells
using GeometryBasics
using Random

rect = Rectangle(Point2(0, 0), Point2(1, 1))
rng = Random.MersenneTwister(1337)
points = [Point2(rand(rng), rand(rng)) for _ in 1:10]


cells1 = voronoicells(points, rect).Cells
cells2 = voronoicells(points, rect).Cells

same = true
for i in eachindex(cells1)
    same = same && all(cells1[i] .== cells2[i])
end

I consistently get different values on at least some of the points and same is false. Granted, these are small differences, e.g. 0.5006658246439755 vs. 0.5006658246439761, but given I have to scale the results quite a lot to get a tessellation of my needed size, these can become quite large. It also prevents my simulation from being reproducible.

Do you have any idea where this might come from? I am willing to take a look as well!

Thanks!

[robertdj]

Well spotted and thanks for reporting it here.

I use the VoronoiDelaunay package to do the actual work of computing the points in the tessellation -- VoronoiCells just capture the results in a convenient manner.

It would be great if you can take a look. I'm quite out of sync with the inner workings of VoronoiDelaunay, but I suppose a starting point could be to check my claim about which package introduce the "randomness".

[skygering]

Thank you for responding! I will work on this when I have some time over the next few weeks. I will update if I have any progress.

[skygering]

I think that I found the issue. Adding in an optional RNG parameter to the call to pull! in VoronoiCells, which calls shuffle! in VoronoiDelaunay seems to do it. I have opened a pull request to VoronoiDelaunay and will make a pull request here as well if that goes through!

[skygering]

@robertdj I see that you're also a contributor on VoronoiDelaunay. Do you think you could take a look at my pull request? It is 3 lines and fully backwards compatible. That will allow me to make my pull request here (I have already written tests and everything) so that the answers won't be non-deterministic anymore. Thank you so much!

[robertdj]

Hi @skygering Sorry for being so slow! I have some time off and is much more responsive the next week.

Great that you found the issue! It's true that I have contributed to VD, but it is a long time ago and I don't think I have the street credit to approve PRs anymore.

But I'll try harder to at least answer here when you post and I'm happy to accept well tested changes :-)

[skygering]

@robertdj Would you have any idea who to contact regarding the PR on VD? It doesn't seem like anyone has worked on the repo for 9 months. This is holding up my progress, so I would like to get it resolved if possible.

Thanks!

[robertdj]

No, sorry. The latest issue also suggests that the repo is unmaintained :-(

If it's only for personal use maybe you can fork it?

[skygering]

It is eventually for public use, so I would really rather not make a copy of a perfectly good piece of software to change 3 lines, one of which is an import 😆 I will ping the geometry channels in the Julia slack and see what can be done

[skygering]

@robertdj Good news! I got my PR merged into VoronoiDelaunay, so I was able to open my pull request here. Let me know if there is anything I can do 😄

[robertdj]

Great! I'll have a look.

[robertdj]

It's merged 🎉 What would be an appropriate version for this change?

[skygering]

Thank you! Can we just bump the patch version? It is backwards compatible so it shouldn't be breaking, so just v0.3.1? Although it does require a higher version of VoronoiDelaunay. Thoughts?

[robertdj]

Hmm. True, but it also adds a new feature. So I'm leaning more towards bumping the minor to 0.4.0

[skygering]

@robertdj That would be fantastic! Thanks for getting it merged and tagged.