Is UMAP sensitive to floating point microarchitecture?

[mcfefa]

I'm trying to run UMAP.jl on two different machines (one is a MacBook Pro and then other is a computational cluster at my institute, both using Julia 1.0) and when I run the same dataset on each of them, I get drastically different results. My MacBook shows clear groupings by communities previously detected in the data (community id was excluded when embeddings were calculated), but when I run the same data on the computational cluster, the UMAP returns a similar overall shape, but there is not longer any clear groupings based on community id (like with my MacBook). I was wondering if UMAP.jl could be sensitive to floating point microarchitecture changes between my MacBook and the computational cluster? And if there is anything I could do to fix this? I have large-ish datasets (20-100k x 12) and they take hours to run on my machine and the cluster, so I want to make sure I have consistency between the computational resources. I've have two smaller datasets (1.5k x 12, which take < 1 min to run) and I've reproduced this discrepancy multiple times across these smaller examples.

[dillondaudert]

Did you figure out the issue?

[mcfefa]

Yes, it looks like I sorted one of the datasets and not the other, which is why things weren't matching up.

…

On Mon, Feb 25, 2019 at 9:11 AM Dillon Daudert ***@***.***> wrote: Did you figure out the issue? — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#7 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AePUPZjApgfVRMCkkzPhPBDe2NeaFhWHks5vQ-8hgaJpZM4bOEPb> .