Umap: Add a seed parameter for repeatability

[ensonario]

It would be great to add seed parameter for repeatability, the way how scikit-learn does it.

[lmcinnes]

I agree, that would make a lot of sense. Ideally it isn't too hard, but has some quirks given how I am currently handling random number generation. It is certainly on my list of things to do (which is unfortunately long).

[ensonario]

It might be a good idea to publish a roadmap, community will be able to contribute!

[lmcinnes]

Sounds like a good plan -- any suggestions for where and how best to do that?

[ensonario]

Well, an issue on github will do. People will add comments, and you will be able to update the issue after each release.

[lmcinnes]

Basic random seed support is now in place via the random_seed parameter which takes an int. Ideally things would work a little differently as per standard sklearn with a random_state that supports more input types (e.g. numpy random states) but that will take a little thought as to the best way to do that.

Edit: and it doesn't actually achieve the desired result :-( Not sure why though. It should provide slightly more consistency though.

[lmcinnes]

Okay, that helps more. I have a nagging feeling there will be more minor things like the eigenvector solver to track down if I want to truly eliminate variability.

[lmcinnes]

Well that was a lot more work than I intended. For the record (since others may face this, and so I will remember in the future) the issue is that numba very cleverly swaps out np.random calls for something lower level (to avoid roundtrips back to python I presume), and this does (may?) not play nice with setting a random seed for numpy. Once I worked out what the issue was and rewrote everything to deal with that the issue resolved itself nicely and we get something repeatable. I believe setting random_state now works and should provide consistent embeddings (with a consistent random state).