Precomputed distances about umap HOT 4 CLOSED

No that's certainly quite feasible, the only catch being that you have to have computed the distance matrix first. I would be happy to work though and merge a pull request if you want to put in the work. Thanks!

from umap.

Sure, should be helpful in the cases where people have precomputed distance matrices due to large computation times and re-use the distance matrix.
I have a working implementation already (results look very nice with UMAP!) and will submit a pull request in the upcoming days.

So far I have added a precomputed flag to make_nn_descent that is set when the given metric is set to the string "precomputed". However, I'm still struggling with what to do with the distance function in that case, because it seems like it expects a numba function in any case.
Would you prefer a flag that changes the behaviour or a dummy distance function named precomputed that is then checked for its name in nn_descent?
Maybe we can come up with an even more elegant way.

If I couldn't make clear what I mean, just wait for my pull request and we can discuss the actual code.

from umap.

I think adding a dummy distance function called precomputed to distance.py is probably the better way to go under those options (would likely have less code duplication).

Since you have the full distance matrix, however, could you not compute the nearest neighbors by an argsort across an index of the distance matrix? That might save you having to mess with nn_descent at all. Then again, if you have something that works I'm not really going to argue.

Looking forward to the PR.

from umap.

You are right, that's more elegant and faster.

from umap.