UMAP performance

[bwang12]

It's great to have a Julia implementation of the UMAP. I have been using the Python one quite a bit and am very impressed with its performance thus far. (https://github.com/lmcinnes/umap)

Since I am somewhat new to Julia, I am wondering how much faster can the Julia version be?

Currently, the Python UMAP takes about 3.2 seconds to run on a randomized 2000 by 2000 matrix.

import numpy as np
import umap
test = np.random.rand(2000, 2000)
UMAP = umap.UMAP(n_components=2)
%timeit UMAP.fit_transform(test)

While Julia UMAP would take about 4.3 seconds to run on a same size randomized matrix.

using BenchmarkTools
using UMAP
test = rand(2000, 2000)
@Btime umap(test)

I'd love to get your take on this @dillondaudert .

[SimonDanisch]

Maybe this can get you into the same ballpark: #2

[kadir-gunel]

Hello, I am trying to use umap for reducing the dimensions of word embeddings (200k x 300) to (200k x 2). And I cannot get any results because it runs on a single thread. I changed the number of threads that Julia is using from 1 to 8 but still UMAP.jl works on a single thread.
What is wrong ?