Results of umap-learn and UMAP by cuML are different

Nobody has claimed this yet.

Hello,

I was comparing the results produced by your version of umap (**umap-learn**) with the results by **cuML** version of UMAP. I used the same (major) parameters for instantiating UMAP models with two different packages (except some parameters such as `dens_lambda`, `dens_frac`, `dens_var_shift` etc. that are not used in the cuML implementation). When I feed them with the same 100 embeddings, the reduced embeddings have quite different values (see screenshot below). May I ask if you have compared your implementation with **cuML** implementation? It might be hard say, but I hope to know which one is the better implementation (without considering the speed, as cuML leverages GPU), and what could be the major reason that causes this difference (the parameters not used in cuML?).


And It would be really great if you could provide GPU support. May I ask if you have any specific schedule for this? (It's quite difficult to setup cuML-UMAP outside of Google Colab) Thanks a lot!

Best regards,
Jingying

Reproduce the comparison with the same 100 embeddings and major UMAP parameters, noting parameters absent from cuML such as dens_lambda, dens_frac, and dens_var_shift. Compare the reduced embeddings and determine whether the discrepancy is expected from implementation or parameter differences; GPU support is a separate request.

Written by the indexing model from the issue text.