search

wilsonjr / humap

Hierarchical Uniform Manifold Approximation and Projection
BSD 3-Clause "New" or "Revised" License
233 stars 8 forks source link

Model persistence #10

Open adsharma opened 5 months ago

adsharma commented 5 months ago

umap allows me to fit the model on a small sample of data and then persist the model to disk.

Subsequently, I can load the model from disk and compute hierarchy for a much larger number of samples that don't fit in memory.

Can I do the same with humap? Even if I can't persist to disk, being able to train on a small number of samples and compute low dimensional embedding (and presumably the same level1 and level2 embeddings) on a larger number of samples is super useful.

wilsonjr commented 1 month ago

This is quite handy, but HUMAP cannot do it.

I believe the simplest solution would be to allow users to save HUMAP levels so that they do not have to fit the hierarchy every time. Transforming to new data is something that I need to give more attention to—perhaps you can help.