wilsonjr / humap Goto Github PK

Semi-related to #4 , but my case is that I want to use HUMAP on a supervised data where I have a training data with labels, and I want to be able to project new test data with the same embeddings. UMAP supports this use case, I was wondering if this would be theoretically possible with HUMAP as well? Would be nice to be able to use HUMAP to interpret classifier decisions.

Hi,

Just putting it out there that you might want to consider putting up your package on conda-forge. Many other packages like numpy, scikit-learn, umap, are all available on conda-forge, and managing them through conda cli makes it easy to be up-to-date and not worry about dependencies like MKL, which pip doesn't handle well.

As a bonus, I see that this package depends on Eigen, which needs to be manually configured on Windows. Conda-forge already has eigen available, which might make this much less error-prone for Windows users, which I assume will be a substantial chunk.

Just as an FYI, here is a link for conda-forge submission process.

Thanks!

Thanks for writing this awesome library, only recently discovered it.
Do you have plans to support semi-supervised umap?
From my first try outs of your library, this is the fastest (h)umap implementation which has nndescent. I would like to use it for semi-supervised learning, too.

Problems when computing hierarchy for small datasets.
I tried to execute HUMAP on Iris dataset using 100, 15, and 10 n_neighbors.

RuntimeError: Some rows contain fewer than n_neighbors distances