ZeroDivisionError when dataset is 4096 elements or more about umap HOT 4 CLOSED

I believe what is happening here is that you have a data point that is identically zero. Since you are using cosine distance it is trying to normalise by the square root of the product of the norms and dying a horrible death. It seems I didn't have sufficient defensive programming on my cosine distance implementation -- fortunately that can be fixed. Thanks for the detailed report -- it made chasing this down much easier.

from umap.

If you have time and can pull from master and rebuild I would appreciate if you can verify if that last commit resolves the issue for you.

from umap.

I'll do so. It's the least I can do for this amazing tool! Will report back.

I did inspect the data to see if there was anything funny at those particular indices (4096) but the vectors are non-zero, no weird unicode characters, nothing obvious. The fact that it happens with two separate datasets at the same index, suspiciously a power of 2, lead me to believe it's something internal.

from umap.

Okay cool, the 0.3dev umap doesn't throw the exception. Fabulous! I'll just keep using that :)

from umap.