TutteInstitute / vectorizers

Vectorizers for a range of different data types
BSD 3-Clause "New" or "Revised" License
97 stars 23 forks source link

Fixes for spherical vectors (e.g. word vectors) #55

Closed lmcinnes closed 3 years ago

lmcinnes commented 3 years ago

There were some catches in the numba handling of spherical vectors which I fixed up, plus a few other minor fixes along the way, and I added a square root for spherical vectors/l2 normalized vectors prior to the SVD for the same reason that we have a fourth root in the word vectors. It does improve performance on downstream tasks in the examples I tried.