btw what did you like most about using the magnitude format? lazy loading seems good, depending on if it takes time to load the sqlite DB at each startup. That'

btw what did you like most about using the magnitude format? lazy loading seems good, depending on if it takes time to load the sqlite DB at each startup. That's one of the major pains of working on anything with word embeddings is that server reloads / hot restarts take forever. The fastText vectors (see below) also use subword embeddings so I think they had potentially better results. I used FT for some Chinese stuff and I think it worked better for that since chinese characters are so much more important than latin scripts. Seems magnitude python API also has some other features like POS tagging, I wonder how that compares to say spaCy.