CU-8696nbm9j: Add module to convert vocab vectors

[mart-r]

Adds a module to convert the vocab vectors from the default (or really anything) to a smaller length.

The default vocab vector length is 300. However, we don't really make use of all this information. Experiments show that we can go quite a lot smaller in vocab size and retain the same performance. See e.g: https://gist.github.com/mart-r/e9db909cde1922464bcc753f54006994 Or (somewhat more comprehensively): https://gist.github.com/mart-r/21460286466d17b9f23719ba3f4dc938

The benefits of using a smaller vocab size mainly boil down to (examples at 50 vector size):

• Smaller saved vocab on disk
  • The vocab size can go from 314MB down to 142MB
  • Or even to 68MB when ditching unigram table (see #503)
  • The CDB size will also go down significantly
  • Because the context vectors stored within it depend on the vectors in the Vocab
  • In a MIMIC-IV trained model it went from 1.7GB to 1.3GB
  • But this effect can be larger if more concepts have been trained on
  • The model pack size will also change accordingly
  • Normal model (MIMIC IV trained) zip was 1.0GB
  • Down sized model (same model) zip was 410MB
• Potentially faster load/save times
  • Since the files will be smaller
  • Though I don't have good evidence for that
  • Loading already unpacked (this can very well be run-to-run variance)
  • Normal: 17.4s
  • Down sized: 17.1s
  • Loading before unpacking
  • Normal: 25.7s
  • Down sized: 22.1s

NOTE: There might be improvements we could do here:

• Should this be in another module?
• Should we add CLI for model pack conversion?

[tomolopolis]

Task linked: CU-8696nbm9j Support changing of vocab vector size