Slim down ML packages #105
Description
Arabic transliteration has a lot of Torch-based dependencies. The Docker image is currently over 3Gb and it takes a long time to build.
Some of this has already been mitigated by splitting off a base image for ML tools, but even pulling the layers from the base image is quite slow.
Ideally, we would want to know what is needed for building a language model (done only once) and what is needed to perform actions against the pre-built model (done for every request). If the size difference is considerable, we could maintain an ad-hoc image just for rebuilding the Arabic LLM, and one for everyday use.