/translation

Primary LanguagePython

Machine Translation Demo

Richard Kuzma, MAY 2022
Built with Streamlit front-end and HuggingFace NLP models

How to run:

  1. Clone git a) Make sure large files (pytorch models) are cloned too b) These are saved here for local running (No internet) c) Two files start with download* that are helpful for downloading models for first time
  2. cd into repo
  3. Build environment with requirements.txt
  4. In terminal, run the command: streamlit run app.py



Demonstration purpose:

  1. Showcase core NLP use cases for customers:
    • Machine translation
    • Topic Classification
    • Entity Extraction
  2. Show peripheral use cases:
    • Translation in multiple languages possible
    • Topics can be added
  3. Streamlit as a low-lift front-end
    • Python wrapper around React front-end components
    • Easy for Python programmers to use
    • Better user experience than dealing with jupyter noteboks
      • Easy to change inputs via buttons, text, drop-downs



Caveats:

  1. Off-the-shelf models, not fine-tuned. Human in the loop required for assessing validity



Time to run:

  • Loading models takes ~4 seconds
  • Inference takes ~2.5 seconds
  • CAVEAT: Running locally on a Macbook Air
    • Models saved locally
    • Running on GPU, distilling models, moving to onyx, quantizing models may all improve performance
    • Horizontal scaling needed for more throughput?