asian_mtl

This repository contains the code and documentation for the machine translation models used for EasierMTL's API.

Improved version of the models in the original repository: EasierMTL/chinese-translation-app

Supported Translators

All translators support dynamic quantization! Our benchmarks indicate that they 2x inference speeds, while losing <1% BLEU.

• ChineseToEnglishTranslator()
• EnglishToChineseTranslator()

Getting Started

pip install asian-mtl

Here's a simple example:

from asian_mtl.models.base import ChineseToEnglishTranslator

translator = ChineseToEnglishTranslator()
# Quantize for better CPU production performance!
translator.quantize()

prediction = translator.predict("我爱ECSE484.")
print(prediction)
# prediction will be:
# "I love ECSE 484."

And you're good to go!

If you are contributing, run:

# https://stackoverflow.com/questions/59882884/vscode-doesnt-show-poetry-virtualenvs-in-select-interpreter-option

poetry config virtualenvs.in-project true

# shows the name of the current environment
poetry env list

poetry install

Usage

When you are using quantized models in this repository, make sure to set torch.set_num_threads(1). This is not set under-the-hood because it could interfere with user setups in an invasive way.

Not doing so will make the quantized models slower than their vanilla counterparts.

Evaluation

See scripts for evaluation scripts.

To run the scripts, simply run:

# Running with CLI and config with BERT
python ./scripts/evaluation/eval.py -c ./scripts/evaluation/configs/helsinki.yaml

Change the config helsinki.yaml to use quantized or your specific use case.

Benchmarks

Here are some basic benchmarks of models in this repository:

Model	Quantized?	N	BLEU	Runtime
Helsinki-NLP/opus-mt-zh-en	No	100	0.319	27s
	Yes	100	0.306	13.5s

The benchmarks described in the docs are a little out-of-date.