Investigate offline translation

[willwade]

See MarianMT

install PyTorch, then pip install sentencepiece and huggingface, the you can run:

from transformers import MarianMTModel, MarianTokenizer
from typing import Sequence

class Translator:
    def __init__(self, source_lang: str, dest_lang: str) -> None:
        self.model_name = f'Helsinki-NLP/opus-mt-{source_lang}-{dest_lang}'
        self.model = MarianMTModel.from_pretrained(self.model_name)
        self.tokenizer = MarianTokenizer.from_pretrained(self.model_name)

    def translate(self, texts: Sequence[str]) -> Sequence[str]:
        tokens = self.tokenizer(list(texts), return_tensors="pt", padding=True)
        translate_tokens = self.model.generate(**tokens)
        return [self.tokenizer.decode(t, skip_special_tokens=True) for t in translate_tokens]

marian_ru_en = Translator('ru', 'en')
marian_ru_en.translate(['что слишком сознавать — это болезнь, настоящая, полная болезнь.'])
# Returns: ['That being too conscious is a disease, a real, complete disease.']

NB:

• Not great at punctuation. Would potentially be wise to strip
• Config would have to download the models in advance.
• Note limited tokens - you need to break it down and parallel run this see https://skeptric.com/python-offline-translation/
• Test CPU/GPU
• Need to test pyfreezing this

[willwade]

This looks promising too https://github.com/facebookresearch/seamless_communication and an on-device model which is around 230Mb https://github.com/facebookresearch/seamless_communication/blob/main/docs/m4t/on_device_README.md

[willwade]

closing this for now. too tricky