Text translation with facebook/nllb-200-3.3B model

nllb200_translate.py

#!/usr/bin/env python3

# Dependencies
# =============================
# pip install nltk transformers

import argparse
import sys
from pathlib import Path

# SEE ALSO: https://huggingface.co/docs/transformers/model_doc/nllb
# facebook/nllb-200-3.3B  # This is a larger model designed for translation tasks. It has a larger size than the distilled version, which might result in better quality translations but at the cost of slower inference
# facebook/nllb-200-distilled-600M # This model is designed for translation tasks and has been updated recently. It's a distilled version of the larger NLLB-200-3.3B model, which means it retains much of the performance of the larger model but with a smaller size
# facebook/nllb-200-1.3B # This model is another variant of the NLLB-200 series. It's designed for translation tasks and has been updated recently huggingface.co.
# michaelfeil/ct2fast-nllb-200-3.3B # This model is also a variant of the NLLB-200 series. It's designed for translation tasks and has been updated recentl
# facebook/mbart-large-50-many-to-many-mmt # This model is designed for many-to-many multilingual machine translation tasks. It has been updated recently


if __name__ == "__main__":
    parser = argparse.ArgumentParser(
        prog='nllb200 translator',
        description='''DESCRIPTION:
    Translate using https://huggingface.co/docs/transformers/model_doc/nllb.''',
        usage="""
SEE FULL language list: https://github.com/facebookresearch/flores/blob/main/flores200/README.md#languages-in-flores-200

| Language                           | FLORES-200 code |
|------------------------------------|-----------------|
| Catalan                            | cat_Latn        |
| Chinese (Simplified)               | zho_Hans        |
| Chinese (Traditional)              | zho_Hant        |
| English                            | eng_Latn        |
| French                             | fra_Latn        |
| German                             | deu_Latn        |
| Portuguese                         | por_Latn        |
| Spanish                            | spa_Latn        |
    """,
    epilog='Model card: https://huggingface.co/docs/transformers/model_doc/nllb')

    parser.add_argument('source',
                        nargs='?',
                        action='store',
                        help='REQUIRED: input file',
                        default='text.txt')

    parser.add_argument('--source-language', '-s',
                        help='source language',
                        action='store',
                        required=True)

    parser.add_argument('--target-language', '-t',
                        help='REQUIRED: target language',
                        action='store',
                        required=True)

    parser.add_argument('--verbose', '-v',
                        action='store_true',
                        help='verbosely list details')

    parser.add_argument('--force-overwrite', '-f',
                        action='store_true',
                        help='force overwrite')

    conf = parser.parse_args(sys.argv[1:])
    print(conf, file=sys.stderr)

    path_in = Path(conf.source)
    text_source = path_in.read_text(encoding='utf-8')

    from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
    from nltk import sent_tokenize

    segments = sent_tokenize(text_source, 'english')
    print(f'Generated {len(segments)} segments', file=sys.stderr, flush=True)

    print('loading model...', end='')

    # https://huggingface.co/facebook/nllb-200-3.3B
    MODEL = "facebook/nllb-200-3.3B"
    

    print(f'Using model: {MODEL}')

    tokenizer = AutoTokenizer.from_pretrained(MODEL,
                                              src_lang=conf.source_language)  # BCP-47 code
    # MORE LANGUAGE codes:
    # https://github.com/facebookresearch/flores/blob/main/flores200/README.md#languages-in-flores-200

    model = AutoModelForSeq2SeqLM.from_pretrained(MODEL)
    print('Ok')

    translated = []

    for idx, segment in enumerate(segments):
        if not len(segment) or segment[0].isdigit():
            decoded = segment
        else:

            inputs = tokenizer(segment, return_tensors="pt")

            #  SEE LANGUAGE CODES in transformers_fb_nllb200distilled600M_text.md
            translated_tokens = model.generate(
                **inputs,
                forced_bos_token_id=tokenizer.lang_code_to_id[conf.target_language],
                max_length=1024
            )
            decoded = tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]

        translated.append(decoded)
        print(decoded)

    path_out = path_in.with_suffix(f'.{MODEL.split("/")[1]}.{conf.source_language}-{conf.target_language}.txt')
    path_out.write_text('\n'.join(translated))
    print(f'Wrote {path_out.absolute()}')

hi

@meiyou313, there is a better model, check out madlad400 (400+ languages!) — there is also a rust-based CPU-optimized version that can be run via this fork (the official original repo is https://github.com/huggingface/candle, examples/quantized-t5); the fork allows taking an input file as a parameter (uses stdout for output, so the result needs to be redirected via the shell, the full command line would be something like

cargo run --example quantized-t5v \
  --release \
  --features accelerate -- \
  --model-id "jbochi/madlad400-3b-mt" \
  --weight-file "model-q4k.gguf" 
  --target-language es
  --file-path input.txt \
  --temperature 0 > result.txt

quantized-t5v is the modified version of the quantized-t5 in the original repo

all languages supported by madlad-400:

• https://huggingface.co/google/madlad400-10b-mt/blob/main/README.md

cuz huggingface/transformers#31348
you will get AttributeError: 'NllbTokenizerFast' object has no attribute 'lang_code_to_id' error
you can downgrade transformers to 4.37.0 or replace
forced_bos_token_id=tokenizer.lang_code_to_id[conf.target_language],
to
forced_bos_token_id=tokenizer.convert_tokens_to_ids(conf.target_language),