A translation library for 200 languages built on Huggingface transformers

💻 GitHub Repository
📚 Documentation
🐍 PyPi project
🧪 Colab Demo / Kaggle Demo

Install the library with pip:

pip install dl-translate

To translate some text:

import dl_translate as dlt

mt = dlt.TranslationModel()  # Slow when you load it for the first time

text_hi = "संयुक्त राष्ट्र के प्रमुख का कहना है कि सीरिया में कोई सैन्य समाधान नहीं है"
mt.translate(text_hi, source=dlt.lang.HINDI, target=dlt.lang.ENGLISH)

Above, you can see that dlt.lang contains variables representing each of the 200 available languages with auto-complete support. Alternatively, you can specify the language (e.g. "Arabic") or the language code (e.g. "fr" for French):

text_ar = "الأمين العام للأمم المتحدة يقول إنه لا يوجد حل عسكري في سوريا."
mt.translate(text_ar, source="Arabic", target="fr")

If you want to verify whether a language is available, you can check it:

print(mt.available_languages())  # All languages that you can use
print(mt.available_codes())  # Code corresponding to each language accepted
print(mt.get_lang_code_map())  # Dictionary of lang -> code

When you load the model, you can specify the device:

mt = dlt.TranslationModel(device="auto")

By default, the value will be device="auto", which means it will use a GPU if possible. You can also explicitly set device="cpu" or device="gpu", or some other strings accepted by torch.device(). In general, it is recommend to use a GPU if you want a reasonable processing time.

By default, the m2m100 model will be used. However, there are a few options:

• mBART-50 Large: Allows translations across 50 languages.
• m2m100: Allows translations across 100 languages.
• nllb-200 (New in v0.3): Allows translations across 200 languages, and is faster than m2m100 (On RTX A6000, we can see speed up of 3x).

Here's an example:

# The default approval
mt = dlt.TranslationModel("m2m100")  # Shorthand
mt = dlt.TranslationModel("facebook/m2m100_418M")  # Huggingface repo

# If you want to use mBART-50 Large
mt = dlt.TranslationModel("mbart50")
mt = dlt.TranslationModel("facebook/mbart-large-50-many-to-many-mmt")

# Or NLLB-200 (faster and has 200 languages)
mt = dlt.TranslationModel("nllb200")
mt = dlt.TranslationModel("facebook/nllb-200-distilled-600M")

Note that the language code will change depending on the model family. To find out the correct language codes, please read the doc page on available languages or run mt.available_codes().

By default, dlt.TranslationModel will download the model from the huggingface repo for mbart50, m2m100, or nllb200 and cache it. It's possible to load the model from a path or a model with a similar format, but you will need to specify the model_family:

mt = dlt.TranslationModel("/path/to/model/directory/", model_family="mbart50")
mt = dlt.TranslationModel("facebook/m2m100_1.2B", model_family="m2m100")
mt = dlt.TranslationModel("facebook/nllb-200-distilled-600M", model_family="nllb200")

Notes:

• Make sure your tokenizer is also stored in the same directory if you load from a file.
• The available languages will change if you select a different model, so you will not be able to leverage dlt.lang or dlt.utils.

It is not recommended to use extremely long texts as it takes more time to process. Instead, you can try to break them down into sentences with the help of nltk. First install the library with pip install nltk, then run:

import nltk

nltk.download("punkt")

text = "Mr. Smith went to his favorite cafe. There, he met his friend Dr. Doe."
sents = nltk.tokenize.sent_tokenize(text, "english")  # don't use dlt.lang.ENGLISH
" ".join(mt.translate(sents, source=dlt.lang.ENGLISH, target=dlt.lang.FRENCH))

It's possible to set a batch size (i.e. the number of elements processed at once) for mt.translate and whether you want to see the progress bar or not:

# ...
mt = dlt.TranslationModel()
mt.translate(text, source, target, batch_size=32, verbose=True)

If you set batch_size=None, it will compute the entire text at once rather than splitting into "chunks". We recommend lowering batch_size if you do not have a lot of RAM or VRAM and run into CUDA memory error. Set a higher value if you are using a high-end GPU and the VRAM is not fully utilized.

An alternative to mt.available_languages() is the dlt.utils module. You can use it to find out which languages and codes are available:

print(dlt.utils.available_languages('mbart50'))  # All languages that you can use
print(dlt.utils.available_codes('m2m100'))  # Code corresponding to each language accepted
print(dlt.utils.get_lang_code_map('nllb200'))  # Dictionary of lang -> code

Unlike the Google translate or MSFT Translator APIs, this library can be fully used offline. However, you will need to first download the packages and models, and move them to your offline environment to be installed and loaded inside a venv.

First, run in your terminal:

mkdir dlt
cd dlt
mkdir libraries
pip download -d libraries/ dl-translate

Once all the required packages are downloaded, you will need to use huggingface hub to download the files. Install it with pip install huggingface-hub. Then, run inside Python:

import shutil
import huggingface_hub as hub

dirname = hub.snapshot_download("facebook/m2m100_418M")
shutil.copytree(dirname, "cached_model_m2m100")  # Copy to a permanent folder

Now, move everything in the dlt directory to your offline environment. Create a virtual environment and run the following in terminal:

pip install --no-index --find-links libraries/ dl-translate

Now, run inside Python:

import dl_translate as dlt

mt = dlt.TranslationModel("cached_model_m2m100", model_family="m2m100")

If you have knowledge of PyTorch and Huggingface Transformers, you can access advanced aspects of the library for more customization:

• Saving and loading: If you wish to accelerate the loading time the translation model, you can use save_obj and reload it later with load_obj. This method is only recommended if you are familiar with huggingface and torch; please read the docs for more information.
• Interacting with underlying model and tokenizer: When initializing model, you can pass in arguments for the underlying BART model and tokenizer with model_options and tokenizer_options respectively. You can also access the underlying transformers with mt.get_transformers_model().
• Keyword arguments for the generate() method: When running mt.translate, you can also give generation_options that is passed to the generate() method of the underlying transformer model.

For more information, please visit the advanced section of the user guide.

dl-translate is built on top of Huggingface's implementation of two models created by Facebook AI Research.

1. The multilingual BART finetuned on many-to-many translation of over 50 languages, which is documented here The original paper was written by Tang et. al from Facebook AI Research; you can find it here and cite it using the following:

   @article{tang2020multilingual,
       title={Multilingual translation with extensible multilingual pretraining and finetuning},
       author={Tang, Yuqing and Tran, Chau and Li, Xian and Chen, Peng-Jen and Goyal, Naman and Chaudhary, Vishrav and Gu, Jiatao and Fan, Angela},
       journal={arXiv preprint arXiv:2008.00401},
       year={2020}
   }
   

2. The transformer model published in Beyond English-Centric Multilingual Machine Translation by Fan et. al, which supports over 100 languages. You can cite it here:

   @misc{fan2020englishcentric,
        title={Beyond English-Centric Multilingual Machine Translation}, 
        author={Angela Fan and Shruti Bhosale and Holger Schwenk and Zhiyi Ma and Ahmed El-Kishky and Siddharth Goyal and Mandeep Baines and Onur Celebi and Guillaume Wenzek and Vishrav Chaudhary and Naman Goyal and Tom Birch and Vitaliy Liptchinsky and Sergey Edunov and Edouard Grave and Michael Auli and Armand Joulin},
        year={2020},
        eprint={2010.11125},
        archivePrefix={arXiv},
        primaryClass={cs.CL}
    }
   

3. The no language left behind model, which extends NMT to 200+ languages. You can cite it here:

   @misc{nllbteam2022language,
       title={No Language Left Behind: Scaling Human-Centered Machine Translation}, 
       author={NLLB Team and Marta R. Costa-jussà and James Cross and Onur Çelebi and Maha Elbayad and Kenneth Heafield and Kevin Heffernan and Elahe Kalbassi and Janice Lam and Daniel Licht and Jean Maillard and Anna Sun and Skyler Wang and Guillaume Wenzek and Al Youngblood and Bapi Akula and Loic Barrault and Gabriel Mejia Gonzalez and Prangthip Hansanti and John Hoffman and Semarley Jarrett and Kaushik Ram Sadagopan and Dirk Rowe and Shannon Spruit and Chau Tran and Pierre Andrews and Necip Fazil Ayan and Shruti Bhosale and Sergey Edunov and Angela Fan and Cynthia Gao and Vedanuj Goswami and Francisco Guzmán and Philipp Koehn and Alexandre Mourachko and Christophe Ropers and Safiyyah Saleem and Holger Schwenk and Jeff Wang},
       year={2022},
       eprint={2207.04672},
       archivePrefix={arXiv},
       primaryClass={cs.CL}
   }
   

dlt is a wrapper with useful utils to save you time. For huggingface's transformers, the following snippet is shown as an example:

from transformers import MBartForConditionalGeneration, MBart50TokenizerFast

article_hi = "संयुक्त राष्ट्र के प्रमुख का कहना है कि सीरिया में कोई सैन्य समाधान नहीं है"
article_ar = "الأمين العام للأمم المتحدة يقول إنه لا يوجد حل عسكري في سوريا."

model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")

# translate Hindi to French
tokenizer.src_lang = "hi_IN"
encoded_hi = tokenizer(article_hi, return_tensors="pt")
generated_tokens = model.generate(**encoded_hi, forced_bos_token_id=tokenizer.lang_code_to_id["fr_XX"])
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "Le chef de l 'ONU affirme qu 'il n 'y a pas de solution militaire en Syria."

# translate Arabic to English
tokenizer.src_lang = "ar_AR"
encoded_ar = tokenizer(article_ar, return_tensors="pt")
generated_tokens = model.generate(**encoded_ar, forced_bos_token_id=tokenizer.lang_code_to_id["en_XX"])
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "The Secretary-General of the United Nations says there is no military solution in Syria."

With dlt, you can run:

import dl_translate as dlt

article_hi = "संयुक्त राष्ट्र के प्रमुख का कहना है कि सीरिया में कोई सैन्य समाधान नहीं है"
article_ar = "الأمين العام للأمم المتحدة يقول إنه لا يوجد حل عسكري في سوريا."

mt = dlt.TranslationModel()
translated_fr = mt.translate(article_hi, source=dlt.lang.HINDI, target=dlt.lang.FRENCH)
translated_en = mt.translate(article_ar, source=dlt.lang.ARABIC, target=dlt.lang.ENGLISH)

Notice you don't have to think about tokenizers, condition generation, pretrained models, and regional codes; you can just tell the model what to translate!

If you are experienced with huggingface's ecosystem, then you should be familiar enough with the example above that you wouldn't need this library. However, if you've never heard of huggingface or mBART, then I hope using this library will give you enough motivation to learn more about them :)