Skip to content

Latest commit

 

History

History
83 lines (64 loc) · 3.76 KB

README.md

File metadata and controls

83 lines (64 loc) · 3.76 KB
Raw

Docker

We provide a Docker based REST-API for EasyNMT: Send a query with your source text, and the API returns the translated text.

Setup

To start the EasyNMT REST API on port 24080run the following docker command:

docker run -p 24080:80 easynmt/api:2.0-cpu

This uses the CPU image. If you have GPU (CUDA), there are various GPU images available. Have a look at our Docker Hub Page.

Usage

After you started the Docker image, you can visit: http://localhost:24080/translate?target_lang=en&text=Hallo%20Welt

This should yield the following JSON:

{
    "target_lang": "en",
    "source_lang": null,
    "detected_langs": [
        "de"
    ],
    "translated": [
        "Hello world"
    ],
    "translation_time": 0.163145542144775
}

If you have started it with a different port, replace 24080 with the port you chose.

Note, for the first translation, the respective models are downloaded. This might take some time. Consecutive calls will be faster.

Programmatic Usage

Documentation

To get an overview of all REST API endpoints, with all possible parameters and their description, you open the following url: http://localhost:24080/docs

Endpoints

The following endpoints with the GET method are defined (i.e. you can call them like http://localhost:24080/name?param1=val1&param2=val2)

/translate
    Translates the text to the given target language.
    :param text: Text that should be translated
    :param target_lang: Target language
    :param source_lang: Language of text. Optional, if empty: Automatic language detection
    :param beam_size: Beam size. Optional
    :param perform_sentence_splitting: Split longer documents into individual sentences for translation. Optional
    :return:  Returns a json with the translated text

/language_detection
    Detects the language for the provided text
    :param text: A single text for which we want to know the language
    :return: The detected language
    
/get_languages
    Returns the languages the model supports
    :param source_lang: Optional. Only return languages with this language as source
    :param target_lang: Optional. Only return languages with this language as target
    :return:

You can call the /translate and /language_detection also with a POST request, giving you the option to pass a list of multiple texts. Then all texts are translated and returned at once.

Environment Variables

You can control the Docker image using various environment variables:

  • MAX_WORKERS_BACKEND: Number of worker processes for the translation. Default: 1
  • MAX_WORKERS_FRONTEND: Number of worker processes for language detection & model info. Default: 2
  • EASYNMT_MODEL: Which EasyNMT Model to load. Default: opus-mt
  • EASYNMT_MODEL_ARGS: Json encoded string with parameters when loading EasyNMT: Default: {}
  • EASYNMT_MAX_TEXT_LEN: Maximal text length for translation. Default: Not set
  • EASYNMT_MAX_BEAM_SIZE: Maximal beam size for translation. Default: Not set
  • EASYNMT_BATCH_SIZE: Batch size for translation. Default: 16
  • TIMEOUT: Gunicorn timeout. Default: 120

All model files are stored at /cache/. You can mount this path to your host machine if you want to re-use previously downloaded models.