Machine translations

Hypha has the ability to utilize argostranslate for machine translations of submitted application content. This is disabled by default and the dependencies are not installed to prevent unneeded bloat due to PyTorch's large language models.

Installing dependencies

As referenced in the production deployment guide, it is required to install the dependencies needed for machine translation dependencies via

python3 -m pip install -r requirements/translate.txt

or, if you are on a platform that does not support GPU processing:

python3 -m pip install -r requirements/translate-cpu.txt

This requirements file will specifically attempt to install the CPU version of PyTorch if available on the detected platform to play better with heroku (doesn't support GPU processing) and to minimize package bloat (CPU package is ~300MB less than the normal GPU). Depending on your use case, you may want to adjust this.

Installing languages

Argostranslate handles translations via it's own packages - ie. Arabic -> English translation would be one package, while English -> Arabic would be another.

Installing/uninstalling these packages can be done with the management commands install_languages/uninstall_languages respectively, utilizing the format of _. For example, installing the Arabic -> English & French -> English packages would look like:

python3 manage.py install_languages ar_en fr_en

Enabling on the system

To enable machine translations on an instance, the proper configuration variables need to be set. These can be found in the configuration options