Support Offline Translation Between SignWriting and Spoken Language

[AmitMY]

Problem

The application currently relies on a server endpoint for spoken-language to SignWriting and SignWriting to spoken-language text-to-text translation.

This prevents us from performing translation offline and makes us reliant on a server that can crash, with an unknown scaling ability.

Description

Using bergamot (already made model training pipeline), we could train generic/language specific translation models.
There are a few issues with the current code:

• Vocabulary is shared between source and target sequences Support training separate source/target SentencePiece Models mozilla/firefox-translations-training#97
• Only trains on a single GPU, can't leverage multiple ones NCCL error 2 'unhandled system error' only with 4 GPUs marian-nmt/marian-dev#899
• Models do not perform well. Might be overfitting or underfitting, requires further exploration.

We already support bergamot worker inference in our client side and server side, but further work needs to be done on the training code to make the resulting models actually useful.

Alternatives

We could use any other framework than bergamot, probably giving us more accessible editing, on the expense of speed.at

[AmitMY]

The offline translation is supported, however, the current main issue is that the quality is horrendous.

[AmitMY]

Model translation quality has been reportedly fixed, but now the problem is that we can't seem to run the pipeline using snakemake like before.

[AmitMY]

To continue, see: mozilla/firefox-translations-training#586