Support src_lang/tgt_lang in InferenceClient.translation()

[Wauplin]

Originally from @LysandreJik :

For some models (and some models only), it is possible to specify a source and target language in transformers and InferenceAPI. It would be nice to support it in InferenceClient as well. Example: facebook/mbart-large-50-many-to-many-mmt

from huggingface_hub import InferenceClient

data = InferenceClient().post(
    json={
        "inputs": "My name is Sarah Jessica Parker but you can call me Jessica",
        "parameters": {"src_lang": "en_XX", "tgt_lang": "fr_XX"},
    },
    model="https://api-inference.huggingface.co/models/facebook/mbart-large-50-many-to-many-mmt",
)

print(data)

b'[{"translation_text":"Mon nom est Sarah Jessica Parker mais vous pouvez m\'appeler Jessica"}]'

---

TODO:

• add both parameters src_lang: Optional[str] = None and tgt_lang: Optional[str] = None to InferenceClient.translation
• raise a ValueError if only one of them is set but not the other (?)
• send them as "parameters" only if not None
• add test with facebook/mbart-large-50-many-to-many-mmt
• update docstring

[AkshathRaghav]

@Wauplin I'd like to work on this. Could you tell me how I can find all the models which support these params?

[Wauplin]

Hi @AkshathRaghav, thanks for proposing your help on this one! 🙏 Unfortunately, there is no easy way to list models supporting these parameters. This is why it should be made clear in the docstring that it is at the user discretion to select a model that handles them. In general, the information will be found in the Model Card of the model on the Hub (along with the list of support language codes).

Therefore as a first step, I would simply implement the feature in InferenceClient for users that want to use it. Then in a second time, and if there is some demand, we could try to find a way to suggest models that support them but it's not trivial (that's the counterpart of having a versatile transformers library).

[ceferisbarov]

@AkshathRaghav are you still working on this? I have implemented this in my fork. I can open a PR.

[AkshathRaghav]

@ceferisbarov HI, I haven't gotten a chance to work on it. Please go ahead!

[Wauplin]

Closing this issue thanks to #1869 by @ceferisbarov ! 🚀

>>> from huggingface_hub import InferenceClient
>>> client = InferenceClient()
>>> client.translation("My name is Sarah Jessica Parker but you can call me Jessica", model="facebook/mbart-large-50-many-to-many-mmt", src_lang="en_XX", tgt_lang="fr_XX")
"Mon nom est Sarah Jessica Parker mais vous pouvez m'appeler Jessica"
>>> client.translation("My name is Sarah Jessica Parker but you can call me Jessica", model="facebook/mbart-large-50-many-to-many-mmt", src_lang="en_XX", tgt_lang="es_XX")
'Mi nombre es Sarah Jessica Parker pero puedes llamarme Jessica'