I am using facebook/mbart-large-50-many-to-many-mmt model to translate texts in different directions.
Using generate, it looks that the model works fine. However, using pipeline the model return random tokens.
from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
article_en = "When you have a medical appointment, your health provider writes notes on the visit that are available to you"
article_fr = "Les infirmières praticiennes et infirmiers praticiens sont des membres du personnel infirmier autorisé qui possèdent une formation et une expérience plus poussées et qui peuvent poser un diagnostic"
model = MBartForConditionalGeneration.from_pretrained(
"facebook/mbart-large-50-many-to-many-mmt")
tokenizer = MBart50TokenizerFast.from_pretrained(
"facebook/mbart-large-50-many-to-many-mmt")
# using generate
tokenizer.src_lang = "en_XX"
encoded_ar = tokenizer(article_en, return_tensors="pt")
generated_tokens = model.generate(
**encoded_ar,
forced_bos_token_id=tokenizer.lang_code_to_id["fr_XX"],
max_length=512
)
pred = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# it returns:
['Lorsque vous avez un rendez-vous médical, votre professionnel de la santé vous envoie des notes sur la visite qui vous sont disponibles.']
print(pred, "\n", "-"*30)
transl = pipeline(task="translation_en_to_fr",
model=model, tokenizer=tokenizer, device=0, src_lang="en_XX", tgt_lang="fr_XX",
generation_kwargs={"skip_special_tokens": True,
"forced_bos_token_id": tokenizer.lang_code_to_id["fr_XX"]})
pred = transl(article_en)
print(pred)
[{'translation_text': 'fr http http http http http'}]