2.9 KiB
This model was released on 2019-07-15 and added to Hugging Face Transformers on 2020-11-16 and contributed by stas.
FSMT
FSMT models participated in the WMT19 shared news translation task for English <-> German and English <-> Russian. The systems are large BPE-based transformer models trained with Fairseq, utilizing sampled back-translations. This year, experiments included various bitext data filtering schemes and the addition of filtered back-translated data. Models were ensembled and fine-tuned on domain-specific data, with decoding enhanced by noisy channel model reranking. The submissions achieved top rankings in all four directions, with the En->De system outperforming other systems and human translations, improving by 4.5 BLEU points from the WMT'18 submission.
import torch
from transformers import pipeline
pipeline = pipeline(task="text2text-generation", model="facebook/wmt19-en-de", dtype="auto")
pipeline("Plants generate energy through a process known as photosynthesis.")
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
model = AutoModelForSeq2SeqLM.from_pretrained("facebook/wmt19-en-de", dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("facebook/wmt19-en-de")
inputs = tokenizer("Plants generate energy through a process known as photosynthesis.", return_tensors="pt")
outputs = model.generate(**inputs, max_length=50)
print(tokenizer.decode(outputs[0]))
Usage tips
- FSMT uses separate source and target vocabulary pairs. It doesn't share embedding tokens. The tokenizer is similar to [
XLMTokenizer
] and the main model derives from [BartModel
].
FSMTConfig
autodoc FSMTConfig
FSMTTokenizer
autodoc FSMTTokenizer - build_inputs_with_special_tokens - get_special_tokens_mask - create_token_type_ids_from_sequences - save_vocabulary
FSMTModel
autodoc FSMTModel - forward
FSMTForConditionalGeneration
autodoc FSMTForConditionalGeneration - forward