from transformers import FSMTForConditionalGeneration, FSMTTokenizer
mname = "facebook/wmt19-de-en"
tokenizer = FSMTTokenizer.from_pretrained(mname)
model = FSMTForConditionalGeneration.from_pretrained(mname)

input = "Maschinelles Lernen ist großartig, oder?"
input_ids = tokenizer.encode(input, return_tensors="pt")
outputs = model.generate(input_ids)
decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(decoded) # Machine learning is great, isn't it?

Limitations and bias

The original (and this ported model) doesn't seem to handle well inputs with repeated sub-phrases, content gets truncated

Training data

Pretrained weights were left identical to the original model released by fairseq. For more details, please, see the paper.

Eval results

pair	fairseq	transformers
de-en	42.3	41.35

The score is slightly below the score reported by fairseq, since `transformers`` currently doesn't support:

model ensemble, therefore the best performing checkpoint was ported (model4.pt).
re-ranking

The score was calculated using this code:

bash

git clone https://github.com/huggingface/transformers
cd transformers
export PAIR=de-en
export DATA_DIR=data/$PAIR
export SAVE_DIR=data/$PAIR
export BS=8
export NUM_BEAMS=15
mkdir -p $DATA_DIR
sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
echo $PAIR
PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS

note: fairseq reports using a beam of 50, so you should get a slightly higher score if re-run with --num_beams 50.

Data Sources

BibTeX entry and citation info

bibtex

@inproceedings{...,
  year={2020},
  title={Facebook FAIR's WMT19 News Translation Task Submission},
  author={Ng, Nathan and Yee, Kyra and Baevski, Alexei and Ott, Myle and Auli, Michael and Edunov, Sergey},
  booktitle={Proc. of WMT},
}

TODO

port model ensemble (fairseq uses 4 model checkpoints)