| --- |
| language: |
| - en |
| - ru |
| - kz |
|
|
| tags: |
| - translation |
| - wmt19 |
| - facebook |
| license: apache-2.0 |
| datasets: |
| - wmt19 |
| metrics: |
| - bleu |
| thumbnail: https://huggingface.co/front/thumbnails/facebook.png |
| --- |
| |
| # FSMT |
|
|
| ## Model description |
|
|
| This is a ported version of [fairseq wmt19 transformer](https://github.com/pytorch/fairseq/blob/master/examples/wmt19/README.md) for en-ru. |
|
|
| For more details, please see, [Facebook FAIR's WMT19 News Translation Task Submission](https://arxiv.org/abs/1907.06616). |
|
|
| The abbreviation FSMT stands for FairSeqMachineTranslation |
|
|
| All four models are available: |
|
|
| * [wmt19-en-ru](https://huggingface.co/facebook/wmt19-en-ru) |
| * [wmt19-ru-en](https://huggingface.co/facebook/wmt19-ru-en) |
| * [wmt19-en-de](https://huggingface.co/facebook/wmt19-en-de) |
| * [wmt19-de-en](https://huggingface.co/facebook/wmt19-de-en) |
|
|
| ## Intended uses & limitations |
|
|
| #### How to use |
|
|
| ```python |
| from transformers import FSMTForConditionalGeneration, FSMTTokenizer |
| mname = "facebook/wmt19-en-ru" |
| tokenizer = FSMTTokenizer.from_pretrained(mname) |
| model = FSMTForConditionalGeneration.from_pretrained(mname) |
| |
| input = "Machine learning is great, isn't it?" |
| input_ids = tokenizer.encode(input, return_tensors="pt") |
| outputs = model.generate(input_ids) |
| decoded = tokenizer.decode(outputs[0], skip_special_tokens=True) |
| print(decoded) # Машинное обучение - это здорово, не так ли? |
| |
| ``` |
|
|
| #### Limitations and bias |
|
|
| - The original (and this ported model) doesn't seem to handle well inputs with repeated sub-phrases, [content gets truncated](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fissues-with-translating-inputs-containing-repeated-phrases%2F981%3C%2Fspan%3E)%3C!----%3E%3C%2Ftd%3E%3C%2Ftr%3E%3Ctr id="L57"> | |
|
| ## Training data |
|
|
| Pretrained weights were left identical to the original model released by fairseq. For more details, please, see the [paper](https://arxiv.org/abs/1907.06616). |
|
|
| ## Eval results |
|
|
| pair | fairseq | transformers |
| -------|---------|---------- |
| en-ru | [36.4](http://matrix.statmt.org/matrix/output/1914?run_id=6724) | 33.47 |
|
|
| 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=en-ru |
| 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 |
|
|
| - [training, etc.](http://www.statmt.org/wmt19/) |
| - [test set](http://matrix.statmt.org/test_sets/newstest2019.tgz?1556572561) |
|
|
|
|
| ### 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) |
