|
--- |
|
tags: |
|
- ctranslate2 |
|
- translation |
|
license: apache-2.0 |
|
--- |
|
# # Fast-Inference with Ctranslate2 |
|
Speedup inference by 2x-8x using int8 inference in C++ |
|
|
|
quantized version of [Helsinki-NLP/opus-mt-fr-de](https://huggingface.co/Helsinki-NLP/opus-mt-fr-de) |
|
```bash |
|
pip install hf-hub-ctranslate2>=1.0.0 ctranslate2>=3.13.0 |
|
``` |
|
Converted using |
|
``` |
|
ct2-transformers-converter --model Helsinki-NLP/opus-mt-fr-de --output_dir /home/michael/tmp-ct2fast-opus-mt-fr-de --force --copy_files README.md generation_config.json tokenizer_config.json vocab.json source.spm .gitattributes target.spm --quantization float16 |
|
``` |
|
|
|
Checkpoint compatible to [ctranslate2](https://github.com/OpenNMT/CTranslate2) and [hf-hub-ctranslate2](https://github.com/michaelfeil/hf-hub-ctranslate2) |
|
- `compute_type=int8_float16` for `device="cuda"` |
|
- `compute_type=int8` for `device="cpu"` |
|
|
|
```python |
|
from hf_hub_ctranslate2 import TranslatorCT2fromHfHub, GeneratorCT2fromHfHub |
|
from transformers import AutoTokenizer |
|
|
|
model_name = "michaelfeil/ct2fast-opus-mt-fr-de" |
|
# use either TranslatorCT2fromHfHub or GeneratorCT2fromHfHub here, depending on model. |
|
model = TranslatorCT2fromHfHub( |
|
# load in int8 on CUDA |
|
model_name_or_path=model_name, |
|
device="cuda", |
|
compute_type="int8_float16", |
|
tokenizer=AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-fr-de") |
|
) |
|
outputs = model.generate( |
|
text=["How do you call a fast Flan-ingo?", "User: How are you doing?"], |
|
) |
|
print(outputs) |
|
``` |
|
|
|
# Licence and other remarks: |
|
This is just a quantized version. Licence conditions are intended to be idential to original huggingface repo. |
|
|
|
# Original description |
|
|
|
|
|
### opus-mt-fr-de |
|
|
|
* source languages: fr |
|
* target languages: de |
|
* OPUS readme: [fr-de](https://github.com/Helsinki-NLP/OPUS-MT-train/blob/master/models/fr-de/README.md) |
|
|
|
* dataset: opus |
|
* model: transformer-align |
|
* pre-processing: normalization + SentencePiece |
|
* download original weights: [opus-2020-01-09.zip](https://object.pouta.csc.fi/OPUS-MT-models/fr-de/opus-2020-01-09.zip) |
|
* test set translations: [opus-2020-01-09.test.txt](https://object.pouta.csc.fi/OPUS-MT-models/fr-de/opus-2020-01-09.test.txt) |
|
* test set scores: [opus-2020-01-09.eval.txt](https://object.pouta.csc.fi/OPUS-MT-models/fr-de/opus-2020-01-09.eval.txt) |
|
|
|
## Benchmarks |
|
|
|
| testset | BLEU | chr-F | |
|
|-----------------------|-------|-------| |
|
| euelections_dev2019.transformer-align.fr | 26.4 | 0.571 | |
|
| newssyscomb2009.fr.de | 22.1 | 0.524 | |
|
| news-test2008.fr.de | 22.1 | 0.524 | |
|
| newstest2009.fr.de | 21.6 | 0.520 | |
|
| newstest2010.fr.de | 22.6 | 0.527 | |
|
| newstest2011.fr.de | 21.5 | 0.518 | |
|
| newstest2012.fr.de | 22.4 | 0.516 | |
|
| newstest2013.fr.de | 24.2 | 0.532 | |
|
| newstest2019-frde.fr.de | 27.9 | 0.595 | |
|
| Tatoeba.fr.de | 49.1 | 0.676 | |
|
|
|
|
