Property	Value
Original model	Helsinki-NLP/opus-mt-es-de
Architecture	MarianMT (encoder-decoder)
Encoder layers	6
Decoder layers	6
Attention heads	8
d_model	512
FFN dimension	2048
Activation function	swish
Vocabulary size	61301
Max sequence length	512
Beam size	4
Quantization	int8
Format	CTranslate2
Tokenizer	SentencePiece (source.spm / target.spm)

Usage

With CTranslate2

python

import ctranslate2
import sentencepiece as spm

model_path = "mijuanlo/opus-mt-es-de-ct2-int8"
sp_source = spm.SentencePieceProcessor(f"{model_path}/source.spm")
sp_target = spm.SentencePieceProcessor(f"{model_path}/target.spm")

translator = ctranslate2.Translator(model_path)

def translate(text: str) -> str:
    input_ids = sp_source.encode(text, out_type=str)
    results = translator.translate_batch([input_ids], beam_size=4)
    output_ids = results[0].hypotheses[0]
    return sp_target.decode(output_ids)

print(translate("¡Hola, mundo!"))

Load from Hugging Face Hub

python

from huggingface_hub import snapshot_download
import ctranslate2
import sentencepiece as spm

model_id = "mijuanlo/opus-mt-es-de-ct2-int8"
model_path = snapshot_download(repo_id=model_id)
translator = ctranslate2.Translator(model_path)
sp_source = spm.SentencePieceProcessor(f"{model_path}/source.spm")
sp_target = spm.SentencePieceProcessor(f"{model_path}/target.spm")

Conversion details

bash

ct2-transformers-converter \
    --model Helsinki-NLP/opus-mt-es-de \
    --output_dir opus-mt-es-de-ct2-int8 \
    --quantization int8 \
    --force

License

MIT. The original model from Helsinki-NLP is also MIT.

Citation

bibtex

@inproceedings{tiedemann-2020-tatoeba,
    title = {The Tatoeba Translation Challenge – Realistic Data Sets for Low Resource and Multilingual MT},
    author = {Tiedemann, J{\"o}rg},
    booktitle = {Proceedings of the Fifth Conference on Machine Translation (WMT)},
    year = {2020},
    publisher = {Association for Computational Linguistics},
}