You can deploy this model and start earning money today!

tutur90/
SONAR-Text-to-Text

Name: SONAR-Text-to-Text
Availability: InStock

safetensors

m2m_100

2308.11466

Views

No views yet

Model Card

Original weights of SONAR converted for hugging face model.

This is a part of Open SONAR project, an open training pipeline for SONAR. This pipeline could be use to finetune or train from scratch a sonar model.

Model Sources

Repository: https://github.com/facebookresearch/SONAR
Paper: https://arxiv.org/abs/2308.11466

Uses

Examples are avalaible here.

Load the model and tokenizer

Python

SONARForSpeech2Text.from_pretrained("tutur90/SONAR-Text-to-Text")

NllbTokenizer.from_pretrained("tutur90/SONAR-Text-to-Text")

The code of SONARForSpeech2Text avalaible in Open SONAR - Model and NllbTokenizer Open SONAR - Tokenizer

Translation

Python

inputs = tokenizer(sentence, langs=src_lang, return_tensors="pt")

generated = sonar.generate(
        **inputs,
        target_lang_ids=[tokenizer.convert_tokens_to_ids(tgt_lang)],
        max_length=128,
        num_beams=1,
        do_sample=False,
    )
decoded = tokenizer.batch_decode(generated, skip_special_tokens=True)[0]

print(f"SONAR {src_lang} -> {tgt_lang}: {decoded}")

Encode

Python

encoder = SONARForText2Text.from_pretrained("tutur90/SONAR-Text-to-Text")

encoder.set_encoder_only() # Delete decoder to save memory, this options is not needed

inputs = tokenizer(sentence, langs=src_lang, return_tensors="pt")

embeddings = encoder.encode(**inputs)

Decode

Python

decoder = SONARForText2Text.from_pretrained("tutur90/SONAR-Text-to-Text")

decoder.set_decoder_only() # Same

decoded = decoder.decode(
        encoder_outputs,
        target_lang_ids=[tokenizer.convert_tokens_to_ids(tgt_lang)],
        max_length=128,
        num_beams=1,
        do_sample=False,
    )
decoded = tokenizer.batch_decode(decoded, skip_special_tokens=True)[0]