app.py

import spaces
import gradio as gr
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from flores import code_mapping
import platform
import torch
import nltk

nltk.download("punkt")

REMOVED_TARGET_LANGUAGES = {"Ligurian", "Lombard", "Sicilian"}


device = "cpu" if platform.system() == "Darwin" else "cuda"
MODEL_NAME = "facebook/nllb-200-3.3B"

code_mapping = dict(sorted(code_mapping.items(), key=lambda item: item[1]))
flores_codes = list(code_mapping.keys())
target_languages = [language for language in flores_codes if not language in REMOVED_TARGET_LANGUAGES]

def load_model():
    model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME).to(device)
    return model


model = load_model()


def load_tokenizer(src_lang, tgt_lang):
    tokenizer = AutoTokenizer.from_pretrained(
        MODEL_NAME, src_lang=code_mapping[src_lang], tgt_lang=code_mapping[tgt_lang]
    )
    return tokenizer


@spaces.GPU
def translate(text: str, src_lang: str, tgt_lang: str):
    tokenizer = load_tokenizer(src_lang, tgt_lang)

    paragraphs = text.split("\n")
    translated_paragraphs = []

    for paragraph in paragraphs:
        sentences = nltk.sent_tokenize(paragraph)
        translated_sentences = []

        for sentence in sentences:
            input_tokens = (
                tokenizer(sentence, return_tensors="pt")
                .input_ids[0]
                .cpu()
                .numpy()
                .tolist()
            )
            translated_chunk = model.generate(
                input_ids=torch.tensor([input_tokens]).to(device),
                forced_bos_token_id=tokenizer.lang_code_to_id[code_mapping[tgt_lang]],
                max_length=len(input_tokens) + 50,
                num_return_sequences=1,
            )
            translated_chunk = tokenizer.decode(
                translated_chunk[0], skip_special_tokens=True
            )
            translated_sentences.append(translated_chunk)

        translated_paragraph = " ".join(translated_sentences)
        translated_paragraphs.append(translated_paragraph)

    return "\n".join(translated_paragraphs)


description = """
UNESCO, Meta, and Hugging Face have come together to create an accessible, high-quality translation experience in 200 languages. 

This is made possible through an open approach to AI innovation using Meta’s open-sourced No Language Left Behind (NLLB) AI model, hosted on Hugging Face Spaces. 
"""

with gr.Blocks() as demo:
    gr.Markdown("# UNESCO Language Translator, powered by Meta and Hugging Face")
    gr.Markdown(description)
    with gr.Row():
        src_lang = gr.Dropdown(label="Source Language", choices=flores_codes)
        target_lang = gr.Dropdown(label="Target Language", choices=target_languages)
    with gr.Row():
        input_text = gr.Textbox(label="Input Text", lines=6)
    with gr.Row():
        btn = gr.Button("Translate text")
    with gr.Row():
        output = gr.Textbox(label="Output Text", lines=6)
    btn.click(
        translate,
        inputs=[input_text, src_lang, target_lang],
        outputs=output,
    )
demo.launch()