app.py

import gradio as gr
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import fasttext

# Initialize fastText model
model_path = 'lid.323.ftz'
language_model = fasttext.load_model(model_path)

model_path_translation = "anzorq/m2m100_418M_ft_ru-kbd_44K"
tokenizer = AutoTokenizer.from_pretrained(model_path_translation)
model = AutoModelForSeq2SeqLM.from_pretrained(model_path_translation, use_safetensors=True)

def translate(text, num_beams=4, num_return_sequences=4):
    # Detect language
    languages, _ = language_model.predict(text, k=1)
    detected_language = languages[0].replace("__label__", "")
    
    inputs = tokenizer(text, return_tensors="pt")
    num_return_sequences = min(num_return_sequences, num_beams)
    translated_tokens = model.generate(
        **inputs, forced_bos_token_id=tokenizer.lang_code_to_id[tgt_lang], num_beams=num_beams, num_return_sequences=num_return_sequences
    )
    translations = [tokenizer.decode(translation, skip_special_tokens=True) for translation in translated_tokens]
    
    return detected_language, text, translations

title = "Russian-Circassian translator demo"
article = "<p style='text-align: center'>Want to help? Join the <a href='https://discord.gg/cXwv495r' target='_blank'>Discord server</a></p>"

num_beams = gr.inputs.Slider(2, 10, step=1, label="Number of beams", default=4)
num_return_sequences = gr.inputs.Slider(2, 10, step=1, label="Number of returned sentences", default=4)

gr.Interface(
    fn=translate,
    inputs=["text", num_beams, num_return_sequences],
    outputs=["text", "text", gr.Textbox()],
    titles=["Detected Language", "Input", "Translations"],
    title=title,
    article=article).launch()

# import gradio as gr

# title = "Русско-черкесский переводчик"
# description = "Demo of a Russian-Circassian (Kabardian dialect) translator. <br>It is based on Facebook's <a href=\"https://about.fb.com/news/2020/10/first-multilingual-machine-translation-model/\">M2M-100 model</a> machine learning model, and has been trained on 45,000 Russian-Circassian sentence pairs. <br>It can also translate from 100 other languages to Circassian (English, French, Spanish, etc.), but less accurately. <br>The data corpus is constantly being expanded, and we need help in finding sentence sources, OCR, data cleaning, etc. <br>If you are interested in helping out with this project, please contact me at the link below.<br><br>This is only a demo, not a finished product. Translation quality is still low and will improve with time and more data.<br>45,000 sentence pairs is not enough to create an accurate machine translation model, and more data is needed.<br>You can help by finding sentence sources (books, web pages, etc.), scanning books, OCRing documents, data cleaning, and other tasks.<br><br>If you are interested in helping out with this project, contact me at the link below."
# article = """<p style='text-align: center'><a href='https://arxiv.org/abs/1806.00187'>Scaling Neural Machine Translation</a> | <a href='https://github.com/pytorch/fairseq/'>Github Repo</a></p>"""

# examples = [
#     ["Мы идем домой"],
#     ["Сегодня хорошая погода"],
#     ["Дети играют во дворе"],
#     ["We live in a big house"],
#     ["Tu es une bonne personne."],
#     ["أين تعيش؟"],
#     ["Bir şeyler yapmak istiyorum."],
# ]

# gr.Interface.load("models/anzorq/m2m100_418M_ft_ru-kbd_44K", title=title, description=description, article=article, examples=examples).launch()