app.py

import gradio as gr
import torch
from transformers import pipeline
import os
import spaces

MODELS = {
    "Moroccan": "BounharAbdelaziz/ModernBERT-Arabic-base-stage-1-pre-decay-predef-bkpt-ini-1024-mx8192", # "BounharAbdelaziz/ModernBERT-Arabic-base-stage-3-decay-mx8192-ARY",
    "Arabic": "BounharAbdelaziz/ModernBERT-Arabic-base-stage-1-pre-decay-predef-bkpt-ini-1024-mx8192", # "BounharAbdelaziz/ModernBERT-Arabic-base-stage-3-decay-mx8192-MSA", # "BounharAbdelaziz/ModernBERT-Arabic-base-stage-1-pre-decay-ini-8192-mx-8192
    "Egyptian": "BounharAbdelaziz/ModernBERT-Arabic-base-stage-1-pre-decay-predef-bkpt-ini-1024-mx8192", # "BounharAbdelaziz/ModernBERT-Arabic-base-stage-3-decay-mx8192-ARZ",
    "Tunisian": "BounharAbdelaziz/ModernBERT-Arabic-base-stage-1-pre-decay-predef-bkpt-ini-1024-mx8192", # "BounharAbdelaziz/ModernBERT-Arabic-base-stage-3-decay-mx8192-ARY",
    "Algerian": "BounharAbdelaziz/ModernBERT-Arabic-base-stage-1-pre-decay-predef-bkpt-ini-1024-mx8192", # "BounharAbdelaziz/ModernBERT-Arabic-base-stage-3-decay-mx8192-ALGERIAN",
}

EXAMPLES = {
    "Moroccan": [
        "الدار البيضاء [MASK]", 
        "المغرب بلاد [MASK]",
        "كناكل [MASK] فالمغرب",
        "العاصمة د [MASK] هي الرباط",
        "المغرب [MASK] زوين",
        "انا سميتي مريم، و كنسكن ف[MASK] العاصمة دفلسطين"
    ],
    "Arabic": [
        "العاصمة الرسمية لمصر هي [MASK].",
        "أطول نهر في العالم هو نهر [MASK].",
        "الشاعر العربي المشهور [MASK] كتب قصيدة 'أراك عصي الدمع'.",
        "عندما أستيقظ في الصباح، أشرب فنجان من [MASK].",
        "في التأني [MASK] وفي العجلة الندامة.",
        "معركة [MASK] كانت من أهم المعارك في تاريخ الإسلام.",
        "يعتبر [MASK] من أهم العلماء في مجال الفيزياء.",
        "تقع جبال [MASK] في شمال إفريقيا.",
        "يعتبر [MASK] من أركان الإسلام الخمسة."
    ],
    "Egyptian": [
        "القاهرة مدينة [MASK]", 
        "مصر بلاد [MASK]", 
        "بنحب [MASK] فمصر"
    ],
    "Tunisian": [
        "تونس بلاد [MASK]", 
        "المنستير مدينة [MASK]", 
        "عيشتي في [MASK]"
    ],
    "Algerian": [
        "الجزائر بلاد [MASK]", 
        "قسنطينة مدينة [MASK]", 
        "نحبو [MASK] ف الجزائر"
    ],
}

TOKEN = os.environ["HF_TOKEN"]
device = "cuda:0" if torch.cuda.is_available() else "cpu"

def load_model(dialect):
    model_path = MODELS.get(dialect, MODELS["Arabic"])
    return pipeline(task="fill-mask", model=model_path, token=TOKEN, device=device)

pipe = None

@spaces.GPU 
def predict(text, dialect):
    global pipe
    if pipe is None or dialect != predict.current_dialect:  # Reload model if dialect changes
        pipe = load_model(dialect)
        predict.current_dialect = dialect
    outputs = pipe(text)
    scores = [x["score"] for x in outputs]
    tokens = [x["token_str"] for x in outputs]
    return {label: float(prob) for label, prob in zip(tokens, scores)}

# Initialize current dialect
predict.current_dialect = None

# Create Gradio interface
with gr.Blocks() as demo:
    with gr.Row():
        with gr.Column():
            # Dropdown for dialect selection
            dialect_dropdown = gr.Dropdown(
                choices=["Arabic", "Tunisian", "Moroccan", "Algerian", "Egyptian"],
                label="Select Dialect",
                value="Arabic"
            )
            
            # Input text box
            input_text = gr.Textbox(
                label="Input",
                placeholder="Enter text here...",
                rtl=True
            )
            
            # Button row
            with gr.Row():
                clear_btn = gr.Button("Clear")
                submit_btn = gr.Button("Submit", variant="primary")
            
            # Examples section (stateful)
            examples_state = gr.State(EXAMPLES["Arabic"])  # Store examples for the current dialect
            
            example_dropdown = gr.Dropdown(
                choices=EXAMPLES["Arabic"],
                label="Select Example",
                interactive=True
            )
            
            load_example_btn = gr.Button("Load Example")

        with gr.Column():
            # Output probabilities
            output_labels = gr.Label(
                label="Prediction Results",
                show_label=False
            )

    # Function to update examples when dialect changes
    def update_example_choices(dialect):
        return gr.update(choices=EXAMPLES.get(dialect, EXAMPLES["Arabic"]), value=EXAMPLES[dialect][0])

    # Function to load selected example into the text box
    def load_example(selected_example):
        return selected_example

    # Update example dropdown choices dynamically
    dialect_dropdown.change(
        update_example_choices,
        inputs=dialect_dropdown,
        outputs=example_dropdown
    )

    # Load selected example into the input text box
    load_example_btn.click(
        load_example,
        inputs=example_dropdown,
        outputs=input_text
    )

    # Button actions
    submit_btn.click(
        predict,
        inputs=[input_text, dialect_dropdown],
        outputs=output_labels
    )
    
    clear_btn.click(
        lambda: "",
        outputs=input_text
    )

# Launch the app
demo.launch()