app.py

import re
import torch
import gradio as gr
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from deep_translator import GoogleTranslator

# -------------------- Load Kannada Model & Tokenizer --------------------
MODEL_PATH = "Thilak118/indic-bert-toxicity-classifier_kannada"

model = AutoModelForSequenceClassification.from_pretrained(MODEL_PATH)
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()

# -------------------- Translator (English → Kannada) --------------------
translator = GoogleTranslator(source="en", target="kn")

# -------------------- Clean Kannada Text --------------------
def clean_text(text):
    text = re.sub(r"[^\u0C80-\u0CFF\s.,!?]", "", str(text))
    text = re.sub(r"\s+", " ", text).strip()
    return text

# -------------------- Transliterate / Translate to Kannada --------------------
def transliterate_to_kannada(text):
    if text and text.strip():
        try:
            return translator.translate(text)
        except Exception:
            return "Translation failed"
    return ""

# -------------------- Toxicity Prediction --------------------
def predict_toxicity(english_text):
    kannada_text = transliterate_to_kannada(english_text)

    if "failed" in kannada_text.lower():
        return f"Transliterated Kannada Text: {kannada_text}\nPrediction: Failed"

    cleaned_text = clean_text(kannada_text)

    inputs = tokenizer(
        cleaned_text,
        return_tensors="pt",
        padding=True,
        truncation=True,
        max_length=128
    ).to(device)

    with torch.no_grad():
        outputs = model(**inputs)

    prediction = torch.argmax(outputs.logits, dim=1).item()
    probs = torch.softmax(outputs.logits, dim=1)[0]
    confidence = probs[prediction].item() * 100

    label = "Toxic" if prediction == 0 else "Non-Toxic"

    return (
        f"Transliterated Kannada Text: {cleaned_text}\n"
        f"Prediction: {label}\n"
        f"Confidence: {confidence:.2f}%"
    )

# -------------------- Gradio UI --------------------
with gr.Blocks(title="Kannada Text Toxicity Classifier (IndicBERT)") as demo:
    gr.Markdown(
        """
        # Kannada Text Toxicity Classifier (IndicBERT + Transliteration)

        Enter Kannada text written in **English letters**  
        Example:  
        **`ninage ashtondu seen illa le`**

        The app converts it into **Kannada script** and predicts toxicity.
        """
    )

    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(
                label="Enter Kannada Text (English Transliteration)",
                placeholder="e.g., ninage ashtondu seen illa le",
                lines=2
            )
        with gr.Column():
            transliterated_text = gr.Textbox(
                label="Kannada Script (Preview)",
                interactive=False,
                lines=2
            )

    with gr.Row():
        preview_btn = gr.Button("Preview Transliteration")
        predict_btn = gr.Button("Predict Toxicity")

    output_text = gr.Textbox(
        label="Prediction Output",
        interactive=False,
        lines=5
    )

    preview_btn.click(
        fn=transliterate_to_kannada,
        inputs=input_text,
        outputs=transliterated_text
    )

    predict_btn.click(
        fn=predict_toxicity,
        inputs=input_text,
        outputs=output_text
    )

# -------------------- Launch App --------------------
demo.launch()