app.py

# INDICTRANS2

# --- 1. CLEAN UP AND PREPARE THE ENVIRONMENT (CORRECTLY) ---
print("Cleaning up and preparing the environment...")
# This command removes the old directory if it exists, preventing the 'already exists' error.
print("✅ Environment ready.")

# --- 2. INSTALL ALL REQUIRED LIBRARIES FROM PyPI (USING A STABLE TRANSLITERATOR) ---
print("Installing all required libraries from PyPI...")
# Pinning transformers to a stable version to prevent caching errors.
# We are now using 'indic-transliteration' which is stable and maintained.
print("✅ All libraries installed successfully.")

# --- 3. SET UP THE SYSTEM PATH FOR THE TRANSLATION TOOLKIT (THE ONLY CORRECT METHOD) ---
import sys
# This tells Python where to find the IndicTransToolkit module without installation.
sys.path.insert(0, '/content/IndicTrans2/src')
print("✅ IndicTransToolkit added to system path.")

# --- 4. IMPORT ALL PACKAGES ---
import gradio as gr
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from IndicTransToolkit.processor import IndicProcessor
from indic_transliteration import sanscript
from indic_transliteration.sanscript import SchemeMap, SCHEMES, transliterate
import torch
print("✅ All packages imported.")

# --- 5. LOAD BOTH MODELS (TRANSLATION AND TRANSLITERATION) ---
print("Loading models and components...")
device = torch.device("cpu")
# A. Translation Model
translator_model_name = "ai4bharat/indictrans2-indic-en-dist-200M"
translator_tokenizer = AutoTokenizer.from_pretrained(translator_model_name, trust_remote_code=True)
translator_model = AutoModelForSeq2SeqLM.from_pretrained(translator_model_name, trust_remote_code=True).to(device)
ip = IndicProcessor(inference=True)
print("✅ Translation model and IndicProcessor are ready!")

# --- 6. DEFINE THE CORRECT, HIGH-ACCURACY TRANSLATION FUNCTIONS ---
LANG_CODES = {
    "Hindi": {"xlit": sanscript.DEVANAGARI, "indictrans": "hin_Deva"},
    "Tamil": {"xlit": sanscript.TAMIL, "indictrans": "tam_Taml"},
    "Bengali": {"xlit": sanscript.BENGALI, "indictrans": "ben_Beng"},
    "Telugu": {"xlit": sanscript.TELUGU, "indictrans": "tel_Telu"},
    "Kannada": {"xlit": sanscript.KANNADA, "indictrans": "kan_Knda"},
    "Malayalam": {"xlit": sanscript.MALAYALAM, "indictrans": "mal_Mlym"},
    "Gujarati": {"xlit": sanscript.GUJARATI, "indictrans": "guj_Gujr"},
    "Punjabi": {"xlit": sanscript.GURMUKHI, "indictrans": "pan_Guru"},
    "Urdu": {"xlit": sanscript.URDU, "indictrans": "urd_Arab"}
}
# Marathi uses Devanagari script for transliteration
LANG_CODES["Marathi"] = {"xlit": sanscript.DEVANAGARI, "indictrans": "mar_Deva"}


def translate_native_script(native_text, source_language_name):
    """Handles the direct native-to-English workflow."""
    try:
        if not native_text or not native_text.strip(): return "Please enter text."
        src_lang = LANG_CODES[source_language_name]["indictrans"]
        processed_text = ip.preprocess_batch([native_text], src_lang=src_lang, tgt_lang="eng_Latn")
        inputs = translator_tokenizer(processed_text, return_tensors="pt", padding=True).to(device)
        with torch.no_grad():
            translated_tokens = translator_model.generate(**inputs, num_beams=5, max_length=256)
        decoded_translation = translator_tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)
        return ip.postprocess_batch(decoded_translation, lang=src_lang)[0]
    except Exception as e:
        return f"An error occurred: {str(e)}"

def translate_roman_script(roman_text, source_language_name):
    """Performs the high-accuracy two-step transliterate-then-translate process."""
    try:
        if not roman_text or not roman_text.strip(): return "Please enter text."

        # Step 1: Transliterate Roman to Native Script using the stable 'indic-transliteration' library
        target_script = LANG_CODES[source_language_name]["xlit"]
        native_text = transliterate(roman_text, sanscript.ITRANS, target_script)

        # Step 2: Translate the resulting Native Script to English
        return translate_native_script(native_text, source_language_name)

    except Exception as e:
        return f"An error occurred: {str(e)}"

print("✅ High-accuracy translation functions are ready.")

# --- 7. CREATE AND LAUNCH THE SEPARATE UI WITH TABS ---
with gr.Blocks() as demo:
    gr.Markdown("## IndicTrans2: Universal Language Translator (Final Accurate Workflow)")
    gr.Markdown("Translate from both native and romanized Indian languages to English using specialized, high-accuracy workflows.")

    with gr.Tab("🇮🇳 Native Script to English"):
        with gr.Row():
            native_inputs = [
                gr.Textbox(lines=5, label="Native Indian Language Text", placeholder="यहाँ अपना पाठ दर्ज करें..."),
                gr.Dropdown(choices=list(LANG_CODES.keys()), label="Select Source Language", value="Hindi")
            ]
            native_output = gr.Textbox(label="English Translation")
        gr.Button("Translate Native Text").click(fn=translate_native_script, inputs=native_inputs, outputs=native_output)

    with gr.Tab("🔡 Romanized Script to English"):
        with gr.Row():
            roman_inputs = [
                gr.Textbox(lines=5, label="Romanized Indian Language Text", placeholder="Aap kaise hain?"),
                gr.Dropdown(choices=list(LANG_CODES.keys()), label="Select Source Language", value="Hindi")
            ]
            roman_output = gr.Textbox(label="English Translation")
        gr.Button("Translate Romanized Text").click(fn=translate_roman_script, inputs=roman_inputs, outputs=roman_output)

print("🚀 Launching the final, robust, and correct Gradio app...")
demo.launch(share=True)