app.py

import os
import sys
import urllib.request
import torch
import gradio as gr

# ---------- Helper functions ----------

def download_file(url, dest_path):
    """Download a file from a URL to a destination path."""
    if not os.path.exists(dest_path):
        print(f"Downloading {url} to {dest_path} ...")
        urllib.request.urlretrieve(url, dest_path)
    else:
        print(f"File {dest_path} already exists.")

def clone_repo(repo_url, folder_name):
    """Clone the repo if the folder does not exist."""
    if not os.path.exists(folder_name):
        print(f"Cloning repository {repo_url} ...")
        os.system(f"git clone {repo_url}")
    else:
        print(f"Repository {folder_name} already exists.")

# ---------- Setup: Clone repository and download models ----------

# Clone the catt repository (which contains the necessary modules)
clone_repo("https://github.com/abjadai/catt.git", "catt")

# Add the cloned repository to the Python path so we can import modules from it.
if "catt" not in sys.path:
    sys.path.append("catt")

# Create models folder if not exists
if not os.path.exists("models"):
    os.makedirs("models")

# URLs for model checkpoints
url_ed = "https://github.com/abjadai/catt/releases/download/v2/best_ed_mlm_ns_epoch_178.pt"
url_eo = "https://github.com/abjadai/catt/releases/download/v2/best_eo_mlm_ns_epoch_193.pt"

ckpt_path_ed = os.path.join("models", "best_ed_mlm_ns_epoch_178.pt")
ckpt_path_eo = os.path.join("models", "best_eo_mlm_ns_epoch_193.pt")

download_file(url_ed, ckpt_path_ed)
download_file(url_eo, ckpt_path_eo)

# ---------- Import required modules from the cloned repository ----------

from tashkeel_tokenizer import TashkeelTokenizer
from utils import remove_non_arabic
from ed_pl import TashkeelModel as TashkeelModel_ED
from eo_pl import TashkeelModel as TashkeelModel_EO

# ---------- Global model initialization ----------

# Prepare tokenizer (used by both models)
tokenizer = TashkeelTokenizer()

# Determine the device
device = "cuda" if torch.cuda.is_available() else "cpu"
print("Using device:", device)

# Global variables to hold the models
model_ed = None
model_eo = None

def load_models():
    global model_ed, model_eo
    max_seq_len = 1024

    # Load Encoder-Decoder model
    print("Loading Encoder-Decoder model...")
    model_ed = TashkeelModel_ED(tokenizer, max_seq_len=max_seq_len, n_layers=3, learnable_pos_emb=False)
    model_ed.load_state_dict(torch.load(ckpt_path_ed, map_location=device))
    model_ed.eval().to(device)
    print("Encoder-Decoder model loaded.")

    # Load Encoder-Only model
    print("Loading Encoder-Only model...")
    model_eo = TashkeelModel_EO(tokenizer, max_seq_len=max_seq_len, n_layers=6, learnable_pos_emb=False)
    model_eo.load_state_dict(torch.load(ckpt_path_eo, map_location=device))
    model_eo.eval().to(device)
    print("Encoder-Only model loaded.")

# Load the models at startup.
load_models()

# ---------- Inference Function ----------

def diacritize_text(model_type, input_text):
    """
    Process the input Arabic text (removing non-Arabic characters),
    and run the appropriate model for diacritization.
    """
    if not input_text.strip():
        return "Please enter some Arabic text."

    # Clean the input text
    text_clean = remove_non_arabic(input_text)
    x = [text_clean]

    batch_size = 16  # Fixed batch size.
    verbose = False

    try:
        if model_type == "Encoder-Decoder":
            output_lines = model_ed.do_tashkeel_batch(x, batch_size, verbose)
        elif model_type == "Encoder-Only":
            output_lines = model_eo.do_tashkeel_batch(x, batch_size, verbose)
        else:
            return "Unknown model type selected."
    except Exception as e:
        return f"An error occurred during diacritization: {str(e)}"

    return output_lines[0] if output_lines else "No output produced."

# ---------- Gradio Interface ----------

title = "Arabic Diacritization with CATT"
description = (
    "Enter Arabic text (without diacritics) below and select a model to perform "
    "automatic diacritization. The Encoder-Decoder model may offer better accuracy, "
    "while the Encoder-Only model is optimized for faster inference."
)

iface = gr.Interface(
    fn=diacritize_text,
    inputs=[
        gr.Dropdown(
            choices=["Encoder-Decoder", "Encoder-Only"],
            value="Encoder-Only",
            label="Model Selection"
        ),
        gr.Textbox(lines=4, placeholder="Enter Arabic text here...", label="Input Text")
    ],
    outputs=gr.Textbox(label="Diacritized Output"),
    title=title,
    description=description,
    allow_flagging="never"
)

# ---------- Launch the Gradio App ----------

if __name__ == "__main__":
    iface.launch()