app.py

import torch
import gradio as gr
from transformers import logging

# Suppress transformers warnings
logging.set_verbosity_error()

# Import the model and tokenizer classes
from transliterator import EnglishHindiTransliterator, CharacterTokenizer

# Load tokenizers
eng_tokenizer = CharacterTokenizer(is_hindi=False)
hindi_tokenizer = CharacterTokenizer(is_hindi=True)

# Load the trained model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = EnglishHindiTransliterator(
    input_dim=len(eng_tokenizer),
    output_dim=len(hindi_tokenizer),
    emb_dim=256,
    hid_dim=512,
    n_layers=3,
    dropout=0.3
).to(device)

checkpoint = torch.load("best_transliteration_model.pt", map_location=device)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

# Define transliteration function
def transliterate_text(text):
    with torch.no_grad():
        src_tensor = torch.tensor(eng_tokenizer.encode(text)).unsqueeze(0).to(device)
        tgt_tensor = torch.tensor([hindi_tokenizer.char2idx[hindi_tokenizer.sos_token]]).unsqueeze(0).to(device)

        for _ in range(50):  # Max sequence length
            output = model(src_tensor, tgt_tensor)
            pred_token = output.argmax(2)[:, -1]
            tgt_tensor = torch.cat([tgt_tensor, pred_token.unsqueeze(1)], dim=1)

            if pred_token.item() == hindi_tokenizer.char2idx[hindi_tokenizer.eos_token]:
                break

    return hindi_tokenizer.decode(tgt_tensor.squeeze().cpu().numpy())

# Create Gradio Interface
interface = gr.Interface(
    fn=transliterate_text,
    inputs=gr.Textbox(label="Enter English Word"),
    outputs=gr.Textbox(label="Hindi Transliteration"),
    title="English to Hindi Transliteration",
    description="Enter an English word and get its Hindi transliteration using a deep learning model."
)

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