app.py

import os
import torch
import gradio as gr
from train import CharTokenizer, Seq2Seq, Encoder, Decoder, TransformerTransliterator

# ----------------------
# 1️⃣ Load LSTM checkpoint
# ----------------------
lstm_ckpt_path = "lstm_transliterator.pt"
lstm_ckpt = torch.load(lstm_ckpt_path, map_location='cpu')

src_vocab = lstm_ckpt['src_vocab']
tgt_vocab = lstm_ckpt['tgt_vocab']

src_tokenizer = CharTokenizer(vocab=src_vocab)
tgt_tokenizer = CharTokenizer(vocab=tgt_vocab)

# Reconstruct LSTM model architecture
EMBED_DIM = 256
ENC_HIDDEN_DIM = 256
DEC_HIDDEN_DIM = 256
NUM_LAYERS_MODEL = 2
DROPOUT = 0.3

device = 'cuda' if torch.cuda.is_available() else 'cpu'

encoder = Encoder(len(src_tokenizer), EMBED_DIM, ENC_HIDDEN_DIM, NUM_LAYERS_MODEL, DROPOUT)
decoder = Decoder(len(tgt_tokenizer), EMBED_DIM, ENC_HIDDEN_DIM, DEC_HIDDEN_DIM, NUM_LAYERS_MODEL, DROPOUT)
lstm_model = Seq2Seq(encoder, decoder, device=device).to(device)
lstm_model.load_state_dict(lstm_ckpt['model_state_dict'])
lstm_model.eval()

print("✅ LSTM model loaded")

# ----------------------
# 2️⃣ Load Transformer checkpoint
# ----------------------
transformer_ckpt_path = "transformer_transliterator.pt"
transformer_ckpt = torch.load(transformer_ckpt_path, map_location='cpu')

transformer_model = TransformerTransliterator(
    src_vocab_size=len(src_tokenizer),
    tgt_vocab_size=len(tgt_tokenizer),
    d_model=256,
    nhead=8,
    num_encoder_layers=2,
    num_decoder_layers=2,
    dim_feedforward=512,
    dropout=0.1,
    max_len=100
).to(device)
transformer_model.load_state_dict(transformer_ckpt['model_state_dict'])
transformer_model.eval()

print("✅ Transformer model loaded")

# ----------------------
# 3️⃣ Load TinyLLaMA
# ----------------------
from transformers import AutoTokenizer, AutoModelForCausalLM

try:
    llm_model_name = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
    llm_tokenizer = AutoTokenizer.from_pretrained(llm_model_name, trust_remote_code=True)
    llm_model = AutoModelForCausalLM.from_pretrained(
        llm_model_name,
        torch_dtype=torch.float16 if device == "cuda" else torch.float32,
        device_map="auto" if device == "cuda" else None,
        trust_remote_code=True,
    )
    if device != "cuda":
        llm_model = llm_model.to(device)
    llm_model.eval()
    print("✅ TinyLLaMA model loaded")
    has_llm = True
except Exception as e:
    print(f"⚠️ TinyLLaMA loading failed: {e}")
    print("⚠️ Will use only LSTM and Transformer models")
    has_llm = False

# ----------------------
# 4️⃣ Transliteration Function
# ----------------------
@torch.no_grad()
def transliterate(word):
    word = word.strip()
    
    if not word:
        return "❌ Empty input", "❌ Empty input", "❌ Empty input"
    
    try:
        # LSTM prediction
        lstm_pred = lstm_model.translate(word, src_tokenizer, tgt_tokenizer)
    except Exception as e:
        lstm_pred = f"Error: {str(e)[:50]}"
    
    try:
        # Transformer prediction (greedy)
        transformer_pred = transformer_model.translate(
            word, src_tokenizer, tgt_tokenizer, 
            device=device, decoding="greedy"
        )
    except Exception as e:
        transformer_pred = f"Error: {str(e)[:50]}"
    
    # TinyLLaMA prediction
    if has_llm:
        try:
            prompt = f"Transliterate the following English word to Hindi (Devanagari script).\nEnglish word: {word}\nHindi transliteration:"
            inputs = llm_tokenizer(prompt, return_tensors="pt").to(device)
            
            with torch.no_grad():
                output_ids = llm_model.generate(
                    **inputs,
                    max_new_tokens=30,
                    temperature=0.7,
                    top_p=0.9,
                    do_sample=True,
                    pad_token_id=llm_tokenizer.eos_token_id,
                    eos_token_id=llm_tokenizer.eos_token_id,
                )
            
            generated = llm_tokenizer.decode(output_ids[0], skip_special_tokens=True)
            llm_pred = generated.replace(prompt, "").strip()
            llm_pred = ''.join(c for c in llm_pred if not (c.isascii() and c.isalpha()) and c.strip())
        except Exception as e:
            llm_pred = f"Error: {str(e)[:50]}"
    else:
        llm_pred = "TinyLLaMA model not loaded (insufficient memory)"
    
    return lstm_pred, transformer_pred, llm_pred

# ----------------------
# 5️⃣ Gradio Interface
# ----------------------
demo = gr.Interface(
    fn=transliterate,
    inputs=gr.Textbox(
        label="Input Hindi Roman Word",
        placeholder="e.g., namaste, dhanyavaad, bharat",
        lines=1
    ),
    outputs=[
        gr.Textbox(label="LSTM Prediction", interactive=False),
        gr.Textbox(label="Transformer Prediction", interactive=False),
        gr.Textbox(label="TinyLLaMA Prediction", interactive=False)
    ],
    title="Hindi Roman to Devanagari Transliteration",
    description="Compare three models: LSTM, Transformer, and TinyLLaMA.\nEnter a Hindi Roman word to get transliteration predictions.",
    examples=[
        ["namaste"],
        ["dhanyavaad"],
        ["bharat"],
        ["mumbai"],
        ["hindustan"],
        ["pranaam"]
    ],
    allow_flagging="never"
)

demo.launch(
    share=False,
    debug=False,
    server_name="0.0.0.0",
    server_port=7860
)