app.py

import streamlit as st
import torch
import torch.nn.functional as F
import os
import time
import tempfile
from PIL import Image
from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
from transformers import AutoTokenizer, AutoModelForCausalLM, CLIPProcessor, CLIPModel
from diffusers import StableDiffusionPipeline
from rouge_score import rouge_scorer

# --- Device Setup ---
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# --- Load Models ---
translator_model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50-many-to-many-mmt").to(device)
translator_tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
translator_tokenizer.src_lang = "ta_IN"

gen_tokenizer = AutoTokenizer.from_pretrained("gpt2")
gen_model = AutoModelForCausalLM.from_pretrained("gpt2").to(device)
gen_model.eval()

pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-1-5").to(device)
pipe.safety_checker = None

clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").to(device)
clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")

# --- Functions ---
def translate_tamil_to_english(text, reference=None):
    start = time.time()
    inputs = translator_tokenizer(text, return_tensors="pt").to(device)
    outputs = translator_model.generate(**inputs, forced_bos_token_id=translator_tokenizer.lang_code_to_id["en_XX"])
    translated = translator_tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
    duration = round(time.time() - start, 2)

    rouge_l = None
    if reference:
        scorer = rouge_scorer.RougeScorer(['rougeL'], use_stemmer=True)
        score = scorer.score(reference.lower(), translated.lower())
        rouge_l = round(score["rougeL"].fmeasure, 4)

    return translated, duration, rouge_l

def generate_creative_text(prompt, max_length=100):
    start = time.time()
    input_ids = gen_tokenizer.encode(prompt, return_tensors="pt").to(device)
    output = gen_model.generate(input_ids, max_length=max_length, do_sample=True, top_k=50, temperature=0.9)
    text = gen_tokenizer.decode(output[0], skip_special_tokens=True)
    duration = round(time.time() - start, 2)

    tokens = text.split()
    repetition_rate = sum(t1 == t2 for t1, t2 in zip(tokens, tokens[1:])) / len(tokens) if len(tokens) > 1 else 0

    with torch.no_grad():
        input_ids = gen_tokenizer.encode(text, return_tensors="pt").to(device)
        outputs = gen_model(input_ids, labels=input_ids)
        loss = outputs.loss
        perplexity = torch.exp(loss).item()

    return text, duration, len(tokens), round(repetition_rate, 4), round(perplexity, 4)

def generate_image(prompt):
    try:
        start = time.time()
        result = pipe(prompt)
        image = result.images[0].resize((256, 256))
        tmp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".png")
        image.save(tmp_file.name)
        duration = round(time.time() - start, 2)
        return tmp_file.name, duration, image
    except Exception as e:
        return None, 0, f"Image generation failed: {str(e)}"

def evaluate_clip_similarity(text, image):
    inputs = clip_processor(text=[text], images=image, return_tensors="pt", padding=True).to(device)
    with torch.no_grad():
        outputs = clip_model(**inputs)
        logits_per_image = outputs.logits_per_image
        probs = F.softmax(logits_per_image, dim=1)
        similarity_score = logits_per_image[0][0].item()
    return round(similarity_score, 4)

# --- Streamlit UI ---
st.set_page_config(page_title="Tamil → English + AI Art", layout="centered")
st.title("🧠 Tamil → English + 🎨 Creative Text + AI Image")

tamil_input = st.text_area("✍️ Enter Tamil text", height=150)
reference_input = st.text_input("📘 Optional: Reference English translation for ROUGE")

if st.button("🚀 Generate Output"):
    if not tamil_input.strip():
        st.warning("Please enter Tamil text.")
    else:
        with st.spinner("🔄 Translating Tamil to English..."):
            english_text, t_time, rouge_l = translate_tamil_to_english(tamil_input, reference_input)

        st.success(f"✅ Translated in {t_time}s")
        st.markdown(f"**📝 English Translation:** `{english_text}`")
        if rouge_l is not None:
            st.markdown(f"📊 ROUGE-L Score: `{rouge_l}`")

        with st.spinner("🖼️ Generating image from text..."):
            image_path, img_time, image_obj = generate_image(english_text)

        if isinstance(image_obj, Image.Image):
            st.success(f"🖼️ Image generated in {img_time}s")
            st.image(Image.open(image_path), caption="AI-Generated Image", use_column_width=True)

            with st.spinner("🔎 Evaluating CLIP similarity..."):
                clip_score = evaluate_clip_similarity(english_text, image_obj)
                st.markdown(f"🔍 **CLIP Text-Image Similarity:** `{clip_score}`")
        else:
            st.error(image_obj)

        with st.spinner("💡 Generating creative text..."):
            creative, c_time, tokens, rep_rate, ppl = generate_creative_text(english_text)

        st.success(f"✨ Creative text in {c_time}s")
        st.markdown(f"**🧠 Creative Output:** `{creative}`")
        st.markdown(f"📌 Tokens: `{tokens}`, 🔁 Repetition Rate: `{rep_rate}`, 📉 Perplexity: `{ppl}`")

st.markdown("---")
st.caption("Built by Sureshkumar R | MBart + GPT-2 + Stable Diffusion + CLIP")