File size: 6,100 Bytes
cbc840b
b88f708
cbc840b
f837ee9
b88f708
cbc840b
b88f708
 
cbc840b
b88f708
f837ee9
 
 
b88f708
 
 
f837ee9
 
 
 
 
 
 
b88f708
 
f837ee9
 
 
cbc840b
b88f708
f837ee9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f67d206
c3b581c
f67d206
b88f708
c3b581c
cbc840b
 
b88f708
cbc840b
b88f708
 
 
cbc840b
 
 
 
 
 
 
 
b88f708
cbc840b
 
f837ee9
b88f708
cbc840b
b88f708
c3b581c
f837ee9
 
 
 
 
c3b581c
cbc840b
 
 
 
f837ee9
cbc840b
f837ee9
f67d206
f837ee9
f67d206
 
f837ee9
f67d206
f837ee9
b88f708
f837ee9
b88f708
cbc840b
 
 
 
 
 
f837ee9
f67d206
f837ee9
 
 
f67d206
 
 
f837ee9
f67d206
c3b581c
cbc840b
f837ee9
b88f708
 
cbc840b
f837ee9
b88f708
f837ee9
c3b581c
b88f708
cbc840b
b88f708
 
 
f837ee9
cbc840b
 
f837ee9
cbc840b
 
f837ee9
cbc840b
c3b581c
f837ee9
cbc840b
f837ee9
 
cbc840b
f837ee9
cbc840b
f837ee9
b88f708
cbc840b
f837ee9
b88f708
f837ee9
c3b581c
f837ee9
b88f708
 
3d2a9f9
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
import streamlit as st
import torch
from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
from transformers import AutoTokenizer, AutoModelForCausalLM, GPT2LMHeadModel
from diffusers import StableDiffusionPipeline
from rouge_score import rouge_scorer
from PIL import Image
import tempfile
import os
import time
import torch.nn.functional as F
import clip  # from OpenAI CLIP repo
from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize

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

# Load MBart model
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"

# Load GPT-2
gen_model = GPT2LMHeadModel.from_pretrained("gpt2").to(device)
gen_model.eval()
gen_tokenizer = AutoTokenizer.from_pretrained("gpt2")

# Try loading SD-2.1, fallback to lightweight
try:
    pipe = StableDiffusionPipeline.from_pretrained(
        "stabilityai/stable-diffusion-2-1",
        torch_dtype=torch.float32,
        use_auth_token=os.getenv("HF_TOKEN")
    ).to(device)
    pipe.safety_checker = None
    model_loaded = "stabilityai/stable-diffusion-2-1"
except Exception as e:
    st.warning("⚠️ SD-2.1 failed. Using lightweight fallback model.")
    pipe = StableDiffusionPipeline.from_pretrained(
        "OFA-Sys/small-stable-diffusion-v0",
        torch_dtype=torch.float32
    ).to(device)
    pipe.safety_checker = None
    model_loaded = "OFA-Sys/small-stable-diffusion-v0"

# Load CLIP for image-text similarity
clip_model, clip_preprocess = clip.load("ViT-B/32", device=device)

# Translation function
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

# Creative text generator with evaluation
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()
    rep_rate = sum(t1 == t2 for t1, t2 in zip(tokens, tokens[1:])) / len(tokens) if len(tokens) > 1 else 0

    # Calculate perplexity
    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(rep_rate, 4), round(perplexity, 4)

# Generate image and CLIP similarity
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)

        # Compute CLIP similarity
        image_input = clip_preprocess(Image.open(tmp_file.name)).unsqueeze(0).to(device)
        text_input = clip.tokenize([prompt]).to(device)
        with torch.no_grad():
            image_features = clip_model.encode_image(image_input)
            text_features = clip_model.encode_text(text_input)
            similarity = F.cosine_similarity(image_features, text_features).item()

        return tmp_file.name, duration, round(similarity, 4)
    except Exception as e:
        return None, None, f"Image generation failed: {str(e)}"

# Streamlit UI
st.set_page_config(page_title="Tamil β†’ English + AI Art", layout="centered")
st.title("🧠 Tamil β†’ English + 🎨 Creative Text + πŸ–ΌοΈ AI Image")

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

if st.button("πŸš€ Generate Output"):
    if not tamil_input.strip():
        st.warning("Please enter Tamil text.")
    else:
        with st.spinner("πŸ”„ Translating..."):
            english_text, t_time, rouge_l = translate_tamil_to_english(tamil_input, reference_input)

        st.success(f"βœ… Translated in {t_time}s")
        st.markdown(f"**πŸ“ English Translation:** `{english_text}`")
        if rouge_l is not None:
            st.markdown(f"πŸ“Š ROUGE-L Score: `{rouge_l}`")

        with st.spinner("πŸ–ΌοΈ Generating image..."):
            image_path, img_time, clip_score = generate_image(english_text)

        if image_path:
            st.success(f"πŸ–ΌοΈ Image generated in {img_time}s using `{model_loaded}`")
            st.image(Image.open(image_path), caption="AI-Generated Image", use_column_width=True)
            st.markdown(f"πŸ” **CLIP Text-Image Similarity:** `{clip_score}`")
        else:
            st.error(clip_score)

        with st.spinner("πŸ’‘ 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"πŸ“Œ Tokens: `{tokens}`, πŸ” Repetition Rate: `{rep_rate}`, πŸ“‰ Perplexity: `{ppl}`")

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