app.py

import gradio as gr
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import os
from datetime import datetime
from functools import lru_cache
import torch
import requests

# Language codes
LANGUAGE_CODES = {
    "English": "eng_Latn", "Korean": "kor_Hang", "Japanese": "jpn_Jpan", "Chinese": "zho_Hans",
    "Spanish": "spa_Latn", "French": "fra_Latn", "German": "deu_Latn", "Russian": "rus_Cyrl",
    "Portuguese": "por_Latn", "Italian": "ita_Latn", "Burmese": "mya_Mymr", "Thai": "tha_Thai"
}

# Translation history
class TranslationHistory:
    def __init__(self):
        self.history = []
    def add(self, src, translated, src_lang, tgt_lang):
        self.history.insert(0, {
            "source": src, "translated": translated,
            "src_lang": src_lang, "tgt_lang": tgt_lang,
            "timestamp": datetime.now().isoformat()
        })
        if len(self.history) > 100:
            self.history.pop()
    def get(self): return self.history
    def clear(self): self.history = []

history = TranslationHistory()

# Translation model
model_name = "facebook/nllb-200-distilled-600M"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

@lru_cache(maxsize=512)
def cached_translate(text, src_lang, tgt_lang, max_length=128, temperature=0.7):
    if not text.strip(): return ""
    src_code = LANGUAGE_CODES.get(src_lang, src_lang)
    tgt_code = LANGUAGE_CODES.get(tgt_lang, tgt_lang)
    input_tokens = tokenizer(text, return_tensors="pt", padding=True)
    input_tokens = {k: v.to(device) for k, v in input_tokens.items()}
    forced_bos_token_id = tokenizer.convert_tokens_to_ids(tgt_code)
    output = model.generate(
        **input_tokens,
        forced_bos_token_id=forced_bos_token_id,
        max_length=max_length, temperature=temperature,
        num_beams=5, early_stopping=True
    )
    result = tokenizer.decode(output[0], skip_special_tokens=True)
    history.add(text, result, src_lang, tgt_lang)
    return result

def translate_file(file, src_lang, tgt_lang, max_length, temperature):
    try:
        lines = file.decode("utf-8").splitlines()
        translated = [cached_translate(line, src_lang, tgt_lang, max_length, temperature) for line in lines if line.strip()]
        return "\n".join(translated)
    except Exception as e:
        return f"File translation error: {e}"

# Summarizer API
API_URL = "https://api-inference.huggingface.co/models/facebook/bart-large-cnn"
HF_API_KEY = os.environ.get("HF_API_KEY")
headers = {"Authorization": f"Bearer {HF_API_KEY}"}

def summarize_text(text, max_length):
    if not text.strip(): return ""
    min_length = max(10, max_length // 4)
    response = requests.post(API_URL, headers=headers, json={
        "inputs": text,
        "parameters": {"min_length": min_length, "max_length": max_length}
    })
    result = response.json()
    return result[0]["summary_text"] if isinstance(result, list) else "Error: " + str(result)

# Paraphraser
paraphrase_tokenizer = AutoTokenizer.from_pretrained("tuner007/pegasus_paraphrase")
paraphrase_model = AutoModelForSeq2SeqLM.from_pretrained("tuner007/pegasus_paraphrase")
paraphrase_model.to(device)

def paraphrase_text(input_text, num_return_sequences, num_beams):
    batch = paraphrase_tokenizer([input_text], truncation=True, padding="longest", max_length=60, return_tensors="pt").to(device)
    translated = paraphrase_model.generate(**batch, max_length=60, num_beams=num_beams, num_return_sequences=num_return_sequences, temperature=1.5)
    return paraphrase_tokenizer.batch_decode(translated, skip_special_tokens=True)

# Grammar Corrector
grammar_model_name = "pszemraj/flan-t5-large-grammar-synthesis"
grammar_tokenizer = AutoTokenizer.from_pretrained(grammar_model_name)
grammar_model = AutoModelForSeq2SeqLM.from_pretrained(grammar_model_name)
grammar_model.to(device)

def correct_grammar(text):
    input_ids = grammar_tokenizer(f"grammar: {text}", return_tensors="pt", truncation=True).input_ids.to(device)
    output_ids = grammar_model.generate(input_ids, max_length=256, num_beams=5)
    return grammar_tokenizer.decode(output_ids[0], skip_special_tokens=True)

# UI Style
gradio_style = """
.gr-button { border-radius: 12px !important; padding: 10px 20px !important; font-weight: bold; }
textarea, input[type=text] { border: 2px solid #00ADB5 !important; border-radius: 10px; transition: 0.2s; }
textarea:focus, input[type=text]:focus { border-color: #FF5722 !important; box-shadow: 0 0 8px #FF5722 !important; }
"""

# Gradio UI
with gr.Blocks(css=gradio_style, theme=gr.themes.Soft()) as demo:
    gr.Markdown("## 🤖 AI Toolbox: Translate, Summarize, Paraphrase, Correct Grammar")

    with gr.Tab("🌐 Translator"):
        src_lang = gr.Dropdown(list(LANGUAGE_CODES.keys()), label="From", value="English")
        swap = gr.Button("⇄")
        tgt_lang = gr.Dropdown(list(LANGUAGE_CODES.keys()), label="To", value="Korean")
        input_text = gr.Textbox(lines=3, label="Input Text")
        output_text = gr.Textbox(lines=3, label="Translated Output", interactive=False)
        translate_btn = gr.Button("🚀 Translate")
        clear_btn = gr.Button("🧽 Clear")
        max_length = gr.Slider(10, 512, value=128, label="Max Length")
        temperature = gr.Slider(0.1, 2.0, value=0.7, label="Temperature")
        translate_btn.click(cached_translate, [input_text, src_lang, tgt_lang, max_length, temperature], output_text)
        clear_btn.click(lambda: ("", ""), None, [input_text, output_text])

    with gr.Tab("📁 File Translator"):
        file_input = gr.File(label="Upload .txt File")
        file_src = gr.Dropdown(list(LANGUAGE_CODES.keys()), label="From", value="English")
        file_tgt = gr.Dropdown(list(LANGUAGE_CODES.keys()), label="To", value="Korean")
        f_max_length = gr.Slider(10, 512, value=128, label="Max Length")
        f_temp = gr.Slider(0.1, 2.0, value=0.7, label="Temperature")
        file_btn = gr.Button("Translate File")
        file_result = gr.Textbox(lines=10, label="File Output", interactive=False)
        file_btn.click(lambda file, src, tgt, ml, temp: translate_file(file.read(), src, tgt, ml, temp),
                       [file_input, file_src, file_tgt, f_max_length, f_temp], file_result)

    with gr.Tab("📝 Summarizer"):
        summary_input = gr.Textbox(lines=5, label="Enter text to summarize")
        summary_length = gr.Slider(32, 512, value=128, step=8, label="Max Length")
        summary_output = gr.Textbox(lines=5, label="Summary", interactive=False)
        summarize_btn = gr.Button("Summarize")
        summarize_btn.click(summarize_text, [summary_input, summary_length], summary_output)

    with gr.Tab("🔁 Paraphraser"):
        para_input = gr.Textbox(lines=4, label="Enter text to paraphrase")
        num_outputs = gr.Slider(1, 5, value=3, step=1, label="Number of Paraphrases")
        beam_width = gr.Slider(1, 10, value=5, step=1, label="Beam Width")
        para_output = gr.Textbox(label="Paraphrased Sentences", lines=6)
        para_btn = gr.Button("Paraphrase")
        para_btn.click(lambda text, num, beams: "\n\n".join(paraphrase_text(text, num, beams)),
                       [para_input, num_outputs, beam_width], para_output)

    with gr.Tab("🛠 Grammar Corrector"):
        grammar_input = gr.Textbox(lines=5, label="Enter sentence to correct")
        grammar_output = gr.Textbox(label="Corrected Sentence", lines=5)
        grammar_btn = gr.Button("Correct Grammar")
        grammar_btn.click(correct_grammar, grammar_input, grammar_output)

    gr.Markdown(f"""
    ### ℹ️ Info
    - Translator: `{model_name}` on `{device}`
    - Paraphraser: `tuner007/pegasus_paraphrase`
    - Summarizer: `facebook/bart-large-cnn`
    - Grammar Corrector: `{grammar_model_name}`
    - API Token: {'✅ Found' if HF_API_KEY else '❌ Not Found'}
    """)

if __name__ == "__main__":
    demo.launch(share=True)