app.py

import gradio as gr
import re
import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline

# Load T5 paraphrase model
model_name = "Vamsi/T5_Paraphrase_Paws"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name, torch_dtype=torch.float16)  # Use fp16 for speed

# Move model to CPU
model.to("cpu")

# Initialize paraphrase pipeline
paraphrase_pipeline = pipeline(
    "text2text-generation",
    model=model,
    tokenizer=tokenizer,
    truncation=True
)

def split_sentences(text):
    """Split text into sentences using regex (faster than nltk)."""
    return re.split(r'(?<=[.!?])\s+', text.strip())

def paraphrase_text(text):
    """Paraphrases input text while maintaining sentence structure."""
    if not text.strip():
        return "⚠️ Please enter some text to paraphrase."

    sentences = split_sentences(text)

    # Apply T5 paraphrasing with controlled creativity
    paraphrased_results = paraphrase_pipeline(
        [f"paraphrase: {sentence} </s>" for sentence in sentences if sentence], 
        max_length=80,
        do_sample=True,  
        temperature=0.7,  # More controlled creativity
        top_p=0.85,  # Reduce randomness
        top_k=50,
        repetition_penalty=1.2,  # Avoid excessive repetition
        num_return_sequences=1,  # Only one paraphrase per sentence
        batch_size=8
    )

    # Extract and join paraphrased sentences
    paraphrased_sentences = [result['generated_text'] for result in paraphrased_results]
    return " ".join(paraphrased_sentences)

# Define Gradio Interface
demo = gr.Interface(
    fn=paraphrase_text,
    inputs=gr.Textbox(label="Enter text", placeholder="Type your text to paraphrase...", lines=10),
    outputs=gr.Textbox(label="Paraphrased Text", lines=10),
    title="🎨 Controlled T5 Paraphraser",
    description="Enter text and get a well-structured paraphrased version without randomness!",
    theme="huggingface"
)

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