app.py

import torch
from transformers import PegasusForConditionalGeneration, PegasusTokenizer
import gradio as gr

# Load the tokenizer and model once when the app starts
model_name = 'tuner007/pegasus_paraphrase'
torch_device = 'cuda' if torch.cuda.is_available() else 'cpu'

# Initialize tokenizer and model
tokenizer = PegasusTokenizer.from_pretrained(model_name)
model = PegasusForConditionalGeneration.from_pretrained(model_name).to(torch_device)

def get_response(input_text, num_return_sequences=1, num_beams=3):
    """
    Generate paraphrased text for a given input_text using the Pegasus model.
    
    Args:
        input_text (str): The text to paraphrase.
        num_return_sequences (int): Number of paraphrased sequences to return.
        num_beams (int): Number of beams for beam search.
        
    Returns:
        list: A list containing paraphrased text strings.
    """
    # Tokenize the input text
    batch = tokenizer(
        [input_text],
        truncation=True,
        padding='longest',
        max_length=60,
        return_tensors="pt"
    ).to(torch_device)
    
    # Generate paraphrased outputs
    translated = model.generate(
        **batch,
        max_length=60,
        num_beams=num_beams,
        num_return_sequences=num_return_sequences,
        temperature=0.7
    )
    
    # Decode the generated tokens
    tgt_text = tokenizer.batch_decode(translated, skip_special_tokens=True)
    return tgt_text

def split_text_by_fullstop(text):
    """
    Split the input text into sentences based on full stops.
    
    Args:
        text (str): The text to split.
        
    Returns:
        list: A list of sentences.
    """
    sentences = [sentence.strip() for sentence in text.split('.') if sentence]
    return sentences

def process_text_by_fullstop(text, num_return_sequences=1, num_beams=3):
    """
    Process the input text by splitting it into sentences and paraphrasing each sentence.
    
    Args:
        text (str): The text to paraphrase.
        num_return_sequences (int): Number of paraphrased sequences per sentence.
        num_beams (int): Number of beams for beam search.
        
    Returns:
        str: The paraphrased text.
    """
    sentences = split_text_by_fullstop(text)
    paraphrased_sentences = []
    
    for sentence in sentences:
        # Ensure each sentence ends with a period
        sentence = sentence + '.' if not sentence.endswith('.') else sentence
        paraphrases = get_response(sentence, num_return_sequences, num_beams)
        paraphrased_sentences.extend(paraphrases)
    
    # Join all paraphrased sentences into a single string
    return ' '.join(paraphrased_sentences)

def paraphrase(text, num_beams, num_return_sequences):
    """
    Interface function to paraphrase input text based on user parameters.
    
    Args:
        text (str): The input text to paraphrase.
        num_beams (int): Number of beams for beam search.
        num_return_sequences (int): Number of paraphrased sequences to return.
        
    Returns:
        str: The paraphrased text.
    """
    return process_text_by_fullstop(text, num_return_sequences, num_beams)

# Define the Gradio interface
iface = gr.Interface(
    fn=paraphrase,
    inputs=[
        gr.components.Textbox(
            lines=10, 
            placeholder="Enter text here...", 
            label="Input Text"
        ),
        gr.components.Slider(
            minimum=1, 
            maximum=10, 
            step=1, 
            value=3, 
            label="Number of Beams"
        ),
        gr.components.Slider(
            minimum=1, 
            maximum=5, 
            step=1, 
            value=1, 
            label="Number of Return Sequences"
        )
    ],
    outputs=gr.components.Textbox(
        lines=10, 
        label="Paraphrased Text"
    ),
    title="Text Paraphrasing App",
    description="Enter your text and adjust the parameters to receive paraphrased versions using the Pegasus model.",
    allow_flagging="never"
)

# Launch the app
if __name__ == "__main__":
    iface.launch()