app.py

import numpy as np 
import pandas as pd 
import re
import torch

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("humarin/chatgpt_paraphraser_on_T5_base")
model = AutoModelForSeq2SeqLM.from_pretrained("humarin/chatgpt_paraphraser_on_T5_base")

tokenizer_gen_title = AutoTokenizer.from_pretrained("Ateeqq/news-title-generator")
model_gen_title = AutoModelForSeq2SeqLM.from_pretrained("Ateeqq/news-title-generator")

def generate_title(input_text): #Generate a title for input text with Ateeq model

  input_ids = tokenizer_gen_title.encode(input_text, return_tensors="pt") #Tokenize input text
  input_ids = input_ids.to('cuda') #Send tokenized inputs to gpu
  output = model_gen_title.generate(input_ids, 
                          max_new_tokens=100, 
                          do_sample=True, 
                          temperature=0.8,
                          top_k = 20
                        )
  decoded_text = tokenizer_gen_title.decode(output[0], skip_special_tokens=True)
  return decoded_text



def split_into_sentences(paragraph): #For paraphraser - return a list of sentences from input para
    # Split sentences after period. Retains \n if part of the text, but not included in model output
    
    sentence_endings = r'(?<!\w\.\w.)(?<![A-Z][a-z]\.)(?<=\.|\?|\!)\s'
    sentences = re.split(sentence_endings, paragraph)    
    return sentences

def paraphrase(
    question,
    #num_beams=10,
    #num_beam_groups=10,
    #num_return_sequences=1,
    #repetition_penalty=1.0,
    #diversity_penalty=1.0,
    no_repeat_ngram_size=3,
    temperature=0.8,
    max_length=128
):
    
    sentence_list = split_into_sentences(question) #feed input para into sentence splitter
    output = [] #List to hold the individual rephrased sentences obtained from the model

    for sentence in sentence_list:
    
        input_ids = tokenizer(
            f'paraphrase: {sentence}', #Using paraphrase prompt for T5
            return_tensors="pt", padding="longest",
            #max_length=max_length,
            #truncation=True,
        ).input_ids
        
        outputs = model.generate(
            input_ids,
            do_sample=True,
            temperature=temperature,
            max_length=max_length, 
            no_repeat_ngram_size=no_repeat_ngram_size
        )
    
        res = tokenizer.batch_decode(outputs, skip_special_tokens=True)
        output.append(res[0]) #Add rephrased sentence to list

    paraphrased_text = "" #to hold the combined sentence output made from generated list

    for sentence in output: #Join all new reworded sentences together
        paraphrased_text += sentence + " "   

    for title in range (1,4): #Print 3 titles by calling Ateeq model fn - generate_title
        print (f"Title {title}: {generate_title (paraphrased_text)}")
        print ("")
    
    return paraphrased_text # Return paraphrased text after printing three titles above