app.py

# -*- coding: utf-8 -*-
"""Ai Re-Phraser.py
 
Automatically generated by Colaboratory.

Original file is located at
    https://colab.research.google.com/drive/18bvmXQqMIkk7G0gY_1dUolI08RK6Ajrf

"""  

# importing the libraries
import os
import pandas as pd
from parrot import Parrot  #Parrot offers knobs to control Adequacy, Fluency and Diversity metrics
import torch
import warnings
import nltk
import warnings
warnings.filterwarnings("ignore")


from sentence_splitter import SentenceSplitter, split_text_into_sentences
splitter = SentenceSplitter(language='en')
from parrot import Parrot
from transformers import PegasusForConditionalGeneration, PegasusTokenizer
from transformers import AutoTokenizer
from transformers import AutoModelForSeq2SeqLM
from parrot.filters import Adequacy
from parrot.filters import Fluency
from parrot.filters import Diversity

# Adding the metrics

adequacy_score = Adequacy()
fluency_score  = Fluency()
diversity_score= Diversity()
device= "cuda:0"
adequacy_threshold = 0.99
fluency_threshold = 0.90 # Fluency (Is the paraphrase fluent English?)
diversity_ranker="levenshtein"
do_diverse=True # Diversity (Lexical / Phrasal / Syntactical) (How much has the paraphrase changed the original sentence?)
#num_beam_groups (int) — Number of groups to divide num_beams into in order to ensure diversity among different groups of beams
# adding the model

model_name = 'tuner007/pegasus_paraphrase'
torch_device = 'cuda' if torch.cuda.is_available() else 'cpu'
tokenizer = PegasusTokenizer.from_pretrained(model_name)
model_pegasus = PegasusForConditionalGeneration.from_pretrained(model_name).to(torch_device)

def get_max_str(lst):
    return max(lst, key=len)
def get_response(input_text):
    batch = tokenizer.prepare_seq2seq_batch([input_text],truncation=True,padding='longest',max_length=60, return_tensors='pt').to(torch_device)
    translated = model_pegasus.generate(**batch,max_length=60,num_beams=15, num_return_sequences=15, temperature=1.5)
    #num_beam_groups=num_beams, diversity_penalty=0.5
    tgt_text = tokenizer.batch_decode(translated, skip_special_tokens=True)
    try:
        adequacy_filtered = adequacy_score.filter(input_text,tgt_text, adequacy_threshold, device)
        if len(adequacy_filtered) > 0 :
            fluency_filtered = fluency_score.filter(adequacy_filtered, fluency_threshold, device )
            if len(fluency_filtered) > 0 :
                diversity_scored = diversity_score.rank(input_text, fluency_filtered, diversity_ranker)
                return get_max_str(diversity_scored)
            else:
                return get_max_str(fluency_filtered)
        else:
            return get_max_str(adequacy_filtered)
    except:
        return(get_max_str(tgt_text))
        
# Deploying the model

import gradio as gr

def get_fun(Input_txt):
  tokens = splitter.split(text=Input_txt)

  txt_paraphrase=''
  for phrase in tokens:
    tmp=get_response(phrase)
    txt_paraphrase=txt_paraphrase+' '+tmp
  return txt_paraphrase

iface = gr.Interface(fn=get_fun, inputs="text", outputs="text", title = " Ai Re-Phraser Q'Hackday")
iface.launch(inline=False)

"""# New Section"""