app.py

import transformers
import streamlit as st
import nltk
from nltk import sent_tokenize
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import json
import numpy as np
from sentence_transformers import SentenceTransformer

nltk.download('punkt')
with open('testbook.json') as f:
    test_book = json.load(f)
tokenizer = AutoTokenizer.from_pretrained("UNIST-Eunchan/bart-dnc-booksum")

@st.cache_resource
def load_model(model_name):
    nltk.download('punkt')
    sentence_transformer_model = SentenceTransformer("sentence-transformers/all-roberta-large-v1")
    model = AutoModelForSeq2SeqLM.from_pretrained("UNIST-Eunchan/bart-dnc-booksum")
    return sentence_transformer_model, model

sentence_transformer_model, model = load_model("UNIST-Eunchan/bart-dnc-booksum")

def infer(input_ids, max_length, temperature, top_k, top_p):

    output_sequences = model.generate(
        input_ids=input_ids,
        max_length=max_length,
        temperature=temperature,
        top_k=top_k,
        top_p=top_p,
        do_sample=True,
        num_return_sequences=1,
        num_beams=4,
        no_repeat_ngram_size=2
    )
    
    return output_sequences


def cos_similarity(v1, v2):
    dot_product = np.dot(v1, v2)
    l2_norm = (np.sqrt(sum(np.square(v1))) * np.sqrt(sum(np.square(v2))))
    similarity = dot_product / l2_norm     
    
    return similarity


@st.cache_data
def chunking(book_text):
    sentences = sent_tokenize(book_text)
    segments = [] 
    token_lens = []

    for sent_i_th in sentences:
        token_lens.append(len(tokenizer.tokenize(sent_i_th)))
    #sentences, token_lens
    current_segment = ""    
    total_token_lens = 0
    for i in range(len(sentences)):
        
        if total_token_lens < 512:
            total_token_lens += token_lens[i]
            current_segment += (sentences[i] + " ")
            
        elif total_token_lens > 768:
            segments.append(current_segment)
            current_segment = sentences[i]
            total_token_lens = token_lens[i]
    
        else:
            #make next_pseudo_segment
            next_pseudo_segment = ""
            next_token_len = 0
            for t in range(10):
                if (i+t < len(sentences)) and (next_token_len + token_lens[i+t] < 512):
                    next_token_len += token_lens[i+t] 
                    next_pseudo_segment += sentences[i+t] 
                
            embs = sentence_transformer_model.encode([current_segment, next_pseudo_segment, sentences[i]]) # current, next, sent
            if cos_similarity(embs[1],embs[2]) > cos_similarity(embs[0],embs[2]):
                segments.append(current_segment)
                current_segment = sentences[i]
                total_token_lens = token_lens[i]
            else: 
                total_token_lens += token_lens[i]
                current_segment += (sentences[i] + " ")

    return segments


book_index = 0
_book = test_book[book_index]['book']

#prompts
st.title("Book Summarization 📚")
st.write("The almighty king of text generation, GPT-2 comes in four available sizes, only three of which have been publicly made available. Feared for its fake news generation capabilities, it currently stands as the most syntactically coherent model. A direct successor to the original GPT, it reinforces the already established pre-training/fine-tuning killer duo. From the paper: Language Models are Unsupervised Multitask Learners by Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei and Ilya Sutskever.")

#book_index = st.sidebar.slider("Select Book Example", value = 0,min_value = 0, max_value=4)
sent = st.text_area("Text", _book[:512], height = 550)
max_length = st.sidebar.slider("Max Length", value = 512,min_value = 10, max_value=1024)
temperature = st.sidebar.slider("Temperature", value = 1.0, min_value = 0.0, max_value=1.0, step=0.05)
top_k = st.sidebar.slider("Top-k", min_value = 0, max_value=5, value = 0)
top_p = st.sidebar.slider("Top-p", min_value = 0.0, max_value=1.0, step = 0.05, value = 0.92)


chunked_segments = chunking(_book)


def generate_output(test_samples):
    inputs = tokenizer(
        test_samples,
        padding=max_length,
        truncation=True,
        max_length=1024,
        return_tensors="pt",
    )
    
    input_ids = inputs.input_ids
    
    attention_mask = inputs.attention_mask
    
    outputs = model.generate(input_ids,
                             max_length = 256,
                             min_length=32,
                             top_p = 0.92,
                             num_beams=5,
                             no_repeat_ngram_size=2,
                             attention_mask=attention_mask)
    output_str = tokenizer.batch_decode(outputs, skip_special_tokens=True)
    return outputs, output_str


chunked_segments = chunking(test_book[0]['book'])


for segment in range(len(chunked_segments)):
        
    summaries = generate_output(segment)
    st.write(summaries[-1])