functions.py

import os
import random

import requests
import torch
from bs4 import BeautifulSoup
from datasets import Dataset
from langchain.docstore.document import Document
from langchain.llms import HuggingFacePipeline
from langchain.text_splitter import CharacterTextSplitter
from peft import PeftConfig, PeftModel
from transformers import (AutoModel, AutoModelForCausalLM, AutoTokenizer,
                          GenerationConfig, pipeline)

# os.environ["CUDA_VISIBLE_DEVICES"] = "0"


generation_config = GenerationConfig(temperature=.8,
                                     top_p=0.75,
                                     top_k=40)
device = 'cuda'

shared = {
    'answer_context': None,
    'embeddings_dataset': None,
    'full_text': None,
}

text_splitter = CharacterTextSplitter()


def get_nearest_examples(question: str, k: int):
    print(['get_nearest_examples', 'start'])
    question_embedding = get_embeddings([question]).cpu().detach().numpy()
    embeddings_dataset = shared['embeddings_dataset']
    scores, samples = embeddings_dataset.get_nearest_examples(
        "embeddings", question_embedding, k)
    print(['get_nearest_examples', 'scores and samples'])
    print(scores)
    print(samples['id'])
    print(['get_nearest_examples', 'end'])
    return samples


def get_embeddings(text):
    print(['get_embeddings', 'start'])
    encoded_input = emb_tokenizer(text,
                                  padding=True,
                                  truncation=True,
                                  return_tensors="pt")
    encoded_input = {k: v.to('cuda') for k, v in encoded_input.items()}
    model_output = emb_model(**encoded_input)
    model_output = model_output.last_hidden_state[:, 0]
    print(['get_embeddings', 'end'])
    return model_output


def build_faiss_index(text):
    print(['build_faiss_index', 'start'])
    text_list = split_text(text)
    emb_list = []
    for i, item in enumerate(text_list):
        emb_list.append({
            "embeddings": get_embeddings(item).cpu().detach().numpy()[0],
            'id': i
        })
    dataset = Dataset.from_list(emb_list)
    dataset.add_faiss_index(column="embeddings")
    shared['embeddings_dataset'] = dataset
    print(['build_faiss_index', 'end'])


def extract_text(url: str):
    print(['extract_text', 'start'])
    if url is None or url.strip() == '':
        return ''
    response = requests.get(url)
    soup = BeautifulSoup(response.text, "html.parser")
    text = '\n\n'.join(map(lambda p: p.text, soup.find_all('p')))
    shared['full_text'] = text
    print(['extract_text', 'end'])
    return text


def split_text(text: str):
    lines = text.split('\n')
    lines = [line.strip() for line in lines if line.strip()]
    return lines


def remove_prompt(text: str) -> str:
    output_prompt = 'Output: '
    idx = text.index(output_prompt)
    res = text[idx + len(output_prompt):].strip()
    res = res.replace('Input: ', '')
    return res


def summarize_text(text: str) -> str:
    print(['summarize_text', 'start'])

    print(['summarize_text', 'splitting text'])
    texts = text_splitter.split_text(text)
    docs = [Document(page_content=t) for t in texts]
    prompts = [f'<s>Instruction: Elabora un resume del siguiente texto.\nInput: {d.page_content}\nOutput: '
               for d in docs]

    print(['summarize_text', 'generating'])
    cleaned_summaries = [remove_prompt(
        s['generated_text']) for s in pipe(prompts)]
    summaries = '\n\n'.join(cleaned_summaries)

    print(['summarize_text', 'end'])
    return summaries


def summarize_text_v1(text: str):
    print(['summarize_text', 'start'])
    input_text = f'<s>Instruction: Elabora un resume del siguiente texto.\nInput: {text}\nOutput: '
    batch = tokenizer(input_text, return_tensors='pt')
    batch = batch.to(device)
    print(['summarize_text', 'generating'])
    with torch.cuda.amp.autocast():
        output_tokens = model.generate(**batch,
                                       max_new_tokens=512,
                                       generation_config=generation_config
                                       )
    output = tokenizer.decode(output_tokens[0], skip_special_tokens=True)
    output = output.replace(input_text, '')
    print(['summarize_text', 'end'])
    return output


def generate_question(text: str):
    print(['generate_question', 'start'])
    # Get a random section of the whole text to generate a question
    fragments = split_text(text)
    rnd_text = random.choice(fragments)
    shared['answer_context'] = rnd_text

    input_text = f'<s>Instruction: Dado el siguiente texto quiero que generes una pregunta cuya respuesta se encuentre en él.\nInput: {rnd_text}\nOutput: '
    batch = tokenizer(input_text, return_tensors='pt')
    print(['generate_question', 'generating'])
    with torch.cuda.amp.autocast():
        output_tokens = model.generate(**batch,
                                       max_new_tokens=256,
                                       generation_config=generation_config)
    output = tokenizer.decode(output_tokens[0], skip_special_tokens=True)
    output = output.replace(input_text, '')
    print(['generate_question', 'end'])
    return output


def get_answer_context():
    return shared['answer_context']


def answer_question(question: str):
    print(['answer_question', 'start'])
    full_text = shared['full_text']

    if not shared['embeddings_dataset']:
        build_faiss_index(full_text)
    top_k_samples = get_nearest_examples(question, k=3)

    index_text = {}
    for i, t in enumerate(split_text(full_text)):
        index_text[i] = t

    context = '\n'.join([index_text[id] for id in top_k_samples['id']])

    input_text = f"""<s>Instruction: Te voy a proporcionar un texto del cual deseo que me respondas una pregunta. 
    El texto es el siguiente: `{context}`\nInput: {question}\nOutput: """
    batch = tokenizer(input_text, return_tensors='pt')
    print(['answer_question', 'generating'])
    with torch.cuda.amp.autocast():
        output_tokens = model.generate(**batch,
                                       max_new_tokens=256,
                                       generation_config=generation_config)
    output = tokenizer.decode(output_tokens[0], skip_special_tokens=True)
    output = output.replace(input_text, '')
    print(['answer_question', 'end'])
    return output


def load_model(peft_model_id):
    print(['load_model', 'start'])
    config = PeftConfig.from_pretrained(peft_model_id)
    print(['load_model', 'loading model'])
    model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path,
                                                 return_dict=True,
                                                 load_in_8bit=True,
                                                 device_map='auto')
    print(['load_model', 'loading tokenizer'])
    tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
    model = PeftModel.from_pretrained(model, peft_model_id)
    model.config.use_cache = True
    print(['load_model', 'end'])
    return model, tokenizer


def load_embeddings_model(model_ckpt: str):
    print(['load_embeddings_model', 'start'])
    print(['load_embeddings_model', 'loading tokenizer'])
    tokenizer = AutoTokenizer.from_pretrained(model_ckpt)
    print(['load_embeddings_model', 'loading model'])
    model = AutoModel.from_pretrained(model_ckpt)
    model = model.to(device)
    print(['load_embeddings_model', 'end'])
    return model, tokenizer


model, tokenizer = load_model(
    "hackathon-somos-nlp-2023/opt-6.7b-lora-sag-t3000-v300-v2")
pipe = pipeline("text2text-generation", model=model,
                tokenizer=tokenizer, max_new_tokens=100)
llm = HuggingFacePipeline(pipeline=pipe)

# Sentence Transformers models
# - paraphrase-multilingual-MiniLM-L12-v2
# - multi-qa-mpnet-base-dot-v1
emb_model, emb_tokenizer = load_embeddings_model("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")