app.py

import faiss
import numpy as np
import pandas as pd
import streamlit as st
import torch
from torch import Tensor
from transformers import AutoModel, AutoTokenizer

import os

os.environ['KMP_DUPLICATE_LIB_OK']='True'


def average_pool(last_hidden_states: Tensor,
                 attention_mask: Tensor) -> Tensor:
    last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
    return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
        
        
@st.cache_resource
def load_model_and_tokenizer():
    tokenizer = AutoTokenizer.from_pretrained('intfloat/multilingual-e5-large')
    model = AutoModel.from_pretrained('intfloat/multilingual-e5-large')
    model.eval()
    
    return model, tokenizer


@st.cache_resource
def load_title_data():
    title_df = pd.read_csv('anlp2024.tsv', names=["pid", "title"], sep="\t")
    
    return title_df


@st.cache_resource
def load_title_embeddings():
    npz_comp = np.load("anlp2024.npz")
    title_embeddings = npz_comp["arr_0"]

    return title_embeddings


@st.cache_data
def get_retrieval_results(index, input_text, top_k, tokenizer, title_df):
    batch_dict = tokenizer(f"query: {input_text}", max_length=512, padding=True, truncation=True, return_tensors='pt')
    with torch.no_grad():
        outputs = model(**batch_dict)
        query_embeddings = average_pool(outputs.last_hidden_state, batch_dict['attention_mask'])
        query_embeddings = F.normalize(query_embeddings, p=2, dim=1)
    
    _, ids = index.search(x=np.array([query_embeddings]), k=top_k)
    retrieved_titles = []
    retrieved_pids = []

    for id in ids[0]:
        retrieved_titles.append(title_df.loc[id, "title"])
        retrieved_pids.append(title_df.loc[id, "pid"])

    df = pd.DataFrame({"pids": retrieved_pids, "paper": retrieved_titles})
    
    return df
    

if __name__ == "__main__":
    model, tokenizer = load_model_and_tokenizer()
    title_df = load_title_data()
    title_embeddings = load_title_embeddings()

    index = faiss.IndexFlatL2(768)
    index.add(title_embeddings)
    
    st.markdown("## NLP2024 類似論文検索")
    input_text = st.text_input('input', '', placeholder='ここに論文のタイトルを入力してください')
    top_k = st.number_input('top_k', min_value=1, value=10, step=1)
    
    if st.button('検索'):
        stripped_input_text = input_text.strip()
        df = get_retrieval_results(index, stripped_input_text, top_k, tokenizer, title_df)
        st.table(df)