app.py

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

import os

os.environ['KMP_DUPLICATE_LIB_OK']='True'
        
        
@st.cache(allow_output_mutation=True)
def load_model_and_tokenizer():
    tokenizer = AutoTokenizer.from_pretrained("cl-tohoku/bert-base-japanese-v2")
    model = AutoModel.from_pretrained("kaisugi/anlp_embedding_model")
    model.eval()
    
    return model, tokenizer


@st.cache(allow_output_mutation=True)
def load_title_data():
    title_df = pd.read_csv("anlp2023.csv")
    
    return title_df


@st.cache(allow_output_mutation=True)
def load_title_embeddings():
    npz_comp = np.load("anlp_title_embeddings.npz")
    title_embeddings = npz_comp["arr_0"]

    return title_embeddings


@st.cache
def get_retrieval_results(index, input_text, top_k, tokenizer, title_df):
    with torch.no_grad():
        inputs = tokenizer.encode_plus(
            input_text, 
            padding=True,
            truncation="only_second",
            return_tensors="pt",
            max_length=512,
        )
        outputs = model(**inputs)
        query_embeddings = outputs.last_hidden_state[:, 0, :][0]
        query_embeddings = query_embeddings.detach().cpu().numpy()
    
    _, 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("## NLP2023 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)