app.py

import streamlit as st
import time

# model part

import json
import torch
from torch import nn
from transformers import AutoTokenizer, AutoModelForSequenceClassification

with open('categories_with_names.json', 'r') as f:
    cat_with_names = json.load(f)
with open('categories_from_model.json', 'r') as f:
    categories_from_model = json.load(f)

@st.cache_resource
def load_models_and_tokenizer():
    tokenizer = AutoTokenizer.from_pretrained("oracat/bert-paper-classifier-arxiv")

    model_titles = AutoModelForSequenceClassification.from_pretrained(
        "powerful_model_titles/checkpoint-13472", num_labels=len(categories_from_model), problem_type="multi_label_classification"
    )
    model_titles.eval()
    model_abstracts = AutoModelForSequenceClassification.from_pretrained(
        "powerful_model_abstracts/checkpoint-13472", num_labels=len(categories_from_model), problem_type="multi_label_classification"
    )
    model_abstracts.eval()
    
    return model_titles, model_abstracts, tokenizer

model_titles, model_abstracts, tokenizer = load_models_and_tokenizer()

def categorize_text(title: str | None = None, abstract: str | None = None, progress_bar = None):
    if title is None and abstract is None:
        raise ValueError('title is None and abstract is None')

    models_to_run = 2 if (title is not None and abstract is not None) else 1
        
    proba_title = None
    if title is not None:
        progresses = (10, 30) if models_to_run == 2 else (20, 60)
        my_bar.progress(progresses[0], text='computing titles')
        input_tok = tokenizer(title, return_tensors='pt')
        with torch.no_grad():
            logits = model_titles(**input_tok)['logits']
            proba_title = torch.sigmoid(logits)[0]
        my_bar.progress(progresses[1], text='computed titles')
            
    proba_abstract = None
    if abstract is not None:
        progresses = (40, 70) if models_to_run == 2 else (20, 60)
        my_bar.progress(progresses[0], text='computing abstracts')
        input_tok = tokenizer(abstract, return_tensors='pt')
        with torch.no_grad():
            logits = model_abstracts(**input_tok)['logits']
            proba_abstract = torch.sigmoid(logits)[0]
        my_bar.progress(progresses[0], text='computed abstracts')
            
    if title is None:
        proba = proba_abstract
    elif abstract is None:
        proba = proba_title
    else:
        proba = proba_title * 0.1 + proba_abstract * 0.9

    progresses = (80, 90) if models_to_run == 2 else (70, 90)

    my_bar.progress(progresses[0], text='computed proba')
    
    sorted_proba, indices = torch.sort(proba, descending=True)
    my_bar.progress(progresses[1], text='sorted proba')
    to_take = 1
    while sorted_proba[:to_take].sum() < 0.95 and to_take < len(categories_from_model):
        to_take += 1
    output = [(cat_with_names[categories_from_model[index]], proba[index].item())
            for index in indices[:to_take]]
    my_bar.progress(100, text='generated output')
    return output

# front part

st.markdown("<h1 style='text-align: center;'>Classify your paper!</h1>", unsafe_allow_html=True)


if "title" not in st.session_state:
    st.session_state.title = ""
if "abstract" not in st.session_state:
    st.session_state.abstract = ""
if "title_input_key" not in st.session_state:
    st.session_state.title_input_key = ""
if "abstract_input_key" not in st.session_state:
    st.session_state.abstract_input_key = ""
if "model_type" not in st.session_state:
    st.session_state.model_type = []

def input_error():
    if not st.session_state.model_type:
        return 'you have to select title or abstract'
    if 'Title' in model_type and not st.session_state.title:
        return 'Title is empty'
    if 'Abstract' in model_type and not st.session_state.abstract:
        return 'Abstract is empty'
    return ''


def clear_input():
    st.session_state.title = title.title()
    st.session_state.abstract = abstract.title()
    if not input_error():
        if "Title" in st.session_state.model_type:
           st.session_state.title_input_key = ""
        if "Abstract" in st.session_state.model_type:
           st.session_state.abstract_input_key = ""


title = st.text_input(r"$\textsf{\Large Title}$", key="title_input_key")

abstract = st.text_input(r"$\textsf{\Large Abstract}$", key="abstract_input_key")


model_type = st.multiselect(
    r"$\textsf{\large Classify by:}$",
    ['Title', 'Abstract'],
)

st.session_state.model_type = model_type

if(st.button('Submit', on_click=clear_input)):
    if input_error():
        st.error(input_error())
    else:
        send_time = time.localtime(time.time())
        #st.success(f"Submitted {(' and '.join(st.session_state.model_type)).lower()} on {time.strftime('%d.%m %H:%M:%S', send_time)}")
        model_input = dict()
        if 'Title' in st.session_state.model_type:
            model_input['title'] = st.session_state.title
        if 'Abstract' in st.session_state.model_type:
            model_input['abstract'] = st.session_state.abstract
        #st.success(f'{model_input=}')
        my_bar = st.progress(0, text='starting model')
        model_result = categorize_text(**model_input, progress_bar=my_bar)
        st.markdown("<h1 style='text-align: center;'>Classification completed!</h1>", unsafe_allow_html=True)
        small_categories = []
        cat, proba = model_result[0]
        st.write(r"$\textsf{\Large " + f'{cat}: {round(100*proba)}' + r"\%}$")
        for cat, proba in model_result[1:]:
            if proba < 0.1:
                small_categories.append(f'{cat}: {round(100*proba, 1)}' + r"\%")
            else:
                st.write(r"$\textsf{\large " + f'{cat}: {round(100*proba)}' + r"\%}$")
        if small_categories:
            st.write(', '.join(small_categories))