app.py

import streamlit as st
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import torch

import tokenizers 
import transformers
from transformers import TextClassificationPipeline, AutoTokenizer, AutoModelForSequenceClassification  

@st.cache(suppress_st_warning=True, hash_funcs={tokenizers.Tokenizer: lambda _: None})
def load_tok_and_model(): 
    tokenizer = AutoTokenizer.from_pretrained('distilbert-base-uncased')
    model = AutoModelForSequenceClassification.from_pretrained(".")
    return tokenizer, model
    
    
    
tag = ['Cs', 'Econ', 'EESS', 'Math', 'Physics', 'Q-bio', 'Q-fin', 'Stat']

inv_map = {3: 'Math', 4: 'Physics', 5: 'Q-bio', 0: 'Cs', 6: 'Q-fin', 7: 'Stat', 2: 'EESS', 1: 'Econ'}


@st.cache(suppress_st_warning=True, hash_funcs={tokenizers.Tokenizer: lambda _: None})  
def predict_label(title, summary, tokenizer, model, inv_map):
    abstract = title.lower() + '. ' + summary.lower()
    token_text = tokenizer.encode(abstract)
    
    with torch.no_grad():
        logits = model(torch.as_tensor([token_text]))[0]
        probs = torch.softmax(logits[-1, :], dim=-1).data.numpy()
    
    idx_label = np.argsort(probs)[::-1]

    sum_probs = 0
    prediction_probs = []
    prediction_classes = []

    idx = 0
    while sum_probs < 0.95:
        cur_predict = inv_map[idx_label[idx]]
        cur_probs = probs[idx_label[idx]]
    
        sum_probs += cur_probs
    
        prediction_probs.append(int(100 * cur_probs))
        prediction_classes.append(cur_predict)
    
        idx += 1
    
    return prediction_classes, prediction_probs, probs

st.title("Classifier of possible topics of articles 📄")

st.markdown("Please insert the summary and/or title of the article below")

tokenizer, model = load_tok_and_model()

title = st.text_area(label='Title', height=50)
abstract = st.text_area(label='Summary', height=150)

if st.button('Start classifier'):
    if title == '' and abstract == '':
        st.markdown("Summary and title should be filled in in the text area above")
    else:
        prediction_classes, prediction_probs, probs = predict_label(title, abstract, tokenizer, model, inv_map)
        
        data = pd.DataFrame({'Categories' : tag, 'Probs' : probs})
        data = data.sort_values(by='Probs', ascending=False)
        
        fig, ax = plt.subplots()
        ax.bar(data['Categories'], data['Probs'])
        ax.bar(prediction_classes, prediction_probs)
        
      
        data_answer = pd.DataFrame({'Categories' : prediction_classes, 'Probs, %' : prediction_probs})
    
    
        st.pyplot(fig)
        st.write('top-95%')
        st.write(data_answer)