app.py

import streamlit as st
import numpy as np
import pandas as pd
import torch
import transformers
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from scipy.special import softmax



def load_model():
    model = AutoModelForSequenceClassification.from_pretrained('model_distilbert_trained')
    tokenizer = AutoTokenizer.from_pretrained(
        'distilbert-base-cased', do_lower_case=True)
    model.eval()
    return model, tokenizer
    

def get_predictions(logits, indexes):
    sum = 0
    ind = []
    probs = []
    for i in indexes:
        sum += logits[i]
        ind.append(i)
        probs.append(indexes[i])
        if sum >= 0.95:
            return ind, probs


def return_pred_name(name_dict, ind):
    out = []
    for i in ind:
        out.append(name_dict[i])
    return out


def predict(title, summary, model, tokenizer):
    text = title + '.' + summary
    tokens = tokenizer.encode(text)
    with torch.no_grad():
        logits = model(torch.as_tensor([tokens]))[0]
        probs = torch.softmax(logits[-1, :], dim=-1).data.cpu().numpy()
        
    classes = np.flip(np.argsort(probs))
    sum_probs = 0
    ind = 0
    prediction = []
    prediction_probs = []
    while sum_probs < 0.95:
        prediction.append(name_dict[classes[ind]])
        prediction_probs.append(str("{:.2f}".format(100 * probs[classes[ind]])) + "%")
        sum_probs += probs[classes[ind]]
        ind += 1
    
    return prediction, prediction_probs


def get_results(prediction, prediction_probs):
    frame =  pd.DataFrame({'Category': prediction, 'Confidence': prediction_probs})
    frame.index = np.arange(1, len(frame) + 1)
    return frame
    
name_dict = {4: 'cs',
 19: 'stat',
 1: 'astro-ph',
 16: 'q-bio',
 6: 'eess',
 3: 'cond-mat',
 12: 'math',
 15: 'physics',
 18: 'quant-ph',
 17: 'q-fin',
 7: 'gr-qc',
 13: 'nlin',
 2: 'cmp-lg',
 5: 'econ',
 8: 'hep-ex',
 11: 'hep-th',
 14: 'nucl-th',
 10: 'hep-ph',
 9: 'hep-lat',
 0: 'adap-org'}



st.title("Find out the topic of the article without reading!")
st.markdown("<h1 style='text-align: center;'><img width=320px src = 'https://upload.wikimedia.org/wikipedia/ru/8/81/Sheldon_cooper.jpg'>",
    unsafe_allow_html=True)
# ^-- можно показывать пользователю текст, картинки, ограниченное подмножество html - всё как в jupyter

title = st.text_area(label='Title',
                            value='',
                            height=30,
                            help='If you know a title type it here')
                            

summary = st.text_area(label='Summary',
                            value='',
                            height=200,
                            help='If you have a summary enter it here')
                            
                       
button = st.button(label='Get the theme!')

if button:
    if (title == '' and summary == ''):
        st.write('There is nothing to analyze...')
        st.write('Fill at list one of the fields')
    else:
        model, tokenizer = load_model()
        prediction, prediction_probs = predict(title, summary, model, tokenizer)
        ans = get_results(prediction, prediction_probs)
        st.write('Result')
        st.write(ans)