app.py

import numpy as np
import pandas as pd
import transformers
import torch
import tokenizers
import streamlit as st

NUM_LABELS = 15

labels_names = {
    0: 'Astrophysics',
    1: 'Condensed Matter',
    2: 'Computer Science',
    3: 'Economics',
    4: 'Electrical Engineering and Systems Science',
    5: 'General Relativity and Quantum Cosmology',
    6: 'High Energy Physics',
    7: 'Mathematics',
    8: 'Nonlinear Sciences',
    9: 'Nuclear Theory',
    10: 'General Physics',
    11: 'Quantitative Biology',
    12: 'Quantitative Finance',
    13: 'Quantum Physics',
    14: 'Statistics',
}

@st.cache(hash_funcs={tokenizers.Tokenizer: lambda _: None}, suppress_st_warning=True)
def get_model(model_name, model_path):
    tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)
    model = transformers.AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=NUM_LABELS)
    model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))
    model.eval()
    return model, tokenizer
    
@st.cache(hash_funcs={tokenizers.Tokenizer: lambda _: None}, suppress_st_warning=True)
def predict(text, tokenizer, model, temperature = 1):
    tokens = tokenizer.encode(text)
    with torch.no_grad():
        logits = model.cpu()(torch.as_tensor([tokens]))[0]
        probs = torch.softmax(logits[-1, :] / temperature, dim=-1).data.cpu().numpy()
    
    indexes_descending = np.argsort(probs)[::-1]
    percents = 0
    preds = []
    pred_probs = []
    for index in indexes_descending:
        preds.append(labels_names[index])
        pred_prob = 100 * probs[index]
        pred_probs.append(f"{pred_prob:.1f}%")

        percents += pred_prob
        if percents >= 95:
          break

    result = pd.DataFrame({'Probability': pred_probs})
    result.index = preds
    
    return result

model, tokenizer = get_model('distilbert-base-cased', 'distilbert-checkpoint-10983.bin')

st.title("Yandex School of Data Analysis. ML course")
st.title("Laboratory work 2: classifier of categories of scientific papers")
st.markdown("<img width=200px src='https://m.media-amazon.com/images/I/71XOMSKx8NL._AC_SL1500_.jpg'>", unsafe_allow_html=True)

st.markdown("\n")

st.markdown("Enter the title of the article and its abstract (although, if you really don't want to, you can do with just the title)")

title = st.text_area(label='Title of the article', height=100)
abstract = st.text_area(label='Abstract of the article', height=200)
button = st.button('Go')

if button:
    try:
        text = ' [ABSTRACT] '.join([title, abstract])
        result = predict(text, tokenizer, model)
        if len(text) > 10:
            st.subheader('Bumblebee thinks, this paper related to')
            st.write(result)
        else:
            st.error("Enter some more info please")
    except Exception:
       st.error("Ooooops, something went wrong. Try again please and report to me, tg: @vladyur")