app.py

import streamlit as st
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import datasets
import torch

model_name = 'distilbert-base-cased'

def load_model():
     return AutoTokenizer.from_pretrained(model_name), AutoModelForSequenceClassification.from_pretrained('./')    

if 'tokenizer' not in globals():
    tokenizer, model = load_model()

title = st.text_area('Title')
abstract = st.text_area('Abstract')


label_to_topic_dict = dict(enumerate(['Computer Science',
                                 'Economics',
                                'Electrical Engineering and Systems Science',
                                'Mathematics',
                                'Physics',
                                'Quantitative Biology',
                                'Quantitative Finance',
                                'Statistics']))
topic_to_label_dict = {label_to_topic_dict[key]: key for key in label_to_topic_dict.keys()}

device='cuda:0' if torch.cuda.is_available() else 'cpu'
def predict(title, abstract):
    d = {'title': [title], 'abstract': [abstract]}
    d = datasets.Dataset.from_dict(d)
    d = tokenizer(d["title"], d['abstract'], padding="max_length", truncation=True, return_tensors='pt')
    logits = model(input_ids=d['input_ids'].to(device), attention_mask=d['attention_mask'].to(device))['logits']
    p = torch.nn.functional.softmax(logits)[0].cpu().detach()
    preds = []
    proba = 0
    for index in p.argsort(descending=True).tolist():
        preds.append((label_to_topic_dict[index], p[index].item()))
        proba += p[index]
        if proba > .95:
            break
    return preds
    
if len(title) == 0 and len(abstract) == 0:
    pass
else:  
    output = predict(title, abstract)
    st.text("Top 95% topics:")
    for topic, proba in output:
        st.text(f"{topic}: {proba*100:.0f}%")