# app.py
import streamlit as st
from transformers import DistilBertTokenizer, DistilBertForSequenceClassification
import torch
import numpy as np

@st.cache(allow_output_mutation=True)
def load_model():
    tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-cased')
    model = DistilBertForSequenceClassification.from_pretrained('model/')
    return tokenizer, model

tokenizer, model = load_model()

st.title('arXiv Article Classifier')
title = st.text_input('Title')
abstract = st.text_area('Abstract')
text = title + ' ' + abstract if abstract else title

if st.button('Predict'):
    if not text.strip():
        st.error('Please enter at least a title.')
    else:
        inputs = tokenizer(
            text,
            truncation=True,
            padding=True,
            max_length=512,
            return_tensors='pt'
        )
        with torch.no_grad():
            logits = model(**inputs).logits
        probs = torch.nn.functional.softmax(logits, dim=1).numpy()[0]
        sorted_indices = np.argsort(-probs)
        
        cumulative = 0
        result = []
        for idx in sorted_indices:
            cumulative += probs[idx]
            result.append((model.config.id2label[idx], probs[idx]))
            if cumulative >= 0.95:
                break
        for tag, prob in result:
            st.write(f'{tag}: {prob:.2%}')