src/streamlit_app.py

import streamlit as st
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from pathlib import Path

MODEL_PATH = str(Path(__file__).resolve().parent.parent / "model")
TOKENIZER_NAME = "oracat/bert-paper-classifier-arxiv"


@st.cache_resource
def load_model():
    tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_NAME)
    model = AutoModelForSequenceClassification.from_pretrained(MODEL_PATH)
    model.eval()
    return tokenizer, model


st.set_page_config(page_title="Arxiv Paper Classifier", layout="centered")
st.title("Arxiv Paper Classifier")
st.markdown("Classify an academic paper into an arxiv category by its title and/or abstract.")

title = st.text_input("Paper title (optional)", placeholder="e.g. Attention Is All You Need")
abstract = st.text_area("Abstract (optional)", placeholder="Paste the paper abstract here...", height=200)

if st.button("Classify", type="primary"):
    if not title and not abstract:
        st.error("Please enter at least a title or an abstract.")
    else:
        tokenizer, model = load_model()

        if title and abstract:
            inputs = tokenizer(title, abstract, truncation=True, max_length=256, return_tensors="pt")
        elif title:
            inputs = tokenizer(title, truncation=True, max_length=256, return_tensors="pt")
        else:
            inputs = tokenizer(abstract, truncation=True, max_length=256, return_tensors="pt")

        with torch.no_grad():
            logits = model(**inputs).logits

        probs = F.softmax(logits, dim=-1).squeeze()
        sorted_indices = torch.argsort(probs, descending=True)

        id2label = model.config.id2label
        cumulative = 0.0
        results = []
        for idx in sorted_indices:
            idx = idx.item()
            prob = probs[idx].item()
            cumulative += prob
            results.append((id2label[idx], prob))
            if cumulative >= 0.95:
                break

        st.subheader("Predicted categories (top-95%)")
        for label, prob in results:
            col1, col2 = st.columns([3, 7])
            with col1:
                st.markdown(f"**{label}**")
            with col2:
                st.progress(prob, text=f"{prob:.1%}")