src/streamlit_app.py

import streamlit as st
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import json
import os

BASE_DIR = os.path.dirname(os.path.abspath(__file__))
MODEL_DIR = os.path.join(BASE_DIR, "arxiv_dir")

st.set_page_config(
    page_title="Arxiv Classifier",
    page_icon="🚀",
    layout="wide",
    initial_sidebar_state="collapsed"
)

st.title("Arxiv Classifier")

@st.cache_resource
def load_model():
    model = AutoModelForSequenceClassification.from_pretrained(MODEL_DIR)
    tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR)
    with open(os.path.join(MODEL_DIR, "id2tag.json")) as f:
        id2tag = {int(k): v for k, v in json.load(f).items()}
    with open(os.path.join(MODEL_DIR, "tag2name.json")) as f:
        tag2name = json.load(f)
    model.eval()
    return model, tokenizer, id2tag, tag2name
    
model, tokenizer, id2tag, tag2name = load_model()

def predict_top95(title, summary=None):
    text = title
    if summary:
        text += " [SEP] " + summary

    tokens = tokenizer(text, truncation=True, max_length=512, return_tensors="pt")
    tokens = {k: v.to(model.device) for k, v in tokens.items()}

    model.eval()
    with torch.no_grad():
        logits = model(**tokens).logits
        probs = torch.softmax(logits, dim=-1)[0]

    sorted_probs, sorted_idx = probs.sort(descending=True)
    cumsum = sorted_probs.cumsum(dim=-1)
    mask = (cumsum - sorted_probs) < 0.95

    results = []
    for prob, idx in zip(sorted_probs[mask], sorted_idx[mask]):
        results.append((id2tag[idx.item()], prob.item()))

    return results

def colored_bar(prob):
    if prob > 0.4:
        color = "#2ecc71"
    elif prob > 0.15:
        color = "#f39c12"
    else:
        color = "#e74c3c"
    st.markdown(f"""
    <div style="background:#e0e0e0;border-radius:4px;height:10px;margin:4px 0">
      <div style="background:{color};width:{prob*100:.1f}%;height:100%;border-radius:4px"></div>
    </div>
    """, unsafe_allow_html=True)

col1, col2 = st.columns([1, 1])

with col1:
    name = st.text_input("Название статьи")
    abstract = st.text_area("Abstract", height=200)
    clicked = st.button("Классифицировать")

with col2:
    if clicked:
        if not name and not abstract:
            st.warning("Введите название или abstract")
        else:
            results = predict_top95(name, abstract if abstract else None)

            st.markdown(f"### Результаты — {len(results)} {'класс' if len(results) == 1 else 'класса' if len(results) < 5 else 'классов'}")

            visible = results[:10]
            hidden = results[10:]

            for tag, prob in visible:
                label = tag2name.get(tag, tag)
                st.markdown(f"**{label}** `{tag}`")
                colored_bar(prob)
                st.caption(f"{prob:.1%}")

            if hidden:
                with st.expander(f"Показать ещё {len(hidden)}"):
                    for tag, prob in hidden:
                        label = tag2name.get(tag, tag)
                        st.markdown(f"**{label}** `{tag}`")
                        colored_bar(prob)
                        st.caption(f"{prob:.1%}")