app.py

import streamlit as st
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer
from model_SingleLabelClassifier import SingleLabelClassifier
from safetensors.torch import load_file
import json


MODEL_NAME = "allenai/scibert_scivocab_uncased"
CHECKPOINT_PATH = "checkpoint-23985"
NUM_CLASSES = 65
MAX_LEN = 250

# Загрузка меток
with open("label_mappings.json", "r") as f:
    mappings = json.load(f)
abel2id = mappings["label2id"]
id2label = {int(k): v for k, v in mappings["id2label"].items()}

# Загрузка модели и токенизатора
@st.cache_resource
def load_model_and_tokenizer():
    tokenizer = AutoTokenizer.from_pretrained(CHECKPOINT_PATH)
    model = SingleLabelClassifier(MODEL_NAME, num_labels=NUM_CLASSES)
    state_dict = load_file(f"{CHECKPOINT_PATH}/model.safetensors")
    model.load_state_dict(state_dict)
    model.eval()
    return model, tokenizer

model, tokenizer = load_model_and_tokenizer()

# Функция предсказания
def predict(title, summary, model, tokenizer, id2label, max_length=320, top_k=3):
    model.eval()
    text = title + ". " + summary

    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        padding="max_length",
        max_length=max_length
    )

    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs["logits"]
        probs = F.softmax(logits, dim=1).squeeze().numpy()

    top_indices = probs.argsort()[::-1][:top_k]
    return [(id2label[i], round(probs[i], 3)) for i in top_indices]

# Интерфейс Streamlit
st.title("ArXiv Tag Predictor")
st.write("Вставьте заголовок и аннотацию статьи!")

title = st.text_input("**Title**")
summary = st.text_area("**Summary**", height=200)

if st.button("Предсказать тег"):
    if not title or not summary:
        st.warning("Пожалуйста, введите и заголовок, и аннотацию!")
    else:
        preds = predict(title, summary, model, tokenizer, id2label)
        st.subheader("Предсказанные теги:")
        for tag, prob in preds:
            st.write(f"**{tag}** — вероятность: {prob:.3f}")