app.py

import streamlit as st
import torch
from transformers import DistilBertTokenizer, DistilBertForSequenceClassification
import pickle

# Скачать и загрузить токенизатор
tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-cased')

# Загружаем квантизированную модель
def load_quantized_model(model_path, num_labels):
    model = DistilBertForSequenceClassification.from_pretrained('distilbert-base-cased', num_labels=num_labels)
    model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu'))['model_state_dict'])
    return model

model_path = "epoch_2.pt"
num_labels = 126  # или другое количество меток в зависимости от вашей задачи
model = load_quantized_model(model_path, num_labels)
model.eval()  # Перевод модели в режим оценки

# Установите CUDA, если доступно, иначе используйте CPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# Streamlit интерфейс
st.title('Arxiv tag classification')

title = st.text_input('Title:', '')
summary = st.text_area('Summary:', '')

if st.button('Predict'):
    # Подготовка данных
    combined_text = f"{title} {summary}"
    inputs = tokenizer(
        combined_text,
        padding=True,
        truncation=True,
        max_length=512,
        return_tensors='pt'
    ).to(device)

    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits
        probs = torch.nn.functional.softmax(logits, dim=-1).squeeze()

    # Отсортируем вероятности и классы
    sorted_indices = torch.argsort(probs, descending=True)
    cumulative_probs = 0.0
    selected_indices = []

    # Выбираем классы пока сумма вероятностей не станет >= 0.95
    for idx in sorted_indices:
        cumulative_probs += probs[idx].item()
        selected_indices.append((idx.item(), probs[idx].item()))
        if cumulative_probs >= 0.95:
            break

    with open('label_encoder.pkl', 'rb') as f:
        label_encoder = pickle.load(f)

    # Отображение результата
    st.write("Predicted classes by probability up to 95%:")
    for class_id, probability in selected_indices:
        st.write(f'Class : {label_encoder.inverse_transform([class_id])[0]}, Probability: {probability * 100:.2f}%')