app.py

import streamlit as st
import torch
from torch import nn
import csv
from transformers import AutoModel, AutoTokenizer
from huggingface_hub import hf_hub_download
from model import ClassificationModel

st.set_page_config(page_title="Article Theme Classifier", layout="centered")

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

MAX_LENGTH = 512

@st.cache_resource
def get_model():
    base_model = AutoModel.from_pretrained("distilbert-base-cased")
    class_model = ClassificationModel(base_model)
    
    weights_path = hf_hub_download(
        repo_id="MostoHF/TunedDistillBertCased", 
        filename="pytorch_model.bin"
    )

    state_dict = torch.load(weights_path, map_location=device)
    class_model.load_state_dict(state_dict)
    class_model.to(device)
    class_model.eval()
    
    return class_model

@st.cache_resource
def get_tokenizer():
    return AutoTokenizer.from_pretrained("distilbert-base-cased")

@st.cache_resource
def get_ind_to_cat():
    ind_to_category_copy = {}
    with open('ind_to_category.csv', mode='r', newline='') as f:
        reader = csv.reader(f)
        next(reader)  # skip header
        for key, value in reader:
            ind_to_category_copy[int(key)] = value  # ключи — int
    return ind_to_category_copy

class_model = get_model()
tokenizer = get_tokenizer()
ind_to_category = get_ind_to_cat()

def inference(title, abstract, threshold=0.95):
    cur_elem = title + '@' + abstract

    encoding = tokenizer(cur_elem, padding="max_length", truncation=True, max_length=MAX_LENGTH, return_tensors="pt")
    input_ids = encoding["input_ids"].to(device)
    attention_mask = encoding["attention_mask"].to(device)

    with torch.no_grad():
        res_probs = torch.exp(class_model(input_ids, attention_mask))
    
    probs = res_probs.squeeze(0)  # (8,)
    sorted_probs, sorted_indices = torch.sort(probs, descending=True)

    total = 0.0
    selected_indices = []
    selected_probs = []
    
    for prob, idx in zip(sorted_probs, sorted_indices):
        total += prob.item()
        selected_indices.append(idx.item())
        selected_probs.append(prob.item())
        if total >= threshold:
            break

    ans_themes = [ind_to_category[idx] for idx in selected_indices]
    return ans_themes, selected_probs


# ------------------- Streamlit UI -------------------

st.title("📄 Article Theme Classifier")

title = st.text_input("Title", value="Введите title...")
abstract = st.text_input("Abstract", value="Введите abstract...")
threshold = st.slider("Выберите cumulative probability threshold", 0.0, 1.0, step=0.01, value=0.95)

if st.button("Submit"):
    if title or abstract:
        st.success(f"✅ Title")
        st.info(f"📑 Abstract")
        themes, probs = inference(title, abstract, threshold)
        st.subheader("Predicted Themes:")
        for i in range(len(themes)):
            st.write(f"**{themes[i]}** — {probs[i]:.4f}")
    else:
        st.warning("❌ Please fill in at least one of the fields.")