app.py

import streamlit as st
from transformers import AutoModelForSequenceClassification
from transformers import AutoTokenizer
import torch
import pandas as pd
import numpy as np


st.markdown("# Arxiv Papers Classifier")
st.markdown("<img width=200px src='https://blog.arxiv.org/files/2021/02/arxiv-logo.svg'>", unsafe_allow_html=True)

st.markdown("После обработки и фильтрации датасета у каждой статьи остался один или несколько классов из 9:")
st.markdown("""
1) ai

2) cs

3) cv

4) lg

5) math

6) ml

7) phys

8) q-bio

9) stat

""")

id2label = {
    0: "ai",
    1: "cs",
    2: "cv",
    3: "lg",
    4: "math",
    5: "ml",
    6: "phys",
    7: "q-bio",
    8: "stat"
  }

title_text = st.text_input("ENTER TITLE HERE")
summary_text = st.text_area("ENTER SUMMARY HERE")
text = title_text + " " + summary_text


# 1
@st.cache
def load_first_model():
    loaded_tokenizer = AutoTokenizer.from_pretrained("multi_class_model")
    loaded_model = AutoModelForSequenceClassification.from_pretrained("multi_class_model")
    return loaded_tokenizer, loaded_model
tokenizer_1, model_1 = load_first_model()
# loaded_tokenizer = AutoTokenizer.from_pretrained("multi_class_model")
# loaded_model = AutoModelForSequenceClassification.from_pretrained("multi_class_model")

st.markdown("## multi-class classification")

text_input = tokenizer_1(text, padding="max_length", truncation=True, return_tensors='pt')
with torch.no_grad():
    text_res = model_1(**text_input)
text_probs = torch.softmax(text_res.logits, dim=1).cpu().numpy()[0]

order = np.argsort(text_probs)[::-1]
ordered_text_probs = text_probs[order]
idxs = order[np.cumsum(ordered_text_probs) <= 0.95]
st.markdown("Топ-95 классов: " + ", ".join([id2label[i] for i in idxs]))

chart_data = pd.DataFrame(
    text_probs,
    columns=['class probability'])
chart_data["index"] = np.array(list(id2label.values()))
chart_data = chart_data.set_index("index")
st.bar_chart(chart_data)


# 2
@st.cache
def load_first_model():
    loaded_tokenizer = AutoTokenizer.from_pretrained("multi_label_model")
    loaded_model = AutoModelForSequenceClassification.from_pretrained("multi_label_model")
    return loaded_tokenizer, loaded_model
tokenizer_2, model_2 = load_first_model()
# loaded_tokenizer = AutoTokenizer.from_pretrained("multi_label_model")
# loaded_model = AutoModelForSequenceClassification.from_pretrained("multi_label_model")

st.markdown("## multi-label classification")

text_input = tokenizer_2(text, padding="max_length", truncation=True, return_tensors='pt')
with torch.no_grad():
    text_res = model_2(**text_input)

text_probs = torch.sigmoid(torch.Tensor(text_res.logits)).cpu().numpy()[0]

probs = np.stack([text_probs, 1 - text_probs], axis=1)

chart_data = pd.DataFrame(
    probs,
    columns=['belong', "not belong"])
chart_data["index"] = np.array(list(id2label.values()))
chart_data = chart_data.set_index("index")
st.markdown("Probabilities for each class")
st.bar_chart(chart_data)