app.py

import logging

import numpy as np
import streamlit as st
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

MAPPING_FROM_TAG_TO_CATEGORY = {
    "cs.AI": ["Computer Science", "Artificial Intelligence"],
    "cs.CL": ["Computer Science", "Computation and Language"],
    "cs.CV": ["Computer Science", "Computer Vision and Pattern Recognition"],
    "cs.NE": ["Computer Science", "Neural and Evolutionary Computing"],
    "stat.ML": ["Statistics", "Machine Learning"],
    "cs.LG": ["Computer Science", "Machine Learning"],
    "physics.soc-ph": ["Physics", "Physics and Society"],
    "stat.AP": ["Statistics", "Applications"],
    "cs.RO": ["Computer Science", "Robotics"],
    "cs.MA": ["Computer Science", "Multiagent Systems"],
    "math.OC": ["Mathematics", "Optimization and Control"],
    "cs.IR": ["Computer Science", "Information Retrieval"],
    "stat.ME": ["Statistics", "Methodology"],
    "cs.DC": ["Computer Science", "Distributed, Parallel, and Cluster Computing"],
    "stat.CO": ["Statistics", "Computation"],
    "q-bio.NC": ["Quantitative Biology", "Neurons and Cognition"],
    "cs.GT": ["Computer Science", "Computer Science and Game Theory"],
    "cs.MM": ["Computer Science", "Multimedia"],
    "cs.CR": ["Computer Science", "Cryptography and Security"],
    "cs.HC": ["Computer Science", "Human-Computer Interaction"],
    "cs.SD": ["Computer Science", "Sound"],
    "cs.GR": ["Computer Science", "Graphics"],
    "cs.CY": ["Computer Science", "Computers and Society"],
    "math.ST": ["Mathematics", "Statistics Theory"],
    "stat.TH": ["Statistics", "Statistics Theory"],
    "cs.IT": ["Computer Science", "Information Theory"],
    "math.IT": ["Mathematics", "Information Theory"],
    "cs.SI": ["Computer Science", "Social and Information Networks"],
    "cs.DB": ["Computer Science", "Databases"],
    "cs.LO": ["Computer Science", "Logic in Computer Science"],
    "cs.SY": ["Computer Science", "Systems and Control"],
    "q-bio.QM": ["Quantitative Biology", "Quantitative Methods"],
    "cs.DS": ["Computer Science", "Data Structures and Algorithms"],
    "cs.NA": ["Computer Science", "Numerical Analysis"],
    "cs.CE": ["Computer Science", "Computational Engineering, Finance, and Science"],
}

MODEL_PATH = "minemile/arxiv-tag-classifier"
MODEL_MAX_LENGTH = 512
CUM_PROB_THRESHOLD = 0.95


@st.cache_resource
def load_model_and_tokenizer(model_path):
    logger.info("Loading model and tokenizer from %s", model_path)
    try:
        tokenizer = AutoTokenizer.from_pretrained(model_path)
        model = AutoModelForSequenceClassification.from_pretrained(
            model_path, device_map="cpu"
        )
        model.eval()
        logger.info("Model and tokenizer loaded successfully.")
        return model, tokenizer
    except Exception as e:
        st.error(f"An error occurred during prediction: {e}")
        return None, None


def inference(model, tokenizer, title, summary=None):
    if not model or not tokenizer:
        st.error("Model or tokenizer not loaded. Cannot predict.")
        return None, None

    combined_text = title
    if summary is not None:
        combined_text += " " + summary

    logger.info("Predicting for text: %s", combined_text)

    try:
        tokenized_inputs = tokenizer(
            combined_text,
            return_tensors="pt",
            truncation=True,
            padding=True,
            max_length=MODEL_MAX_LENGTH,
        )

        with torch.no_grad():
            logits = model(**tokenized_inputs).logits
            probabilities = torch.softmax(logits, dim=-1).squeeze().cpu().numpy()
        return probabilities

    except Exception as e:
        st.error(f"An error occurred during prediction: {e}")
        return None, None


def prepare_output(probabilities, id2label):
    top_indices = np.argsort(probabilities)[::-1]
    cum_sum = 0.0
    top_tags = []
    top_category = []
    top_subcategory = []
    top_probas = []
    for indx in top_indices:
        tag = id2label[indx]
        if tag not in MAPPING_FROM_TAG_TO_CATEGORY:
            top_tags.append(f"{tag}")
            top_category.append("Unknown")
            top_subcategory.append("Unknown")
            logger.warning("Tag %s not found in mapping from tag to category.", tag)
        else:
            top_tags.append(f"{tag}")
            top_category.append(MAPPING_FROM_TAG_TO_CATEGORY[tag][0])
            top_subcategory.append(MAPPING_FROM_TAG_TO_CATEGORY[tag][1])
        top_probas.append(f"{probabilities[indx]*100:.2f}%")
        cum_sum += probabilities[indx]
        if cum_sum >= CUM_PROB_THRESHOLD or len(top_tags) >= 5:
            break
    return {
        "Tag": top_tags,
        "Category": top_category,
        "Subcategory": top_subcategory,
        "Probability": top_probas,
    }


def main():
    st.set_page_config(page_title="ArXiv Category Tag Classifier", layout="wide")
    st.title("ArXiv Category Tag Classifier")
    with st.spinner("Loading model and tokenizer..."):
        model, tokenizer = load_model_and_tokenizer(MODEL_PATH)
    if model is None or tokenizer is None:
        st.error("Failed to load model/tokenizer.")
        return

    st.markdown(
        f"Enter the title (required) and summary (abstract) of an ArXiv paper to predict its "
        f"ArXiv category using a transformer model. There are {len(model.config.id2label)} available categories."
    )
    st.divider()
    col1, col2 = st.columns(2)
    with col1:
        title = st.text_input(
            "Title",
            placeholder="Enter the title of the paper (Required)",
        )
    with col2:
        paper_summary = st.text_area(
            "Paper Summary (Optional):",
            placeholder="Paste the paper's abstract here. It can increase the accuracy of the prediction.",
            height=180,
        )

    predict_button = st.button("Predict Category", type="primary")

    st.markdown("---")

    if predict_button:
        if not title:
            st.error("Title of the paper is required!")
        else:
            with st.spinner("Predicting category..."):
                class_probabilities = inference(model, tokenizer, title, paper_summary)

            if class_probabilities is not None:
                st.subheader("Top Tags Predictions:")
                output = prepare_output(class_probabilities, model.config.id2label)
                logger.info("Output: %s", output)
                st.dataframe(output, use_container_width=True)
    st.markdown("---")


if __name__ == "__main__":
    main()