app.py

import streamlit as st
import torch
import os
import torch.nn as nn
from safetensors import safe_open
from transformers import BertPreTrainedModel, BertModel, BertTokenizer, BertConfig

st.set_page_config(page_title="Paper Classifier", layout="wide")

class BERTClass(BertPreTrainedModel):
    def __init__(self, config, p=0.3):
        super().__init__(config)
        self.bert = BertModel(config)
        self.dropout = nn.Dropout(p)
        self.linear = nn.Linear(config.hidden_size, config.num_labels)
        self.init_weights()

    def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, labels=None):
        outputs = self.bert(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            return_dict=True
        )
        pooled_output = outputs.pooler_output
        pooled_output = self.dropout(pooled_output)
        logits = self.linear(pooled_output)
        loss = None
        if labels is not None:
            loss_fct = nn.BCEWithLogitsLoss()
            loss = loss_fct(logits, labels)
        return {"loss": loss, "logits": logits}

MODEL_PATH =  "."
LABELS = ['astro-ph', 'cond-mat', 'cs', 'eess', 'gr-qc',
        'hep-ex', 'hep-lat', 'hep-ph', 'hep-th', 'math', 'math-ph', 'nlin',
        'nucl-ex', 'nucl-th', 'physics', 'q-bio', 'quant-ph', 'stat']
MAX_LEN = 512

@st.cache_resource
def load_model():
    try:
        config = BertConfig.from_pretrained("bert-base-cased")
        config.num_labels = len(LABELS)
        model = BERTClass(config)

        with safe_open(f"{MODEL_PATH}/model.safetensors", framework="pt") as f:
            state_dict = {key: f.get_tensor(key) for key in f.keys()}

        model.load_state_dict(state_dict)
        tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
        return model.eval(), tokenizer
    
    except Exception as e:
        st.error(f"Model loading failed: {str(e)}")
        st.stop()


@st.cache_data
def predict(title, abstract):
    if not title.strip() and not abstract.strip():
        raise ValueError("Bro, do you want me to guess?) Give me at least the title!")
    
    text = f"{title.strip()}. {abstract.strip()}".strip()
    if len(text) < 10:
        raise ValueError("Too short text to say anything sensible")

    device = next(model.parameters()).device
    inputs = tokenizer.encode_plus(
        text,
        max_length=MAX_LEN,
        padding="max_length",
        truncation=True,
        return_tensors="pt"
    ).to(device)
    
    with torch.no_grad():
        outputs = model(**inputs)
    
    logits = outputs['logits']
    probs = torch.sigmoid(logits).cpu().numpy()[0]
    return {label: float(probs[i]) for i, label in enumerate(LABELS)}

model, tokenizer = load_model()

with st.sidebar:
    st.header("Display Settings")
    display_mode = st.radio(
        "Result filtering mode",
        ["Top-k categories", "Top-% confidence"],
        index=0
    )
    
    if display_mode == "Top-k categories":
        top_k = st.slider(
            "Number of categories to show",
            min_value=1,
            max_value=10,
            value=3,
            help="Select how many top categories to display"
        )
    else:
        selected_percent = st.selectbox(
            "Confidence threshold",
            ["50%", "75%", "95%"],
            index=2,
            help="Display categories until reaching this cumulative confidence"
        )

st.title("📄 Academic Paper Classifier")

with st.form("input_form"):
    title = st.text_input("Paper Title", placeholder="Enter paper title...")
    abstract = st.text_area("Abstract", placeholder="Paste paper abstract here...", height=200)
    submitted = st.form_submit_button("Classify")

if submitted:
    with st.spinner("Analyzing paper..."):
        try:
            full_predictions = predict(title, abstract)
            sorted_preds = sorted(full_predictions.items(), 
                                key=lambda x: x[1], 
                                reverse=True)

            if display_mode == "Top-k categories":
                filtered = dict(sorted_preds[:top_k])
            else:
                threshold = {"50%": 0.5, "75%": 0.75, "95%": 0.95}[selected_percent]
                total = sum(score for _, score in sorted_preds)
                cumulative = 0
                filtered = {}
                
                for label, score in sorted_preds:
                    cumulative += score
                    filtered[label] = score
                    if cumulative >= threshold: 
                        break
                    if len(filtered) >= 10:
                        break

            if not filtered:
                st.warning("No categories meet the selected criteria")
            else:
                top_class = max(filtered, key=filtered.get)
                st.success(f"Most likely category: **{top_class}**")
                
                st.subheader("Category Confidence Scores:")
                total_shown = sum(filtered.values())
                
                for label, score in filtered.items():
                    relative_score = score / total_shown
                    st.progress(
                        relative_score,
                        text=f"{label}: {score:.1%}"
                    )
                
                st.caption(f"Coverage: {sum(filtered.values()):.1%} of total confidence")

        except Exception as e:
            st.error(f"Error: {str(e)}")

with st.sidebar:
    st.header("About")
    st.markdown(f"""
    This tool predicts the arxiv tag of research papers by their title and abstarct via fine-tuned BERT.
    - Enter title and abstract
    - Enjoy the magnificent classification results
    """)