app.py

import streamlit as st
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModel
import numpy as np
import requests
import os
from PyPDF2 import PdfReader

st.set_page_config(
    page_title="Citation Impact Predictor",
    page_icon="📊",
    layout="wide",
    initial_sidebar_state="expanded",
)

MODEL_NAME      = "allenai/specter2_base"
CHECKPOINT_PATH = "best_model.pt"
N_META          = 5
N_CLASSES       = 4
THRESHOLDS_5Y   = [1.5, 3.5, 5.5]
DEVICE          = "cuda" if torch.cuda.is_available() else "cpu"

CLASS_NAMES  = ["🗑️ Low (0)", "📄 Medium (1)", "📈 High (2)", "🏆 Top (3)"]
CLASS_LABELS = ["Low", "Medium", "High", "Top"]
CLASS_COLORS = ["#e74c3c", "#f39c12", "#3498db", "#2ecc71"]

class CitationPredictor(nn.Module):
    def __init__(self, model_name: str, n_meta: int):
        super().__init__()
        self.encoder   = AutoModel.from_pretrained(model_name)
        self.meta_proj = nn.Sequential(nn.Linear(n_meta, 64), nn.GELU())
        self.layer     = nn.Sequential(
            nn.Linear(768 + 64, 256),
            nn.GELU(),
            nn.LayerNorm(256),
            nn.Dropout(0.2),
        )
        self.head = nn.Linear(256, 1)

    def forward(self, input_ids, attention_mask, meta):
        cls_emb = self.encoder(
            input_ids=input_ids, attention_mask=attention_mask
        ).last_hidden_state[:, 0]
        m_emb = self.meta_proj(meta)
        feat  = self.layer(torch.cat([cls_emb, m_emb], dim=-1))
        return self.head(feat).squeeze(-1)


def to_class(pred: float) -> int:
    if pred < THRESHOLDS_5Y[0]: return 0
    if pred < THRESHOLDS_5Y[1]: return 1
    if pred < THRESHOLDS_5Y[2]: return 2
    return 3

def noise_score(text: str) -> float:
    """Доля букв в тексте — простая метрика осмысленности"""
    letters = sum(c.isalpha() for c in text)
    return letters / max(len(text), 1)

def compute_meta_from_inputs(
    publication_year: int,
    abstract: str,
    title: str,
    author_count: int,
) -> torch.Tensor:
    text = (title + " " + abstract).strip()
    meta = [
        float(publication_year) / 2026,
        float(np.log1p(len(abstract))),
        float(np.log1p(len(title))),
        float(np.log1p(min(author_count, 200))),
        noise_score(text)  # осмысленность текста
    ]
    return torch.tensor([meta], dtype=torch.float)


def fetch_openalex_by_doi(doi: str) -> dict | None:
    clean = doi.strip().replace("https://doi.org/", "").replace("http://doi.org/", "")
    url   = f"https://api.openalex.org/works/doi:{clean}"
    params = {
        "select": "title,abstract_inverted_index,publication_year,authorships",
        "mailto": "demo@example.com",
    }
    try:
        r = requests.get(url, params=params, timeout=15)
        if r.status_code == 200:
            return r.json()
    except Exception:
        pass
    return None


def decode_abstract(inv_idx: dict) -> str:
    if not inv_idx:
        return ""
    words = []
    for word, positions in inv_idx.items():
        for pos in positions:
            words.append((pos, word))
    return " ".join(w for _, w in sorted(words)).strip()


@st.cache_resource(show_spinner="Loading model weights…")
def load_model():
    if not os.path.exists(CHECKPOINT_PATH):
        st.error(
            f"`{CHECKPOINT_PATH}` not found. "
            "Make sure it is uploaded to the Space root directory."
        )
        st.stop()

    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
    model     = CitationPredictor(MODEL_NAME, N_META).to(DEVICE)
    state     = torch.load(CHECKPOINT_PATH, map_location=DEVICE)
    model.load_state_dict(state)
    model.eval()
    return model, tokenizer


model, tokenizer = load_model()

def extract_text_from_pdf(file) -> str:
    try:
        reader = PdfReader(file)
        text = ""
        for page in reader.pages:
            text += page.extract_text() or ""
        return text.strip()
    except Exception:
        return ""

def predict(title: str, abstract: str, meta_tensor: torch.Tensor):
    text = f"{title} [SEP] {abstract}"
    enc  = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        max_length=512,
        padding="max_length",
    )
    with torch.no_grad():
        raw = model(
            enc["input_ids"].to(DEVICE),
            enc["attention_mask"].to(DEVICE),
            meta_tensor.to(DEVICE),
        )
    score           = raw.item()
    pred_class      = to_class(score)
    est_citations   = float(np.expm1(max(score, 0)))   # inverse of log1p
    return pred_class, score, est_citations

st.title("📊 Citation Impact Predictor")
st.markdown("""
### 🤔 Зачем это нужно?
Узнать заранее по названию абстракту, числу авторов, году выхода и наличию открытого доступа стоит ли вообще тратить время на изучение статьи
Мы делим работы на 4 категории:
- 🗑️ **Мусор** — не стоит читать  
- 📄 **Середняк** — можно читать, если это ваша область и более сильных работ сейчас нет 
- 📈 **Сильная работа** — стоит обратить внимание  
- 🏆 **Топ** — читать обязательно  
💡 Это не заменяет экспертную оценку —  
но помогает быстро отфильтровать поток научных работ.
""")
st.divider()

st.sidebar.header("📥 Paper Input")
input_mode = st.sidebar.radio(
    "Input method",
    ["Manual text", "Fetch by DOI", "Upload PDF"],
    help="Choose how to provide the paper.",
)

title    = ""
abstract = ""
pub_year = 2020

# DOI input stays in sidebar; text input moves to main area
if input_mode == "Fetch by DOI":
    doi_input = st.sidebar.text_input("DOI", placeholder="10.1234/example")
else:
    doi_input = ""

st.sidebar.divider()
st.sidebar.header("🔢 Metadata")

pub_year      = st.sidebar.number_input("Publication year", 2000, 2024, 2020)
author_count  = st.sidebar.number_input("Author count", min_value=1, max_value=200, value=3)

# ── Main panel: wide left for input, narrow right for button ──────────────────
col_left, col_right = st.columns([4, 1])

with col_left:
    if input_mode == "Manual text":
        title    = st.text_input("Title", placeholder="e.g. Attention Is All You Need")
        abstract = st.text_area("Abstract", height=250, placeholder="Paste the abstract here…")

    elif input_mode == "Fetch by DOI":
        if doi_input:
            with st.spinner("Fetching metadata from OpenAlex…"):
                paper = fetch_openalex_by_doi(doi_input)
            if paper:
                title    = paper.get("title") or ""
                abstract = decode_abstract(paper.get("abstract_inverted_index") or {})
                pub_year = paper.get("publication_year") or 2020
                st.sidebar.success("✅ Paper found!")
                st.success(f"**{title}**")
                st.markdown(abstract[:800] + ("…" if len(abstract) > 800 else ""))
            else:
                st.error("Could not fetch paper. Check the DOI.")
        else:
            st.info("Enter a DOI in the sidebar to fetch paper metadata.")

    elif input_mode == "Upload PDF":
        uploaded_file = st.file_uploader("Upload PDF", type=["pdf"])

        if uploaded_file is not None:
            with st.spinner("Extracting text from PDF…"):
                text = extract_text_from_pdf(uploaded_file)

            if text:
                lines = text.split("\n")
                title = lines[0][:300]

                abstract = " ".join(lines[1:])[:3000]

                st.success("✅ PDF processed")
                st.markdown(f"**{title}**")
                st.markdown(abstract[:800] + ("…" if len(abstract) > 800 else ""))
            else:
                st.error("Could not extract text from PDF.")

with col_right:
    st.markdown("<br><br>", unsafe_allow_html=True)
    run = st.button("🔍 Predict", use_container_width=True, type="primary")

if run:
    if not title and not abstract:
        st.warning("Please provide at least a title or abstract.")
    else:
        text = (title + " " + abstract).strip()
        meta_tensor = compute_meta_from_inputs(
            publication_year=int(pub_year),
            abstract=abstract,
            title=title,
            author_count=int(author_count)
        )

        with st.spinner("Running inference…"):
            pred_class, raw_score, est_citations = predict(title, abstract, meta_tensor)

        st.divider()
        st.subheader("📊 Prediction Results")

        # Main result badge
        color = CLASS_COLORS[pred_class]
        label = CLASS_LABELS[pred_class]
        st.markdown(
            f"""
            <div style="
                background:{color}22;
                border-left: 6px solid {color};
                padding: 1rem 1.5rem;
                border-radius: 8px;
                margin-bottom: 1rem;
            ">
                <h2 style="margin:0; color:{color}">Class {pred_class} — {label}</h2>
                <p style="margin:0.4rem 0 0; color:#555; font-size:1.1rem;">
                    Estimated citations in the first 5 years:
                    <strong style="font-size:1.3rem;">~{est_citations:.0f}</strong>
                </p>
                <p style="margin:0.15rem 0 0; color:#aaa; font-size:0.85rem;">
                    (raw log-score: {raw_score:.3f} → e^score − 1 = {est_citations:.1f})
                </p>
            </div>
            """,
            unsafe_allow_html=True,
        )

        st.markdown("**Score vs. class thresholds**")
        thresh_cols = st.columns(4)
        boundaries = [0, THRESHOLDS_5Y[0], THRESHOLDS_5Y[1], THRESHOLDS_5Y[2], 8]
        for i, (col, name, color) in enumerate(zip(thresh_cols, CLASS_LABELS, CLASS_COLORS)):
            lo, hi = boundaries[i], boundaries[i + 1]
            active = pred_class == i
            col.markdown(
                f"""<div style="
                    background:{'#2222' if not active else color+'33'};
                    border:2px solid {color if active else '#ccc'};
                    border-radius:6px; padding:0.5rem; text-align:center;">
                    <b style="color:{color}">{name}</b><br>
                    <small style="color:#888">{lo:.1f} – {hi:.1f}</small>
                    {"<br>✅" if active else ""}
                </div>""",
                unsafe_allow_html=True,
            )

        st.divider()
        interpretations = {
            0: "This paper is predicted to receive **very few citations** in its first 5 years — typical of niche, incremental, or low-visibility work.",
            1: "This paper is predicted to receive a **moderate number of citations** — solid work with a reasonable audience.",
            2: "This paper is predicted to receive a **high number of citations** — likely a meaningful contribution to its field.",
            3: "This paper is predicted to be a **top-cited paper** — potentially a landmark contribution with broad impact.",
        }
        st.markdown(f"💡 **Interpretation:** {interpretations[pred_class]}")

st.divider()
st.caption(
    "Model: fine-tuned `allenai/specter2_base` · "
    "Classes defined by log1p(5-year citations) thresholds [1.5, 3.5, 5.5] · "
    "© 2026 Citation Predictor"
)