File size: 9,745 Bytes
ba415d6 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 |
"""
PosterSentry — Multimodal Scientific Poster Classifier
=======================================================
Architecture:
┌──────────┐ ┌──────────────┐ ┌───────────────┐
│ PDF text │ │ PDF → image │ │ PDF structure │
└────┬─────┘ └──────┬───────┘ └───────┬───────┘
│ │ │
model2vec 15 visual 15 structural
→ 512-d emb features features
│ │ │
└────────┬────────┴────────────────────┘
│
concat → 542-d input
│
LogisticRegression
│
poster / non_poster
Single linear classifier on the concatenated feature vector.
Same paradigm as PubGuard — lightweight, CPU-only, fast.
"""
import logging
import time
from pathlib import Path
from typing import Any, Dict, List, Optional
import numpy as np
from .features import (
VisualFeatureExtractor,
PDFStructuralExtractor,
N_VISUAL_FEATURES,
N_STRUCTURAL_FEATURES,
)
logger = logging.getLogger(__name__)
class PosterSentry:
"""
Multimodal poster classifier.
Combines:
- model2vec text embedding (512-d)
- 15 visual features (color, edge, FFT, whitespace)
- 15 structural features (page geometry, fonts, text blocks)
into a single 542-d feature vector for logistic regression.
"""
def __init__(
self,
model_name: str = "minishlab/potion-base-32M",
models_dir: Optional[Path] = None,
):
self.model_name = model_name
self.models_dir = models_dir or self._default_models_dir()
self.models_dir = Path(self.models_dir)
self.text_model = None
self.W: Optional[np.ndarray] = None
self.b: Optional[np.ndarray] = None
self.scaler_mean: Optional[np.ndarray] = None
self.scaler_scale: Optional[np.ndarray] = None
self.labels = ["non_poster", "poster"]
self.visual_extractor = VisualFeatureExtractor()
self.structural_extractor = PDFStructuralExtractor()
self._initialized = False
@staticmethod
def _default_models_dir() -> Path:
import os
if env := os.environ.get("POSTER_SENTRY_MODELS_DIR"):
return Path(env)
home = Path.home() / ".poster_sentry" / "models"
home.mkdir(parents=True, exist_ok=True)
return home
# ── Initialization ──────────────────────────────────────────
def initialize(self) -> bool:
if self._initialized:
return True
logger.info("Initializing PosterSentry...")
t0 = time.time()
self._load_text_model()
self._load_head()
self._initialized = True
logger.info(f"PosterSentry initialized in {time.time()-t0:.1f}s")
return True
def _load_text_model(self):
from model2vec import StaticModel
cache = self.models_dir / "poster-sentry-embedding"
if cache.exists():
self.text_model = StaticModel.from_pretrained(str(cache))
else:
self.text_model = StaticModel.from_pretrained(self.model_name)
cache.parent.mkdir(parents=True, exist_ok=True)
self.text_model.save_pretrained(str(cache))
def _load_head(self):
path = self.models_dir / "poster_sentry_head.npz"
if path.exists():
data = np.load(path, allow_pickle=True)
self.W = data["W"]
self.b = data["b"]
if "labels" in data:
self.labels = list(data["labels"])
if "scaler_mean" in data and "scaler_scale" in data:
self.scaler_mean = data["scaler_mean"]
self.scaler_scale = data["scaler_scale"]
logger.info(f" Loaded classifier head: {path}")
else:
logger.warning(f" Head not found: {path} — run training first")
def save_head(self, path: Optional[Path] = None):
path = path or (self.models_dir / "poster_sentry_head.npz")
path.parent.mkdir(parents=True, exist_ok=True)
np.savez(path, W=self.W, b=self.b, labels=np.array(self.labels))
# ── Feature extraction ──────────────────────────────────────
def extract_text(self, pdf_path: str, max_chars: int = 4000) -> str:
"""Extract and clean text from first page of PDF."""
try:
import fitz
doc = fitz.open(pdf_path)
if len(doc) == 0:
doc.close()
return ""
text = doc[0].get_text()
doc.close()
# Basic cleanup
import re
text = re.sub(r"\s+", " ", text).strip()
return text[:max_chars]
except Exception:
return ""
def embed_texts(self, texts: List[str]) -> np.ndarray:
"""Encode texts with model2vec, L2-normalize."""
embeddings = self.text_model.encode(texts)
norms = np.linalg.norm(embeddings, axis=1, keepdims=True)
norms = np.where(norms == 0, 1, norms)
return (embeddings / norms).astype("float32")
def build_feature_vector(
self,
text_emb: np.ndarray,
visual_feats: np.ndarray,
structural_feats: np.ndarray,
) -> np.ndarray:
"""Concatenate all features: [512 text + 15 visual + 15 structural] = 542."""
return np.concatenate([text_emb, visual_feats, structural_feats])
# ── Inference ───────────────────────────────────────────────
def classify(self, pdf_path: str) -> Dict[str, Any]:
"""Classify a single PDF as poster or non-poster."""
if not self._initialized:
self.initialize()
return self.classify_batch([pdf_path])[0]
def classify_batch(self, pdf_paths: List[str]) -> List[Dict[str, Any]]:
"""Classify a batch of PDFs."""
if not self._initialized:
self.initialize()
texts = []
visual_vecs = []
structural_vecs = []
for p in pdf_paths:
texts.append(self.extract_text(p))
img = self.visual_extractor.pdf_to_image(p)
if img is not None:
vf = self.visual_extractor.extract(img)
else:
vf = {n: 0.0 for n in self.visual_extractor.FEATURE_NAMES}
visual_vecs.append(self.visual_extractor.to_vector(vf))
sf = self.structural_extractor.extract(p)
structural_vecs.append(self.structural_extractor.to_vector(sf))
# Embed text
text_embs = self.embed_texts(texts)
visual_arr = np.array(visual_vecs, dtype="float32")
struct_arr = np.array(structural_vecs, dtype="float32")
# Concatenate
X = np.concatenate([text_embs, visual_arr, struct_arr], axis=1)
# Scale features (critical for balanced text vs structural signal)
if self.scaler_mean is not None and self.scaler_scale is not None:
X = (X - self.scaler_mean) / np.where(self.scaler_scale == 0, 1, self.scaler_scale)
# Predict
if self.W is None:
return [{"path": p, "is_poster": False, "confidence": 0.0,
"error": "Model not trained"} for p in pdf_paths]
logits = X @ self.W + self.b
e = np.exp(logits - logits.max(axis=-1, keepdims=True))
probs = e / e.sum(axis=-1, keepdims=True)
results = []
for i, p in enumerate(pdf_paths):
poster_prob = float(probs[i, 1])
results.append({
"path": str(p),
"is_poster": poster_prob > 0.5,
"confidence": round(poster_prob, 4),
"text_score": round(float(probs[i, 1]), 4),
})
return results
# ── Text-only classification (for PubGuard integration) ─────
def classify_text(self, text: str) -> Dict[str, Any]:
"""Classify from text alone (no PDF needed). Used by PubGuard."""
return self.classify_texts([text])[0]
def classify_texts(self, texts: List[str]) -> List[Dict[str, Any]]:
"""Classify from text alone (batch)."""
if not self._initialized:
self.initialize()
if self.W is None:
return [{"is_poster": False, "confidence": 0.0}] * len(texts)
text_embs = self.embed_texts(texts)
# Zero-fill visual and structural features
zeros_visual = np.zeros((len(texts), N_VISUAL_FEATURES), dtype="float32")
zeros_struct = np.zeros((len(texts), N_STRUCTURAL_FEATURES), dtype="float32")
X = np.concatenate([text_embs, zeros_visual, zeros_struct], axis=1)
# Scale
if self.scaler_mean is not None and self.scaler_scale is not None:
X = (X - self.scaler_mean) / np.where(self.scaler_scale == 0, 1, self.scaler_scale)
logits = X @ self.W + self.b
e = np.exp(logits - logits.max(axis=-1, keepdims=True))
probs = e / e.sum(axis=-1, keepdims=True)
return [{"is_poster": float(probs[i, 1]) > 0.5,
"confidence": round(float(probs[i, 1]), 4)}
for i in range(len(texts))]
|