File size: 14,767 Bytes
e37bd2f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 | """Build a stratified evaluation sample with full observables per figure.
Sampling strategy:
- Select DOMAIN_COUNT Locarno chapters (by patent count)
- Within each chapter, sample N_PER_DOMAIN patents stratified by text richness
- Within each patent, include all figures
- Compute all observables: text features, structural features, image complexity
Produces: data/sample/eval_sample.parquet (figures)
data/sample/eval_sample_patents.parquet (per-patent aggregates)
Usage:
python scripts/eval/build_sample.py \
--enriched data/enriched/enriched_2022.parquet \
--images /tmp/patent_sample/2022 \
--out data/sample \
--domains 8 \
--per-domain 100
"""
import argparse
import re
from pathlib import Path
import numpy as np
import pandas as pd
from PIL import Image
from tqdm import tqdm
# ── Locarno chapter labels ────────────────────────────────────────────────────
CHAPTER_LABELS = {
'D14': 'Screens/electronics',
'D24': 'Medical',
'D12': 'Construction',
'D23': 'Fluid/sanitation',
'D21': 'Games/sports',
'D13': 'Transport',
'D26': 'Lighting',
'D29': 'Fire/accident prevention',
'D32': 'Graphic symbols/logos',
'D15': 'Machines',
'D10': 'Clocks/watches',
'D28': 'Pharma/cosmetics',
'D11': 'Ornamental articles',
'D27': 'Tobacco/recreation',
'D8': 'Tools/hardware',
'D9': 'Packaging',
'D16': 'Photo/optical',
}
# ── viewpoint parsing ─────────────────────────────────────────────────────────
VP_BUCKETS = [
('perspective', lambda v: 'perspective' in v),
('front_elev', lambda v: 'front elev' in v or v in ('front elevation','front plan','front view')),
('rear_elev', lambda v: 'rear elev' in v or 'rear elevation' in v),
('side_elev', lambda v: 'side elev' in v or 'side elevation' in v),
('top_plan', lambda v: 'top plan' in v or v == 'top view'),
('bottom_plan', lambda v: 'bottom plan' in v or 'bottom' in v),
('cross_section', lambda v: 'cross' in v or 'section' in v),
('enlarged', lambda v: 'enlarg' in v or 'detail' in v),
('exploded', lambda v: 'explod' in v),
('reference', lambda v: 'reference' in v),
('unknown', lambda v: True),
]
DIFFICULTY = {
'perspective': 'baseline',
'front_elev': 'easy',
'rear_elev': 'easy',
'side_elev': 'easy',
'top_plan': 'medium',
'bottom_plan': 'medium',
'enlarged': 'hard',
'cross_section': 'very_hard',
'exploded': 'hard',
'reference': 'baseline',
'unknown': 'unknown',
}
def parse_vp(desc: str, fig_num: int) -> str:
pat = re.compile(
rf"FIG\.\s*{fig_num + 1}\s+is\s+(?:a\s+|an\s+)?(.{{5,80}}?)\s*(?:view|thereof|;|\n|$)",
re.IGNORECASE,
)
m = pat.search(desc or "")
return m.group(1).strip().lower() if m else ""
def bucket_vp(vp: str) -> str:
for name, fn in VP_BUCKETS:
if fn(vp):
return name
return "unknown"
# ── image observables ─────────────────────────────────────────────────────────
def find_image(images_dir: Path, image_filename: str) -> Path | None:
parts = image_filename.split("-D0")
if len(parts) < 2:
return None
p = images_dir / parts[0] / image_filename
return p if p.exists() else None
def image_observables(img_path: Path) -> dict:
"""Compute pixel-level features from a bilevel patent TIF."""
try:
img = Image.open(img_path)
arr = np.array(img) # mode 1: True=white, False=black
h, w = arr.shape
total = h * w
ink_frac = (~arr).sum() / total
# Edge transitions (white→black per row, sampled)
transitions = int(sum(
np.sum(row[1:] < row[:-1])
for row in arr[::4] # sample every 4th row
))
# Perceptual hash proxy: 8×8 mean block comparison
from PIL import Image as PILImage
small = img.resize((32, 32), PILImage.LANCZOS).convert("L")
small_arr = np.array(small)
phash_var = float(small_arr.var())
return {
"ink_frac": float(ink_frac),
"edge_transitions": transitions,
"img_width": w,
"img_height": h,
"img_aspect": float(w / h) if h > 0 else 1.0,
"phash_var": phash_var, # higher = more detail/complexity
}
except Exception:
return {
"ink_frac": None,
"edge_transitions": None,
"img_width": None,
"img_height": None,
"img_aspect": None,
"phash_var": None,
}
# ── text observables ──────────────────────────────────────────────────────────
def text_observables(row: pd.Series) -> dict:
draw = str(row.get("drawing_description") or "")
detail = str(row.get("detailed_description") or "")
claims = str(row.get("claims") or "")
summary = str(row.get("brief_summary") or "")
caption = str(row.get("caption") or "")
# Count FIG. references in draw_desc
n_fig_refs = len(re.findall(r"FIG\.\s*\d+", draw, re.IGNORECASE))
return {
"draw_desc_chars": len(draw),
"draw_desc_words": len(draw.split()),
"draw_desc_fig_refs": n_fig_refs,
"detail_desc_chars": len(detail),
"claims_chars": len(claims),
"summary_chars": len(summary),
"caption_chars": len(caption),
# Text richness quartile (filled in at patent level)
}
# ── main ──────────────────────────────────────────────────────────────────────
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--enriched", default="data/enriched/enriched_2022.parquet")
parser.add_argument("--images", default="/tmp/patent_sample/2022")
parser.add_argument("--out", default="data/sample")
parser.add_argument("--domains", type=int, default=8,
help="Number of top Locarno chapters to include")
parser.add_argument("--per-domain", type=int, default=100,
help="Max patents per domain")
parser.add_argument("--seed", type=int, default=42)
args = parser.parse_args()
rng = np.random.default_rng(args.seed)
images_dir = Path(args.images)
out_dir = Path(args.out)
out_dir.mkdir(parents=True, exist_ok=True)
print("Loading enriched data...")
df = pd.read_parquet(args.enriched)
# Parse chapter and viewpoint
df["locarno_chapter"] = (
df["class"].astype(str)
.str.extract(r"^(D\s*\d{1,2})", expand=False)
.str.replace(" ", "", regex=False)
.fillna("other")
)
df["vp_raw"] = df.apply(
lambda r: parse_vp(r.get("drawing_description", ""), r["figure_number"]),
axis=1,
)
df["vp_bucket"] = df["vp_raw"].apply(bucket_vp)
df["difficulty"] = df["vp_bucket"].map(DIFFICULTY)
# Text features at figure level
df["draw_desc_chars"] = df["drawing_description"].fillna("").str.len()
df["draw_desc_words"] = df["drawing_description"].fillna("").str.split().str.len()
# Image path
df["img_path"] = df["image_filename"].apply(lambda fn: find_image(images_dir, fn))
# Per-patent aggregates for richness quartile
patent_agg = df.groupby("patent_id").agg(
draw_len=("draw_desc_chars", "first"),
n_figs=("figure_id", "count"),
n_vp_types=("vp_bucket", "nunique"),
has_perspective=("vp_bucket", lambda x: "perspective" in x.values),
has_front=("vp_bucket", lambda x: "front_elev" in x.values),
has_cross=("vp_bucket", lambda x: "cross_section" in x.values),
parse_rate=("vp_bucket", lambda x: (x != "unknown").mean()),
chapter=("locarno_chapter", "first"),
).reset_index()
patent_agg["text_richness_q"] = pd.qcut(
patent_agg["draw_len"], 4,
labels=["sparse", "moderate", "rich", "very_rich"]
).astype(str)
# Select top domains
top_chapters = (
patent_agg[patent_agg["chapter"] != "other"]
.groupby("chapter")["patent_id"].count()
.sort_values(ascending=False)
.head(args.domains)
.index.tolist()
)
print(f"\nSelected domains: {top_chapters}")
# Stratified sample: within each domain, stratify by text richness
sampled_pids = []
for chapter in top_chapters:
chap_patents = patent_agg[patent_agg["chapter"] == chapter]
n = min(args.per_domain, len(chap_patents))
# Stratified by text richness (equal from each quartile if possible)
per_q = n // 4
chunk = []
for q in ["sparse", "moderate", "rich", "very_rich"]:
pool = chap_patents[chap_patents["text_richness_q"] == q]["patent_id"].tolist()
rng.shuffle(pool)
chunk.extend(pool[: per_q])
# Top up to n from the full pool
remaining = set(chap_patents["patent_id"]) - set(chunk)
remaining = list(remaining)
rng.shuffle(remaining)
chunk.extend(remaining[: n - len(chunk)])
sampled_pids.extend(chunk[:n])
label = CHAPTER_LABELS.get(chapter, chapter)
print(f" {chapter} ({label}): {len(chunk[:n])} patents sampled")
print(f"\nTotal patents sampled: {len(sampled_pids)}")
# Filter to sampled patents
sample_df = df[df["patent_id"].isin(set(sampled_pids))].copy()
# Merge patent-level features
sample_df = sample_df.merge(
patent_agg[["patent_id", "text_richness_q", "n_vp_types",
"has_perspective", "has_front", "has_cross", "parse_rate"]],
on="patent_id", how="left"
)
sample_df["domain_label"] = sample_df["locarno_chapter"].map(CHAPTER_LABELS).fillna(sample_df["locarno_chapter"])
print(f"Total figures in sample: {len(sample_df):,}")
print(f"Images available: {sample_df['img_path'].notna().sum():,}")
# Compute image observables on all available images
print("\nComputing image observables...")
img_obs_list = []
for _, row in tqdm(sample_df.iterrows(), total=len(sample_df)):
p = row["img_path"]
obs = image_observables(p) if p is not None else {}
obs["figure_id"] = row["figure_id"]
img_obs_list.append(obs)
img_obs_df = pd.DataFrame(img_obs_list)
sample_df = sample_df.merge(img_obs_df, on="figure_id", how="left")
# Summary
print("\n=== SAMPLE SUMMARY ===")
print(f"{'Domain':<6} {'Label':<22} {'Patents':>8} {'Figs':>8}")
for chap in top_chapters:
g = sample_df[sample_df["locarno_chapter"] == chap]
label = CHAPTER_LABELS.get(chap, chap)
print(f" {chap:<6} {label:<22} {g['patent_id'].nunique():>8} {len(g):>8}")
print()
print("=== VIEWPOINT DISTRIBUTION IN SAMPLE ===")
vp_dist = sample_df["vp_bucket"].value_counts()
for vp, n in vp_dist.items():
print(f" {vp:<16} {n:>6,} ({n/len(sample_df):.1%})")
print()
print("=== TEXT RICHNESS IN SAMPLE ===")
for q, g in sample_df.groupby("text_richness_q", observed=True):
print(f" {q:<12} {g['patent_id'].nunique():>5} patents "
f"draw_len median: {g['draw_desc_chars'].median():.0f} chars")
# Save
cols_to_save = [
"figure_id", "patent_id", "figure_number", "n_figures_in_patent",
"sibling_figure_ids", "patent_title", "caption",
"drawing_description", "detailed_description", "brief_summary", "claims",
"vp_raw", "vp_bucket", "difficulty",
"locarno_chapter", "domain_label", "text_richness_q",
"draw_desc_chars", "draw_desc_words",
"detail_desc_chars" if "detail_desc_chars" in sample_df.columns else None,
"claims_chars" if "claims_chars" in sample_df.columns else None,
"n_vp_types", "has_perspective", "has_front", "has_cross", "parse_rate",
"ink_frac", "edge_transitions", "img_width", "img_height",
"img_aspect", "phash_var",
"image_filename", "year",
]
cols_to_save = [c for c in cols_to_save if c and c in sample_df.columns]
sample_df[cols_to_save].to_parquet(out_dir / "eval_sample.parquet", index=False)
# Per-patent summary
patent_summary = sample_df.groupby("patent_id").agg(
n_figs=("figure_id", "count"),
locarno_chapter=("locarno_chapter", "first"),
domain_label=("domain_label", "first"),
text_richness_q=("text_richness_q", "first"),
draw_desc_chars=("draw_desc_chars", "first"),
n_vp_types=("n_vp_types", "first"),
has_perspective=("has_perspective", "first"),
has_front=("has_front", "first"),
has_cross=("has_cross", "first"),
parse_rate=("parse_rate", "first"),
median_ink_frac=("ink_frac", "median"),
median_edge_transitions=("edge_transitions", "median"),
median_phash_var=("phash_var", "median"),
patent_title=("patent_title", "first"),
).reset_index()
patent_summary.to_parquet(out_dir / "eval_sample_patents.parquet", index=False)
print(f"\nSaved:")
print(f" {out_dir}/eval_sample.parquet ({len(sample_df):,} figures)")
print(f" {out_dir}/eval_sample_patents.parquet ({len(patent_summary):,} patents)")
print()
print("=== OBSERVABLES CAPTURED ===")
obs_fields = {
"TEXT (draw_desc)": ["draw_desc_chars", "draw_desc_words", "text_richness_q"],
"TEXT (missing — need re-download)": ["detail_desc_chars", "claims_chars", "summary_chars"],
"STRUCTURAL": ["n_figures_in_patent", "n_vp_types", "has_perspective", "has_front", "has_cross", "parse_rate"],
"IMAGE": ["ink_frac", "edge_transitions", "img_aspect", "phash_var"],
"METADATA": ["locarno_chapter", "domain_label", "year"],
}
for group, fields in obs_fields.items():
available = [f for f in fields if f in sample_df.columns and sample_df[f].notna().any()]
missing = [f for f in fields if f not in available]
print(f" {group}:")
if available:
print(f" ✓ {', '.join(available)}")
if missing:
print(f" ✗ {', '.join(missing)} (not in current data)")
if __name__ == "__main__":
main()
|