File size: 10,149 Bytes
d126496 | 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 | """Figure completion retrieval benchmark.
Leave-one-out: for each patent, mask one view (by default the top plan view,
or the hardest available). Given embeddings of N-1 sibling views as context,
retrieve the correct masked view from a pool of 100 candidates.
Baselines:
random — chance (1%)
single — embed only the perspective view, retrieve
multi — average CLIP embeddings of all N-1 context views, retrieve
vlm — (future) use VLM to describe missing view, embed description
Usage:
python scripts/eval/retrieval_eval.py \
--embeddings data/embeddings/embeddings_2022_vitl14.parquet \
--enriched data/enriched/enriched_2022.parquet \
--n 500 \
--pool-size 100 \
--out results/retrieval_eval.json
"""
import argparse
import json
import random
import re
from collections import defaultdict
from pathlib import Path
import faiss
import numpy as np
import pandas as pd
from tqdm import tqdm
# ── viewpoint helpers ─────────────────────────────────────────────────────────
def parse_viewpoint(drawing_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(drawing_desc or "")
return m.group(1).strip().lower() if m else ""
TARGET_PRIORITY = [
# (label_fragment, difficulty)
("cross-sectional", "very_hard"),
("cross section", "very_hard"),
("enlarged", "hard"),
("detail", "hard"),
("top plan", "hard"),
("bottom plan", "medium"),
("rear elevation", "medium"),
("rear elev", "medium"),
("side elev", "easy"),
("front elev", "easy"),
("perspective", "baseline"),
]
def pick_target_view(viewpoints: list[str]) -> tuple[int, str]:
"""Pick the highest-priority masking target. Returns (index, difficulty)."""
for frag, difficulty in TARGET_PRIORITY:
for i, vp in enumerate(viewpoints):
if frag in vp:
return i, difficulty
# Fallback: pick any non-first view
return 1 if len(viewpoints) > 1 else 0, "unknown"
# ── retrieval ─────────────────────────────────────────────────────────────────
def build_faiss_index(vectors: np.ndarray) -> faiss.IndexFlatIP:
"""Build an inner-product FAISS index (cosine sim on L2-normed vectors)."""
dim = vectors.shape[1]
index = faiss.IndexFlatIP(dim)
index.add(vectors.astype(np.float32))
return index
def retrieve(
query_vec: np.ndarray, # (dim,)
candidate_indices: list[int], # indices into the full embedding matrix
all_vectors: np.ndarray,
correct_idx: int, # index into candidate_indices
) -> dict:
"""Score retrieval: rank correct candidate among candidates by cosine sim."""
cand_vecs = all_vectors[candidate_indices].astype(np.float32)
q = query_vec.astype(np.float32).reshape(1, -1)
sims = (cand_vecs @ q.T).squeeze()
ranks = np.argsort(-sims) # descending
rank_of_correct = int(np.where(ranks == correct_idx)[0][0]) + 1 # 1-indexed
return {
"rank": rank_of_correct,
"r1": int(rank_of_correct <= 1),
"r5": int(rank_of_correct <= 5),
"r10": int(rank_of_correct <= 10),
"sim_correct": float(sims[correct_idx]),
"sim_top1": float(sims[ranks[0]]),
}
# ── main eval ─────────────────────────────────────────────────────────────────
def run_eval(
embeddings_path: str,
enriched_path: str,
n: int,
pool_size: int,
out_path: str,
seed: int = 42,
):
rng = random.Random(seed)
print("Loading embeddings...")
emb_df = pd.read_parquet(embeddings_path)
fig_id_to_idx = {fid: i for i, fid in enumerate(emb_df["figure_id"])}
all_vecs = np.vstack(emb_df["embedding"].tolist()).astype(np.float32)
# Ensure unit norm for cosine sim
norms = np.linalg.norm(all_vecs, axis=1, keepdims=True)
all_vecs = all_vecs / np.maximum(norms, 1e-8)
print(f"Embeddings: {all_vecs.shape}")
print("Loading enriched metadata...")
df = pd.read_parquet(enriched_path)
df["viewpoint_parsed"] = df.apply(
lambda r: parse_viewpoint(r.get("drawing_description", ""), r["figure_number"]),
axis=1,
)
# Keep only figures with embeddings
df = df[df["figure_id"].isin(fig_id_to_idx)].copy()
df["_vec_idx"] = df["figure_id"].map(fig_id_to_idx)
# Group by patent
patent_groups = {
pid: g.sort_values("figure_number")
for pid, g in df.groupby("patent_id")
if len(g) >= 3
}
# Build Locarno-class → figure_id list for distractor sampling
class_to_fids = defaultdict(list)
for _, row in df.iterrows():
cls = row.get("class") or row.get("locarno_class") or "unknown"
class_to_fids[str(cls)].append(row["figure_id"])
# Sample eligible patents
all_pids = list(patent_groups.keys())
rng.shuffle(all_pids)
eval_pids = all_pids[:n]
print(f"Evaluating {len(eval_pids)} patents (pool_size={pool_size})")
by_difficulty = defaultdict(lambda: {"r1": 0, "r5": 0, "r10": 0, "n": 0})
results = []
for pid in tqdm(eval_pids):
group = patent_groups[pid]
fids = group["figure_id"].tolist()
vps = group["viewpoint_parsed"].tolist()
vec_idxs = group["_vec_idx"].tolist()
cls = str(group["class"].iloc[0] if "class" in group.columns else "unknown")
# Pick target view to mask
target_pos, difficulty = pick_target_view(vps)
target_fid = fids[target_pos]
target_vec_idx = vec_idxs[target_pos]
context_idxs = [vi for i, vi in enumerate(vec_idxs) if i != target_pos]
if not context_idxs:
continue
# Build query: average of context embeddings
query_vec = all_vecs[context_idxs].mean(axis=0)
query_vec /= max(np.linalg.norm(query_vec), 1e-8)
# Build candidate pool: target + (pool_size - 1) distractors from same class
distractors_pool = [
f for f in class_to_fids.get(cls, [])
if f not in set(fids) and f in fig_id_to_idx
]
rng.shuffle(distractors_pool)
distractors = distractors_pool[: pool_size - 1]
if len(distractors) < pool_size - 1:
# Fill from any other patent
other_fids = [
f for f in df["figure_id"].tolist()
if f not in set(fids) and f not in set(distractors) and f in fig_id_to_idx
]
rng.shuffle(other_fids)
distractors += other_fids[: pool_size - 1 - len(distractors)]
if len(distractors) < 3:
continue
candidates = distractors[: pool_size - 1]
# Insert correct answer at random position
insert_pos = rng.randint(0, len(candidates))
candidates.insert(insert_pos, target_fid)
candidate_vec_idxs = [fig_id_to_idx[f] for f in candidates]
# Multi-view retrieval
score = retrieve(query_vec, candidate_vec_idxs, all_vecs, insert_pos)
# Single-view retrieval (perspective only, if available)
persp_pos = next((i for i, v in enumerate(vps) if "perspective" in v and i != target_pos), None)
if persp_pos is not None:
single_score = retrieve(all_vecs[vec_idxs[persp_pos]], candidate_vec_idxs, all_vecs, insert_pos)
else:
single_score = score # fallback
for bucket in [difficulty, "all"]:
by_difficulty[bucket]["r1"] += score["r1"]
by_difficulty[bucket]["r5"] += score["r5"]
by_difficulty[bucket]["r10"] += score["r10"]
by_difficulty[bucket]["n"] += 1
results.append({
"patent_id": pid,
"target_view": vps[target_pos],
"difficulty": difficulty,
"pool_size": len(candidates),
"multi_view": score,
"single_view": single_score,
})
# Summary
print("\n" + "=" * 60)
print("RETRIEVAL EVAL RESULTS")
print(f"{'Difficulty':<16} {'N':>5} {'R@1':>6} {'R@5':>6} {'R@10':>6} {'Chance R@1':>10}")
print("-" * 60)
for diff in ["all", "baseline", "easy", "medium", "hard", "very_hard"]:
b = by_difficulty[diff]
if b["n"] == 0:
continue
chance = 100.0 / pool_size
print(
f"{diff:<16} {b['n']:>5} "
f"{b['r1']/b['n']:>5.1%} "
f"{b['r5']/b['n']:>5.1%} "
f"{b['r10']/b['n']:>5.1%} "
f"{chance:>9.1f}%"
)
output = {
"summary": {d: {k: v/b["n"] if k != "n" else v for k, v in b.items()}
for d, b in by_difficulty.items()},
"pool_size": pool_size,
"n_patents": len(results),
"results": results,
}
Path(out_path).parent.mkdir(parents=True, exist_ok=True)
with open(out_path, "w") as f:
json.dump(output, f, indent=2)
print(f"\nFull results → {out_path}")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--embeddings", default="data/embeddings/embeddings_2022_vitl14.parquet")
parser.add_argument("--enriched", default="data/enriched/enriched_2022.parquet")
parser.add_argument("--n", type=int, default=500)
parser.add_argument("--pool-size", type=int, default=100)
parser.add_argument("--out", default="results/retrieval_eval.json")
parser.add_argument("--seed", type=int, default=42)
args = parser.parse_args()
run_eval(args.embeddings, args.enriched, args.n, args.pool_size, args.out, args.seed)
if __name__ == "__main__":
main()
|