File size: 13,665 Bytes
eac1c63 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 | #!/usr/bin/env python3
"""
Comprehensive multi-model benchmark (benchmark-only; no production wiring).
Models (via dghs-imgutils where available):
- WD SwinV2 v3, EVA02 Large, ViT Large, ConvNext v3
- ML-Danbooru
- Camie (initial + refined, macro_opt mode)
- PixAI v0.9
Metrics on local SFW Safebooru samples:
- latency (warm + infer)
- known-tag recall for site tags 1girl/solo @ threshold
- taxonomy evidence-tag coverage (vocab)
- taxonomy evidence hit-rate: how often any evidence tag for each
destination appears above threshold on images that carry that tag
in the Safebooru tag string (when present)
Usage (from backend/):
../.venv/Scripts/python.exe scripts/bench_models_comprehensive.py
"""
from __future__ import annotations
import json
import sys
import time
import traceback
from pathlib import Path
ROOT = Path(__file__).resolve().parents[2]
BACKEND = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(BACKEND))
SAMPLE_DIR = ROOT / "sample_data" / "sfw_safebooru"
META_PATH = SAMPLE_DIR / "manifest.json"
TAXONOMY_PATH = BACKEND / "app" / "data" / "taxonomy.json"
OUT_PATH = ROOT / "sample_data" / "bench_models_report.json"
GENERAL_THRESHOLD = 0.35
def _normalize(tag: str) -> str:
text = tag.strip().lower()
parts: list[str] = []
for ch in text:
if ch.isalnum():
parts.append(ch)
elif ch in {" ", "-", ".", "/", "_"}:
parts.append("_")
return "".join(parts).strip("_")
def _as_general_dict(raw: object) -> dict[str, float]:
scores: dict[str, float] = {}
if isinstance(raw, dict):
for tag, score in raw.items():
scores[_normalize(str(tag))] = float(score)
return scores
if isinstance(raw, (list, tuple)):
for part in raw:
if isinstance(part, dict):
for tag, score in part.items():
key = _normalize(str(tag))
scores[key] = max(scores.get(key, 0.0), float(score))
return scores
def load_taxonomy_evidence() -> dict[str, list[str]]:
tax = json.loads(TAXONOMY_PATH.read_text(encoding="utf-8"))
out: dict[str, list[str]] = {}
for bucket in tax["buckets"]:
out[bucket["folder"]] = [
_normalize(e["tag"]) for e in bucket.get("evidence", [])
]
return out
def score_with_model(model_id: str, path: Path) -> dict[str, float]:
if model_id.startswith("wd_"):
from imgutils.tagging import get_wd14_tags
mapping = {
"wd_swinv2_v3": "SwinV2_v3",
"wd_eva02_large": "EVA02_Large",
"wd_vit_large": "ViT_Large",
"wd_convnext_v3": "ConvNext_v3",
}
raw = get_wd14_tags(
str(path),
model_name=mapping[model_id],
general_threshold=GENERAL_THRESHOLD,
no_underline=False,
drop_overlap=False,
fmt="general",
)
return _as_general_dict(raw)
if model_id == "ml_danbooru":
from imgutils.tagging import get_mldanbooru_tags
raw = get_mldanbooru_tags(
str(path),
threshold=GENERAL_THRESHOLD,
size=448,
keep_ratio=True,
drop_overlap=False,
use_real_name=False,
)
return _as_general_dict(raw)
if model_id.startswith("camie_"):
from imgutils.tagging import get_camie_tags
# camie_initial_macro_opt / camie_refined_balanced
name = None
mode = None
for candidate in ("initial", "refined"):
prefix = f"camie_{candidate}_"
if model_id.startswith(prefix):
name = candidate
mode = model_id[len(prefix) :]
break
if name is None or not mode:
raise ValueError(f"Bad camie model id: {model_id}")
raw = get_camie_tags(
str(path),
model_name=name,
mode=mode, # type: ignore[arg-type]
no_underline=False,
drop_overlap=False,
fmt="general",
)
return _as_general_dict(raw)
if model_id == "pixai_v0_9":
from imgutils.tagging import get_pixai_tags
raw = get_pixai_tags(
str(path),
model_name="v0.9",
thresholds=GENERAL_THRESHOLD,
fmt="general",
)
return _as_general_dict(raw)
raise ValueError(f"unknown model_id={model_id}")
def vocab_for_model(model_id: str) -> set[str] | None:
"""Return tag vocabulary when cheaply available; None if unknown."""
try:
if model_id.startswith("wd_"):
from huggingface_hub import hf_hub_download
import pandas as pd
repos = {
"wd_swinv2_v3": "SmilingWolf/wd-swinv2-tagger-v3",
"wd_eva02_large": "SmilingWolf/wd-eva02-large-tagger-v3",
"wd_vit_large": "SmilingWolf/wd-vit-large-tagger-v3",
"wd_convnext_v3": "SmilingWolf/wd-convnext-tagger-v3",
}
path = hf_hub_download(repos[model_id], "selected_tags.csv")
return {_normalize(t) for t in pd.read_csv(path)["name"].tolist()}
if model_id == "ml_danbooru":
from huggingface_hub import hf_hub_download
import pandas as pd
path = hf_hub_download(
"deepghs/imgutils-models", "mldanbooru/mldanbooru_tags.csv"
)
return {_normalize(t) for t in pd.read_csv(path)["name"].tolist()}
if model_id.startswith("camie_"):
from huggingface_hub import hf_hub_download
import pandas as pd
name = "refined" if "refined" in model_id else "initial"
path = hf_hub_download(
"deepghs/camie_tagger_onnx", f"{name}/selected_tags.csv"
)
return {_normalize(t) for t in pd.read_csv(path)["name"].tolist()}
if model_id == "pixai_v0_9":
from imgutils.tagging import pixai
tags_df, _ips = pixai._open_tags("v0.9")
return {_normalize(str(t)) for t in tags_df["name"].tolist()}
except Exception as err:
print(f" vocab_error {model_id}: {err}", flush=True)
return None
return None
def warm_model(model_id: str, path: Path) -> float:
t0 = time.perf_counter()
score_with_model(model_id, path)
return time.perf_counter() - t0
def main() -> int:
# Match production DLL search so imgutils ORT sessions can use CUDA.
from app.providers import ensure_nvidia_dll_search_path, preload_onnx_runtime_dlls
ensure_nvidia_dll_search_path()
preload_onnx_runtime_dlls()
if not META_PATH.exists():
print(
f"ERROR: missing {META_PATH}. Run scripts/bench_sfw_sample.py first.",
file=sys.stderr,
)
return 2
manifest = json.loads(META_PATH.read_text(encoding="utf-8"))
paths = [Path(row["path"]) for row in manifest]
evidence = load_taxonomy_evidence()
all_evidence = sorted({t for tags in evidence.values() for t in tags})
models = [
"wd_swinv2_v3",
"wd_eva02_large",
"wd_vit_large",
"wd_convnext_v3",
"ml_danbooru",
"camie_initial_macro_opt",
"camie_refined_macro_opt",
"pixai_v0_9",
]
report: dict[str, object] = {
"general_threshold": GENERAL_THRESHOLD,
"sample_count": len(paths),
"sample_dir": str(SAMPLE_DIR),
"models": {},
"notes": [
"Camie v2 (Camais03/camie-tagger-v2) is not in dghs-imgutils 0.19; "
"benchmarked ONNX initial/refined mirrors instead.",
"Published F1 numbers are not directly comparable across vendors "
"(different splits/thresholds/tag sets).",
"SFW-only images; taxonomy NSFW folders measured via vocab coverage "
"and evidence-tag presence on images that carry those tags.",
],
}
print(f"samples={len(paths)} models={len(models)}", flush=True)
for model_id in models:
print(f"\n=== {model_id} ===", flush=True)
entry: dict[str, object] = {"ok": False}
try:
vocab = vocab_for_model(model_id)
if vocab is not None:
covered = [t for t in all_evidence if t in vocab]
missing = [t for t in all_evidence if t not in vocab]
per_folder = {}
for folder, tags in evidence.items():
hit = [t for t in tags if t in vocab]
per_folder[folder] = {
"coverage": len(hit) / max(1, len(tags)),
"present": hit,
"missing": [t for t in tags if t not in vocab],
}
entry["vocab_size"] = len(vocab)
entry["taxonomy_evidence_coverage"] = len(covered) / max(
1, len(all_evidence)
)
entry["taxonomy_missing"] = missing
entry["taxonomy_per_folder"] = per_folder
print(
f" vocab={len(vocab)} evidence_coverage="
f"{entry['taxonomy_evidence_coverage']:.1%}",
flush=True,
)
warm_s = warm_model(model_id, paths[0])
print(f" warm={warm_s:.2f}s", flush=True)
must_hits = 0
latencies: list[float] = []
per_image: list[dict] = []
# taxonomy evidence recall on images whose site tags include evidence
tax_denom = {f: 0 for f in evidence}
tax_numer = {f: 0 for f in evidence}
for row, path in zip(manifest, paths):
t0 = time.perf_counter()
scores = score_with_model(model_id, path)
dt = time.perf_counter() - t0
latencies.append(dt)
must = {_normalize(t) for t in row.get("must_have", [])}
known = {_normalize(t) for t in row.get("known_general", [])}
hit = must <= set(scores)
must_hits += int(hit)
top = sorted(scores.items(), key=lambda x: -x[1])[:8]
for folder, ev_tags in evidence.items():
# Only evaluate folders whose evidence appears in site tags
if not (known & set(ev_tags)):
continue
tax_denom[folder] += 1
if any(scores.get(t, 0.0) >= GENERAL_THRESHOLD for t in ev_tags):
tax_numer[folder] += 1
per_image.append(
{
"id": row["id"],
"must_hit": hit,
"latency_s": dt,
"top": top,
"score_count": len(scores),
}
)
print(
f" #{row['id']} must_hit={hit} n={len(scores)} "
f"{dt*1000:.0f}ms top={[t for t,_ in top[:5]]}",
flush=True,
)
tax_recall = {
f: (tax_numer[f] / tax_denom[f] if tax_denom[f] else None)
for f in evidence
}
entry.update(
{
"ok": True,
"warm_s": warm_s,
"infer_total_s": sum(latencies),
"ms_per_image": (sum(latencies) / len(latencies)) * 1000.0,
"must_have_recall": must_hits / len(paths),
"taxonomy_evidence_recall_on_labeled": tax_recall,
"taxonomy_evidence_support": tax_denom,
"per_image": per_image,
}
)
print(
f" SUMMARY recall={entry['must_have_recall']:.0%} "
f"ms/img={entry['ms_per_image']:.1f}",
flush=True,
)
except Exception as err:
entry["ok"] = False
entry["error"] = f"{type(err).__name__}: {err}"
entry["traceback"] = traceback.format_exc()
print(f" FAILED: {entry['error']}", flush=True)
report["models"][model_id] = entry
# Ranking: prioritize must_have recall, then taxonomy coverage, then speed
ranked = []
for model_id, entry in report["models"].items():
if not entry.get("ok"):
continue
ranked.append(
{
"model": model_id,
"must_have_recall": entry.get("must_have_recall"),
"taxonomy_evidence_coverage": entry.get("taxonomy_evidence_coverage"),
"ms_per_image": entry.get("ms_per_image"),
"vocab_size": entry.get("vocab_size"),
}
)
ranked.sort(
key=lambda r: (
-(r["must_have_recall"] or 0),
-(r["taxonomy_evidence_coverage"] or 0),
r["ms_per_image"] or 1e9,
)
)
report["ranking"] = ranked
OUT_PATH.parent.mkdir(parents=True, exist_ok=True)
OUT_PATH.write_text(json.dumps(report, indent=2), encoding="utf-8")
print(f"\nWrote {OUT_PATH}", flush=True)
print("\n=== RANKING ===", flush=True)
for i, row in enumerate(ranked, 1):
print(
f"{i}. {row['model']}: recall={row['must_have_recall']:.0%} "
f"tax_cov={row['taxonomy_evidence_coverage']:.0%} "
f"ms/img={row['ms_per_image']:.1f} vocab={row['vocab_size']}",
flush=True,
)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|