v15_benchmark / scripts /offline_low_score_ensemble.py
Sumail's picture
Upload folder using huggingface_hub (part 11)
f789875 verified
Raw
History Blame Contribute Delete
15.3 kB
"""Offline ensemble for low-score images from a previous online benchmark.
Reads a list of low-score indices, runs multiple attack/budget configurations,
keeps the best-scoring adversarial PNG per image, and compares with the online
baseline.
Usage:
# Pilot on the worst 108 images (score < 0.90)
python scripts/offline_low_score_ensemble.py \
--online-dir /workspace/Perturb/benchmark_2000_v15 \
--indices-file /workspace/Perturb/benchmark_2000_v15/low_score_090_indices.json \
--workers 4
# Full run on all score < 0.95 images
python scripts/offline_low_score_ensemble.py \
--online-dir /workspace/Perturb/benchmark_2000_v15 \
--indices-file /workspace/Perturb/benchmark_2000_v15/low_score_095_indices.json \
--configs v15:60 v13:60 v15:120 \
--workers 4
"""
from __future__ import annotations
import argparse
import base64
import io
import json
import os
import subprocess
import sys
import time
from pathlib import Path
from types import SimpleNamespace
from typing import Any
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
import numpy as np
import torch
from PIL import Image
from perturbnet import constants as C
from perturbnet.attacks import ATTACKS
from perturbnet.image_io import decode_image_b64, encode_image_b64
from perturbnet.imagenet100_bootstrap import load_imagenet100
from perturbnet.model import load_efficientnet_v2_l, predict_label
from neurons.validator import ChallengeSpec, PerturbValidator
def build_validator_stub(*, device: torch.device, model: torch.nn.Module) -> PerturbValidator:
config = SimpleNamespace(
perturb=SimpleNamespace(
min_linf_delta=C.MIN_LINF_DELTA,
max_linf_delta=C.MAX_LINF_DELTA,
min_ssim=C.MIN_SSIM,
min_psnr_db=C.MIN_PSNR_DB,
linf_component_weight=C.LINF_COMPONENT_WEIGHT,
rmse_component_weight=C.RMSE_COMPONENT_WEIGHT,
analyze_bucket_margin_weight=C.ANALYZE_BUCKET_MARGIN_WEIGHT,
analyze_bucket_novelty_weight=C.ANALYZE_BUCKET_NOVELTY_WEIGHT,
analyze_bucket_novelty_target_pixels=C.ANALYZE_BUCKET_NOVELTY_TARGET_PIXELS,
)
)
stub: PerturbValidator = object.__new__(PerturbValidator)
stub.config = config
stub.device = device
stub.model = model
return stub
def load_clean_b64_and_label(dataset: Any, ds_idx: int, device: torch.device, model: torch.nn.Module) -> tuple[str, str]:
example = dataset[ds_idx]
buffer = io.BytesIO()
example["image"].convert("RGB").save(buffer, format="JPEG", quality=95)
clean_b64 = base64.b64encode(buffer.getvalue()).decode("utf-8")
clean = decode_image_b64(clean_b64).to(device)
true_label = predict_label(model=model, image_chw=clean)
return clean_b64, true_label
def score_adv(stub: PerturbValidator, *, clean_b64: str, adv_b64: str, true_label: str, ds_idx: int) -> Any:
challenge = ChallengeSpec(
task_id=f"offline-{ds_idx:07d}",
image_id=str(ds_idx),
model_name=C.MODEL_NAME,
clean_image_b64=clean_b64,
true_label=true_label,
epsilon=C.MAX_LINF_DELTA,
norm_type="Linf",
)
return PerturbValidator.verify_and_score(stub, challenge=challenge, perturbed_image_b64=adv_b64)
def tensor_to_pil(image_chw: torch.Tensor) -> Image.Image:
arr = (image_chw.detach().clamp(0, 1).cpu().numpy() * 255).astype(np.uint8)
arr = np.transpose(arr, (1, 2, 0))
return Image.fromarray(arr, mode="RGB")
def save_tensor_png(path: Path, image_chw: torch.Tensor) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
tensor_to_pil(image_chw).save(path, format="PNG")
def parse_config(s: str) -> dict:
"""Parse 'attack:budget' or 'attack:budget:retries'."""
parts = s.split(":")
if len(parts) == 2:
return {"attack": parts[0], "budget": float(parts[1]), "retries": 0}
if len(parts) == 3:
return {"attack": parts[0], "budget": float(parts[1]), "retries": int(parts[2])}
raise ValueError(f"Invalid config '{s}', expected attack:budget or attack:budget:retries")
def summarize(records: list[dict]) -> dict:
n = len(records)
succ = [r for r in records if r["reason"] == "success"]
scores = [r["score"] for r in records]
mean_score = sum(scores) / n if n else 0.0
return {
"n": n,
"success_rate": len(succ) / n if n else 0.0,
"avg_score": mean_score,
"score_std": (sum((s - mean_score) ** 2 for s in scores) / n) ** 0.5 if n else 0.0,
"min_score": min(scores, default=0.0),
"max_score": max(scores, default=0.0),
"p50_score": float(np.median(scores)) if scores else 0.0,
"avg_time": sum(r.get("time", r.get("total_time", 0.0)) for r in records) / n if n else 0.0,
}
def run_offline_ensemble(
*,
indices: list[int],
configs: list[dict],
output_dir: Path,
device: torch.device,
worker_id: int,
workers: int,
early_stop_score: float = 0.965,
) -> dict:
"""Run all configs for a shard of indices and keep the best per image."""
dataset = load_imagenet100()
model = load_efficientnet_v2_l(device=device)
stub = build_validator_stub(device=device, model=model)
offline_dir = output_dir / "offline_ensemble"
offline_dir.mkdir(parents=True, exist_ok=True)
records: list[dict] = []
prefix = f"[w{worker_id}] "
for i, ds_idx in enumerate(indices):
json_path = offline_dir / f"{ds_idx:07d}.json"
if json_path.exists():
record = json.loads(json_path.read_text())
records.append(record)
print(f"{prefix}[{i+1}/{len(indices)}] idx={ds_idx} (cached)")
continue
clean_b64, true_label = load_clean_b64_and_label(dataset, ds_idx, device, model)
clean = decode_image_b64(clean_b64).to(device)
best_score = -1.0
best_adv = None
best_cfg_name = None
best_result = None
attempts: list[dict] = []
for cfg in configs:
for attempt in range(1 + cfg["retries"]):
t0 = time.time()
adv, info = ATTACKS[cfg["attack"]](model, clean, device, time_budget=cfg["budget"])
elapsed = time.time() - t0
adv_b64 = encode_image_b64(adv)
result = score_adv(stub, clean_b64=clean_b64, adv_b64=adv_b64, true_label=true_label, ds_idx=ds_idx)
attempts.append({
"attack": cfg["attack"],
"budget": cfg["budget"],
"attempt": attempt,
"score": result.score,
"reason": result.reason,
"rmse": result.rmse,
"norm": result.norm,
"margin": result.margin,
"ssim": result.ssim,
"time": elapsed,
"path": info.get("attack", cfg["attack"]),
})
if result.score > best_score:
best_score = result.score
best_adv = adv
best_cfg_name = f"{cfg['attack']}@{cfg['budget']}s"
best_result = result
if result.score >= early_stop_score:
break
if best_score >= early_stop_score:
break
adv_path = offline_dir / f"{ds_idx:07d}_adv.png"
save_tensor_png(adv_path, best_adv)
record = {
"ds_idx": ds_idx,
"true_label": true_label,
"score": best_result.score,
"reason": best_result.reason,
"prediction": best_result.model_prediction,
"norm": best_result.norm,
"rmse": best_result.rmse,
"ssim": best_result.ssim,
"psnr_db": best_result.psnr_db,
"margin": best_result.margin,
"best_config": best_cfg_name,
"attempts": attempts,
"total_time": sum(a["time"] for a in attempts),
"adv_path": str(adv_path),
}
json_path.write_text(json.dumps(record, indent=2, default=str))
records.append(record)
print(
f"{prefix}[{i+1}/{len(indices)}] idx={ds_idx} best={best_result.score:.4f} "
f"config={best_cfg_name} tried={len(attempts)}"
)
summary = summarize(records)
summary.update({"worker_id": worker_id, "workers": workers, "configs": configs})
summary_path = offline_dir / f"summary_worker{worker_id}.json"
summary_path.write_text(json.dumps(summary, indent=2, default=str))
return summary
def compare_with_online(online_dir: Path, offline_dir: Path, indices: list[int]) -> dict:
online_records = {}
online_json_dir = online_dir / "v15"
for idx in indices:
path = online_json_dir / f"{idx:07d}.json"
if path.exists():
online_records[idx] = json.loads(path.read_text())
offline_records = {}
offline_json_dir = offline_dir / "offline_ensemble"
for path in offline_json_dir.glob("[0-9][0-9][0-9][0-9][0-9][0-9][0-9].json"):
rec = json.loads(path.read_text())
offline_records[rec["ds_idx"]] = rec
common = [idx for idx in indices if idx in online_records and idx in offline_records]
if not common:
return {}
online_scores = [online_records[idx]["score"] for idx in common]
offline_scores = [offline_records[idx]["score"] for idx in common]
gains = [offline_records[idx]["score"] - online_records[idx]["score"] for idx in common]
improved = sum(1 for g in gains if g > 0)
worsened = sum(1 for g in gains if g < 0)
unchanged = sum(1 for g in gains if g == 0)
return {
"n": len(common),
"online_avg": sum(online_scores) / len(common),
"offline_avg": sum(offline_scores) / len(common),
"avg_gain": sum(gains) / len(common),
"max_gain": max(gains),
"max_loss": min(gains),
"improved": improved,
"worsened": worsened,
"unchanged": unchanged,
}
def main() -> None:
parser = argparse.ArgumentParser(description="Offline ensemble for low-score images")
parser.add_argument("--online-dir", type=Path, required=True, help="Directory with existing online benchmark")
parser.add_argument("--indices-file", type=Path, required=True, help="JSON file with list of low-score indices")
parser.add_argument("--output-dir", type=Path, default=Path("offline_ensemble_output"), help="Output directory")
parser.add_argument("--configs", type=str, nargs="+", default=["v15:60", "v13:60", "v15:120"], help="Configs like 'v15:60' or 'v15:120:2'")
parser.add_argument("--workers", type=int, default=4, help="Number of parallel workers")
parser.add_argument("--early-stop-score", type=float, default=0.965, help="Stop trying more configs if this score reached")
parser.add_argument("--indices-file-internal", type=Path, default=None, help=argparse.SUPPRESS)
args = parser.parse_args()
if args.indices_file_internal:
indices = json.loads(args.indices_file_internal.read_text())
is_top_level = False
else:
indices = json.loads(args.indices_file.read_text())
if isinstance(indices, dict):
indices = indices.get("indices", [])
is_top_level = True
configs = [parse_config(c) for c in args.configs]
print(f"[info] {'top-level' if is_top_level else 'worker'}: {len(indices)} images, configs={configs}")
args.output_dir.mkdir(parents=True, exist_ok=True)
if args.workers == 1:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
run_offline_ensemble(
indices=indices,
configs=configs,
output_dir=args.output_dir,
device=device,
worker_id=0,
workers=1,
early_stop_score=args.early_stop_score,
)
else:
log_dir = args.output_dir / "logs"
log_dir.mkdir(parents=True, exist_ok=True)
(args.output_dir / "indices.json").write_text(
json.dumps({"n": len(indices), "indices": indices}, indent=2)
)
processes = []
for worker_id in range(args.workers):
shard = [idx for i, idx in enumerate(indices) if i % args.workers == worker_id]
if not shard:
continue
shard_file = args.output_dir / f"shard_worker{worker_id}.json"
shard_file.write_text(json.dumps(shard, indent=2))
cmd = [
sys.executable,
str(Path(__file__).resolve()),
"--online-dir", str(args.online_dir),
"--indices-file", str(args.indices_file),
"--output-dir", str(args.output_dir),
"--early-stop-score", str(args.early_stop_score),
"--workers", "1",
"--indices-file-internal", str(shard_file),
"--configs", *args.configs,
]
log_path = log_dir / f"worker{worker_id}.log"
log_file = open(log_path, "w")
proc = subprocess.Popen(cmd, stdout=log_file, stderr=subprocess.STDOUT)
processes.append(proc)
print(f"[info] worker {worker_id} PID={proc.pid} shard={len(shard)}")
for proc in processes:
proc.wait()
if is_top_level:
comparison = compare_with_online(args.online_dir, args.output_dir, indices)
if comparison:
print("\n" + "=" * 64)
print("OFFLINE vs ONLINE COMPARISON")
print("=" * 64)
print(json.dumps(comparison, indent=2))
(args.output_dir / "comparison.json").write_text(json.dumps(comparison, indent=2))
# Also show what the overall 2000-image average would be if we merged offline bests.
online_json_dir = args.online_dir / "v15"
offline_json_dir = args.output_dir / "offline_ensemble"
merged_scores = []
for path in online_json_dir.glob("[0-9][0-9][0-9][0-9][0-9][0-9][0-9].json"):
rec = json.loads(path.read_text())
idx = rec["ds_idx"]
offline_path = offline_json_dir / f"{idx:07d}.json"
if offline_path.exists():
offline_rec = json.loads(offline_path.read_text())
merged_scores.append(max(rec["score"], offline_rec["score"]))
else:
merged_scores.append(rec["score"])
if merged_scores:
original_scores = []
for path in online_json_dir.glob("[0-9][0-9][0-9][0-9][0-9][0-9][0-9].json"):
rec = json.loads(path.read_text())
original_scores.append(rec["score"])
original_avg = sum(original_scores) / len(original_scores)
merged_avg = sum(merged_scores) / len(merged_scores)
print("\n" + "=" * 64)
print("PROJECTED OVERALL 2000-IMAGE AVERAGE")
print("=" * 64)
print(f"original online avg: {original_avg:.4f}")
print(f"after offline merge: {merged_avg:.4f}")
print(f"improvement: +{merged_avg - original_avg:.4f}")
print(f"\n[info] outputs saved to {args.output_dir}")
if __name__ == "__main__":
main()