File size: 15,946 Bytes
5fed0fc |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 |
#!/usr/bin/env python3
"""
VDB Design Evaluator
Evaluates custom ANN index implementations by benchmarking them on SIFT1M dataset.
The solution.py should export a class with the following interface:
- __init__(dim: int, **kwargs)
- add(xb: np.ndarray) -> None
- search(xq: np.ndarray, k: int) -> tuple[np.ndarray, np.ndarray]
"""
import argparse
import importlib.util
import json
import os
import signal
import sys
import time
from pathlib import Path
from types import ModuleType
import numpy as np
try:
from faiss.contrib.datasets_fb import DatasetSIFT1M
except ImportError:
from faiss.contrib.datasets import DatasetSIFT1M
CONFIG_PATH = Path(__file__).with_name("score_config.json")
# TO BE DETERMINED
TIMEOUT_SECONDS = 3600
class TimeoutError(Exception):
pass
def timeout_handler(signum, frame):
raise TimeoutError(f"Solver execution exceeded {TIMEOUT_SECONDS} seconds")
def with_timeout(func):
def wrapper(*args, **kwargs):
old_handler = signal.signal(signal.SIGALRM, timeout_handler)
signal.alarm(TIMEOUT_SECONDS)
try:
result = func(*args, **kwargs)
return result
finally:
signal.alarm(0)
signal.signal(signal.SIGALRM, old_handler)
return wrapper
def load_score_config() -> dict:
"""Load scoring configuration from JSON, returning defaults if missing."""
default = {
"baseline": {
"name": "baseline",
"recall_at_1": 0.99,
"avg_query_time_ms": 5.0,
},
"scoring": {
"scale": 60.0,
"recall_weight": 2.0,
"latency_weight": 1.0,
"max_score": 100.0,
"min_score": 0.0,
"epsilon": 1e-9,
},
"pareto": {
"recall_margin": 0.0,
"latency_margin_ms": 0.0,
},
}
if CONFIG_PATH.exists():
try:
with CONFIG_PATH.open("r", encoding="utf-8") as fh:
loaded = json.load(fh)
# Merge loaded values onto defaults without mutating default structure
def merge(base: dict, new: dict) -> dict:
result = dict(base)
for key, value in new.items():
if isinstance(value, dict) and isinstance(result.get(key), dict):
result[key] = merge(result[key], value)
else:
result[key] = value
return result
return merge(default, loaded)
except (OSError, json.JSONDecodeError) as exc:
print(f"[evaluator] WARNING: Failed to load score_config.json ({exc}); using defaults", file=sys.stderr)
return default
SCORE_CONFIG = load_score_config()
def load_solution_module(solution_path: Path) -> ModuleType:
"""Load the solution module from the given path."""
if not solution_path.exists():
raise FileNotFoundError(f"solution.py not found at {solution_path}")
spec = importlib.util.spec_from_file_location("submitted_solution", solution_path)
if spec is None or spec.loader is None:
raise ImportError(f"Failed to load spec for {solution_path}")
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module
def find_solution_class(module: ModuleType):
"""
Find the custom ANN index class in the solution module.
Looks for a class with add() and search() methods.
"""
for name in dir(module):
if name.startswith('_'):
continue
obj = getattr(module, name)
if not isinstance(obj, type):
continue
if hasattr(obj, 'add') and hasattr(obj, 'search'):
return obj
raise AttributeError(
"solution.py must define a class with add(xb) and search(xq, k) methods"
)
def evaluate_index(index, xq: np.ndarray, gt: np.ndarray, k: int) -> dict:
"""
Evaluate index performance on query set.
Returns metrics: recall@1, total_time_ms, avg_query_time_ms
"""
nq = xq.shape[0]
# Batched search timing
t0 = time.time()
D, I = index.search(xq, k)
t1 = time.time()
# Calculate metrics
recall_at_1 = (I[:, :1] == gt[:, :1]).sum() / float(nq)
total_ms = (t1 - t0) * 1000.0
per_query_ms = total_ms / float(nq)
# Single-request latency sample (measure first 100 queries)
num_samples = min(100, nq)
single_total = 0.0
for i in range(num_samples):
q = xq[i:i+1]
s0 = time.time()
index.search(q, k)
s1 = time.time()
single_total += (s1 - s0)
single_avg_ms = (single_total / num_samples) * 1000.0
return {
'recall_at_1': float(recall_at_1),
'total_time_ms': float(total_ms),
'avg_query_time_ms': float(per_query_ms),
'single_query_avg_ms': float(single_avg_ms),
'num_queries': int(nq),
}
def _safe_ratio(numerator: float, denominator: float, epsilon: float) -> float:
"""Return a stable ratio, clamping to epsilon on the denominator."""
denom = max(denominator, epsilon)
return max(numerator / denom, epsilon)
def compute_score(metrics: dict, config: dict = SCORE_CONFIG) -> float:
"""
Compute score based on scoring mode in config.
Modes:
- "latency_gated_recall": If latency meets threshold, score based on recall.
If recall > baseline_recall: 100 points
Otherwise: (recall - lower_threshold) / (baseline_recall - lower_threshold) * 100
- "balanced" (default): Balances recall and latency improvements
"""
scoring_cfg = config["scoring"]
baseline_cfg = config["baseline"]
max_score = float(scoring_cfg.get("max_score", 100.0))
min_score = float(scoring_cfg.get("min_score", 0.0))
# Check if using new latency_gated_recall mode
mode = scoring_cfg.get("mode", "balanced")
if mode == "latency_gated_recall":
# New scoring: latency is a gate, recall determines score
latency_threshold = float(scoring_cfg.get("latency_threshold_ms", 3.85))
recall_lower_threshold = float(scoring_cfg.get("recall_lower_threshold", 0.9914))
actual_latency = float(metrics["avg_query_time_ms"])
actual_recall = float(metrics["recall_at_1"])
baseline_recall = float(baseline_cfg["recall_at_1"])
# Check if latency meets threshold
if actual_latency > latency_threshold:
# Latency gate not met - score 0
return min_score
# Latency gate met - score based on recall
if actual_recall > baseline_recall:
# Exceeds baseline recall - perfect score
return max_score
else:
# Score proportional to recall vs lower threshold
recall_range = baseline_recall - recall_lower_threshold
if recall_range <= 0:
# Edge case: lower threshold >= baseline
return max_score if actual_recall >= recall_lower_threshold else min_score
recall_proportion = (actual_recall - recall_lower_threshold) / recall_range
# Clamp to [0, 1] and scale to [min_score, max_score]
recall_proportion = max(0.0, min(1.0, recall_proportion))
score = min_score + (max_score - min_score) * recall_proportion
return float(score)
else:
# Original balanced mode
recall_ratio = _safe_ratio(metrics["recall_at_1"], baseline_cfg["recall_at_1"], scoring_cfg.get("epsilon", 1e-9))
latency_ratio = _safe_ratio(baseline_cfg["avg_query_time_ms"], metrics["avg_query_time_ms"], scoring_cfg.get("epsilon", 1e-9))
recall_weight = float(scoring_cfg.get("recall_weight", 2.0))
latency_weight = float(scoring_cfg.get("latency_weight", 1.0))
scale = float(scoring_cfg.get("scale", 60.0))
threshold = max(0.0, float(scoring_cfg.get("improvement_threshold", 0.0)))
raw = (recall_ratio ** recall_weight) * (latency_ratio ** latency_weight)
if raw <= 1.0:
score = scale * raw
else:
improvement = raw - 1.0
if improvement <= threshold:
score = scale
else:
target_cfg = config.get("targets", {})
target_recall = max(
float(target_cfg.get("aspirational_recall", baseline_cfg["recall_at_1"])),
float(baseline_cfg["recall_at_1"]),
)
target_latency = min(
float(target_cfg.get("aspirational_avg_query_time_ms", baseline_cfg["avg_query_time_ms"])),
float(baseline_cfg["avg_query_time_ms"]),
)
target_raw = (
_safe_ratio(target_recall, baseline_cfg["recall_at_1"], scoring_cfg.get("epsilon", 1e-9)) ** recall_weight
) * (
_safe_ratio(baseline_cfg["avg_query_time_ms"], target_latency, scoring_cfg.get("epsilon", 1e-9)) ** latency_weight
)
effective_raw = 1.0 + (improvement - threshold)
target_effective = max(1.0 + (target_raw - 1.0 - threshold), 1.0 + scoring_cfg.get("epsilon", 1e-9))
effective_clamped = min(effective_raw, target_effective)
growth = max(
0.0,
min(1.0, (effective_clamped - 1.0) / max(target_effective - 1.0, scoring_cfg.get("epsilon", 1e-9))),
)
score = scale + (max_score - scale) * growth
return float(min(max_score, max(min_score, score)))
def assess_pareto(metrics: dict, config: dict = SCORE_CONFIG) -> dict:
"""
Compare solution metrics against the configured baseline to determine Pareto status.
Returns:
Dict containing:
- reference: name of baseline
- pareto_frontier: bool (True if not dominated by reference)
- dominates_reference: bool
- dominated_by_reference: bool
- status: human-readable label
- deltas: recall/time deltas vs reference
"""
baseline = config["baseline"]
margins = config.get("pareto", {})
recall_margin = float(margins.get("recall_margin", 0.0))
latency_margin = float(margins.get("latency_margin_ms", 0.0))
recall = float(metrics["recall_at_1"])
latency = float(metrics["avg_query_time_ms"])
base_recall = float(baseline["recall_at_1"])
base_latency = float(baseline["avg_query_time_ms"])
recall_delta = recall - base_recall
latency_delta = latency - base_latency
dominates = (
recall >= base_recall - recall_margin
and latency <= base_latency + latency_margin
and (recall > base_recall + recall_margin or latency < base_latency - latency_margin)
)
dominated = (
recall <= base_recall + recall_margin
and latency >= base_latency - latency_margin
and (recall < base_recall - recall_margin or latency > base_latency + latency_margin)
)
if dominates:
status = "dominates_reference"
elif dominated:
status = "dominated_by_reference"
else:
status = "tradeoff"
return {
"reference": baseline.get("name", "baseline"),
"pareto_frontier": not dominated,
"dominates_reference": dominates,
"dominated_by_reference": dominated,
"status": status,
"deltas": {
"recall_at_1": recall_delta,
"avg_query_time_ms": latency_delta,
},
"reference_metrics": {
"recall_at_1": base_recall,
"avg_query_time_ms": base_latency,
},
}
@with_timeout
def evaluate(solution_path: Path, k: int = 1) -> dict:
"""
Main evaluation function.
Args:
solution_path: Path to solution.py
k: Number of nearest neighbors to retrieve (default: 1)
Returns:
Dictionary with evaluation results and score
Raises:
TimeoutError: If evaluation takes longer than 10 minutes
"""
print("[evaluator] Loading solution module...", file=sys.stderr)
module = load_solution_module(solution_path)
print("[evaluator] Finding solution class...", file=sys.stderr)
IndexClass = find_solution_class(module)
print(f"[evaluator] Using class: {IndexClass.__name__}", file=sys.stderr)
print("[evaluator] Loading SIFT1M dataset...", file=sys.stderr)
ds = DatasetSIFT1M()
xb = ds.get_database() # 1M base vectors
xq = ds.get_queries() # 10k query vectors
gt = ds.get_groundtruth() # Ground truth
xt = ds.get_train() # Training vectors (optional)
d = xb.shape[1] # dimension (128 for SIFT1M)
print(f"[evaluator] Dataset loaded: {xb.shape[0]} base vectors, dim={d}", file=sys.stderr)
print("[evaluator] Building index...", file=sys.stderr)
index = IndexClass(d)
print("[evaluator] Adding vectors to index...", file=sys.stderr)
t0 = time.time()
index.add(xb)
build_time = time.time() - t0
print(f"[evaluator] Index built in {build_time:.2f}s", file=sys.stderr)
print("[evaluator] Running evaluation...", file=sys.stderr)
metrics = evaluate_index(index, xq, gt, k)
print(f"[evaluator] Recall@1: {metrics['recall_at_1']:.4f}", file=sys.stderr)
print(f"[evaluator] Avg query time: {metrics['avg_query_time_ms']:.3f}ms", file=sys.stderr)
# Compute final score
score = compute_score(metrics)
pareto = assess_pareto(metrics)
print(f"[evaluator] Final score: {score:.2f}", file=sys.stderr)
print(f"[evaluator] Pareto status vs {pareto['reference']}: {pareto['status']}", file=sys.stderr)
return {
'score': score,
'metrics': metrics,
'build_time_seconds': build_time,
'index_class': IndexClass.__name__,
'k': k,
'pareto': pareto,
}
def main() -> None:
parser = argparse.ArgumentParser(description="Evaluate VDB design solution")
parser.add_argument(
"--solution",
default="../../execution_env/solution_env/solution.py",
help="Path to solution.py",
)
parser.add_argument(
"--out",
default="results.json",
help="Output JSON file path",
)
parser.add_argument(
"--k",
type=int,
default=1,
help="Number of nearest neighbors to retrieve",
)
args = parser.parse_args()
solution_path = Path(args.solution).resolve()
output_path = Path(args.out)
try:
results = evaluate(solution_path, k=args.k)
# Write results to file
with output_path.open('w') as f:
json.dump(results, f, indent=2)
print(f"[evaluator] Results written to {output_path}", file=sys.stderr)
# Print score to stdout for evaluate.sh to capture
print(json.dumps(results))
except TimeoutError as e:
print(f"[evaluator] TIMEOUT FAILURE: {e}", file=sys.stderr)
timeout_payload = {
'score': 0.0,
'error': str(e),
'error_type': 'TimeoutError',
'timeout_seconds': TIMEOUT_SECONDS,
}
with output_path.open('w') as f:
json.dump(timeout_payload, f, indent=2)
print(json.dumps(timeout_payload))
sys.exit(1)
except Exception as e:
error_payload = {
'score': 0.0,
'error': str(e),
'error_type': type(e).__name__,
}
with output_path.open('w') as f:
json.dump(error_payload, f, indent=2)
print(f"[evaluator] ERROR: {e}", file=sys.stderr)
print(json.dumps(error_payload))
sys.exit(1)
if __name__ == "__main__":
main()
|