Spaces:
Sleeping
Sleeping
File size: 18,584 Bytes
dbc6675 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 | """Frozen transition-model feature extraction for candidate plans."""
from __future__ import annotations
import json
import math
from dataclasses import dataclass
from pathlib import Path
import numpy as np
from code.downstream.families import SourceFamily
from code.downstream.plan_utils import ProblemContext, load_problem_context, rollout_plan
from code.tokenization.factory import create_tokenizer
from code.tokenization.multidomain import MultiDomainUnionTokenizer, load_tokenizer_from_manifest
SERIES_NAMES = [
"residual_l2",
"residual_l1_mean",
"cosine_distance",
"pred_norm",
"pred_delta_norm",
"candidate_delta_norm",
"current_goal_l2",
"pred_goal_l2",
"goal_progress_l2",
]
SUMMARY_STATS = ["mean", "std", "min", "max", "final"]
SCALAR_FEATURES = [
"plan_len",
"log_plan_len",
"plan_to_budget_ratio",
"initial_goal_l2",
"final_goal_l2",
"final_minus_initial_goal_l2",
"hidden_h_final_norm",
"hidden_c_final_norm",
"hidden_h_abs_mean",
"hidden_c_abs_mean",
]
FEATURE_NAMES = [
f"{series}_{stat}"
for series in SERIES_NAMES
for stat in SUMMARY_STATS
] + SCALAR_FEATURES
@dataclass
class LoadedSource:
"""Loaded frozen model/tokenizer artifacts for one family/domain/seed."""
family: SourceFamily
domain: str
model: object
tokenizer: object
mode: str
model_kind: str
input_dim: int
device: object | None = None
class FrozenTransitionFeatureExtractor:
"""Extract fixed-size consistency summaries from frozen transition models."""
def __init__(
self,
*,
run_root: str | Path,
source_data_dir: str | Path,
family: SourceFamily,
seed: int,
device: str = "cpu",
xgb_n_jobs: int = 1,
) -> None:
self.run_root = Path(run_root)
self.source_data_dir = Path(source_data_dir)
self.family = family
self.seed = seed
self.device_arg = device
self.xgb_n_jobs = xgb_n_jobs
self._source_cache: dict[str, LoadedSource] = {}
self._problem_cache: dict[tuple[str, str, str], ProblemContext] = {}
self._wl_problem_cache: dict[tuple[str, str, str], object] = {}
@property
def feature_names(self) -> list[str]:
return list(FEATURE_NAMES)
def extract(self, candidate: dict) -> np.ndarray:
"""Extract one feature vector for a candidate-plan record."""
domain = candidate["domain"]
split = candidate["split"]
problem = candidate["problem"]
plan = [str(action) for action in candidate.get("plan", [])]
source = self._load_source(domain)
problem_context = self._load_problem_context(domain, split, problem)
states = rollout_plan(problem_context, plan)
goal_vec = self._embed_goal(source, problem_context)
initial_vec = self._embed_state(source, problem_context, states[0])
final_vec = self._embed_state(source, problem_context, states[-1])
initial_goal_l2 = _l2(initial_vec - goal_vec)
final_goal_l2 = _l2(final_vec - goal_vec)
series_values = {name: [] for name in SERIES_NAMES}
hidden_summary = {
"hidden_h_final_norm": 0.0,
"hidden_c_final_norm": 0.0,
"hidden_h_abs_mean": 0.0,
"hidden_c_abs_mean": 0.0,
}
hidden = None
for current_atoms, next_atoms in zip(states, states[1:]):
current_vec = self._embed_state(source, problem_context, current_atoms)
next_vec = self._embed_state(source, problem_context, next_atoms)
pred_next_vec, pred_delta_vec, hidden = self._predict_next(
source,
current_vec,
goal_vec,
hidden,
)
candidate_delta = next_vec - current_vec
residual = pred_next_vec - next_vec
current_goal_l2 = _l2(current_vec - goal_vec)
next_goal_l2 = _l2(next_vec - goal_vec)
pred_goal_l2 = _l2(pred_next_vec - goal_vec)
series_values["residual_l2"].append(_l2(residual))
series_values["residual_l1_mean"].append(float(np.mean(np.abs(residual))))
series_values["cosine_distance"].append(_cosine_distance(pred_next_vec, next_vec))
series_values["pred_norm"].append(_l2(pred_next_vec))
series_values["pred_delta_norm"].append(_l2(pred_delta_vec))
series_values["candidate_delta_norm"].append(_l2(candidate_delta))
series_values["current_goal_l2"].append(current_goal_l2)
series_values["pred_goal_l2"].append(pred_goal_l2)
series_values["goal_progress_l2"].append(current_goal_l2 - next_goal_l2)
if source.model_kind == "lstm" and hidden is not None:
hidden_summary.update(_summarize_lstm_hidden(hidden))
values: list[float] = []
for series in SERIES_NAMES:
values.extend(_summarize_series(series_values[series]))
plan_len = float(len(plan))
transition_budget = float(max(100, 10 * max(1, len(problem_context.objects))))
values.extend(
[
plan_len,
math.log1p(plan_len),
plan_len / transition_budget,
initial_goal_l2,
final_goal_l2,
final_goal_l2 - initial_goal_l2,
hidden_summary["hidden_h_final_norm"],
hidden_summary["hidden_c_final_norm"],
hidden_summary["hidden_h_abs_mean"],
hidden_summary["hidden_c_abs_mean"],
]
)
return np.nan_to_num(np.asarray(values, dtype=np.float32))
def _load_problem_context(self, domain: str, split: str, problem: str) -> ProblemContext:
key = (domain, split, problem)
cached = self._problem_cache.get(key)
if cached is None:
cached = load_problem_context(
data_dir=self.source_data_dir,
domain=domain,
split=split,
problem=problem,
)
self._problem_cache[key] = cached
return cached
def _load_source(self, domain: str) -> LoadedSource:
cached = self._source_cache.get(domain)
if cached is not None:
return cached
if self.family.model == "xgboost":
source = self._load_xgb_source(domain)
elif self.family.model == "lstm":
source = self._load_lstm_source(domain)
else:
raise ValueError(f"Unsupported source model: {self.family.model}")
self._source_cache[domain] = source
return source
def _load_xgb_source(self, domain: str) -> LoadedSource:
import xgboost as xgb
checkpoint_dir = self._checkpoint_dir(domain)
model_name = domain if self.family.is_domain_dependent else "all_domains"
model_path = checkpoint_dir / f"{model_name}_xgb.json"
meta_path = checkpoint_dir / f"{model_name}_xgb_meta.json"
if not model_path.exists():
raise FileNotFoundError(f"Missing XGBoost source model: {model_path}")
model = xgb.XGBRegressor(device="cpu", n_jobs=self.xgb_n_jobs)
model.load_model(str(model_path))
meta = _read_json(meta_path)
tokenizer = self._load_tokenizer(domain)
input_dim = int(meta.get("output_dim", tokenizer.get_embedding_dim()))
return LoadedSource(
family=self.family,
domain=domain,
model=model,
tokenizer=tokenizer,
mode=str(meta.get("delta", True) and "delta" or "state"),
model_kind="xgboost",
input_dim=input_dim,
)
def _load_lstm_source(self, domain: str) -> LoadedSource:
import torch
from code.modeling.models import StateCentricLSTM, StateCentricLSTM_Delta
checkpoint_dir = self._checkpoint_dir(domain)
model_name = domain if self.family.is_domain_dependent else "all_domains"
checkpoint_path = checkpoint_dir / f"{model_name}_lstm_best.pt"
meta_path = checkpoint_dir / f"{model_name}_lstm_meta.json"
if not checkpoint_path.exists():
raise FileNotFoundError(f"Missing LSTM source model: {checkpoint_path}")
meta = _read_json(meta_path)
tokenizer = self._load_tokenizer(domain)
input_dim = int(meta.get("input_dim", tokenizer.get_embedding_dim()))
hidden_dim = int(meta.get("hidden_dim", 256))
use_projection = not bool(meta.get("no_projection", False))
mode = str(meta.get("mode", self.family.mode))
device = _resolve_torch_device(self.device_arg)
if mode == "delta":
model = StateCentricLSTM_Delta(
input_dim,
hidden_dim=hidden_dim,
use_projection=use_projection,
).to(device)
else:
model = StateCentricLSTM(
input_dim,
hidden_dim=hidden_dim,
use_projection=use_projection,
).to(device)
model.load_state_dict(torch.load(checkpoint_path, map_location=device))
model.eval()
return LoadedSource(
family=self.family,
domain=domain,
model=model,
tokenizer=tokenizer,
mode=mode,
model_kind="lstm",
input_dim=input_dim,
device=device,
)
def _checkpoint_dir(self, domain: str) -> Path:
root = self.run_root / self.family.regime / "checkpoints" / f"seed_{self.seed}"
root = root / self.family.tokenizer / self.family.model_mode_dir
if self.family.is_domain_dependent:
root = root / domain
return root
def _load_tokenizer(self, domain: str):
domain_pddl = self.source_data_dir / "pddl" / domain / "domain.pddl"
if self.family.is_all_domains:
manifest = (
self.run_root
/ "all_domains"
/ "tokenizers"
/ self.family.tokenizer
/ f"all_domains_{self.family.tokenizer}.json"
)
tokenizer = load_tokenizer_from_manifest(str(manifest))
if isinstance(tokenizer, MultiDomainUnionTokenizer):
tokenizer.set_active_domain(domain, str(domain_pddl))
elif hasattr(tokenizer, "set_domain"):
tokenizer.set_domain(str(domain_pddl))
return tokenizer
tokenizer = create_tokenizer(self.family.tokenizer)
vocab_path = self._domain_vocab_path(domain)
if not vocab_path.exists():
raise FileNotFoundError(f"Missing tokenizer vocabulary: {vocab_path}")
tokenizer.load_vocabulary(str(vocab_path))
if hasattr(tokenizer, "set_domain"):
tokenizer.set_domain(str(domain_pddl))
return tokenizer
def _domain_vocab_path(self, domain: str) -> Path:
model_dir = (
self.run_root
/ "domain_dependent"
/ "data"
/ "encodings"
/ "models"
)
if self.family.tokenizer == "wl":
return model_dir / f"{domain}_wl_tok.json"
return model_dir / f"{domain}_{self.family.tokenizer}.json"
def _embed_state(
self,
source: LoadedSource,
context: ProblemContext,
atoms: frozenset[str],
) -> np.ndarray:
tokenizer = source.tokenizer
wl_prob = self._wl_problem(source, context)
state_atoms = list(atoms)
try:
vec = tokenizer.transform_state(
state_atoms,
list(context.goal_atoms),
list(context.objects),
problem_pddl_path=str(context.problem_path),
_wl_prob=wl_prob,
)
except TypeError:
try:
vec = tokenizer.transform_state(
state_atoms,
list(context.goal_atoms),
list(context.objects),
problem_pddl_path=str(context.problem_path),
)
except TypeError:
vec = tokenizer.transform_state(
state_atoms,
list(context.goal_atoms),
list(context.objects),
)
return np.asarray(vec, dtype=np.float32).reshape(-1)
def _embed_goal(self, source: LoadedSource, context: ProblemContext) -> np.ndarray:
tokenizer = source.tokenizer
wl_prob = self._wl_problem(source, context)
try:
vec = tokenizer.transform_goal(
list(context.goal_atoms),
list(context.objects),
problem_pddl_path=str(context.problem_path),
_wl_prob=wl_prob,
)
except TypeError:
try:
vec = tokenizer.transform_goal(
list(context.goal_atoms),
list(context.objects),
problem_pddl_path=str(context.problem_path),
)
except TypeError:
vec = tokenizer.transform_goal(list(context.goal_atoms), list(context.objects))
return np.asarray(vec, dtype=np.float32).reshape(-1)
def _predict_next(
self,
source: LoadedSource,
current_vec: np.ndarray,
goal_vec: np.ndarray,
hidden,
) -> tuple[np.ndarray, np.ndarray, object | None]:
if source.model_kind == "xgboost":
x = np.hstack([current_vec.reshape(1, -1), goal_vec.reshape(1, -1)])
pred = np.asarray(source.model.predict(x), dtype=np.float32).reshape(-1)
if source.mode == "delta":
return current_vec + pred, pred, hidden
return pred, pred - current_vec, hidden
return self._predict_lstm(source, current_vec, goal_vec, hidden)
def _predict_lstm(
self,
source: LoadedSource,
current_vec: np.ndarray,
goal_vec: np.ndarray,
hidden,
) -> tuple[np.ndarray, np.ndarray, object | None]:
import torch
device = source.device
state_tensor = (
torch.tensor(current_vec, dtype=torch.float32, device=device)
.reshape(1, 1, -1)
)
goal_tensor = torch.tensor(goal_vec, dtype=torch.float32, device=device).reshape(1, -1)
with torch.inference_mode():
pred, next_hidden = source.model(state_tensor, goal_tensor, hidden=hidden)
pred_vec = pred.detach().cpu().numpy().reshape(-1).astype(np.float32)
if source.mode == "delta":
return current_vec + pred_vec, pred_vec, next_hidden
return pred_vec, pred_vec - current_vec, next_hidden
def _wl_problem(self, source: LoadedSource, context: ProblemContext):
if source.family.tokenizer != "wl" and not isinstance(source.tokenizer, MultiDomainUnionTokenizer):
return None
key = (context.domain, context.split, context.problem)
if key not in self._wl_problem_cache:
try:
from wlplan.planning import parse_problem as wl_parse_problem
self._wl_problem_cache[key] = wl_parse_problem(
str(context.domain_path),
str(context.problem_path),
)
except Exception:
self._wl_problem_cache[key] = None
return self._wl_problem_cache[key]
def save_feature_matrix(
*,
path: str | Path,
candidates: list[dict],
features: np.ndarray,
feature_names: list[str],
) -> None:
"""Save feature arrays and row metadata in one compressed artifact."""
path = Path(path)
path.parent.mkdir(parents=True, exist_ok=True)
np.savez_compressed(
path,
X=features.astype(np.float32),
y=np.asarray([int(row["label_valid"]) for row in candidates], dtype=np.int64),
candidate_ids=np.asarray([row["candidate_id"] for row in candidates], dtype=object),
domains=np.asarray([row["domain"] for row in candidates], dtype=object),
splits=np.asarray([row["split"] for row in candidates], dtype=object),
problems=np.asarray([row["problem"] for row in candidates], dtype=object),
corruption_types=np.asarray(
[row["corruption_type"] for row in candidates],
dtype=object,
),
feature_names=np.asarray(feature_names, dtype=object),
)
def load_feature_matrix(path: str | Path) -> dict:
"""Load a feature matrix artifact with object metadata enabled."""
return dict(np.load(path, allow_pickle=True))
def _read_json(path: Path) -> dict:
if not path.exists():
return {}
with open(path, "r", encoding="utf-8") as f:
return json.load(f)
def _resolve_torch_device(device_arg: str):
import torch
if device_arg == "auto":
if torch.cuda.is_available():
return torch.device("cuda")
if torch.backends.mps.is_available():
return torch.device("mps")
return torch.device("cpu")
if device_arg == "cuda" and not torch.cuda.is_available():
return torch.device("cpu")
if device_arg == "mps" and not torch.backends.mps.is_available():
return torch.device("cpu")
return torch.device(device_arg)
def _summarize_series(values: list[float]) -> list[float]:
if not values:
return [0.0 for _ in SUMMARY_STATS]
arr = np.asarray(values, dtype=np.float32)
return [
float(np.mean(arr)),
float(np.std(arr)),
float(np.min(arr)),
float(np.max(arr)),
float(arr[-1]),
]
def _summarize_lstm_hidden(hidden) -> dict[str, float]:
h, c = hidden
h_cpu = h.detach().float().cpu()
c_cpu = c.detach().float().cpu()
return {
"hidden_h_final_norm": float(h_cpu[-1].norm().item()),
"hidden_c_final_norm": float(c_cpu[-1].norm().item()),
"hidden_h_abs_mean": float(h_cpu.abs().mean().item()),
"hidden_c_abs_mean": float(c_cpu.abs().mean().item()),
}
def _l2(vec: np.ndarray) -> float:
return float(np.linalg.norm(np.asarray(vec, dtype=np.float32).reshape(-1)))
def _cosine_distance(left: np.ndarray, right: np.ndarray) -> float:
u = np.asarray(left, dtype=np.float32).reshape(-1)
v = np.asarray(right, dtype=np.float32).reshape(-1)
denom = float(np.linalg.norm(u) * np.linalg.norm(v))
if denom == 0.0:
return 1.0
return float(1.0 - np.dot(u, v) / denom)
|