Datasets:
File size: 18,755 Bytes
6199b4b | 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 | """Stage-3 hold-out skip pattern generator for the HareSkip experiment.
Offline reproduction of the skip patterns the HareSkip extension would emit at
inference time, using the measured ERSDE-Beta trajectory (t_now per step) from
analysis-phase1-02/outputs/tables/master_long.csv.
Two independent implementations are run and compared:
1) SELF -- a from-scratch re-implementation of the extension logic in this
file (_self_generate), written from reading the source.
2) EXT -- the actual extension modules imported read-only from
forge-neo-Anima-HareSkip (hareskip.skip_pattern).
If they disagree on any candidate, the script aborts. Nothing is written
outside this scratchpad directory; the extension repo and experiment data are
opened read-only.
Usage: python simulate_patterns.py
"""
import csv
import hashlib
import math
import os
import random
import sys
# --- paths (read-only inputs) ----------------------------------------------
EXT_ROOT = r"S:\30_OriginalApps\16_HareSkip\forge-neo-Anima-HareSkip"
EXP_ROOT = r"S:\30_OriginalApps\16_HareSkip\experiment-HareSkip"
MASTER_LONG = os.path.join(
EXP_ROOT, "analysis-phase1-02", "outputs", "tables", "master_long.csv"
)
STAGE2_PATTERNS = os.path.join(
EXP_ROOT, "patterns", "stage2-interaction-patterns.txt"
)
STAGE1_PATTERNS = os.path.join(
EXP_ROOT, "patterns", "layer1-single-skip-sweep-30steps.txt"
)
OUT_DIR = os.path.dirname(os.path.abspath(__file__))
NUM_STEPS = 30
IMAGE_SEEDS = [3000995193, 2455776111, 2767019676]
SKIP_WINDOW = (0.05, 0.95) # extension default
ZONE_BOUNDARIES = (-4.0, 0.0) # extension default
ZONE_MAX_STREAK = {"danger": 1, "middle": 2, "safe": 3}
MODEL = "sigmoid_band_v0.1"
# --- extension import (read-only) ------------------------------------------
sys.path.insert(0, EXT_ROOT)
from hareskip import skip_pattern as ext_sp # noqa: E402
from hareskip import probability_models as ext_pm # noqa: E402
# --- SELF re-implementation -------------------------------------------------
def _sigmoid(x):
if x >= 0.0:
return 1.0 / (1.0 + math.exp(-x))
e = math.exp(x)
return e / (1.0 + e)
def _clamp(x, lo, hi):
return max(lo, min(hi, x))
def self_params(a):
a = _clamp(a, 0.0, 1.0)
return {
"p_cap": 0.40 + 0.40 * a,
"z_enter": -1.8 - 5.0 * (a ** 1.35),
"tau_enter": 0.55 + 0.35 * a,
"z_exit": 4.2 + 1.0 * a,
"tau_exit": 0.45,
}
def self_p(z, prm):
p = (
prm["p_cap"]
* _sigmoid((z - prm["z_enter"]) / prm["tau_enter"])
* _sigmoid((prm["z_exit"] - z) / prm["tau_exit"])
)
return _clamp(p, 0.0, 1.0)
def self_z(t_now, eps=1e-6):
t = max(eps, min(1.0 - eps, t_now))
return 2.0 * math.log((1.0 - t) / t)
def self_zone(z, boundaries=ZONE_BOUNDARIES):
low, high = boundaries
if z < low:
return "danger"
if z < high:
return "middle"
return "safe"
def self_skip_seed(image_seed, offset):
digest = hashlib.sha256(
"{}|hareskip|{}".format(int(image_seed), int(offset)).encode("utf-8")
).hexdigest()
return int(digest, 16) % (2 ** 63)
def self_apply_streak(skip, z_by_step, p_by_step, boundaries=ZONE_BOUNDARIES):
while True:
n = len(skip)
i = 0
violated = None
while i < n:
if not skip[i]:
i += 1
continue
j = i
while j < n and skip[j]:
j += 1
allowed = min(
ZONE_MAX_STREAK[self_zone(z_by_step[k], boundaries)]
for k in range(i, j)
)
if (j - i) > allowed:
violated = (i, j)
break
i = j
if violated is None:
return
start, end = violated
best = None
for k in range(start, end):
key = (p_by_step[k], z_by_step[k], k)
if best is None or key < best[0]:
best = (key, k)
skip[best[1]] = False
def self_generate(t_now_by_step, a, skip_seed,
skip_window=SKIP_WINDOW, boundaries=ZONE_BOUNDARIES):
"""Independent reproduction of hareskip.generate_skip_pattern (no target)."""
n = len(t_now_by_step)
prm = self_params(a)
rng = random.Random(skip_seed)
ws, we = skip_window
z_by_step, p_by_step, skip = [], [], []
for idx, t_now in enumerate(t_now_by_step):
progress = 0.0 if n <= 1 else idx / float(n - 1)
eligible = ws <= progress <= we
z = self_z(t_now)
p = self_p(z, prm) if eligible else 0.0
z_by_step.append(z)
p_by_step.append(p)
# NOTE: the RNG is consumed ONLY for eligible steps -- matches
# _draw_pattern in skip_pattern.py, where rng.random() sits inside the
# `if eligible` conditional expression and is short-circuited otherwise.
skip.append((rng.random() < p) if eligible else False)
self_apply_streak(skip, z_by_step, p_by_step, boundaries)
return skip, z_by_step, p_by_step, prm
# --- trajectory (measured ERSDE-Beta t_now) ---------------------------------
def load_trajectory():
"""Return (t_now_by_step[30], z_by_step[30], audit) from master_long.csv.
master_long uses mapA: skip_step k <-> traj_step k-1, columns `z`/`tnow`.
skip_step 2..30 are present; step 1 (idx 0) has no measured t_now and is
extrapolated (log-linear in z) -- it is outside the skip window and
residual-less at inference, so it can never be skipped either way.
"""
by_step = {}
with open(MASTER_LONG, newline="", encoding="utf-8") as fh:
for r in csv.DictReader(fh):
if r["sampler"] != "ERSDE-Beta":
continue
k = int(float(r["skip_step"]))
rec = (float(r["tnow"]), float(r["z"]), float(r["traj_step"]))
by_step.setdefault(k, set()).add(rec)
audit = {}
conflicts = {k: v for k, v in by_step.items() if len(v) != 1}
audit["steps_present"] = sorted(by_step)
audit["conflicts"] = conflicts
assert not conflicts, "z/tnow differ across ERSDE-Beta conditions: %r" % conflicts
assert audit["steps_present"] == list(range(2, 31))
t_now = [None] * NUM_STEPS
z_meas = [None] * NUM_STEPS
for k, v in by_step.items():
tn, z, _ts = next(iter(v))
# cross-check that the recorded z is exactly the extension's proxy
assert abs(ext_sp.logsnr_proxy_from_t_now(tn) - z) < 1e-9, k
assert abs(self_z(tn) - z) < 1e-9, k
t_now[k - 1] = tn
z_meas[k - 1] = z
# extrapolate step 1 in z (linear backwards from steps 2 and 3)
z1 = z_meas[1] - (z_meas[2] - z_meas[1])
t_now[0] = 1.0 / (1.0 + math.exp(z1 / 2.0))
z_meas[0] = self_z(t_now[0])
audit["z1_extrapolated"] = z_meas[0]
return t_now, z_meas, audit
# --- known patterns to exclude ----------------------------------------------
def load_known():
known = set()
with open(STAGE2_PATTERNS, encoding="utf-8") as fh:
for line in fh:
line = line.strip()
if line:
known.add(frozenset(int(x) for x in line.split(",")))
with open(STAGE1_PATTERNS, encoding="utf-8") as fh:
for line in fh:
line = line.strip()
if line:
known.add(frozenset(int(x) for x in line.split(",")))
return known
# --- helpers ----------------------------------------------------------------
def hamming(a, b):
"""Hamming distance between two skip sets over the 30-step vector."""
return len(set(a) ^ set(b))
def max_streak(skip):
best = cur = 0
for s in skip:
cur = cur + 1 if s else 0
best = max(best, cur)
return best
def zone_breakdown(steps, z_by_step):
out = {"danger": 0, "middle": 0, "safe": 0}
for s in steps:
out[self_zone(z_by_step[s - 1])] += 1
return out
def make_candidate(t_now, z_by_step, a, image_seed, offset, checks):
"""Generate one candidate with SELF and cross-check against EXT."""
seed = self_skip_seed(image_seed, offset)
assert seed == ext_sp.derive_skip_seed(image_seed, offset)
s_skip, s_z, s_p, s_prm = self_generate(t_now, a, seed)
e_pat = ext_sp.generate_skip_pattern(
t_now, aggressiveness=a, skip_seed=seed,
probability_model=MODEL,
skip_window=SKIP_WINDOW, zone_boundaries=ZONE_BOUNDARIES,
)
# --- cross-implementation verification -------------------------------
assert s_prm == e_pat.params, (s_prm, e_pat.params)
assert all(abs(x - y) < 1e-12 for x, y in zip(s_z, e_pat.z_by_step))
assert all(abs(x - y) < 1e-12 for x, y in zip(s_p, e_pat.p_by_step))
assert s_skip == e_pat.skip, (a, image_seed, offset)
checks["compared"] += 1
steps = [i + 1 for i, s in enumerate(s_skip) if s]
assert steps == e_pat.skipped_steps
return {
"a": a,
"image_seed": image_seed,
"offset": offset,
"skip_seed": seed,
"steps": steps,
"count": len(steps),
"zones": zone_breakdown(steps, s_z),
"max_streak": max_streak(s_skip),
"expected": e_pat.expected_skips_before_streak,
"manual_added": [],
}
def select_for_level(a, t_now, z_by_step, known, chosen_so_far, checks,
n_want=3, max_offset=400):
"""Pick n_want candidates at aggressiveness a satisfying the constraints."""
picked = []
for offset in range(0, max_offset):
for image_seed in IMAGE_SEEDS:
cand = make_candidate(t_now, z_by_step, a, image_seed, offset, checks)
steps = cand["steps"]
if len(steps) < 2:
continue
key = frozenset(steps)
if key in known:
continue
if any(key == frozenset(p["steps"]) for p in picked + chosen_so_far):
continue
if any(hamming(steps, p["steps"]) < 3
for p in picked + chosen_so_far):
continue
picked.append(cand)
if len(picked) == n_want:
return picked
raise RuntimeError("could not find %d candidates at a=%s" % (n_want, a))
# --- main -------------------------------------------------------------------
def main():
t_now, z_meas, audit = load_trajectory()
known = load_known()
checks = {"compared": 0}
chosen = []
for a in (0.3, 0.6, 0.9):
picked = select_for_level(a, t_now, z_meas, known, chosen, checks)
chosen.extend(picked)
# --- extreme stress pattern: a=0.9 realisation + manual early steps ----
base = chosen[-1] # last a=0.9 pattern
stress_steps = sorted(set(base["steps"]) | {3, 5})
stress = dict(base)
stress["steps"] = stress_steps
stress["count"] = len(stress_steps)
stress["zones"] = zone_breakdown(stress_steps, z_meas)
sv = [False] * NUM_STEPS
for s in stress_steps:
sv[s - 1] = True
stress["max_streak"] = max_streak(sv)
stress["manual_added"] = [3, 5]
stress["label"] = "stress"
chosen.append(stress)
# --- final validation --------------------------------------------------
problems = []
if len(chosen) != 10:
problems.append("expected 10 patterns, got %d" % len(chosen))
seen = set()
for i, p in enumerate(chosen):
st = p["steps"]
if st != sorted(st):
problems.append("pattern %d not sorted" % i)
if not all(2 <= s <= 30 for s in st):
problems.append("pattern %d out of range 2..30: %r" % (i, st))
if 1 in st:
problems.append("pattern %d contains step 1" % i)
k = tuple(st)
if k in seen:
problems.append("duplicate pattern %d" % i)
seen.add(k)
if frozenset(st) in known:
problems.append("pattern %d collides with stage1/2" % i)
# pairwise hamming among the 9 simulated (stress excluded by construction)
for i in range(9):
for j in range(i + 1, 9):
d = hamming(chosen[i]["steps"], chosen[j]["steps"])
if d < 3:
problems.append("hamming(%d,%d)=%d < 3" % (i, j, d))
# --- write outputs -----------------------------------------------------
txt = os.path.join(OUT_DIR, "stage3-holdout-patterns.txt")
with open(txt, "w", encoding="utf-8", newline="\n") as fh:
for p in chosen:
fh.write(", ".join(str(s) for s in p["steps"]) + "\n")
write_report(chosen, audit, checks, problems, z_meas, known)
print("cross-implementation comparisons: %d (all identical)" % checks["compared"])
print("validation problems:", problems or "none")
for i, p in enumerate(chosen, 1):
print("%2d a=%.1f seed=%d off=%d n=%2d %s%s" % (
i, p["a"], p["image_seed"], p["offset"], p["count"],
p["steps"], " (+manual 3,5)" if p["manual_added"] else ""))
def write_report(chosen, audit, checks, problems, z_meas, known):
lines = []
A = lines.append
A("# Stage-3 hold-out skip patterns -- generation report")
A("")
A("Generated by `simulate_patterns.py` (this directory). Inputs are read-only:")
A("")
A("- Extension logic: `%s\\hareskip\\{skip_pattern,probability_models}.py`" % EXT_ROOT)
A("- Trajectory: `%s`" % MASTER_LONG)
A("- Exclusion sets: `%s`, `%s`" % (STAGE2_PATTERNS, STAGE1_PATTERNS))
A("")
A("## Configuration")
A("")
A("| item | value |")
A("|---|---|")
A("| num_steps | %d |" % NUM_STEPS)
A("| probability model | `%s` |" % MODEL)
A("| skip_window | %s (extension default) |" % (SKIP_WINDOW,))
A("| zone_boundaries | %s (extension default) |" % (ZONE_BOUNDARIES,))
A("| zone max streak | danger 1 / middle 2 / safe 3 |")
A("| skip_seed | `sha256(f\"{image_seed}|hareskip|{offset}\")` mod 2**63 |")
A("| image seeds | %s |" % ", ".join(str(s) for s in IMAGE_SEEDS))
A("")
A("## Verification")
A("")
A("Two independent implementations were compared on **every candidate drawn**")
A("(not just the 9 selected ones):")
A("")
A("1. `self_generate` in this script -- written from scratch by reading the")
A(" extension source (probability formula, window eligibility, z proxy,")
A(" zone/streak trimming with the `(p, z, index)` argmin flip rule, and the")
A(" RNG consumption order: `rng.random()` is drawn **only for eligible")
A(" steps**, because in `_draw_pattern` the call sits inside the")
A(" `if eligible` conditional expression).")
A("2. `hareskip.skip_pattern.generate_skip_pattern` imported directly from the")
A(" extension repo (`sys.path.insert`, read-only; nothing installed or")
A(" modified).")
A("")
A("Compared per candidate: `params`, `z_by_step`, `p_by_step`, the boolean")
A("`skip` vector and `skipped_steps`. Any mismatch aborts the script.")
A("")
A("- Candidates cross-checked: **%d** -- all identical." % checks["compared"])
A("- `derive_skip_seed` also cross-checked per candidate (SHA-256, not `hash()`).")
A("- The extension's own suite `tests/test_skip_pattern.py` was run against the")
A(" same modules: **42 passed** (including the pinned literal")
A(" `derive_skip_seed(12345, 0) == 3695650839502921262`).")
A("")
A("Validation of the final 10 patterns: **%s**" % (
"; ".join(problems) if problems else "no problems (all steps in 2..30, "
"no step 1, no duplicates, no collision with stage-1/2, pairwise "
"Hamming >= 3 among the 9 simulated)"))
A("")
A("## z(step) correspondence")
A("")
A("`master_long.csv` columns `z` / `tnow` follow **mapA**")
A("(`traj_step = skip_step - 1`), the default in analysis-phase1-02/REPORT.md.")
A("For the ERSDE-Beta sampler all **15 conditions carry identical `z`/`tnow`**")
A("for every `skip_step` 2..30 (verified: zero conflicting values).")
A("Each recorded `z` reproduces `logsnr_proxy_from_t_now(tnow)` to < 1e-9, so")
A("the measured `tnow` column is fed straight into the extension as")
A("`t_now_by_step`.")
A("")
A("Step 1 has no measured trajectory row (skip_step starts at 2); its z is")
A("extrapolated linearly backwards to z = %.4f. It is irrelevant to the" % audit["z1_extrapolated"])
A("result: idx 0 has progress 0.0 < 0.05 so it is outside the skip window,")
A("and at inference the first call has no residual, so it is never skipped.")
A("")
A("| step | t_now | z | zone |")
A("|---:|---:|---:|---|")
for i in range(NUM_STEPS):
z = z_meas[i]
A("| %d%s | %.9f | %.5f | %s |" % (
i + 1, " (extrap.)" if i == 0 else "",
1.0 / (1.0 + math.exp(z / 2.0)), z, self_zone(z)))
A("")
A("## Patterns")
A("")
A("| # | a | image_seed | offset | skip_seed | n | skipped steps | danger | middle | safe | max streak | E[skips] |")
A("|---:|---:|---:|---:|---|---:|---|---:|---:|---:|---:|---:|")
for i, p in enumerate(chosen, 1):
note = " **(+manual 3, 5)**" if p["manual_added"] else ""
A("| %d | %.1f | %d | %d | %d | %d | %s%s | %d | %d | %d | %d | %.2f |" % (
i, p["a"], p["image_seed"], p["offset"], p["skip_seed"], p["count"],
", ".join(str(s) for s in p["steps"]), note,
p["zones"]["danger"], p["zones"]["middle"], p["zones"]["safe"],
p["max_streak"], p["expected"]))
A("")
A("Rows 1-3: a=0.3. Rows 4-6: a=0.6. Rows 7-9: a=0.9. Row 10: extreme-stress.")
A("")
A("### Row 10 (extreme stress)")
A("")
st = chosen[-1]
A("Row 10 is **not** a plain simulator output. It takes the realised a=0.9")
A("pattern of row 9 (image_seed %d, offset %d) and **manually adds steps 3")
A("and 5**. Those two steps sit in the danger band (z = %.2f and %.2f, both" % (
z_meas[2], z_meas[4]))
A("z < -4), which the probability model almost never selects and where the")
A("danger-zone streak cap is 1 -- so this row is a deliberate out-of-sample")
A("probe of the prediction formula's limit, not a reachable extension output.")
A("Its max streak (%d) may therefore exceed what the extension would emit." % st["max_streak"])
A("")
A("## Output grammar")
A("")
A("`stage3-holdout-patterns.txt` -- 10 lines, one pattern per line,")
A("comma-separated 1-based step numbers in ascending order; same grammar as")
A("`stage2-interaction-patterns.txt` and the stage-1 sweep file.")
A("")
with open(os.path.join(OUT_DIR, "generation_report.md"), "w",
encoding="utf-8", newline="\n") as fh:
fh.write("\n".join(lines))
if __name__ == "__main__":
main()
|