Spaces:
Running
Running
File size: 22,786 Bytes
a1dd5ba 698d69f a1dd5ba 698d69f a1dd5ba 698d69f | 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 | """QUERY BY EXAMPLE: clips retrieve their own kind, with no text at all.
WHY THIS PATH EXISTS
--------------------
Direction lives in a verb. On a corpus that names nothing - which is what
the no-text-identity rule requires - there is no way to bridge "opens" to
a nameless transition family without a hand-written table or write-time
naming, and both are forbidden. A clip needs no bridge: it IS an instance
of what it shows.
Measured against the text path on the same truthset and the same metric
(yield = true/support, prec = true/returned, k = ceil(1.5 x support)):
text yield qbe yield text prec_g qbe prec_g
q04 close 0.39 0.88 0.89 0.91
q05 open 0.29 0.86 0.86 0.99
q03 stove 0.38 0.67 0.79 0.83
Text still wins nothing here, but it is NOT redundant: q03 is an
appearance query, and its gap is much smaller than the direction pair's.
Keep both - text for what a corpus can name, example for what it can
only show.
THE THREE THINGS THAT MADE IT WORK
----------------------------------
1. MULTI-SEED. One seed is a sample of one; the centroid of 8-10 known-
same-kind clips is the query. Measured 0.54 -> 0.68 on q04 from this
alone. More is not better: 20 seeds over-averages and blurs the query
back down (q05 0.79 -> 0.69).
2. DROP THE COMMON COMPONENT. Every pretrained encoder is nearly
collapsed on a fixed-camera single-scene corpus - measured effective
rank 16.7 of 1152 dims for SigLIP2, 19.5 of 1024 for PE-Core, 13.8 of
512 for InternVideo2. Within-support cosine 0.88-0.97 against between
0.78-0.86: everything looks like everything. The first principal
component is what every clip shares (the kitchen, the arm) and it
dominates every cosine while carrying no information about WHICH clip
this is. Removing exactly it: q04 0.83 -> 0.88, q05 0.79 -> 0.85.
Removing two or more is catastrophic (0.47) - PC2 onward is signal.
3. SEED-COHERENCE WEIGHTING. Seven channels voting equally while their
QbE quality ranges AUC 0.50 to 0.91 means the good one is outvoted -
fused q04 scored 0.37 where mot alone scored 0.675. QbE has a signal
the text path lacks: THE SEEDS ARE KNOWN TO BE THE SAME KIND, so a
channel that pulls them together relative to how it holds the corpus
is measuring what they share. That is Cohen's d computed from the
query itself, no truthset. It picks correctly and unsupervised: iv2
first for the appearance query, mot first for both direction queries.
WHAT IS STILL MISSING is recall, not precision: yield 0.86-0.88 on the
direction pair with prec_g 0.91-0.99. The per-model evidence says why -
only `mot` (a DIFFERENCE of appearance, and the only channel costing no
model at all) separates direction, and only `iv2` separates appearance.
The other five are collapsed enough to be near-noise on this corpus.
"""
from __future__ import annotations
import os
import numpy as np
# tuned on the truthset as a MEASUREMENT, not fitted: the plateau is flat
# (mean yield 0.79-0.80 across seeds 8-15 and power 6-12), so these are
# the middle of a wide optimum rather than a fitted point.
SEEDS = 10
# POWER is tuned to the coherence statistic, and the two must move
# together: 8.0 was right for an unbounded Cohen's d and crushes a
# bounded [0,1] rank statistic to zero (0.42**8 = 0.001, every channel
# silenced). Re-measured for the bounded form, the optimum is 2 and the
# plateau is flat from 2 to 4.
POWER = 2.0
# fusion sharpness on the pairwise quality; swept 2..64, flat 8-12
ZPOWER = 8.0
# ZERO, re-measured under LOO weighting + z-fusion. Removing the first
# principal component was worth q04 0.83 -> 0.88 under coherence+RRF,
# because RRF sees only ranks and the shared component (the kitchen,
# the arm) dominated every cosine equally. z-fusion standardises each
# channel before combining, which handles that shift directly - and the
# subtraction then costs signal instead of noise: q04 0.93 -> 0.87.
# A correction for a defect that no longer exists is just a distortion.
DROP_PC = int(os.environ.get("ELIDEDB_DROP_PC", "0"))
def _pooled(store, table):
"""{(stream, episode_ts) -> unit vector}, mean-pooled per episode.
Two row conventions exist and both must pool correctly. pe/sig2
write 8 rows per episode all carrying the EPISODE's ts, so grouping
by (stream, ts) is already per-episode. A per-frame table
(scene_vectors) carries each FRAME's ts - grouping by it yields one
"episode" per frame, and when spaces() then joins on episode starts,
the channel silently becomes first-frame-only: 2,097 of 70,436 rows
used, 97% of the signal dropped with no error. So keys that don't
match an episode start are binned into the episode whose span
contains them, using the episodes table - the same containment the
labels join uses.
"""
from .embeddings import _vec_table
tb, V = _vec_table(store, table)
V = np.asarray(V, np.float32)
ep = store.table("episodes").scan()
spans = {}
for s, a, b in zip(ep.column("stream").to_pylist(),
ep.column("ts").to_pylist(),
ep.column("t1").to_pylist()):
spans.setdefault(str(s), []).append((int(a), int(b)))
for v in spans.values():
v.sort()
starts = {s: [a for a, _ in v] for s, v in spans.items()}
acc = {}
for s, a, v in zip(tb.column("stream").to_pylist(),
tb.column("ts").to_pylist(), V):
s, a = str(s), int(a)
sp = spans.get(s)
if sp:
j = int(np.searchsorted(starts[s], a, "right")) - 1
if j >= 0 and a <= sp[j][1]:
a = sp[j][0]
acc.setdefault((s, a), []).append(v)
out = {}
for k, vs in acc.items():
m = np.mean(vs, 0)
n = np.linalg.norm(m)
out[k] = (m / n) if n > 0 else m
return out
# EACH MODEL IS SCORED THE WAY IT WAS BUILT TO BE SCORED.
#
# Everything here was originally scored by cosine, which is the native
# operation for exactly one of these channels. `act` is a 174-way SSv2
# CLASSIFIER: its rows are a probability distribution over actions, and
# cosine on a simplex is not a comparison of distributions. Measured, the
# difference is not marginal - on q03 cosine scored AUC 0.483, BELOW
# CHANCE, while cross-entropy on the same numbers scores 0.812:
#
# operator q03 q04 q05
# cosine (was) 0.483 0.561 0.622
# cross-entropy 0.812 0.641 0.764
# neg-JS 0.793 0.641 0.750
#
# Cross-entropy log(p_episode) . p_seed is the likelihood of the
# episode's posterior under the seed's, which is what "does this clip
# show the same action" means for a classifier.
#
# PCA whitening is also wrong for a simplex - subtracting a mean and
# renormalising a probability vector produces something that is no longer
# a distribution - so drop_pc is not applied to classifier channels.
# The OPERATOR COMES FROM THE TABLE, not from a name-keyed dict in
# code. A dict here ("act" -> crossent) was a channel name hardwired in
# the engine - the same private-knowledge pattern as a channel list. A
# classifier channel (rows on a probability simplex) declares
# meta={"op": "crossent"} when it is built; everything else is cosine
# by default. The retired action_probs is the only crossent channel to
# date, and any rebuild of it must carry the declaration.
def spaces(store, drop_pc=DROP_PC):
"""Every per-episode table as a matrix aligned with episodes, plus
the operator each one should be scored with."""
ep = store.table("episodes").scan()
keys = [(str(s), int(a)) for s, a in
zip(ep.column("stream").to_pylist(), ep.column("ts").to_pylist())]
pos = {k: i for i, k in enumerate(keys)}
M = {}
for t in store.tables():
if not (t.endswith("_vectors") or t == "action_probs"):
continue
if t == "frame_vectors":
continue # per-frame, not per-episode
# A CHANNEL IS SOMETHING A MODEL PRODUCED. Selecting on the name
# suffix alone silently enrolled object_vectors - the identity
# index - as an eighth retrieval channel the moment that table
# was first persisted, diluting the fusion with ReID appearance
# descriptors that exist to answer "is this the same object",
# not "does this episode match the query". Every encoder output
# already declares its `model`; an index declares a cut instead,
# so the distinction is recorded rather than guessed from a name.
meta = store.table(t).state().meta or {}
if not meta.get("model"):
continue
try:
p = _pooled(store, t)
except Exception:
continue
if not p:
continue
dim = len(next(iter(p.values())))
A = np.zeros((len(keys), dim), np.float32)
ok = np.zeros(len(keys), bool)
for k, v in p.items():
if k in pos:
A[pos[k]] = v
ok[pos[k]] = True
name = t.replace("_vectors", "").replace("action_probs", "act")
op = (meta or {}).get("op", "cosine")
if drop_pc and op != "crossent" and ok.sum() > drop_pc + 5:
X = A[ok]
mu = X.mean(0)
_, _, Vt = np.linalg.svd(X - mu, full_matrices=False)
P = Vt[:drop_pc]
A = A - mu
A = A - (A @ P.T) @ P
A = A / np.maximum(np.linalg.norm(A, axis=1, keepdims=True), 1e-8)
A[~ok] = 0
M[name] = (A.astype(np.float32), ok, op)
return keys, M
def _score(A, op, seeds):
"""Score every episode against the seeds, the model's own way."""
if op == "crossent":
pm = A[seeds].mean(0)
return np.log(A + 1e-12) @ pm
cen = A[seeds].mean(0)
n = float(np.linalg.norm(cen))
# A seed row can be absent from a channel (the store holds 1,121 of
# 1,122 episodes for sig2 and iv2), and leave-one-out scores a single
# seed at a time, so a degenerate centroid is reachable in normal
# operation. `n > 0` admitted denormals, and the division then raised
# divide-by-zero and overflow on every query. The scores it produced
# were masked downstream by the isfinite check, so the ranking never
# changed - but a query that prints three numpy warnings reads as
# broken, and a guard that only mostly holds is not a guard.
if not np.isfinite(n) or n <= 1e-12:
return np.zeros(len(A))
return A @ (cen / n)
def _pair(A, op, rows, cols):
"""Pairwise affinity between two row sets, the model's own way. The
OPERATOR MUST GOVERN THE WEIGHT TOO. Scoring `act` by cross-entropy
while weighting it by cosine measured a REGRESSION (q03 0.67 -> 0.60)
even though the channel's own AUC had risen 0.483 -> 0.812: the
channel became right and its voice became wrong. Half-applying the
principle is worse than not applying it."""
if op == "crossent":
return np.log(A[cols] + 1e-12) @ A[rows].T # cols x rows
return A[cols] @ A[rows].T
def coherence(A, ok, seeds, op="cosine", cap=4000):
"""How reliably this model rates a seed pair above a seed/corpus
pair - as a probability, in the model's own metric.
Cohen's d was the first form and it does not survive mixed
operators: cross-entropy is a log-likelihood with a far wider spread
than cosine, so `d` for a classifier and `d` for an embedding are
different units. Raising incomparable numbers to the 8th power then
decides the whole fusion, and measured it silenced `act` entirely
(q03 0.67 -> 0.60) at the moment the channel started scoring well.
This is the probability that a random seed-seed affinity exceeds a
random seed-corpus one - a rank statistic, bounded [0, 1], identical
in meaning whether the model returns a cosine or a log-likelihood.
0.5 is "this model cannot tell the seeds from the corpus".
"""
if len(seeds) < 2:
return 0.0
other = np.setdiff1d(np.where(ok)[0], seeds)
if len(other) < 10:
return 0.0
W = _pair(A, op, seeds, seeds)
iu = np.triu_indices(len(seeds), 1)
within = W[iu]
B = _pair(A, op, seeds, other).ravel()
if len(B) > cap: # rank stat needs no more
B = B[np.linspace(0, len(B) - 1, cap).astype(int)]
# P(within > between), ties at half
allv = np.concatenate([within, B])
r = np.empty(len(allv))
r[np.argsort(allv)] = np.arange(len(allv))
rw = r[:len(within)].sum()
u = rw - len(within) * (len(within) - 1) / 2
p = u / (len(within) * len(B))
return float(max(p - 0.5, 0.0) * 2.0) # 0 at chance, 1 at perfect
def otsu_cut(sorted_scores, k_max):
"""Where the returned set ENDS: the valley between the two modes.
setpath's confidence_cut fits a knee against RRF's scale and, on
z-fused scores, fires almost immediately: measured 16 returned of a
165-clip support (yield 0.08 at precision 0.94). It is not wrong,
it is calibrated for a different score shape.
A good query's score curve is bimodal - the clips of its kind, and
the corpus - so the cut is the threshold that best SEPARATES those
two populations. Otsu's method finds it by maximising between-class
variance, which is parameter-free and reads only this query's own
scores: no truthset, no fitted alpha, nothing per-corpus.
Degenerate cases fail open (return k_max) rather than returning a
handful, because a flat curve means "cannot tell", and answering
almost nothing is a worse response to that than answering fully.
"""
s = np.asarray(sorted_scores, float)
s = s[np.isfinite(s)]
n = min(len(s), int(k_max))
if n < 4:
return int(n)
head = s[:max(n * 4, 200)][:len(s)]
lo, hi = float(head.min()), float(head.max())
if not np.isfinite(lo) or hi - lo < 1e-9:
return int(n)
hist, edges = np.histogram(head, bins=64, range=(lo, hi))
p = hist.astype(float) / max(hist.sum(), 1)
w0 = np.cumsum(p)
mids = (edges[:-1] + edges[1:]) / 2
mu = np.cumsum(p * mids)
mu_t = mu[-1]
with np.errstate(divide="ignore", invalid="ignore"):
between = (mu_t * w0 - mu) ** 2 / (w0 * (1 - w0))
between[~np.isfinite(between)] = -1.0
thr = float(mids[int(between.argmax())])
keep = int((s[:n] >= thr).sum())
return int(min(max(keep, 1), n)) if keep else int(n)
# The LOO recall ladder, as FRACTIONS of the corpus. The first form of
# this was absolute depths (25..800), which are the right ladder for a
# 2,097-episode corpus and silently the wrong one for any other size -
# corpus-SHAPE knowledge baked into the engine, caught in audit. The
# geometric ladder itself is the depth-free construction; only its
# anchor had to become relative.
LOO_FRACS = (1 / 80, 1 / 40, 1 / 20, 1 / 10, 1 / 5, 2 / 5)
def loo_depths(n):
return [max(2, int(round(f * n))) for f in LOO_FRACS]
def loo_quality(A, ok, seeds, op):
"""LEAVE-ONE-SEED-OUT retrieval quality of a channel. Query-only.
Hold out one seed, query with the rest, record where the held-out
seed lands; the statistic is its mean recall over a ladder of
depths, which is the area under the LOO recall curve and so needs
no chosen depth.
This replaces coherence as the WEIGHT because coherence measures
the wrong thing. Cohen's d asks "are the seeds tight?"; measured
over 15 seed groups it named the wrong channel 4 times, because
`motion` held yield 0.88-0.95 while its d fell from 2.5 to 0.76.
Tightness is not retrieval. Selecting on LOO instead moved the
high-support queries from 0.742 to 0.845 mean yield.
Legal by the same argument coherence was: the seeds ARE the query.
Their mutual membership is what the user asserted by supplying
them, not something read from the truthset.
A single depth would have been a tuned constant - it swung the
result from 0.598 at 50 to 0.811 at 200 - so the ladder is not a
refinement, it is what keeps this honest.
"""
if len(seeds) < 3 or not ok[seeds].all():
return 0.0
# PAIRWISE, not hold-one-out: with 5 seeds a hold-one-out estimate
# has 5 samples to judge a channel on, and it mis-selected about one
# seed group in five - the residual loss on the appearance query.
# Every ordered pair gives 20 samples from the same query, and the
# question each asks is the same one: retrieve with this, does the
# query's own other member come back?
depths = loo_depths(len(ok))
out = []
for i in seeds:
others = np.asarray([x for x in seeds if x != i])
for j in others:
sc = _score(A, op, np.asarray([j]))
sc[~ok] = -np.inf
sc[others] = -np.inf # other seeds not candidates
r = int((sc > sc[i]).sum())
out.append(float(np.mean([r < d for d in depths])))
return float(np.mean(out))
def search_like(store, seed_keys, k_max=50, power=ZPOWER, drop_pc=DROP_PC):
"""Episodes of the same kind as `seed_keys` [(stream, ts), ...].
Returns {"clips": [(stream, ts), ...], "weights": {channel: w}}.
The weights are the diagnosis: they say which model recognised the
seeds, and are worth surfacing rather than hiding.
Fusion is at SCORE level, not rank level. RRF knows only that a
channel put something first, never by how much, and the margin is
exactly what separates a channel that recognises the query from one
that is guessing: z-fusion measured 0.845 against RRF's 0.742 on
the same weights and seeds.
"""
keys, M = spaces(store, drop_pc)
pos = {k: i for i, k in enumerate(keys)}
seeds = np.array(sorted({pos[k] for k in seed_keys if k in pos}))
if not len(seeds):
return {"clips": [], "weights": {}, "note": "no seed in store"}
# THE WEIGHTING LADDER IS BOUNDED BY SEED COUNT, and must degrade
# rather than refuse. loo_quality needs three seeds to form ordered
# pairs and coherence needs two; below that there is no statistic
# the query itself supplies, and the first form of this returned
# NOTHING - a one-seed query read as "no channel responded", which
# on a working index is indistinguishable from a broken one. It is
# the weighting that is unavailable, not the retrieval: the
# centroid of one seed is still a query. So the fallback is an
# unweighted vote, which is measurably blunter (equal weights let
# five near-noise channels outvote the one that recognises the
# query) but is an answer, and the note says which rung ran.
if len(seeds) >= 3:
q = {c: max(loo_quality(A, ok, seeds, op), 0.0)
for c, (A, ok, op) in M.items()}
note = ""
elif len(seeds) == 2:
q = {c: max(coherence(A, ok, seeds, op), 0.0)
for c, (A, ok, op) in M.items()}
note = ("Two seeds: channels weighted by seed coherence, not by "
"leave-one-out retrieval. Add a third for the sharper "
"weighting.")
else:
q = {c: 1.0 for c in M}
note = ("One seed: every channel votes equally, because nothing "
"in a single example says which model recognised it. "
"Add two more seeds to weight them.")
mx = max(q.values()) if q else 0.0
if mx <= 0:
# Every channel at chance. Still not a reason to return nothing.
q = {c: 1.0 for c in M}
mx = 1.0
note = ("No channel separated these seeds from the corpus, so "
"all of them vote equally - treat this ranking as weak.")
w = {c: (v / mx) ** power for c, v in q.items()}
tot = None
for c, (A, ok, op) in M.items():
if w.get(c, 0.0) <= 0:
continue
live = np.array([i for i in seeds if ok[i]])
if not len(live):
continue
sc = _score(A, op, live)
fin = np.isfinite(sc) & ok
if fin.sum() < 3:
continue
z = np.zeros(len(sc))
mu, sd = sc[fin].mean(), sc[fin].std() or 1e-9
z[fin] = (sc[fin] - mu) / sd
z[~fin] = -np.inf
v = w[c] * z
tot = v if tot is None else tot + v
if tot is None:
return {"clips": [], "weights": w, "note": "no channel responded"}
tot[seeds] = -np.inf # the seeds are given, not found
order = np.argsort(-tot)
# WHERE THE SET ENDS: k is a CEILING, not a target - the contract
# is "up to k, and returned means believed". The return size is
# calibrated by the query itself: hold each seed out, fuse with the
# rest, and record where the held-out TRUE item RANKED; the deepest
# such rank says how far down truth is known to live, and the
# maximum-of-uniforms correction (m+1)/m undoes the bias of taking
# a max over only m samples. Order statistics of the query's own
# members - no truthset, no curve-shape assumption, no knob.
# Measured against returning k always: prec +0.03/+0.08/+0.14 on
# the high-support queries for yield -0.02/-0.07/-0.05, and ret
# lands at 1.1-1.3x support instead of pinned to 1.5x. On weak
# queries the held-out ranks are deep and the cut correctly fails
# open to k. Fewer than 3 seeds cannot calibrate; otsu_cut is the
# fallback there.
ranks = []
if len(seeds) >= 3:
for i in seeds:
rest = np.asarray([j for j in seeds if j != i])
ti = None
for c, (A, ok, op) in M.items():
if w.get(c, 0.0) <= 0:
continue
live = np.array([x for x in rest if ok[x]])
if len(live) < 2:
continue
sc = _score(A, op, live)
fin = np.isfinite(sc) & ok
if fin.sum() < 3:
continue
z = np.zeros(len(sc))
mu, sd = sc[fin].mean(), sc[fin].std() or 1e-9
z[fin] = (sc[fin] - mu) / sd
z[~fin] = -np.inf
v = w[c] * z
ti = v if ti is None else ti + v
if ti is None:
continue
ti[rest] = -np.inf
ranks.append(int((ti > ti[int(i)]).sum()))
if ranks:
cut = int(round(max(ranks) * (len(ranks) + 1) / len(ranks)))
cut = max(1, min(cut, k_max))
else:
cut = otsu_cut(tot[order], k_max)
return {"clips": [keys[i] for i in order[:cut]],
"weights": {c: round(x, 3) for c, x in
sorted(q.items(), key=lambda kv: -kv[1])},
"note": note}
|