Spaces:
Running
Running
File size: 23,098 Bytes
a1dd5ba | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 | """Post-training cellular turnover for the context tower β PPO + reverse
attention, ported from FDNN_BrainModel/models/pruning.py.
WHY PRUNE AT ALL
----------------
Because the only thing this index is for is answering fast. Every channel the
tower carries is multiplied over T timesteps, for every window, at every
rebuild. Capacity that is not earning its keep is latency the user pays for
forever. And with a few hundred labelled windows, an over-provisioned tower
is also memorising β so apoptosis buys accuracy and speed at the same time,
which is exactly the regime cellular turnover was designed for.
WHAT CHANGED FROM FDNN, AND WHY
-------------------------------
FDNN prunes per (layer, neuron): every layer has its own aliveness mask. That
is right for a plain stack. This tower is a RESIDUAL stack, and in a residual
stack channel j is a wire that runs the whole depth β masking it in layer 2
alone does not remove it, because the skip connection keeps carrying it. So
the prune unit here is the residual-stream CHANNEL, decided once and applied
to every block.
That is not a weakening of the idea, it is the correct form of it for this
topology (the same structured channel pruning people apply to ResNets), and
it has a property per-layer masking does not: a dead channel can be
physically deleted. `compact()` rebuilds the tower without it, so the mask
turns into real wall-clock speed instead of a multiply by zero. Pruning that
does not shrink the matmul is not pruning, it is decoration.
THE THREE MEASURES (unchanged in spirit from FDNN)
--------------------------------------------------
utilization Ξ validation RETRIEVAL loss when the channel is silenced.
Not weight magnitude β the thing we actually care about.
reverse attention softmax(-utilization): mass on the channels the network
attends AWAY from, i.e. the prune candidates.
PPO a clipped-PPO contextual bandit over per-channel features
that samples masks, scores them by
-(loss ratio) - λ·kept_fraction, and learns which channels
can actually go.
"""
from __future__ import annotations
import mlx.core as mx
import mlx.nn as nn
import mlx.optimizers as optim
import numpy as np
# ===========================================================================
# 0. Channel mask plumbing
# ===========================================================================
def n_channels(model):
return model.blocks[0].max_out_d
def get_channel_mask(model):
return np.array(model.blocks[0].mask).copy()
def set_channel_mask(model, mask):
"""One decision, applied to every block β see module docstring."""
m = np.asarray(mask, dtype=np.float32)
for b in model.turnover_layers():
b.set_active_mask(m)
# ===========================================================================
# 1. Utilization ("consumption")
# ===========================================================================
def measure_utilization(model, loss_of, verbose=False):
"""Per-channel ablation importance + the cheap descriptive stats.
`loss_of(model)` must return the validation retrieval loss as a float.
Ablation is the increase in that loss when a single channel is silenced
everywhere β the most directly interpretable notion of how much the tower
relies on it.
"""
model.set_training(False)
base = loss_of(model)
mask = get_channel_mask(model)
n = len(mask)
abl = np.zeros(n, dtype=np.float32)
for c in np.where(mask == 1.0)[0]:
probe = mask.copy()
probe[c] = 0.0
set_channel_mask(model, probe)
abl[c] = max(loss_of(model) - base, 0.0)
set_channel_mask(model, mask)
return {"mask": mask, "ablation": abl, "base_loss": base}
def descriptive_stats(model, X):
"""Activation magnitude per channel (averaged over blocks) and the
downstream weight norm each channel feeds."""
model.set_training(False)
outs = []
h = mx.array(X)
for b in model.turnover_layers():
outs.append(np.array(mx.mean(mx.abs(b.neuron_outputs(h)), axis=(0, 1))))
h = b(h)
act = np.mean(np.stack(outs), axis=0)
n = n_channels(model)
wf = np.array(model.gru_f.Wx) # (3*hidden, n_hidden)
wb = np.array(model.gru_b.Wx)
hw = np.array(model.head.weight) # (d_out, n_hidden + 2*gru_hidden)
pw = np.array(model.pool.proj.weight)
down = (np.linalg.norm(wf, axis=0) + np.linalg.norm(wb, axis=0)
+ np.linalg.norm(hw[:, :n], axis=0)
+ np.linalg.norm(pw[:, :n], axis=0))
return act.astype(np.float32), down.astype(np.float32)
# ===========================================================================
# 2. Reverse attention
# ===========================================================================
def reverse_attention(importance, mask, temperature=1.0):
"""Forward attention concentrates on useful units; reverse attention
inverts it so the LEAST useful surface as prune candidates. Dead channels
score 0; scores over the live set sum to 1."""
importance = np.asarray(importance, dtype=np.float64)
alive = mask == 1.0
r = np.zeros_like(importance)
if alive.sum() == 0:
return r.astype(np.float32)
imp = importance[alive]
imp = (imp - imp.mean()) / (imp.std() + 1e-8)
logits = -imp / max(temperature, 1e-6)
logits -= logits.max()
e = np.exp(logits)
r[alive] = e / (e.sum() + 1e-12)
return r.astype(np.float32)
# ===========================================================================
# 3. PPO agent
# ===========================================================================
class PrunePolicy(nn.Module):
def __init__(self, feat_dim, hidden=32):
super().__init__()
self.fc1 = nn.Linear(feat_dim, hidden)
self.fc2 = nn.Linear(hidden, hidden)
self.pi_head = nn.Linear(hidden, 1)
self.v_head = nn.Linear(hidden, 1)
def __call__(self, feats):
h = nn.relu(self.fc1(feats))
h = nn.relu(self.fc2(h))
return self.pi_head(h)[:, 0], self.v_head(h)[:, 0]
def _bernoulli_logp(actions, logits):
return actions * logits - mx.logaddexp(mx.zeros_like(logits), logits)
def _entropy(logits):
p = mx.sigmoid(logits)
return nn.softplus(logits) - p * logits
class PPOPruner:
"""Contextual bandit over prune masks of the residual stream.
One step per episode: observe a fixed per-channel feature matrix from the
frozen tower, emit a keep-probability per channel, sample a binary mask,
receive one scalar reward. Maximising it means "prune as much as possible
without inflating retrieval loss".
"""
def __init__(self, model, X, loss_of, sparsity_coef=0.35,
loss_tolerance=1.0, hidden=32, lr=3e-3, clip_eps=0.2,
entropy_coef=0.01, value_coef=0.5, seed=0):
self.model, self.loss_of = model, loss_of
self.sparsity_coef = sparsity_coef
self.loss_tolerance = loss_tolerance
self.clip_eps, self.entropy_coef = clip_eps, entropy_coef
self.value_coef = value_coef
np.random.seed(seed)
mx.random.seed(seed)
self.base_mask = get_channel_mask(model)
util = measure_utilization(model, loss_of)
self.base_loss = util["base_loss"]
self.ablation = util["ablation"]
act, down = descriptive_stats(model, X)
ra = reverse_attention(self.ablation, self.base_mask)
omega = model.blocks[0].omegas_per_neuron
self.candidates = np.where(self.base_mask == 1.0)[0]
f = np.stack([self.ablation[self.candidates], act[self.candidates],
down[self.candidates], ra[self.candidates],
omega[self.candidates]], axis=1).astype(np.float32)
mu, sd = f.mean(0, keepdims=True), f.std(0, keepdims=True) + 1e-6
f[:, :4] = (f[:, :4] - mu[:, :4]) / sd[:, :4]
self.feats = mx.array(f)
self.reverse_attn = ra
self.policy = PrunePolicy(f.shape[1], hidden=hidden)
self.opt = optim.Adam(learning_rate=lr)
self.history = {"iter": [], "reward": [], "kept_frac": [],
"loss_ratio": [], "entropy": []}
def _apply(self, action):
m = self.base_mask.copy()
m[self.candidates] = action
set_channel_mask(self.model, m)
def _reward(self, action):
self._apply(action)
loss = self.loss_of(self.model)
set_channel_mask(self.model, self.base_mask)
kept = float(action.mean()) if len(action) else 0.0
ratio = loss / (self.base_loss + 1e-12)
return (-self.loss_tolerance * ratio - self.sparsity_coef * kept,
kept, ratio)
def train(self, n_iters=40, episodes_per_iter=12, ppo_epochs=4,
verbose=True):
for it in range(n_iters):
logits, _ = self.policy(self.feats)
probs = np.array(mx.sigmoid(logits))
acts, rews, kfs, lrs = [], [], [], []
for _ in range(episodes_per_iter):
a = (np.random.uniform(size=probs.shape) < probs).astype(np.float32)
r, kf, lr_ = self._reward(a)
acts.append(a); rews.append(r); kfs.append(kf); lrs.append(lr_)
acts = np.array(acts, np.float32)
rews = np.array(rews, np.float32)
adv = (rews - rews.mean()) / (rews.std() + 1e-6)
acts_mx, adv_mx, rews_mx = mx.array(acts), mx.array(adv), mx.array(rews)
old_logits, _ = self.policy(self.feats)
old_logp = mx.array(np.array(
mx.sum(_bernoulli_logp(acts_mx, old_logits[None, :]), axis=1)))
for _ in range(ppo_epochs):
_, grads = self._loss_and_grad(acts_mx, old_logp, adv_mx, rews_mx)
self.opt.update(self.policy, grads)
mx.eval(self.policy.parameters(), self.opt.state)
ent = float(mx.mean(_entropy(logits)).item())
self.history["iter"].append(it)
self.history["reward"].append(float(rews.mean()))
self.history["kept_frac"].append(float(np.mean(kfs)))
self.history["loss_ratio"].append(float(np.mean(lrs)))
self.history["entropy"].append(ent)
if verbose and (it % 5 == 0 or it == n_iters - 1):
print(f" PPO {it:3d} | reward {rews.mean():+.3f} | kept "
f"{np.mean(kfs):.2f} | loss x{np.mean(lrs):.3f} | "
f"H {ent:.3f}", flush=True)
return self.history
def _loss_and_grad(self, actions, old_logp, adv, rewards):
def loss_fn(policy):
logits, values = policy(self.feats)
v = mx.mean(values)
logp = mx.sum(_bernoulli_logp(actions, logits[None, :]), axis=1)
ratio = mx.exp(logp - old_logp)
pl = -mx.mean(mx.minimum(ratio * adv,
mx.clip(ratio, 1 - self.clip_eps,
1 + self.clip_eps) * adv))
vl = mx.mean(mx.square(v - rewards))
return pl + self.value_coef * vl \
- self.entropy_coef * mx.mean(_entropy(logits))
return nn.value_and_grad(self.policy, loss_fn)(self.policy)
def apply_greedy(self, keep_threshold=0.5):
logits, _ = self.policy(self.feats)
probs = np.array(mx.sigmoid(logits))
m = self.base_mask.copy()
m[self.candidates] = (probs >= keep_threshold).astype(np.float32)
set_channel_mask(self.model, m)
return m, probs
# ===========================================================================
# 4. Neurogenesis β revive dead channels with fresh, spectrally diverse ones
# ===========================================================================
def neurogenesis(model, n_revive, seed=0):
"""Rebirth. New channels start near-silent (tiny w2) so they cannot shock
the forward pass, and are re-initialised with FINER's WIDE bias range β
without that they spawn as low-frequency clones and the population slowly
loses the spectral diversity that made the KAN basis worth having."""
rng = np.random.default_rng(seed)
mask = get_channel_mask(model)
dead = np.where(mask == 0.0)[0]
if len(dead) == 0 or n_revive <= 0:
return 0
revive = dead[:min(n_revive, len(dead))]
for b in model.turnover_layers():
w1 = np.array(b.w1) # (kernel, in_d, out*k)
b1 = np.array(b.b1)
w2 = np.array(b.w2)
ph = np.array(b.phases)
gs = np.array(b.gabor_s)
la = np.array(b.log_alpha)
om_full = np.repeat(b.omegas_per_neuron, b.k)
mean_om = float(b.omegas_per_neuron.mean())
lim = float(np.sqrt(6.0 / (b.kernel * b.in_d)) / mean_om)
for c in revive:
s, e = c * b.k, (c + 1) * b.k
w1[:, :, s:e] = rng.uniform(-lim, lim, (b.kernel, b.in_d, b.k))
b1[s:e] = rng.uniform(-b.bias_range, b.bias_range, b.k)
ph[s:e] = rng.uniform(0, 2 * np.pi, b.k)
gs[s:e] = rng.uniform(0.3, 1.5, b.k)
lw = np.log(np.clip(om_full[s:e], 1e-3, None))
la[s:e] = rng.uniform(np.zeros_like(lw), lw)
w2[c, :] = rng.standard_normal(b.k) * 1e-6
b.w1 = mx.array(w1.astype(np.float32))
b.b1 = mx.array(b1.astype(np.float32))
b.w2 = mx.array(w2.astype(np.float32))
b.phases = mx.array(ph.astype(np.float32))
b.gabor_s = mx.array(gs.astype(np.float32))
b.log_alpha = mx.array(la.astype(np.float32))
mask[revive] = 1.0
set_channel_mask(model, mask)
return len(revive)
# ===========================================================================
# 5. Compaction β where the mask becomes actual speed
# ===========================================================================
def compact(model):
"""Physically delete dead channels: return a smaller, equivalent tower.
Every dead channel is removed from each block's output slots, from every
consumer's input slots (both GRUs, the attention pool, the head), and from
the residual stream itself. Output is numerically equivalent to the masked
model but the matmuls are genuinely smaller.
"""
from .ctxtower import ContextTower
keep = np.where(get_channel_mask(model) == 1.0)[0]
n_new = len(keep)
if n_new == 0:
raise RuntimeError("every channel was pruned")
if n_new == n_channels(model):
return model, keep
cfg = dict(model.cfg)
cfg["n_hidden"] = int(n_new)
new = ContextTower(**cfg)
k = model.blocks[0].k
sub = np.concatenate([np.arange(c * k, (c + 1) * k) for c in keep])
for ob, nb in zip(model.blocks, new.blocks):
w1 = np.array(ob.w1)
if not ob.use_residual: # first block: in_d is d_in
nb.w1 = mx.array(w1[:, :, sub])
else: # in_d is the stream: slice both
nb.w1 = mx.array(w1[:, keep][:, :, sub])
nb.b1 = mx.array(np.array(ob.b1)[sub])
nb.phases = mx.array(np.array(ob.phases)[sub])
nb.gabor_s = mx.array(np.array(ob.gabor_s)[sub])
nb.log_alpha = mx.array(np.array(ob.log_alpha)[sub])
nb.w2 = mx.array(np.array(ob.w2)[keep])
# omegas is TRAINED, so it must be sliced from the live array β not
# rebuilt from omegas_per_neuron, which is the frozen design band and
# is stale the moment training starts. Rebuilding it silently reverted
# every learned frequency and moved val loss by 0.015.
nb.omegas = mx.array(np.array(ob.omegas)[sub])
nb.omegas_per_neuron = ob.omegas_per_neuron[keep] # design band only
nb.basis_types = mx.array(np.array(ob.basis_types)[sub])
nb.freeze(keys=["basis_types"], recurse=False)
nb.set_active_mask(np.ones(n_new, np.float32))
n_old = n_channels(model)
g = model.cfg["gru_hidden"]
for og, ng in ((model.gru_f, new.gru_f), (model.gru_b, new.gru_b)):
ng.Wx = mx.array(np.array(og.Wx)[:, keep])
ng.Wh = mx.array(np.array(og.Wh))
ng.b = mx.array(np.array(og.b))
ng.bhn = mx.array(np.array(og.bhn))
# consumers of `cat` = [stream | gru_f | gru_b]: keep the stream slice,
# carry the two GRU slices through untouched.
tail = np.arange(n_old, n_old + 2 * g)
cols = np.concatenate([keep, tail])
new.pool.proj.weight = mx.array(np.array(model.pool.proj.weight)[:, cols])
new.pool.proj.bias = mx.array(np.array(model.pool.proj.bias))
new.pool.score.weight = mx.array(np.array(model.pool.score.weight))
new.pool.score.bias = mx.array(np.array(model.pool.score.bias))
new.head.weight = mx.array(np.array(model.head.weight)[:, cols])
new.head.bias = mx.array(np.array(model.head.bias))
new.log_gamma = mx.array(np.array(model.log_gamma))
new.set_training(False)
return new, keep
# ===========================================================================
# 6. Re-settle
# ===========================================================================
def finetune(model, loss_and_grad_fn, epochs=150, lr=5e-4,
weight_decay=1e-3, loss_of=None, patience=40):
"""Short re-settling pass with aliveness frozen (see the mask gotcha).
Validation-aware on purpose. Survivors re-settle on the TRAIN set, and
with a couple of hundred windows an unguarded 400-epoch re-settle simply
re-memorises: measured, it took a pruned tower from 0.311 val back up to
0.357, undoing everything apoptosis had just gained. Keeping the best
validation state means re-settling can only help, never hurt.
"""
from .ctxtower import _clone_params, _load_params
opt = optim.AdamW(learning_rate=lr, weight_decay=weight_decay)
frozen = get_channel_mask(model)
best, best_w, since = float("inf"), None, 0
if loss_of is not None:
best, best_w = loss_of(model), _clone_params(model)
for _ in range(epochs):
model.set_training(True)
loss, grads = loss_and_grad_fn(model)
opt.update(model, grads)
set_channel_mask(model, frozen)
mx.eval(model.parameters(), opt.state)
if loss_of is None:
continue
v = loss_of(model)
if v < best - 1e-5:
best, best_w, since = v, _clone_params(model), 0
else:
since += 1
if since >= patience:
break
if best_w is not None:
_load_params(model, best_w)
set_channel_mask(model, frozen)
mx.eval(model.parameters())
model.set_training(False)
return best
# ===========================================================================
# 7. The full cycle
# ===========================================================================
def run_pruning_cycle(model, X, loss_of, loss_and_grad_fn, ppo_iters=40,
episodes_per_iter=12, sparsity_coef=0.35,
keep_threshold=0.5, finetune_epochs=150,
rebirth_fraction=0.5, select_tolerance=0.02,
seed=0, verbose=True):
"""apoptosis β re-settle β neurogenesis β re-settle, on a frozen tower.
The cycle is a SEARCH, so it returns the best point it visited, not the
last one. Rebirth is a bet that the data can support more capacity; on a
small corpus that bet often loses, and silently shipping the final state
would hand back a tower worse than the one we started from.
"""
from .ctxtower import _clone_params, _load_params
rec = {"stages": []}
saved = [] # (stage, loss, alive, params, mask) for every stage
def snap(stage):
u = measure_utilization(model, loss_of)
alive = int(u["mask"].sum())
if verbose:
print(f"[{stage}] val loss {u['base_loss']:.4f} | "
f"alive {alive}/{len(u['mask'])}", flush=True)
rec["stages"].append({"stage": stage, "loss": u["base_loss"],
"alive": alive,
"ablation": u["ablation"].tolist(),
"mask": u["mask"].tolist()})
saved.append((stage, u["base_loss"], alive,
_clone_params(model), u["mask"].copy()))
return u
snap("before")
if verbose:
print("\n== PPO reverse-attention pruning search ==", flush=True)
pruner = PPOPruner(model, X, loss_of, sparsity_coef=sparsity_coef, seed=seed)
rec["ppo_history"] = pruner.train(n_iters=ppo_iters,
episodes_per_iter=episodes_per_iter,
verbose=verbose)
_, probs = pruner.apply_greedy(keep_threshold)
rec["keep_probs"] = probs.tolist()
rec["reverse_attention"] = pruner.reverse_attn.tolist()
snap("after_apoptosis")
if finetune_epochs:
if verbose:
print("\n== re-settling survivors ==", flush=True)
finetune(model, loss_and_grad_fn, epochs=finetune_epochs,
loss_of=loss_of)
snap("after_finetune")
if rebirth_fraction > 0:
dead = int((get_channel_mask(model) == 0.0).sum())
born = neurogenesis(model, int(round(rebirth_fraction * dead)), seed=seed)
if verbose:
print(f"\n== neurogenesis: {born} channels reborn ==", flush=True)
if born:
snap("after_rebirth")
if finetune_epochs:
finetune(model, loss_and_grad_fn,
epochs=finetune_epochs, loss_of=loss_of)
snap("after_rebirth_finetune")
# ---- operating-point selection ----------------------------------------
# Not "lowest loss" β that rule can never prune, because the unpruned
# tower is usually the most accurate one and the whole exercise then
# returns its own input. The question a database actually asks is: what
# is the CHEAPEST tower I can serve without giving up measurable quality?
# So: take the best loss seen, allow `select_tolerance` relative slack,
# and among everything inside that band keep the fewest channels.
floor = min(s[1] for s in saved)
budget = floor * (1.0 + select_tolerance)
eligible = [s for s in saved if s[1] <= budget]
stage, loss, alive, params, mask = min(eligible, key=lambda s: (s[2], s[1]))
_load_params(model, params)
set_channel_mask(model, mask)
mx.eval(model.parameters())
if verbose:
print(f"\n== operating point: {stage} β {alive} channels, val loss "
f"{loss:.4f} (best seen {floor:.4f}, "
f"budget +{select_tolerance:.0%}) ==", flush=True)
rec["selected"] = {"stage": stage, "loss": loss, "channels": alive,
"best_loss_seen": floor,
"tolerance": select_tolerance,
"candidates": [{"stage": s[0], "loss": s[1],
"channels": s[2]} for s in saved]}
return rec
|