File size: 21,398 Bytes
ec0d725 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 | """Recursive latent planner over the frozen LeWM PushT world model.
Implements the HRM/TRM-style nested recursion of ``recursive_planner_design.pdf``
section 9, placed inside an MPC rollout::
for k = 1 .. H: # outer - imagined environment steps
for j = 1 .. T: # middle - answer (action) improvement
for i = 1 .. n: # inner - latent reasoning refinement
z = f(z, h_G, h_{k-1}, y, c)
y = g(y, z, h_G)
c = consequence(M(C, psi(y)), h_G) # Change 2
b_k = psi(y); h_k = M(h_{k-1}, b_k) # frozen model advances
``f`` and ``g`` are weight-tied across every ``i``, ``j`` and ``k`` — the
recursion buys depth and compute, not parameters.
Three things in here are load-bearing and easy to get silently wrong:
* **The gradient policy (Change 1).** Cycles ``1..T-1`` run under ``no_grad``;
only cycle ``T`` is differentiated, and ``(y, z)`` are detached when crossing
from horizon step ``k`` to ``k+1``. The chain of world-model states ``h_k`` is
*never* detached — that chain is the entire planning signal, and cutting it
leaves a loss that still falls while the controller quietly becomes greedy.
* **Action/frame alignment.** Block ``k`` is the block *leaving* context frame
``k``. With ``N`` context frames there are ``N-1`` past blocks between them,
and the current frame pairs with the first block of the plan. This mirrors
``lejepa_control.rollout.rollout_plan`` and ``LeWM.rollout`` exactly.
* **Recursion stability (Change 11).** RMSNorm on ``z`` and ``y`` before every
``f``/``g``, a bounded residual gate on each update, and ``h_{k-1}``/``h_G``
re-injected at *every* ``f`` application rather than only the first.
"""
import torch
from torch import nn
__all__ = [
'ActionEmbedding',
'FTheta',
'GTheta',
'RMSNorm',
'RecursivePlanner',
'consequence_features',
'm_step',
]
class RMSNorm(nn.Module):
"""Root-mean-square layer norm, no mean subtraction.
Change 11's first guard. A weight-tied map applied ~90 times has no reason
to be norm-preserving; a mild 5% growth per step compounds to 80x over one
training step's recursion, and long before that ``f`` sees inputs outside
the range its weights were fit for.
"""
def __init__(self, dim, eps=1e-6):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(dim))
def forward(self, x):
scale = torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
return x * scale * self.weight
class ActionEmbedding(nn.Module):
"""The ``phi`` / ``psi`` pair: action block <-> answer space.
``phi`` encodes a real 10-dim block into the ``W``-dim answer space, ``psi``
decodes an answer back to a tanh-bounded block. Pre-train the pair as a
plain autoencoder on dataset blocks (Change 7), then keep both trainable
with the round-trip anchor holding them consistent.
Args:
block_dim: ``frameskip * action_dim`` (10 for PushT).
width: Answer-space width ``W``.
hidden: MLP hidden width.
action_dim: Native env action dim (2).
frameskip: Env actions per world-model transition (5).
action_center / action_scale: Per-dim tanh bounds expressed in the
*normalized* action units the world model was trained on. For raw
PushT actions in ``[-1, 1]`` these are ``-mean/std`` and ``1/std``.
"""
def __init__(
self,
block_dim=10,
width=256,
hidden=256,
action_dim=2,
frameskip=5,
action_center=0.0,
action_scale=1.0,
):
super().__init__()
self.block_dim = block_dim
self.width = width
self.action_dim = action_dim
self.frameskip = frameskip
center = torch.as_tensor(action_center).float().expand(action_dim)
scale = torch.as_tensor(action_scale).float().expand(action_dim)
self.register_buffer('action_center', center.clone())
self.register_buffer('action_scale', scale.clone())
self.encode_net = nn.Sequential(
nn.Linear(block_dim, hidden),
nn.GELU(),
nn.Linear(hidden, width),
)
self.decode_net = nn.Sequential(
RMSNorm(width),
nn.Linear(width, hidden),
nn.GELU(),
nn.Linear(hidden, block_dim),
)
def encode(self, block):
"""``phi``: ``(..., A)`` normalized block -> ``(..., W)`` answer."""
return self.encode_net(block)
def bound(self, raw):
"""Map pre-tanh activations to a valid normalized action block."""
r = raw.unflatten(-1, (self.frameskip, self.action_dim))
return (self.action_center + self.action_scale * torch.tanh(r)).flatten(-2)
def decode(self, y, return_raw=False):
"""``psi``: ``(..., W)`` answer -> ``(..., A)`` bounded block.
``return_raw`` also yields the pre-tanh activations, which is what the
saturation barrier (Change 6) penalizes — the barrier has to act before
the tanh or it cannot reach a dimension that has already frozen.
"""
raw = self.decode_net(y)
block = self.bound(raw)
return (block, raw) if return_raw else block
def round_trip(self, y):
"""``phi(psi(y))`` — the manifold anchor's prediction of ``y``."""
return self.encode(self.decode(y))
class _GatedUpdate(nn.Module):
"""Shared body for ``f`` and ``g``: normalize, condition, gated residual.
The update is ``x <- x + sigmoid(eta) * Delta(...)`` with ``Delta``'s output
layer initialized small, so the recursion starts near-identity and cannot
destroy a good answer in early training (Change 11's second guard).
"""
def __init__(self, width, cond_dim, hidden, gate_init=0.0, out_std=0.01):
super().__init__()
self.norm_state = RMSNorm(width)
self.norm_other = RMSNorm(width)
self.cond_proj = nn.Linear(cond_dim, width)
self.net = nn.Sequential(
nn.Linear(3 * width, hidden),
nn.GELU(),
nn.Linear(hidden, hidden),
nn.GELU(),
)
self.out = nn.Linear(hidden, width)
nn.init.normal_(self.out.weight, std=out_std)
nn.init.zeros_(self.out.bias)
self.gate = nn.Parameter(torch.tensor(float(gate_init)))
# counts applications made with grad enabled; the Change-1 detach
# schedule is verified against this, see RecursivePlanner.grad_calls
self.grad_calls = 0
def forward(self, state, other, cond):
if torch.is_grad_enabled():
self.grad_calls += 1
x = torch.cat(
[
self.norm_state(state),
self.norm_other(other),
self.cond_proj(cond),
],
dim=-1,
)
return state + torch.sigmoid(self.gate) * self.out(self.net(x))
def gate_value(self):
with torch.no_grad():
return torch.sigmoid(self.gate).item()
class FTheta(nn.Module):
"""Inner-loop latent reasoning update ``z <- f(z, h_prev, h_G, y, c)``.
Conditioning is ``[h_{k-1}, h_G, c]`` where ``c`` is the consequence
feature from Change 2 — ``[h_hat, h_hat - h_G, ||h_hat - h_G||^2 / D]``.
All of it is re-supplied at every application, not just the first, so the
recursion cannot drift away from the question it was asked.
"""
def __init__(self, width=256, latent_dim=192, hidden=512, gate_init=0.0):
super().__init__()
self.latent_dim = latent_dim
# h_prev (D) + h_G (D) + consequence (2D + 1)
cond_dim = 4 * latent_dim + 1
self.body = _GatedUpdate(width, cond_dim, hidden, gate_init)
def forward(self, z, h_prev, h_goal, y, consequence):
cond = torch.cat([h_prev, h_goal, consequence], dim=-1)
return self.body(z, y, cond)
class GTheta(nn.Module):
"""Middle-loop answer update ``y <- g(y, z, h_G)``.
``g`` deliberately does not see ``h_{k-1}``: the current state reaches the
answer only through ``z``. That is TRM's convention and it is what makes
``z`` a scratchpad rather than a redundant conditioning path.
"""
def __init__(self, width=256, latent_dim=192, hidden=512, gate_init=0.0):
super().__init__()
self.body = _GatedUpdate(width, latent_dim, hidden, gate_init)
def forward(self, y, z, h_goal):
return self.body(y, z, h_goal)
def consequence_features(h_hat, h_goal):
"""Change 2's feedback vector: ``[h_hat, h_hat - h_G, ||.||^2 / D]``.
Gives ``f`` an error vector in the same space it is trying to shrink, which
is what turns the middle loop from an open-loop guesser into a corrector.
"""
delta = h_hat - h_goal
dist = delta.pow(2).mean(dim=-1, keepdim=True)
return torch.cat([h_hat, delta, dist], dim=-1)
def m_step(model, frames, blocks, num_context):
"""One frozen-world-model transition, matching ``rollout_plan``'s windows.
At rollout step ``t`` the predictor consumes frames ``[t, t+N)`` and the
action blocks leaving those same frames. Because ``frames`` has ``N+t``
entries and ``blocks`` has ``N+t`` entries once the step's block is
appended, both are just the trailing ``N``.
Args:
model: The frozen ``LeWM``. Gradients flow *through* it, never into it.
frames: List of ``(B, D)`` latents, oldest first, length ``N + t``.
blocks: List of ``(B, A)`` normalized blocks, length ``N + t``, where
entry ``i`` is the block leaving ``frames[i]``.
num_context: ``N``.
Returns:
``(B, D)`` the predicted next latent.
"""
n = num_context
assert len(blocks) == len(frames), (
f'alignment: {len(frames)} frames but {len(blocks)} blocks; block i '
f'must be the block leaving frame i'
)
emb_win = torch.stack(frames[-n:], dim=1)
act_win = model.action_encoder(torch.stack(blocks[-n:], dim=1))
return model.predict(emb_win, act_win)[:, -1]
class RecursivePlanner(nn.Module):
"""Three-loop driver: inner ``n``, middle ``T``, outer ``H``.
Args:
latent_dim: World-model latent width ``D`` (192).
num_context: Context frames the predictor consumes ``N`` (3).
action_dim / frameskip: Native action dim and env steps per transition.
width: Recursion width ``W`` (256).
hidden: MLP hidden width inside ``f`` and ``g``.
inner: ``n``, latent refinements per cycle (6).
cycles: ``T``, answer revisions per horizon step (3; 1 during stage A).
horizon: ``H``, imagined lookahead steps (3 -> 5 curriculum).
use_feedback: Change 2. When off there is no per-cycle lookahead, so
no per-cycle distances are produced and deep supervision has
nothing to score.
warm_start: Change 9. Start ``y`` from the last executed block and
carry the answer across horizon steps instead of resetting it.
lambda_z: Change 9's ``lambda_z``. ``0`` means a fresh scratchpad
``z0`` at every horizon step, which is the documented default;
larger values blend in the detached carried state.
learn_lambda_z: Make ``lambda_z`` a learned scalar.
action_center / action_scale: tanh bounds in normalized action units.
"""
def __init__(
self,
latent_dim=192,
num_context=3,
action_dim=2,
frameskip=5,
width=256,
hidden=512,
inner=6,
cycles=3,
horizon=5,
use_feedback=True,
warm_start=True,
lambda_z=0.0,
learn_lambda_z=False,
gate_init=0.0,
action_center=0.0,
action_scale=1.0,
detach_schedule='last-cycle',
):
super().__init__()
assert detach_schedule in ('last-cycle', 'one-step', 'full')
self.detach_schedule = detach_schedule
self.latent_dim = latent_dim
self.num_context = num_context
self.action_dim = action_dim
self.frameskip = frameskip
self.block_dim = frameskip * action_dim
self.width = width
self.inner = inner
self.cycles = cycles
self.horizon = horizon
self.use_feedback = use_feedback
self.warm_start = warm_start
self.f = FTheta(width, latent_dim, hidden, gate_init)
self.g = GTheta(width, latent_dim, hidden, gate_init)
self.action_embed = ActionEmbedding(
block_dim=self.block_dim,
width=width,
hidden=hidden // 2,
action_dim=action_dim,
frameskip=frameskip,
action_center=action_center,
action_scale=action_scale,
)
# fresh scratchpad, and the cold-start answer when warm start is off
self.z0 = nn.Parameter(torch.randn(1, width) * 0.02)
self.y0 = nn.Parameter(torch.randn(1, width) * 0.02)
self._capture = False
self.taps = {'z_norms': [], 'grad_first': [], 'grad_last': []}
if learn_lambda_z:
self.lambda_z = nn.Parameter(torch.tensor(float(lambda_z)))
else:
self.register_buffer(
'lambda_z', torch.tensor(float(lambda_z)), persistent=True
)
# -- diagnostics -------------------------------------------------------
def reset_call_counts(self):
self.f.body.grad_calls = 0
self.g.body.grad_calls = 0
def start_capture(self):
"""Begin collecting the section-11 recursion diagnostics.
Populates ``self.taps`` during the next forward/backward with:
``z_norms`` (should be flat across ``i`` — RMSNorm makes it so) and,
after ``backward()``, ``grad_first`` / ``grad_last``, the gradient
norms at ``f``'s first and last application inside the gradient cycle.
A first/last ratio outside ~10x means the backprop depth is larger
than the detach schedule intends.
"""
self.taps = {'z_norms': [], 'grad_first': [], 'grad_last': []}
self._capture = True
def stop_capture(self):
self._capture = False
def _tap_grad(self, tensor, key):
if tensor.requires_grad:
tensor.register_hook(
lambda g, k=key: self.taps[k].append(g.norm().item())
)
@property
def grad_calls(self):
"""``f`` and ``g`` applications made with grad enabled.
Under the Change-1 schedule this is ``H * (n + 1)`` — one gradient
cycle per horizon step. Without it, it is ``H * T * (n + 1)``, which
is the failure the activation-count test exists to catch.
"""
return self.f.body.grad_calls + self.g.body.grad_calls
def gate_values(self):
return {'f': self.f.body.gate_value(), 'g': self.g.body.gate_value()}
# -- the recursion -----------------------------------------------------
def _initial_answer(self, past_actions, batch):
if self.warm_start:
# Change 9: consecutive optimal blocks are highly correlated, so
# phi of the last executed block is a free head start
return self.action_embed.encode(past_actions[:, -1])
return self.y0.expand(batch, -1)
def forward(
self,
model,
ctx_emb,
past_actions,
goal_emb,
horizon=None,
cycles=None,
inner=None,
):
"""Run one imagined rollout and return everything the loss needs.
Args:
model: Frozen ``LeWM``.
ctx_emb: ``(B, N, D)`` context latents.
past_actions: ``(B, N-1, A)`` normalized executed blocks.
goal_emb: ``(B, D)`` goal latent.
horizon / cycles / inner: Per-call overrides of ``H`` / ``T`` /
``n``, used by the curriculum and the anytime-inference sweep.
Returns:
Dict with ``distances`` ``(B, H)``, ``cycle_distances``
``(B, H, T)`` or ``None``, ``blocks`` ``(B, H, A)``, ``raw``
``(B, H, A)`` pre-tanh, ``contexts`` ``(B, H, N, D)``, ``answers``
``(B, H, W)`` and ``frames`` ``(B, N+H, D)``.
"""
H = horizon or self.horizon
T = cycles or self.cycles
n = inner or self.inner
N = self.num_context
B = ctx_emb.size(0)
assert ctx_emb.size(1) == N, f'expected {N} context frames'
assert past_actions.size(1) == N - 1, (
f'expected {N - 1} past blocks between {N} context frames, got '
f'{past_actions.size(1)}'
)
frames = list(ctx_emb.unbind(dim=1))
blocks = list(past_actions.unbind(dim=1))
y = self._initial_answer(past_actions, B)
z = self.z0.expand(B, -1)
# consequence_features width: [h_hat, h_hat - h_G, ||.||^2/D] = 2D + 1
zero_c = ctx_emb.new_zeros(B, 2 * self.latent_dim + 1)
distances, cycle_d = [], []
out_blocks, out_raw, out_ctx, out_y = [], [], [], []
for _ in range(H):
# the state the recursion plans *from*, read before it advances
h_prev = frames[-1]
context = torch.stack(frames[-N:], dim=1)
# Change 1: no BPTT across horizon steps. The h-chain above stays
# differentiable; only the recursion's own carry is cut.
y = y.detach()
# Change 9: z_k^0 = z0 + lambda_z * sg(z_{k-1}^n); lambda_z = 0 is
# the documented default and gives a fresh scratchpad
z = self.z0.expand(B, -1) + self.lambda_z * z.detach()
c = zero_c
step_cycles = []
# ---- cycles 1..T-1 : forward only ----------------------------
# 'full' keeps the graph across every cycle: the ablation-4 control
# showing why Change 1 exists. It may simply not train.
early_grad = self.detach_schedule == 'full'
with torch.set_grad_enabled(
early_grad and torch.is_grad_enabled()
):
for _ in range(T - 1):
for _ in range(n):
z = self.f(z, h_prev, goal_emb, y, c)
y = self.g(y, z, goal_emb)
if self.use_feedback:
# Change 2: what would this answer actually cause?
b_hat = self.action_embed.decode(y)
h_hat = m_step(model, frames, blocks + [b_hat], N)
c = consequence_features(h_hat, goal_emb)
step_cycles.append(
(h_hat - goal_emb).pow(2).mean(dim=-1)
)
if not early_grad:
# the lookahead distances above are constants; keep them out of
# the graph explicitly rather than relying on no_grad's scope
step_cycles = [d.detach() for d in step_cycles]
# ---- cycle T : the only one that carries gradient -------------
# 'one-step' is HRM's original approximation, kept as ablation 4's
# middle rung: only the final f application carries gradient.
head = n - 1 if self.detach_schedule == 'one-step' else 0
with torch.set_grad_enabled(False):
for _ in range(head):
z = self.f(z, h_prev, goal_emb, y, c)
for i in range(head, n):
z = self.f(z, h_prev, goal_emb, y, c)
if self._capture:
self.taps['z_norms'].append(z.detach().norm(dim=-1).mean().item())
if i == head:
self._tap_grad(z, 'grad_first')
if i == n - 1:
self._tap_grad(z, 'grad_last')
y = self.g(y, z, goal_emb)
# ---- commit the action and advance the frozen world model -----
block, raw = self.action_embed.decode(y, return_raw=True)
h = m_step(model, frames, blocks + [block], N)
frames.append(h)
blocks.append(block)
d = (h - goal_emb).pow(2).mean(dim=-1)
distances.append(d)
step_cycles.append(d)
cycle_d.append(torch.stack(step_cycles, dim=-1))
out_blocks.append(block)
out_raw.append(raw)
out_ctx.append(context)
out_y.append(y)
return {
# pre-stack views. Indexing a stacked tensor produces a node that
# is *downstream* of the committed blocks, so `autograd.grad(d[-1],
# stacked[:, 0])` reports None even on a perfectly intact chain.
# The section-9 invariants are stated on the lists for that reason,
# and the tests assert against these.
'distance_seq': distances,
'block_seq': out_blocks,
'distances': torch.stack(distances, dim=1),
'cycle_distances': (
torch.stack(cycle_d, dim=1) if self.use_feedback else None
),
'blocks': torch.stack(out_blocks, dim=1),
'raw': torch.stack(out_raw, dim=1),
'contexts': torch.stack(out_ctx, dim=1),
'answers': torch.stack(out_y, dim=1),
'frames': torch.stack(frames, dim=1),
}
|