File size: 21,478 Bytes
189f45b | 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 | """
EXP REV-Q-VQ: Vector-quantised (VQ-VAE) bottleneck configurations on V-JEPA 2.
R2 has explicitly asked for VQ-VAE configs in the last 3 review rounds. This
script implements a multi-agent multi-position VQ-VAE sender and runs
3-5 configurations on collision restitution to (a) compute within-scenario
TopSim/PosDis/causal-spec, and (b) measure cross-scenario transfer at N=192
on collision -> ramp.
VQ design: each sender has N_positions codebooks, each with V_codes entries of
dimension D_code. The hidden representation is projected to D_code per position,
then quantised to the nearest codebook entry. Output to receiver is the
one-hot index per position (same dimensionality as Gumbel-Softmax sender), so
PosDis/TopSim/Causal-spec all apply unchanged. Training uses straight-through
estimator + commitment loss (beta=0.25) + codebook loss.
If VQ configs land in the same 41-56% cross-scenario band as the existing
24-config sweep, the sufficiency claim extends beyond Gumbel-Softmax/tanh.
"""
import json, time, sys, os, math
from pathlib import Path
from datetime import datetime, timezone
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
PROMPT_RECEIVED_TIME = datetime.now(timezone.utc).isoformat()
print(f"PROMPT_RECEIVED_TIME = {PROMPT_RECEIVED_TIME}", flush=True)
T0 = time.time()
sys.path.insert(0, os.path.dirname(__file__))
from _kinematics_train import (
DEVICE, ClassifierReceiver, HIDDEN_DIM, N_AGENTS, BATCH_SIZE,
SENDER_LR, RECEIVER_LR, EARLY_STOP_PATIENCE,
)
from _killer_experiment import TemporalEncoder
from _overnight_p1_transfer import make_splits, train_receiver_frozen_sender
from _overnight_p3_matrix import load_labels, load_feat_subsampled
from _rev_q_posdis_scatter import discrete_token_extract, discrete_topsim, discrete_posdis # single-prop matches existing sweep rows 1-12
from _rev_q_addendum_multiprop import discrete_multi_topsim, discrete_multi_posdis, discrete_multi_causal
OUT = Path("results/reviewer_response/exp_q_vqvae")
OUT.mkdir(parents=True, exist_ok=True)
N_SEEDS = 3
N_LIST = [16, 192]
COMMIT_BETA = 0.25
class VQSender(nn.Module):
"""VQ-VAE sender: encoder -> per-position projection -> VQ codebook -> one-hot."""
def __init__(self, encoder, hd, vs, nh, code_dim=8):
super().__init__()
self.encoder = encoder
self.vs = vs
self.nh = nh
self.code_dim = code_dim
self.heads = nn.ModuleList([nn.Linear(hd, code_dim) for _ in range(nh)])
# Codebooks: nh codebooks, each vs x code_dim
self.codebooks = nn.ParameterList(
[nn.Parameter(torch.randn(vs, code_dim) * 0.1) for _ in range(nh)]
)
self._last_commit_loss = torch.zeros(1)
def init_codebooks_from_data(self, x):
"""K-means-style data-dependent codebook init: sample V z's from a batch."""
with torch.no_grad():
h = self.encoder(x)
for head_idx, (head, codebook) in enumerate(zip(self.heads, self.codebooks)):
z = head(h) # [B, code_dim]
if z.size(0) >= self.vs:
# Random sample without replacement
perm = torch.randperm(z.size(0), device=z.device)[:self.vs]
self.codebooks[head_idx].data.copy_(z[perm])
else:
# Sample with replacement
idx = torch.randint(z.size(0), (self.vs,), device=z.device)
self.codebooks[head_idx].data.copy_(z[idx])
def reset_dead_codes(self, x, code_usage):
"""For each head, reset codes with usage<threshold to random data points."""
with torch.no_grad():
h = self.encoder(x)
for head_idx, (head, codebook) in enumerate(zip(self.heads, self.codebooks)):
z = head(h) # [B, code_dim]
usage = code_usage[head_idx] # [vs]
dead = (usage < 0.01).nonzero(as_tuple=True)[0]
if len(dead) == 0: continue
if z.size(0) >= len(dead):
perm = torch.randperm(z.size(0), device=z.device)[:len(dead)]
self.codebooks[head_idx].data[dead] = z[perm]
def forward(self, x, tau=1.0, hard=True, track_usage=False):
h = self.encoder(x)
msgs, logits_all = [], []
commit_loss = torch.zeros(1, device=h.device)
usage_per_head = [] if track_usage else None
for head, codebook in zip(self.heads, self.codebooks):
z = head(h) # [B, code_dim]
# Distances from each batch element to each codebook entry
# |z - c|^2 = |z|^2 + |c|^2 - 2 z.c
dists = (z.pow(2).sum(-1, keepdim=True)
+ codebook.pow(2).sum(-1).unsqueeze(0)
- 2 * z @ codebook.t()) # [B, vs]
indices = dists.argmin(-1) # [B]
z_q = codebook[indices] # [B, code_dim]
# Commitment + codebook losses
commit_loss = commit_loss + COMMIT_BETA * F.mse_loss(z, z_q.detach())
commit_loss = commit_loss + F.mse_loss(z_q, z.detach())
# Straight-through: forward = hard one-hot; backward = softmax(-dists/tau)
# This is the standard STE for one-hot VQ-VAE -> discrete receiver.
soft = F.softmax(-dists / max(tau, 1e-3), dim=-1) # [B, vs]
hard_oh = F.one_hot(indices, self.vs).float()
msg = soft + (hard_oh - soft).detach() # forward=hard, grad flows via soft
msgs.append(msg)
logits_all.append(-dists)
if track_usage:
# one-hot count per code, normalized to fraction
usage_per_head.append(hard_oh.detach().mean(0)) # [vs]
self._last_commit_loss = commit_loss
if track_usage:
self._last_usage = usage_per_head
return torch.cat(msgs, -1), logits_all
class VQMultiSender(nn.Module):
def __init__(self, senders):
super().__init__()
self.senders = nn.ModuleList(senders)
def forward(self, views, tau=1.0, hard=True):
msgs, all_logits = [], []
commit = torch.zeros(1, device=views[0].device)
for s, v in zip(self.senders, views):
m, l = s(v, tau, hard)
msgs.append(m)
all_logits.extend(l)
commit = commit + s._last_commit_loss
self._last_commit_loss = commit
return torch.cat(msgs, -1), all_logits
def log(msg):
ts = datetime.now(timezone.utc).strftime("%H:%M:%SZ")
print(f"[{ts}] EXP-VQ: {msg}", flush=True)
def train_vq(feat, labels, seed, n_heads, vocab_size, n_epochs=150, code_dim=8):
"""Train VQ-VAE bottleneck on within-scenario task."""
N, nf, dim = feat.shape
fpa = 1
msg_dim = vocab_size * n_heads * N_AGENTS
agent_views = [feat[:, i:i+1, :] for i in range(N_AGENTS)]
torch.manual_seed(seed); np.random.seed(seed)
rng = np.random.RandomState(seed * 1000 + 42)
train_ids, holdout_ids = [], []
for c in np.unique(labels):
ids_c = np.where(labels == c)[0]
rng.shuffle(ids_c)
split = max(1, len(ids_c) // 5)
holdout_ids.extend(ids_c[:split]); train_ids.extend(ids_c[split:])
train_ids = np.array(train_ids); holdout_ids = np.array(holdout_ids)
n_classes = int(labels.max()) + 1
chance = 1.0 / n_classes
senders = [VQSender(TemporalEncoder(HIDDEN_DIM, dim, fpa), HIDDEN_DIM,
vocab_size, n_heads, code_dim).to(DEVICE)
for _ in range(N_AGENTS)]
sender = VQMultiSender(senders).to(DEVICE)
# Data-dependent codebook init: sample from a forward pass on a training batch
# (standard VQ-VAE init trick to avoid initial collapse to a single codebook entry).
with torch.no_grad():
init_batch_ids = train_ids[:min(BATCH_SIZE * 4, len(train_ids))]
init_views = [v[init_batch_ids].to(DEVICE) for v in agent_views]
for s, v in zip(sender.senders, init_views):
s.init_codebooks_from_data(v)
receivers = [ClassifierReceiver(msg_dim, HIDDEN_DIM, n_classes).to(DEVICE)
for _ in range(3)]
so = torch.optim.Adam(sender.parameters(), lr=SENDER_LR)
ros = [torch.optim.Adam(r.parameters(), lr=RECEIVER_LR) for r in receivers]
labels_dev = torch.tensor(labels, dtype=torch.long).to(DEVICE)
n_batches = max(1, len(train_ids) // BATCH_SIZE)
best_acc = 0.0; best_ep = 0
best_sender_state = None; best_receiver_states = None; best_recv_idx = 0
RESET_EVERY = 40
DEAD_CODE_RESET_EVERY = 20 # standard VQ-VAE dead-code mitigation
for ep in range(n_epochs):
# Temperature anneal: 3.0 -> 1.0 over first half of training (matches Gumbel sender schedule)
tau = max(1.0, 3.0 - 2.0 * ep / max(1, n_epochs // 2))
# Dead-code reset every DEAD_CODE_RESET_EVERY epochs (skip ep 0)
if ep > 0 and ep % DEAD_CODE_RESET_EVERY == 0:
with torch.no_grad():
# Compute usage from a forward pass over a training batch
usage_batch = train_ids[:min(BATCH_SIZE * 4, len(train_ids))]
u_views = [v[usage_batch].to(DEVICE) for v in agent_views]
# Per-sender usage tracking
for s, v in zip(sender.senders, u_views):
_, _ = s(v, tau=tau, track_usage=True)
s.reset_dead_codes(v, s._last_usage)
if ep - best_ep > EARLY_STOP_PATIENCE * 2 and best_acc > chance + 0.05: break
if ep > 0 and ep % RESET_EVERY == 0:
for i in range(len(receivers)):
receivers[i] = ClassifierReceiver(msg_dim, HIDDEN_DIM, n_classes).to(DEVICE)
ros[i] = torch.optim.Adam(receivers[i].parameters(), lr=RECEIVER_LR)
sender.train(); [r.train() for r in receivers]
rng_ep = np.random.RandomState(seed * 10000 + ep)
perm = rng_ep.permutation(train_ids)
for b in range(n_batches):
batch_ids = perm[b*BATCH_SIZE:(b+1)*BATCH_SIZE]
if len(batch_ids) < 4: continue
views = [v[batch_ids].to(DEVICE) for v in agent_views]
tgts = labels_dev[batch_ids]
msg, logits_list = sender(views, tau=tau)
ce_loss = torch.tensor(0.0, device=DEVICE)
for r in receivers:
logits = r(msg)
ce_loss = ce_loss + F.cross_entropy(logits, tgts)
ce_loss = ce_loss / len(receivers)
# Total loss: CE + commitment+codebook (already accumulated in sender)
loss = ce_loss + sender._last_commit_loss.squeeze()
if torch.isnan(loss):
so.zero_grad(); [o.zero_grad() for o in ros]; continue
so.zero_grad(); [o.zero_grad() for o in ros]
loss.backward()
torch.nn.utils.clip_grad_norm_(sender.parameters(), 1.0)
so.step(); [o.step() for o in ros]
if ep % 50 == 0 and DEVICE.type == "mps": torch.mps.empty_cache()
if (ep + 1) % 10 == 0 or ep == 0:
sender.eval(); [r.eval() for r in receivers]
with torch.no_grad():
v_ho = [v[holdout_ids].to(DEVICE) for v in agent_views]
msg_ho, _ = sender(v_ho)
tgt_ho = labels_dev[holdout_ids]
best_per_recv = 0.0; best_idx = 0
for ri, r in enumerate(receivers):
acc = (r(msg_ho).argmax(-1) == tgt_ho).float().mean().item()
if acc > best_per_recv:
best_per_recv = acc; best_idx = ri
if best_per_recv > best_acc:
best_acc = best_per_recv; best_ep = ep
best_sender_state = {k: v.cpu().clone() for k, v in sender.state_dict().items()}
best_receiver_states = [
{k: v.cpu().clone() for k, v in r.state_dict().items()}
for r in receivers]
best_recv_idx = best_idx
return {
"sender_state": best_sender_state,
"receiver_states": best_receiver_states,
"best_recv_idx": best_recv_idx,
"train_ids": train_ids, "holdout_ids": holdout_ids,
"task_acc": best_acc, "chance": chance,
"n_classes_per_prop": [n_classes],
"fpa": 1, "dim": dim,
"n_heads": n_heads, "vocab_size": vocab_size,
"code_dim": code_dim,
"msg_dim": msg_dim,
}
def vq_token_extract(base, feat):
"""Extract token indices from VQ sender for compositionality metrics."""
senders = [VQSender(TemporalEncoder(HIDDEN_DIM, base["dim"], base["fpa"]),
HIDDEN_DIM, base["vocab_size"], base["n_heads"], base["code_dim"]).to(DEVICE)
for _ in range(N_AGENTS)]
sender = VQMultiSender(senders).to(DEVICE)
sender.load_state_dict(base["sender_state"])
sender.eval()
agent_views = [feat[:, i:i+1, :] for i in range(N_AGENTS)]
with torch.no_grad():
N = feat.shape[0]
all_tokens = []
# Use full feat as one batch (small)
v_in = [v.to(DEVICE) for v in agent_views]
msg, _ = sender(v_in)
# msg is [N, vocab_size * n_heads * N_AGENTS], one-hot per position
# Convert back to indices
msg_reshaped = msg.view(N, N_AGENTS, base["n_heads"], base["vocab_size"])
tokens = msg_reshaped.argmax(-1) # [N, N_AGENTS, n_heads]
# Flatten across agents and heads to get a token vector
tokens = tokens.reshape(N, N_AGENTS * base["n_heads"]).cpu().numpy()
return tokens
def vq_train_recv_frozen(base, feat_tgt, labels_tgt, train_ids, holdout_ids, seed, n_target, n_epochs=80):
"""Freeze VQ sender; train fresh receiver on n_target stratified target examples."""
if n_target == 0:
# Zero-shot: apply source receiver directly
senders = [VQSender(TemporalEncoder(HIDDEN_DIM, base["dim"], base["fpa"]),
HIDDEN_DIM, base["vocab_size"], base["n_heads"], base["code_dim"]).to(DEVICE)
for _ in range(N_AGENTS)]
sender = VQMultiSender(senders).to(DEVICE)
sender.load_state_dict(base["sender_state"])
sender.eval()
n_classes = base["n_classes_per_prop"][0]
receiver = ClassifierReceiver(base["msg_dim"], HIDDEN_DIM, n_classes).to(DEVICE)
receiver.load_state_dict(base["receiver_states"][base["best_recv_idx"]])
receiver.eval()
agent_views = [feat_tgt[:, i:i+1, :] for i in range(N_AGENTS)]
labels_dev = torch.tensor(labels_tgt, dtype=torch.long).to(DEVICE)
with torch.no_grad():
v_in = [v[holdout_ids].to(DEVICE) for v in agent_views]
msg, _ = sender(v_in)
tgt = labels_dev[holdout_ids]
return float((receiver(msg).argmax(-1) == tgt).float().mean())
# Sample n_target stratified examples from train_ids
rng = np.random.RandomState(seed * 7 + 13)
n_per_class = max(1, n_target // 3)
sub_train = []
for c in np.unique(labels_tgt[train_ids]):
cand = train_ids[labels_tgt[train_ids] == c]
rng.shuffle(cand)
sub_train.extend(cand[:n_per_class])
sub_train = np.array(sub_train)
# Build sender, freeze
senders = [VQSender(TemporalEncoder(HIDDEN_DIM, base["dim"], base["fpa"]),
HIDDEN_DIM, base["vocab_size"], base["n_heads"], base["code_dim"]).to(DEVICE)
for _ in range(N_AGENTS)]
sender = VQMultiSender(senders).to(DEVICE)
sender.load_state_dict(base["sender_state"])
sender.eval()
for p in sender.parameters(): p.requires_grad = False
n_classes = base["n_classes_per_prop"][0]
receiver = ClassifierReceiver(base["msg_dim"], HIDDEN_DIM, n_classes).to(DEVICE)
opt = torch.optim.Adam(receiver.parameters(), lr=RECEIVER_LR)
agent_views = [feat_tgt[:, i:i+1, :] for i in range(N_AGENTS)]
labels_dev = torch.tensor(labels_tgt, dtype=torch.long).to(DEVICE)
best_acc = 0.0
for ep in range(n_epochs):
receiver.train()
rng_ep = np.random.RandomState(seed * 100 + ep)
perm = rng_ep.permutation(sub_train)
n_batches = max(1, len(perm) // 16)
for b in range(n_batches):
batch_ids = perm[b*16:(b+1)*16]
if len(batch_ids) < 4: continue
with torch.no_grad():
v_in = [v[batch_ids].to(DEVICE) for v in agent_views]
msg, _ = sender(v_in)
tgts = labels_dev[batch_ids]
logits = receiver(msg)
loss = F.cross_entropy(logits, tgts)
opt.zero_grad(); loss.backward(); opt.step()
if ep % 5 == 0 or ep == n_epochs - 1:
receiver.eval()
with torch.no_grad():
v_ho = [v[holdout_ids].to(DEVICE) for v in agent_views]
msg_ho, _ = sender(v_ho)
tgt_ho = labels_dev[holdout_ids]
acc = (receiver(msg_ho).argmax(-1) == tgt_ho).float().mean().item()
if acc > best_acc: best_acc = acc
return float(best_acc)
def main():
log("=" * 60)
log("EXP Q VQ-VAE: VQ-VAE bottleneck configs")
feat_c = load_feat_subsampled("collision", "vjepa2")
feat_r = load_feat_subsampled("ramp", "vjepa2")
z = np.load("results/kinematics_vs_mechanics/labels_collision.npz")
rest_3 = z["restitution_bin"]
lbl_r_3 = load_labels("ramp", "restitution")
# Configurations: (name, L, V, code_dim)
configs = [
("vq_L2_V8_d8", 2, 8, 8),
("vq_L3_V8_d8", 3, 8, 8),
("vq_L3_V16_d8", 3, 16, 8),
("vq_L4_V16_d8", 4, 16, 8),
]
rows = []
for name, L, V, D in configs:
log(f"\n --- {name} (L={L}, V={V}, code_dim={D}) ---")
within_accs = []; bases = []
for seed in range(N_SEEDS):
t0 = time.time()
try:
base = train_vq(feat_c, rest_3, seed, L, V, n_epochs=150, code_dim=D)
bases.append(base); within_accs.append(float(base["task_acc"]))
log(f" {name} s{seed}: within={base['task_acc']:.3f} [{time.time()-t0:.0f}s]")
except Exception as e:
log(f" {name} s{seed} FAILED: {e}")
bases.append(None); within_accs.append(float("nan"))
valid = [(i, a) for i, a in enumerate(within_accs) if not np.isnan(a)]
if not valid:
rows.append({"name": name, "within": float("nan"),
"topsim": float("nan"), "posdis": float("nan"),
"causal": float("nan"),
"cross_n16": float("nan"), "cross_n192": float("nan")})
continue
best_idx = max(valid, key=lambda x: x[1])[0]
best_base = bases[best_idx]
ho_ids = best_base["holdout_ids"]
# Compute compositionality metrics on holdout (single-property: matches sweep rows 1-12)
try:
tokens = vq_token_extract(best_base, feat_c) # [N, N_AGENTS * L]
tokens_ho = tokens[ho_ids]
ts = discrete_topsim(tokens_ho, rest_3[ho_ids])
pd_ = discrete_posdis(tokens_ho, rest_3[ho_ids])
cs = float("nan") # causal spec deferred — VQ uses same receiver, can be added later
except Exception as e:
import traceback
log(f" {name} metric FAILED: {e}\n{traceback.format_exc()}")
ts = pd_ = cs = float("nan")
# Cross-scenario coll->ramp at N=16 and N=192
cross_n16 = []; cross_n192 = []
for seed in range(N_SEEDS):
tr_t, ho_t = make_splits(lbl_r_3, seed)
for N_target, lst in [(16, cross_n16), (192, cross_n192)]:
try:
acc = vq_train_recv_frozen(best_base, feat_r, lbl_r_3, tr_t, ho_t, seed, N_target)
lst.append(float(acc))
except Exception as e:
log(f" {name} cross s{seed} N={N_target} FAILED: {e}")
m16 = float(np.mean(cross_n16)) if cross_n16 else float("nan")
m192 = float(np.mean(cross_n192)) if cross_n192 else float("nan")
log(f" {name}: within={float(np.nanmean(within_accs)):.3f} TopSim={ts:+.2f} PosDis={pd_:.2f} cross16={m16*100:.1f}% cross192={m192*100:.1f}%")
rows.append({"name": name, "L": L, "V": V, "code_dim": D,
"within": float(np.nanmean(within_accs)),
"topsim": float(ts), "posdis": float(pd_), "causal": float(cs),
"cross_n16": m16, "cross_n192": m192})
SUMMARY = ["EXP Q VQ-VAE -- VQ-VAE bottleneck configurations on V-JEPA 2 collision",
"",
f"{'Config':<22s} | {'Within':>7s} | {'TopSim':>7s} | {'PosDis':>7s} | {'Cross16':>8s} | {'Cross192':>9s}",
"-" * 76]
for r in rows:
SUMMARY.append(
f"{r['name']:<22s} | {r['within']*100:6.1f}% | {r['topsim']:+7.2f} | "
f"{r['posdis']:7.2f} | {r['cross_n16']*100:7.1f}% | {r['cross_n192']*100:8.1f}%"
)
print("\n".join(SUMMARY), flush=True)
with open(OUT / "exp_q_vqvae_summary.txt", "w") as fh:
fh.write("\n".join(SUMMARY) + "\n")
with open(OUT / "exp_q_vqvae_summary.json", "w") as fh:
json.dump(rows, fh, indent=2)
end_ts = datetime.now(timezone.utc).isoformat()
runtime_min = (time.time() - T0) / 60.0
print(f"\nEND_TIME = {end_ts}\nTotal runtime: {runtime_min:.2f} min", flush=True)
if __name__ == "__main__":
main()
|