File size: 11,527 Bytes
0283577 | 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 | """Controls for the LaWAM-VGGT Stage-1 probes -- the numbers `eval_lam_probes.py`
reports are not interpretable on their own.
Two gaps in the original evaluation, both of which could make a dead `z` look alive:
A. `z2action_r2 = 0.55` has no baseline. LIBERO actions are strongly predictable
from the current frame alone (the arm's position implies where it is going).
If a 32-d PCA of `u_t` scores the same R^2, `z` contributes nothing and the
"z encodes action" claim collapses. We match dimensionality exactly (32 vs 32)
so the comparison is about *content*, not capacity, and fit the PCA on train
episodes only.
B. `delta_z` shuffles `z` across the whole batch, which mixes two effects: wrong
*motion* and wrong *scene/task*. A `z` that only encoded "which LIBERO suite
is this" would still produce a large delta_z. Shuffling *within* an episode
holds scene and task fixed, so only the motion component can explain the gap.
Also characterizes what the decoder does with a wrong `z` (collapse to identity,
or confidently wrong motion?) and adds z=0 / random-z references.
"""
import argparse
import importlib.machinery
import sys
import types
from collections import defaultdict
import numpy as np
import torch
import torch.nn.functional as F
REPO = "/home/ma-user/work/dataset/xxd-dataset/dataset_yhw/WAM/LaWAM_official"
sys.path.insert(0, REPO)
def _stub(name, **attrs):
mod = types.ModuleType(name)
mod.__spec__ = importlib.machinery.ModuleSpec(name, None)
mod.__path__ = []
for k, v in attrs.items():
setattr(mod, k, v)
sys.modules[name] = mod
return mod
class _Callback:
pass
_stub("lightning", LightningModule=torch.nn.Module)
_stub("lightning.pytorch", Callback=_Callback)
_stub("lightning.pytorch.callbacks", Callback=_Callback)
sys.modules["lightning"].pytorch = sys.modules["lightning.pytorch"]
_stub("wandb", Image=lambda *a, **k: None, log=lambda *a, **k: None)
from latent_action_model.core.lam_model import load_latent_action_model # noqa: E402
from latent_action_model.data_loader.lerobot_dataset import LeRobotLAMDataset # noqa: E402
from latent_action_model.data_loader.collate import lam_collate # noqa: E402
from latent_action_model.data_loader.video_aug import gpu_two_view_video_aug # noqa: E402
def standardize(x, mu, sd):
return (x - mu) / sd
def ridge_r2(X, Y, tr, te, tag):
"""Episode-split ridge with a lambda sweep. Returns (r2, lambda)."""
Xm, Xs = X[tr].mean(0), X[tr].std(0) + 1e-8
Ym = Y[tr].mean(0)
Xtr, Xte = (X[tr] - Xm) / Xs, (X[te] - Xm) / Xs
Ytr, Yte = Y[tr] - Ym, Y[te] - Ym
best = (-1e9, None)
for lam in [1e-3, 1e-2, 1e-1, 1.0, 10.0, 100.0, 1e3]:
A = Xtr.T @ Xtr + lam * np.eye(Xtr.shape[1])
W = np.linalg.solve(A, Xtr.T @ Ytr)
P = Xte @ W
r2 = 1 - ((Yte - P) ** 2).sum() / max((Yte ** 2).sum(), 1e-12)
if r2 > best[0]:
best = (r2, lam)
print(f" {tag:36s} R^2 = {best[0]:.4f} (lambda={best[1]})")
return best
@torch.no_grad()
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--ckpt", default="/home/ma-user/work/lam_runs/vggt_vae_libero/checkpoints/epoch=39.ckpt")
ap.add_argument("--config", default="/home/ma-user/work/lam_runs/vggt_vae_libero/version_0/config.yaml")
ap.add_argument("--n-batches", type=int, default=60)
ap.add_argument("--batch-size", type=int, default=16)
ap.add_argument("--val-tail-ratio", type=float, default=0.05)
ap.add_argument("--seed", type=int, default=0)
args = ap.parse_args()
torch.manual_seed(args.seed)
np.random.seed(args.seed)
dev = "cuda"
model = load_latent_action_model(args.ckpt, args.config).to(dev).eval()
print(f"[ctrl] encoder={type(model.vision_encoder).__name__} input_dim={model.input_dim}")
ds = LeRobotLAMDataset(
data_root_dir="/home/ma-user/work/lam_datasets",
data_mix="libero",
num_frames=2,
mode="val",
val_tail_ratio=args.val_tail_ratio,
video_backend="pyav",
image_hw=(256, 256),
frame_dt_sec=1.6,
debug_repeat_batch=False,
)
loader = torch.utils.data.DataLoader(
ds, batch_size=args.batch_size, num_workers=8,
collate_fn=lambda b: lam_collate(b, max_state_dim=32), drop_last=True,
)
all_dec_in, all_tgt, all_recon, all_z, all_traj, all_base = [], [], [], [], [], []
it = iter(loader)
for i in range(args.n_batches):
try:
batch = next(it)
except StopIteration:
print(f"[ctrl] loader exhausted after {i} batches")
break
videos = batch["videos"].to(dev)
v1, v2 = gpu_two_view_video_aug(videos, output_size=(256, 256), training=False, dual_view_aug=False)
states = batch["states"].to(dev)
state_mask = batch["state_mask"].to(dev)
emb = batch["embodiment_ids"].to(dev)
with torch.autocast("cuda", dtype=torch.bfloat16):
recon, dec_in, tgt, _, _, _, _, z, _, _ = model.inference(
v1, states, v2, state_mask=state_mask, embodiment_ids=emb
)
all_dec_in.append(dec_in.float().cpu())
all_tgt.append(tgt.float().cpu())
all_recon.append(recon.float().cpu())
all_z.append(z.float().reshape(z.shape[0], -1).cpu())
all_traj.extend(batch["trajectory_ids"])
all_base.append(batch["base_indices"])
if (i + 1) % 20 == 0:
print(f"[ctrl] pass1 {i+1}/{args.n_batches}")
dec_in = torch.cat(all_dec_in)
tgt = torch.cat(all_tgt)
recon = torch.cat(all_recon)
Z = torch.cat(all_z)
base_idx = torch.cat(all_base).numpy()
eps_all = np.array([int(t) for t in all_traj])
N = dec_in.shape[0]
print(f"\n[ctrl] N={N} samples | z dim={Z.shape[1]} | {len(np.unique(eps_all))} episodes")
flat = tgt.reshape(-1, tgt.shape[-1])
mu, sd = flat.mean(0), flat.std(0).clamp_min(1e-6)
T_ = standardize(tgt, mu, sd)
R_ = standardize(recon, mu, sd)
I_ = standardize(dec_in, mu, sd)
mse = lambda a, b: (a - b).pow(2).mean().item()
mse_pred, mse_identity = mse(R_, T_), mse(I_, T_)
print(f"[ctrl] mse_pred={mse_pred:.4f} mse_identity={mse_identity:.4f} mse_mean={mse(torch.zeros_like(T_), T_):.4f}")
def decode_with(zs):
"""Re-run only the decoder with the given latents. Returns standardized pred."""
outs = []
for s in range(0, N, args.batch_size):
e = min(s + args.batch_size, N)
with torch.autocast("cuda", dtype=torch.bfloat16):
r = model.decoder(dec_in[s:e].to(dev), zs[s:e].to(dev).unsqueeze(1)).float().cpu()
outs.append(standardize(r, mu, sd))
return torch.cat(outs)
# ---- B. delta_z variants ----------------------------------------------
print("\n=== 对照 B: delta_z 的四种 z 替换 ===")
g = torch.Generator().manual_seed(args.seed)
# (b1) shuffle across the whole set -- the original probe
R_glob = decode_with(Z[torch.randperm(N, generator=g)])
# (b2) shuffle WITHIN episode -- holds scene/task fixed, only motion differs
by_ep = defaultdict(list)
for i, e in enumerate(eps_all):
by_ep[int(e)].append(i)
perm_w = np.arange(N)
n_swappable = 0
for e, idxs in by_ep.items():
if len(idxs) < 2:
continue
a = np.array(idxs)
b = a.copy()
# derangement-ish: roll by 1 within the episode
perm_w[a] = np.roll(b, 1)
n_swappable += len(idxs)
R_within = decode_with(Z[torch.from_numpy(perm_w)])
# (b3) z = 0 and (b4) z ~ N(0, I)
R_zero = decode_with(torch.zeros_like(Z))
R_rand = decode_with(torch.randn(Z.shape, generator=g) * Z.std())
rows = [
("真实 z", mse_pred, R_),
("跨全集打乱 z", mse(R_glob, T_), R_glob),
("同 episode 内打乱 z", mse(R_within, T_), R_within),
("z = 0", mse(R_zero, T_), R_zero),
("z ~ 高斯噪声", mse(R_rand, T_), R_rand),
("抄袭 u_t (identity)", mse_identity, I_),
]
print(f" {'替换方式':22s} {'mse':>8s} {'delta_z':>9s} {'与u_t距离':>10s}")
for name, m, arr in rows:
d = m - mse_pred
di = mse(arr, I_)
print(f" {name:22s} {m:8.4f} {d:9.4f} {di:10.4f}")
print(f" 可同-episode 交换的样本: {n_swappable}/{N}")
print(" 注: '与u_t距离'=0 表示解码器退化成原样抄袭输入")
# ---- A. z vs u_t as action predictors ---------------------------------
print("\n=== 对照 A: z 与 u_t 谁能解出动作 (同为32维, 同 episode 切分) ===")
import pyarrow.parquet as pq
tbl = pq.read_table(
"/home/ma-user/work/lam_datasets/libero_merged_no_noops_20hz/data/chunk-000/file-000.parquet",
columns=["episode_index", "frame_index", "action"],
)
ep_arr = np.asarray(tbl.column("episode_index"))
fr_arr = np.asarray(tbl.column("frame_index"))
act_arr = np.stack(tbl.column("action").to_numpy(zero_copy_only=False))
key2row = {(int(e), int(f)): i for i, (e, f) in enumerate(zip(ep_arr, fr_arr))}
STRIDE = 32
keep, Y_rows = [], []
for i in range(N):
e, b = int(eps_all[i]), int(base_idx[i])
chunk, ok = [], True
for k in range(0, STRIDE, 4):
r = key2row.get((e, b + k))
if r is None:
ok = False
break
chunk.append(act_arr[r])
if ok:
keep.append(i)
Y_rows.append(np.concatenate(chunk))
keep = np.array(keep)
Y = np.asarray(Y_rows, dtype=np.float64)
eps_k = eps_all[keep]
print(f" 对齐样本 {len(keep)} | action {Y.shape[1]}维")
uniq = np.unique(eps_k)
rng = np.random.RandomState(0)
rng.shuffle(uniq)
tr_eps = set(uniq[: max(1, int(len(uniq) * 0.7))].tolist())
tr = np.array([e in tr_eps for e in eps_k])
te = ~tr
print(f" episode 级切分: train {tr.sum()} / test {te.sum()}, {len(uniq)} episodes")
Xz = Z[keep].numpy().astype(np.float64)
# u_t -> 32 dims via PCA fit on TRAIN episodes only (no leakage)
U = dec_in[keep, 0].mean(dim=1).numpy().astype(np.float64) # [n, 2048] token-mean
Um = U[tr].mean(0)
Uc = U - Um
_, _, Vt = np.linalg.svd(Uc[tr], full_matrices=False)
Xu = Uc @ Vt[:32].T
# also the delta feature u_t itself at full token resolution, PCA-32
Uf = dec_in[keep, 0].reshape(len(keep), -1).numpy().astype(np.float64)
Ufm = Uf[tr].mean(0)
Ufc = Uf - Ufm
_, _, Vt2 = np.linalg.svd(Ufc[tr], full_matrices=False)
Xuf = Ufc @ Vt2[:32].T
r2_z, _ = ridge_r2(Xz, Y, tr, te, "z (32维, 本模型的latent action)")
r2_u, _ = ridge_r2(Xu, Y, tr, te, "PCA32(u_t token均值) [对照]")
r2_uf, _ = ridge_r2(Xuf, Y, tr, te, "PCA32(u_t 全token) [对照]")
r2_cat, _ = ridge_r2(np.concatenate([Xz, Xuf], 1), Y, tr, te, "z + PCA32(u_t) 拼接")
print("\n=== 汇总 ===")
print(f" z 单独 R^2 = {r2_z:.4f}")
print(f" u_t 单独 (最好) R^2 = {max(r2_u, r2_uf):.4f}")
print(f" 增量 (z - u_t) = {r2_z - max(r2_u, r2_uf):+.4f}")
print(f" 拼接 R^2 = {r2_cat:.4f}")
print(f" delta_z 跨全集 = {mse(R_glob, T_) - mse_pred:.4f}")
print(f" delta_z 同episode内 = {mse(R_within, T_) - mse_pred:.4f} <- 排除场景/任务混淆")
if __name__ == "__main__":
main()
|