File size: 11,706 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 | """LaWAM Stage-1 (VGGT) evaluation: is the latent action `z` actually alive?
A low `val/recon_loss` and a 0.99 cosine only prove the world model can predict
future VGGT features. They do NOT prove `z` carries action information -- the
decoder could be ignoring `z` entirely and just learning "copy the input plus an
average motion prior". VGGT features make this especially easy to fake: ~75% of
every token is a constant shared across all tokens, so even the identity map
scores a high raw cosine.
Two probes decide whether Stage 2 is viable at all:
1. delta_z -- re-run the decoder with `z` shuffled across the batch. If the loss
barely moves, `z` is decorative and no Stage-2 policy can recover actions.
Reported against the trivial baselines (predict-the-mean, predict-identity)
so the numbers are interpretable rather than just "small".
2. z -> action linear probe -- ridge regression from `z` to the ground-truth
action chunk, scored by R^2 on held-out episodes. Stage 2 trains a policy to
emit `z`; if even a *linear* map cannot recover actions from `z`, a learned
policy has nothing to aim at.
All feature-space metrics are computed on standardized features. Raw VGGT
cosines are rank-inverted (cross-episode pairs score *higher* than temporal
pairs) because of that shared constant component, so uncentered numbers would be
actively misleading.
"""
import argparse
import importlib.machinery
import sys
import types
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):
"""Stub a module that `latent_action_model.core` imports but we don't need."""
mod = types.ModuleType(name)
# accelerate probes importlib.util.find_spec("wandb"), which raises if
# __spec__ is None on an already-imported module.
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
@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=40)
ap.add_argument("--batch-size", type=int, default=16)
ap.add_argument("--val-tail-ratio", type=float, default=0.05, help="held-out episode fraction")
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"[eval] ckpt={args.ckpt}")
print(f"[eval] encoder={type(model.vision_encoder).__name__} input_dim={model.input_dim}")
# Held-out episodes. The training run used val_tail_ratio=0.001 (1 episode);
# 5% (~85 episodes) gives probe estimates that aren't dominated by noise.
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,
)
# ---- pass 1: collect features, latents, predictions -------------------
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"[eval] loader exhausted after {i} batches")
break
videos = batch["videos"].to(dev)
# eval mode: no augmentation, video1 == video2 (matches validation_step)
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) % 10 == 0:
print(f"[eval] pass1 {i+1}/{args.n_batches}")
dec_in = torch.cat(all_dec_in) # [N,1,K,D] features at time t
tgt = torch.cat(all_tgt) # [N,1,K,D] features at t+tau (target)
recon = torch.cat(all_recon) # [N,1,K,D] prediction
Z = torch.cat(all_z) # [N, code_dim]
base_idx = torch.cat(all_base).numpy()
N = dec_in.shape[0]
print(f"\n[eval] N={N} samples | z dim={Z.shape[1]}")
# ---- standardize in feature space -------------------------------------
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(R_, T_)
mse_mean = mse(torch.zeros_like(T_), T_) # predict dataset mean == 1.0
mse_identity = mse(I_, T_) # predict "no change"
print("\n=== 特征空间基线 (标准化后) ===")
print(f" mse_pred = {mse_pred:.4f} <- 模型预测")
print(f" mse_mean = {mse_mean:.4f} <- 预测数据集均值 (定义上=1.00)")
print(f" mse_identity = {mse_identity:.4f} <- 预测 u_t (抄袭输入)")
beats = mse_pred < min(mse_mean, mse_identity)
print(f" => 击败所有平凡基线: {'YES' if beats else 'NO'}")
# centered cosines (Fig.10-style curves)
def cos_c(a, b):
a2 = (a - a.mean(dim=-1, keepdim=True)).reshape(-1, a.shape[-1])
b2 = (b - b.mean(dim=-1, keepdim=True)).reshape(-1, b.shape[-1])
return F.cosine_similarity(a2, b2, dim=-1).mean().item()
print(f" cos_pred_gt = {cos_c(R_, T_):.4f} (预测 vs 真实未来)")
print(f" cos_init_gt = {cos_c(I_, T_):.4f} (当前 vs 真实未来)")
print(f" cos_pred_init = {cos_c(R_, I_):.4f} (预测 vs 当前)")
# ---- probe 1: delta_z (shuffle z, re-decode) ---------------------------
# Re-run only the decoder with permuted latents; everything else identical.
print("\n=== Probe 1: delta_z (打乱 z 重新解码) ===")
perm = torch.randperm(N)
shuf_mse = []
bs = args.batch_size
for s in range(0, N, bs):
e = min(s + bs, N)
feats = dec_in[s:e].to(dev)
z_shuf = Z[perm[s:e]].to(dev).unsqueeze(1)
with torch.autocast("cuda", dtype=torch.bfloat16):
r = model.decoder(feats, z_shuf).float().cpu()
shuf_mse.append(standardize(r, mu, sd))
R_shuf = torch.cat(shuf_mse)
mse_shuf = mse(R_shuf, T_)
delta_z = mse_shuf - mse_pred
print(f" mse_pred = {mse_pred:.4f}")
print(f" mse_shuffled = {mse_shuf:.4f}")
print(f" delta_z = {delta_z:.4f} (相对提升 {100*delta_z/max(mse_pred,1e-9):.1f}%)")
print(f" => z 是否在驱动预测: {'YES' if delta_z > 0.02 else 'NO — z 基本是死的'}")
# ---- probe 2: z -> action linear probe --------------------------------
print("\n=== Probe 2: z -> 真实动作 线性探针 (ridge, R^2) ===")
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))
# (episode, frame) -> row
key2row = {(int(e), int(f)): i for i, (e, f) in enumerate(zip(ep_arr, fr_arr))}
# z spans [t, t+32] (frame_dt_sec=1.6 @ 20fps). Use the whole action chunk.
STRIDE = 32
X_rows, Y_rows = [], []
for i in range(N):
ep = int(all_traj[i])
b = int(base_idx[i])
chunk = []
ok = True
for k in range(0, STRIDE, 4): # subsample the 32-step chunk -> 8 x 7 = 56 dims
r = key2row.get((ep, b + k))
if r is None:
ok = False
break
chunk.append(act_arr[r])
if ok:
X_rows.append(Z[i].numpy())
Y_rows.append(np.concatenate(chunk))
if len(X_rows) < 50:
print(f" 跳过: 只对齐上 {len(X_rows)} 个样本")
return
X = np.asarray(X_rows, dtype=np.float64)
Y = np.asarray(Y_rows, dtype=np.float64)
print(f" 对齐样本 {X.shape[0]} | z {X.shape[1]}维 -> action {Y.shape[1]}维")
# split by episode so train/test never share an episode
eps = np.array([int(all_traj[i]) for i in range(N)][: len(X_rows)])
uniq = np.unique(eps)
rng = np.random.RandomState(0)
rng.shuffle(uniq)
n_tr = max(1, int(len(uniq) * 0.7))
tr_eps = set(uniq[:n_tr].tolist())
tr = np.array([e in tr_eps for e in eps])
te = ~tr
if te.sum() < 10 or tr.sum() < 10:
print(f" 跳过: episode 切分后训练/测试太小 ({tr.sum()}/{te.sum()})")
return
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]:
A = Xtr.T @ Xtr + lam * np.eye(Xtr.shape[1])
W = np.linalg.solve(A, Xtr.T @ Ytr)
P = Xte @ W
ss_res = ((Yte - P) ** 2).sum()
ss_tot = (Yte ** 2).sum()
r2 = 1 - ss_res / max(ss_tot, 1e-12)
if r2 > best[0]:
best = (r2, lam)
r2, lam = best
print(f" episode 级切分: train {tr.sum()} / test {te.sum()} 样本, {len(uniq)} episodes")
print(f" best ridge lambda={lam}")
print(f" z2action_r2 = {r2:.4f}")
print(f" => {'YES' if r2 > 0.3 else 'NO — 线性解不出动作'}")
print("\n=== 汇总 ===")
print(f" mse_pred {mse_pred:.4f} vs identity {mse_identity:.4f} vs mean {mse_mean:.4f}")
print(f" delta_z = {delta_z:.4f} (目标 > 0.02, 理想 > 0.1)")
print(f" z2action_r2 = {r2:.4f} (目标 > 0.3)")
if __name__ == "__main__":
main()
|