File size: 16,127 Bytes
436b829 | 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 | """
Per-claim verification of the LPD implementation against paper.tex.
Each `check()` block ties one specific paper claim to the code that
realises it; failures are reported with the missing/incorrect piece. Run as
cd /mnt/sig/pixel-perfect-depth
python -m ppd.lpd.tests.verify_paper
Exits with non-zero status if any check fails. Designed to be a
single-pass audit, not a unit-test suite — it executes real tensor ops
on small inputs to confirm shapes, equations, and gradient flow.
"""
from __future__ import annotations
import os
import sys
import math
import inspect
from typing import Callable
import torch
import torch.nn.functional as F
CHECKS: list[tuple[str, Callable[[], None]]] = []
def check(name: str):
def deco(fn: Callable[[], None]):
CHECKS.append((name, fn))
return fn
return deco
def assert_close(actual: torch.Tensor, expected: torch.Tensor, msg: str, atol: float = 1e-5):
if not torch.allclose(actual, expected, atol=atol):
diff = (actual - expected).abs().max().item()
raise AssertionError(f"{msg}: max diff {diff:.3e}")
# ----------------------------------------------------------------- §3.1 image
@check("§3.1: sparse-prompt encoder pools at scales {4, 8, 16, 32}")
def _():
from ppd.lpd.prompt_encoder import SparsePromptEncoder
enc = SparsePromptEncoder()
assert tuple(enc.scales) == (4, 8, 16, 32), f"scales={enc.scales}"
@check("§3.1: encoder produces both depth and density per scale")
def _():
from ppd.lpd.prompt_encoder import masked_avg_pool
d = torch.zeros(1, 1, 32, 32); d[:, :, ::8, ::8] = 1.0
m = (d > 0).float()
pooled, density = masked_avg_pool(d, m, kernel=4)
assert pooled.shape == density.shape == (1, 1, 8, 8)
# density at a fully-observed cell should be 1/16 (one observation in 4x4)
assert (density.max() - 1 / 16.0).abs() < 1e-5
# mask sum 16; pooled should equal 1 at sampled cells (since masked avg)
assert pooled.max() == 1.0
@check("§3.1: encoder applies a two-layer CNN + linear projection")
def _():
from ppd.lpd.prompt_encoder import SparsePromptEncoder, _SmallCNN
enc = SparsePromptEncoder()
# _SmallCNN: Conv → GELU → Conv (= 2 convs)
n_convs = sum(1 for m in enc.per_scale[0].net if isinstance(m, torch.nn.Conv2d))
assert n_convs == 2, f"expect two convs, got {n_convs}"
# Final projection should be a linear layer.
assert isinstance(enc.fuse, torch.nn.Linear)
@check("§3.1: prompt-aware quantile log normalization produces ~[-0.5, 0.5]")
def _():
from ppd.lpd.prompt_encoder import quantile_log_normalize
d = torch.linspace(0.5, 50.0, 256).reshape(1, 1, 16, 16)
m = torch.ones_like(d)
nd = quantile_log_normalize(d, m)
assert -0.5 - 1e-3 <= nd.min().item() <= 0.05
assert 0.5 - 0.05 <= nd.max().item() <= 1.0
@check("§3.1 Eq.(1): prompt gate computes s_sem + g(p,ρ,t) ⊙ m(s_sem,p,ρ,t)")
def _():
from ppd.lpd.prompt_gate import PromptGate
g = PromptGate(embed_dim=32, timestep_dim=32, hidden=32)
B, T, D = 2, 8, 32
s_sem = torch.randn(B, T, D)
p = torch.randn(B, T, D)
rho = torch.rand(B, T, 1)
t = torch.randn(B, D)
out = g(s_sem, p, rho, t)
# zero-init last layers ⇒ delta=0 and gate output passes sigmoid(0)=0.5,
# so g*delta = 0 and the joint should equal s_sem on init.
assert_close(out, s_sem, "joint should equal s_sem at init", atol=1e-5)
@check("§3.1: m and g are zero-initialized so model starts as pretrained PPD")
def _():
from ppd.lpd.prompt_gate import PromptGate
g = PromptGate(embed_dim=16, timestep_dim=16, hidden=16)
# last layer of mixer must be zero
assert torch.all(g.mixer[-1].weight == 0)
assert torch.all(g.mixer[-1].bias == 0)
# gate's pre-sigmoid linear must be zero (Linear is at index -2 before Sigmoid)
assert torch.all(g.gate[-2].weight == 0)
assert torch.all(g.gate[-2].bias == 0)
@check("§3.1: timestep embedding is projected before entering the gate")
def _():
# LPDDiT calls self.t_embedder(timestep) which contains a 2-layer MLP
from ppd.models.dit import TimestepEmbedder
t_embed = TimestepEmbedder(hidden_size=32)
assert isinstance(t_embed.mlp, torch.nn.Sequential)
assert sum(1 for m in t_embed.mlp if isinstance(m, torch.nn.Linear)) == 2
# ----------------------------------------------------------------- §3.3 score decomp
@check("§3.3 Eq.(5): LiDAR likelihood gradient = -M ⊙ (x - y) / R")
def _():
from ppd.lpd.posterior_projection import posterior_project
x = torch.full((1, 1, 4, 4), 0.5)
y = torch.full((1, 1, 4, 4), 0.0)
M = torch.ones_like(x)
out = posterior_project(
x, sigma_t=torch.tensor(1.0),
sparse_depth=y, sparse_mask=M, R=0.1,
mu_prior=None, P_prior=None, alpha=1.0,
)
# eta = sigma² · alpha = 1, kalman term zero ⇒ x ← x + 1 · (-M⊙(x-y)/R)
expected = x + 1.0 * (-M * (x - y) / 0.1)
assert_close(out, expected, "Eq.(5) projection")
@check("§3.3 Eq.(6): Kalman temporal-prior gradient = -(x - μ) / P")
def _():
from ppd.lpd.posterior_projection import posterior_project
x = torch.full((1, 1, 4, 4), 0.3)
mu = torch.full((1, 1, 4, 4), 0.1)
P = torch.full((1, 1, 4, 4), 0.5)
out = posterior_project(
x, sigma_t=torch.tensor(1.0),
sparse_depth=torch.zeros_like(x),
sparse_mask=torch.zeros_like(x),
R=0.1,
mu_prior=mu, P_prior=P, alpha=1.0,
)
expected = x + 1.0 * (-(x - mu) / P)
assert_close(out, expected, "Eq.(6) Kalman prior gradient")
@check("§3.3 Eq.(7): η_τ = α · σ_τ²")
def _():
from ppd.lpd.posterior_projection import posterior_project
x = torch.full((1, 1, 2, 2), 1.0)
y = torch.zeros_like(x)
M = torch.ones_like(x)
sigma = torch.tensor(0.5)
out = posterior_project(
x, sigma, sparse_depth=y, sparse_mask=M, R=1.0,
mu_prior=None, P_prior=None, alpha=2.0,
)
eta = 2.0 * 0.5 ** 2 # = 0.5
expected = x + eta * (-M * (x - y) / 1.0)
assert_close(out, expected, "Eq.(7) step-size schedule")
# ----------------------------------------------------------------- §3.4 KIL
@check("§3.4 Algorithm 1: Kalman gain K = P / (P + σ²)")
def _():
P = torch.tensor(0.5); sig2 = torch.tensor(0.25)
K = P / (P + sig2)
assert (K - 2 / 3).abs() < 1e-6
@check("§3.4 Algorithm 1: variance update P_τ = (1-K) P_{τ-1} ⇒ monotone decrease")
def _():
P = torch.tensor(1.0)
for sig2 in [1.0, 0.5, 0.25, 0.0625]:
K = P / (P + sig2)
P_new = (1 - K) * P
assert P_new <= P + 1e-9, "variance must not grow"
P = P_new
@check("§3.4: Kalman state μ_τ = μ_{τ-1} + K (x̂_0 - μ_{τ-1})")
def _():
mu = torch.tensor(0.0); P = torch.tensor(1.0)
x_hat = torch.tensor(1.0); sig2 = torch.tensor(1.0)
K = P / (P + sig2)
mu_new = mu + K * (x_hat - mu)
assert (mu_new - 0.5).abs() < 1e-6
@check("§3.4: kalman_in_loop_sample returns (depth, posterior_variance)")
def _():
import inspect
from ppd.lpd.kalman_in_loop import kalman_in_loop_sample
sig = inspect.signature(kalman_in_loop_sample)
assert "x_T" in sig.parameters and "sparse_depth" in sig.parameters
# ----------------------------------------------------------------- §3.5 temporal Kalman
@check("§3.5: predict step warps state and inflates variance by Q")
def _():
from ppd.lpd.temporal_kalman import TemporalKalmanFilter, TemporalKalmanConfig
kf = TemporalKalmanFilter(
shape=(1, 1, 8, 8), device=torch.device("cpu"),
config=TemporalKalmanConfig(Q_base=0.1, alpha=0.0, P_init=0.0, occ_threshold=999.0),
)
kf.mu.fill_(1.0)
kf.P.fill_(0.0)
kf.has_state = True
flow = torch.zeros(1, 2, 8, 8) # zero flow ⇒ identity warp
kf.predict(flow_fwd=flow, flow_bwd=flow)
# variance should grow by Q_base since alpha=0
assert (kf.P - 0.1).abs().max() < 1e-5
@check("§3.5 Eq.(9): forward-backward error ε = ||p + f_fwd + f_bwd(p+f_fwd)||")
def _():
from ppd.lpd.temporal_kalman import forward_backward_error
f_fwd = torch.zeros(1, 2, 8, 8); f_fwd[:, 0] = 2.0 # +2 in x
f_bwd = -f_fwd # exact inverse ⇒ ε ≈ 0
eps = forward_backward_error(f_fwd, f_bwd)
assert eps.max() < 1e-3, f"ε should be ~0, got {eps.max().item()}"
@check("§3.5: occluded pixels (ε > τ_occ) reset variance to P_max")
def _():
from ppd.lpd.temporal_kalman import TemporalKalmanFilter, TemporalKalmanConfig
kf = TemporalKalmanFilter(
shape=(1, 1, 8, 8), device=torch.device("cpu"),
config=TemporalKalmanConfig(P_max=99.0, occ_threshold=0.5),
)
kf.mu.fill_(1.0); kf.P.fill_(0.1); kf.has_state = True
f_fwd = torch.zeros(1, 2, 8, 8); f_fwd[:, 0] = 5.0 # 5px fwd
f_bwd = torch.zeros_like(f_fwd) # no return ⇒ ε = 5
kf.predict(f_fwd, f_bwd)
assert kf.P.max() >= 99.0
@check("§3.5: update step Kalman gain K = P / (P + R) at observed pixels")
def _():
from ppd.lpd.temporal_kalman import TemporalKalmanFilter, TemporalKalmanConfig
kf = TemporalKalmanFilter(
shape=(1, 1, 4, 4), device=torch.device("cpu"),
config=TemporalKalmanConfig(R=0.1, P_init=1.0),
)
sd = torch.full((1, 1, 4, 4), 0.5)
sm = torch.ones_like(sd)
mu, P = kf.update(sd, sm)
K = 1.0 / (1.0 + 0.1)
expected_mu = 0.0 + K * (0.5 - 0.0)
expected_P = (1 - K) * 1.0
assert (mu - expected_mu).abs().max() < 1e-5
assert (P - expected_P).abs().max() < 1e-5
@check("§3.5: at unobserved pixels (mask=0), state passes through unchanged")
def _():
from ppd.lpd.temporal_kalman import TemporalKalmanFilter, TemporalKalmanConfig
kf = TemporalKalmanFilter(
shape=(1, 1, 4, 4), device=torch.device("cpu"),
config=TemporalKalmanConfig(R=0.1, P_init=0.5),
)
kf.mu.fill_(0.7)
sd = torch.zeros(1, 1, 4, 4)
sm = torch.zeros_like(sd) # nothing observed
mu, P = kf.update(sd, sm)
assert (mu - 0.7).abs().max() < 1e-6
assert (P - 0.5).abs().max() < 1e-6
@check("§3.5: metric uncertainty = exp(sqrt(P)) - 1")
def _():
from ppd.lpd.temporal_kalman import TemporalKalmanFilter, TemporalKalmanConfig
kf = TemporalKalmanFilter(
shape=(1, 1, 1, 1), device=torch.device("cpu"),
config=TemporalKalmanConfig(P_init=0.25),
)
expected = math.exp(math.sqrt(0.25)) - 1
actual = kf.metric_uncertainty().item()
assert abs(actual - expected) < 1e-5
# ----------------------------------------------------------------- §3.6 modulation
@check("§3.6 Eq.(8): ρ̃(p) = ρ(p) · (1 + P(p)/max P)")
def _():
from ppd.lpd.uncertainty_modulation import modulate_density
rho = torch.full((1, 4, 1), 0.5)
P_full = torch.tensor([0.0, 0.5, 1.0, 2.0]).reshape(1, 1, 1, 4)
rho_tilde = modulate_density(rho, P_full)
# max P = 2.0; ρ̃ = 0.5 * (1 + P/2.0)
expected = 0.5 * (1 + P_full.squeeze(2).squeeze(1).reshape(1, 4, 1) / 2.0)
assert_close(rho_tilde, expected, "Eq.(8) modulation")
# ----------------------------------------------------------------- §3.7 training
@check("§3.7: anchor loss is L1(x̂_0 - y) over observed pixels")
def _():
from ppd.lpd.losses import anchor_loss
x = torch.tensor([[[[0.0, 0.5, 1.0, 0.0]]]]).float()
y = torch.tensor([[[[0.5, 0.5, 0.5, 0.0]]]]).float()
m = torch.tensor([[[[1.0, 1.0, 1.0, 0.0]]]])
# observed diffs: |0-0.5|+|0.5-0.5|+|1-0.5|=1.0; |M|=3 → 1/3
loss = anchor_loss(x, y, m).item()
assert abs(loss - 1.0 / 3) < 1e-6
@check("§3.7: total training loss combines MSE + λ_a anchor + λ_g grad")
def _():
src = inspect.getsource(__import__("ppd.lpd.lpd_train", fromlist=["LiDARPerfectDepth"]))
assert "lambda_anchor" in src
assert "anchor_loss" in src
assert "multi_scale_grad_loss" in src
@check("§3.7: backbone freeze leaves only prompt-encoder + gate trainable")
def _():
from ppd.lpd.lpd_dit import LPDDiT
m = LPDDiT(hidden_size=128, depth=4, num_heads=4, patch_size=8)
m.freeze_backbone()
for n, p in m.named_parameters():
if p.requires_grad:
assert n.startswith("sparse_prompt_encoder") or n.startswith("prompt_gate"), \
f"unexpected trainable: {n}"
# ----------------------------------------------------------------- §4.1 sparse simulator
@check("§4.1: sparse simulator implements random / scan_line / grid / hybrid")
def _():
from ppd.lpd.sparse_simulator import simulate
d = torch.ones(1, 1, 32, 32); m = torch.ones_like(d, dtype=torch.bool)
for pat in ["random", "scan_line", "grid", "hybrid"]:
sd, sm = simulate(d, m, pattern=pat, density=0.05, n_lines=4,
line_density=0.5, grid_stride=8, min_points=4)
assert sm.sum() > 0, f"{pat} produced no observations"
# depth should be zero where mask is false
assert (sd[~sm] == 0).all().item()
# ----------------------------------------------------------------- §4.4 implementation
@check("§4.4: temporal Kalman defaults — R=0.01, Q_base=0.005, α=0.5, P_max=10, τ_occ=2.0")
def _():
from ppd.lpd.temporal_kalman import TemporalKalmanConfig
c = TemporalKalmanConfig()
assert c.R == 0.01
assert c.Q_base == 0.005
assert c.alpha == 0.5
assert c.P_max == 10.0
assert c.occ_threshold == 2.0
@check("§4.4: posterior projection R_proj defaults to 0.1")
def _():
from ppd.lpd.kalman_in_loop import KalmanInLoopConfig
c = KalmanInLoopConfig()
assert c.R_proj == 0.1
@check("§4.4: PPD weights load with smart partial loading (strict=False)")
def _():
src = inspect.getsource(__import__("ppd.lpd.lpd_train", fromlist=["LiDARPerfectDepth"]))
assert "strict=False" in src
assert "_load_ppd_weights" in src
# ----------------------------------------------------------------- end-to-end shape sanity
@check("end-to-end: LPDDiT forward at training resolution returns (B,1,H,W)")
def _():
from ppd.lpd.lpd_dit import LPDDiT
m = LPDDiT(hidden_size=128, depth=4, num_heads=4, patch_size=8)
B, H, W = 1, 64, 64
x = torch.randn(B, 4, H, W)
sem = torch.randn(B, (H // 16) * (W // 16), 1024)
t = torch.tensor([100.0])
sd = torch.zeros(B, 1, H, W); sd[:, :, ::8, ::8] = 0.3
sm = (sd > 0)
out = m(x, sem, t, sparse_depth=sd, sparse_mask=sm)
assert out.shape == (B, 1, H, W)
@check("end-to-end: KIL sampler produces depth + variance maps with the right shapes")
def _():
from ppd.utils.diffusion.timesteps import Timesteps
from ppd.utils.diffusion.schedule import LinearSchedule
from ppd.utils.diffusion.sampler import EulerSampler
from ppd.lpd.kalman_in_loop import kalman_in_loop_sample, KalmanInLoopConfig
sched = LinearSchedule(T=1000)
ts = Timesteps(T=1000, steps=4, device=torch.device("cpu"))
sampler = EulerSampler(schedule=sched, timesteps=ts, prediction_type="velocity")
B, H, W = 1, 32, 32
x_T = torch.randn(B, 1, H, W)
cond = torch.randn(B, 3, H, W)
sd = torch.zeros(B, 1, H, W); sm = torch.zeros_like(sd)
def predict(x_tau, tau): return torch.zeros_like(x_tau)
out, P = kalman_in_loop_sample(
dit_predict_x0=predict, sampler=sampler,
timesteps=list(ts), x_T=x_T, cond=cond,
semantics_fn=lambda: None,
sparse_depth=sd, sparse_mask=sm,
config=KalmanInLoopConfig(),
)
assert out.shape == (B, 1, H, W)
assert P.shape == (B, 1, H, W)
# ----------------------------------------------------------------- runner
def main() -> int:
ok = fail = 0
width = max(len(name) for name, _ in CHECKS)
for name, fn in CHECKS:
try:
fn()
print(f" ✓ {name.ljust(width)}")
ok += 1
except Exception as e:
print(f" ✗ {name.ljust(width)} → {type(e).__name__}: {e}")
fail += 1
print(f"\n{ok} passed, {fail} failed (of {len(CHECKS)})")
return 0 if fail == 0 else 1
if __name__ == "__main__":
sys.path.insert(0, os.getcwd())
sys.exit(main())
|