File size: 18,213 Bytes
2adb41a | 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 | """
ml/models/geometry/sequence_vae.py
1D-CNN Variational Autoencoder (VAE) for Novel Geometry Anomaly Detection (tc.v1 SOTA).
Features:
1. Dilated 1D-CNN Encoder (captures 1-3m, 3-10m, and 10-30m wavelengths)
2. Dual-Path Anomaly Scoring (Reconstruction MSE + Latent Mahalanobis Distance)
3. Sigmoid Distance Calibration for [0.0, 1.0] confidence mapping
"""
import os
from pathlib import Path
from typing import Dict, Tuple, Optional, Union, Any, List
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from ml.core.schema import CalibratedSignal, SignalType, DefectClass
from ml.core.registry import register_model
from ml.calibration.patchcore_scale import SigmoidDistanceCalibrator
from ml.models.geometry.sequence_vae_enhanced import EnhancedSequenceVAE
class DilatedEncoder(nn.Module):
"""
Multi-Scale Dilated 1D-CNN Encoder capturing short (1-3m), medium (3-10m),
and long (10-30m) spatial track wavelengths simultaneously.
"""
def __init__(self, n_features: int = 5, seq_len: int = 80, latent_dim: int = 16):
super().__init__()
self.n_features = n_features
self.seq_len = seq_len
self.latent_dim = latent_dim
# Parallel multi-scale dilated convolutions
self.branch_short = nn.Conv1d(n_features, 32, kernel_size=3, dilation=1, padding=1) # 1-3m wavelengths
self.branch_medium = nn.Conv1d(n_features, 32, kernel_size=3, dilation=4, padding=4) # 3-10m wavelengths
self.branch_long = nn.Conv1d(n_features, 32, kernel_size=3, dilation=10, padding=10) # 10-30m wavelengths
self.fusion = nn.Sequential(
nn.Conv1d(96, 64, kernel_size=1),
nn.ReLU(),
nn.AdaptiveAvgPool1d(20), # Downsample sequence to 20 spatial bins
)
self.flat_size = 64 * 20 # 1280
# Latent Space Projections
self.fc_mu = nn.Linear(self.flat_size, latent_dim)
self.fc_logvar = nn.Linear(self.flat_size, latent_dim)
def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
# x: [batch, n_features, seq_len]
short = F.relu(self.branch_short(x))
medium = F.relu(self.branch_medium(x))
long = F.relu(self.branch_long(x))
combined = torch.cat([short, medium, long], dim=1) # [B, 96, seq_len]
h = self.fusion(combined).flatten(1) # [B, 1280]
mu = self.fc_mu(h)
logvar = self.fc_logvar(h)
return mu, logvar
class SequenceVAE(nn.Module):
"""
1D-CNN Variational Autoencoder with Dilated Multi-Scale Encoder
and Transposed Conv1D Decoder for track geometry waveform reconstruction.
"""
def __init__(
self,
seq_len: int = 80,
n_features: int = 5,
latent_dim: int = 16,
):
super().__init__()
self.seq_len = seq_len
self.n_features = n_features
self.latent_dim = latent_dim
# Multi-scale dilated encoder
self.encoder = DilatedEncoder(n_features=n_features, seq_len=seq_len, latent_dim=latent_dim)
self.flat_size = self.encoder.flat_size
# Decoder: Reconstructs 5-channel sequence back to 80 spatial bins
self.decoder_input = nn.Linear(latent_dim, self.flat_size)
self.decoder = nn.Sequential(
nn.Unflatten(1, (64, 20)),
nn.ConvTranspose1d(64, 32, kernel_size=3, stride=2, padding=1, output_padding=1), # 20 -> 40
nn.ReLU(),
nn.ConvTranspose1d(32, n_features, kernel_size=3, stride=2, padding=1, output_padding=1), # 40 -> 80
)
def reparameterize(self, mu: torch.Tensor, logvar: torch.Tensor) -> torch.Tensor:
std = torch.exp(0.5 * logvar)
eps = torch.randn_like(std)
return mu + eps * std
def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
if x.dim() == 2:
x = x.unsqueeze(0)
# Transpose [batch, seq_len, features] -> [batch, features, seq_len] for Conv1d
x_trans = x.transpose(1, 2)
mu, logvar = self.encoder(x_trans)
z = self.reparameterize(mu, logvar)
h = F.relu(self.decoder_input(z))
recon_trans = self.decoder(h)
recon_x = recon_trans.transpose(1, 2) # [batch, seq_len, features]
return recon_x, mu, logvar
def loss_function(
self,
recon_x: torch.Tensor,
x: torch.Tensor,
mu: torch.Tensor,
logvar: torch.Tensor,
beta: float = 0.01,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
if recon_x.shape != x.shape:
min_len = min(recon_x.shape[1], x.shape[1])
recon_x = recon_x[:, :min_len, :]
x = x[:, :min_len, :]
recon_loss = F.mse_loss(recon_x, x, reduction="mean")
kld_loss = -0.5 * torch.mean(1 + logvar - mu.pow(2) - logvar.exp())
total_loss = recon_loss + beta * kld_loss
return total_loss, recon_loss, kld_loss
def compute_reconstruction_error(self, x: torch.Tensor) -> torch.Tensor:
self.eval()
with torch.no_grad():
recon_x, _, _ = self.forward(x)
if recon_x.shape != x.shape:
min_len = min(recon_x.shape[1], x.shape[1])
recon_x = recon_x[:, :min_len, :]
x = x[:, :min_len, :]
error = torch.mean((recon_x - x) ** 2, dim=(1, 2))
return error
def compute_anomaly_score(self, x: torch.Tensor) -> torch.Tensor:
"""Alias for compute_reconstruction_error."""
return self.compute_reconstruction_error(x)
@register_model("geometry_sequence_vae")
class SequenceVAEDetector:
"""
Production detector wrapper for SequenceVAE featuring:
- Enhanced multi-scale dilated convolutions matching trained SOTA checkpoint
- Dual-Path Anomaly Scoring (Reconstruction MSE + Latent Mahalanobis Distance)
- EVT & Sigmoid Calibration for calibrated [0.0, 1.0] novelty probabilities
"""
def __init__(
self,
weights_path: Optional[Union[str, Path]] = "artifacts/checkpoints/geometry/sequence_vae_enhanced.pt",
calibrator_path: Optional[Union[str, Path]] = "artifacts/calibration/vae_calibration.json",
device: str = "cpu",
seq_len: int = 80,
n_features: int = 5,
latent_dim: int = 16,
alpha: float = 0.7,
threshold: float = 0.50,
):
self.device = torch.device(device if torch.cuda.is_available() or device == "cpu" else "cpu")
self.seq_len = seq_len
self.n_features = n_features
self.latent_dim = latent_dim
self.alpha = float(alpha)
self.threshold = threshold
# Default model is EnhancedSequenceVAE matching trained SOTA checkpoint ([128, 192, 1] fusion)
self.model: Union[EnhancedSequenceVAE, SequenceVAE] = EnhancedSequenceVAE(
seq_len=seq_len, n_features=n_features, latent_dim=latent_dim
).to(self.device)
# Resolve candidate weight paths in order of preference
candidate_weights = [weights_path] if weights_path else []
candidate_weights.extend([
"artifacts/checkpoints/geometry/sequence_vae_enhanced.pt",
"artifacts/checkpoints/geometry/sequence_vae.pt",
"ml/models/geometry/weights/sequence_vae.pt",
"ml/models/geometry/weights/sequence_vae_enhanced.pt",
])
loaded_checkpoint = False
for wp in candidate_weights:
if wp and os.path.exists(str(wp)):
try:
ckpt = torch.load(wp, map_location=self.device)
state = ckpt["model_state_dict"] if isinstance(ckpt, dict) and "model_state_dict" in ckpt else ckpt
# Auto-detect architecture from state dict
if isinstance(state, dict) and "encoder.fusion.0.weight" in state:
in_channels = state["encoder.fusion.0.weight"].shape[1]
if in_channels == 96:
self.model = SequenceVAE(seq_len=seq_len, n_features=n_features, latent_dim=latent_dim).to(self.device)
else:
self.model = EnhancedSequenceVAE(seq_len=seq_len, n_features=n_features, latent_dim=latent_dim).to(self.device)
self.model.load_state_dict(state, strict=True)
loaded_checkpoint = True
break
except Exception as e:
if weights_path is not None and str(wp) == str(weights_path):
raise RuntimeError(
f"VAE weight load FAILED for {weights_path} — architecture mismatch. "
f"Do NOT fall back to a random model. Error: {e}"
)
self.model.eval()
# Latent distribution baseline for Mahalanobis distance
self.normal_latent_mean: Optional[np.ndarray] = None
self.normal_latent_cov_inv: Optional[np.ndarray] = None
# Sigmoid / EVT threshold calibrator
self.calibrator = SigmoidDistanceCalibrator(threshold=1.65, steepness_k=2.0)
candidate_calibs = [calibrator_path] if calibrator_path else []
candidate_calibs.extend([
"artifacts/calibration/vae_calibration.json",
"artifacts/calibration/sequence_vae_calibration.json",
])
loaded_calib = False
for cp in candidate_calibs:
if cp and os.path.exists(str(cp)):
try:
with open(cp, "r", encoding="utf-8") as f:
cdata = json.load(f)
thresh = float(cdata.get("threshold_evt", cdata.get("threshold_p99", 1.65)))
k_val = float(cdata.get("steepness_k", cdata.get("steepness", 2.0)))
self.calibrator = SigmoidDistanceCalibrator(threshold=thresh, steepness_k=k_val)
if isinstance(self.model, EnhancedSequenceVAE):
self.model.threshold_evt = thresh
self.model.steepness_k = k_val
loaded_calib = True
break
except Exception:
pass
if not loaded_calib:
# Auto-initialize baseline distribution on nominal noise
synth_baseline = np.random.normal(0.0, 0.1, (30, self.seq_len, self.n_features)).astype(np.float32)
self.fit_latent_distribution(synth_baseline)
synth_scores = [self.compute_anomaly_score(s) for s in synth_baseline]
self.calibrator.fit(synth_scores, percentile=99.0)
self.calibrator.threshold = max(self.calibrator.threshold, 0.5)
def _format_input(self, geometry_window: Union[np.ndarray, Dict[str, np.ndarray], torch.Tensor]) -> torch.Tensor:
if isinstance(geometry_window, dict):
keys = [
("twist_3m", "twist_3m_mm"),
("versine_10m", "versine_10m_mm"),
("versine_20m", "versine_20m_mm"),
("unevenness_10m", "unevenness_10m_mm"),
("cant", "cant_mm"),
]
cols = []
for k1, k2 in keys:
if k1 in geometry_window:
cols.append(np.asarray(geometry_window[k1]))
elif k2 in geometry_window:
cols.append(np.asarray(geometry_window[k2]))
else:
cols.append(np.zeros(self.seq_len))
arr = np.column_stack(cols)
elif isinstance(geometry_window, torch.Tensor):
arr = geometry_window.cpu().numpy()
else:
arr = np.asarray(geometry_window)
if arr.ndim == 2:
arr = np.expand_dims(arr, axis=0) # [1, seq_len, 5]
b, t, c = arr.shape
if t != self.seq_len:
fixed = np.zeros((b, self.seq_len, c), dtype=np.float32)
copy_len = min(t, self.seq_len)
fixed[:, :copy_len, :] = arr[:, :copy_len, :]
arr = fixed
return torch.tensor(arr, dtype=torch.float32).to(self.device)
def fit_latent_distribution(self, normal_sequences: Union[np.ndarray, torch.Tensor, List[np.ndarray]]):
"""Fits baseline latent Gaussian distribution (mean and regularized inverse covariance)."""
if isinstance(self.model, EnhancedSequenceVAE):
if isinstance(normal_sequences, list):
normal_tensor = torch.stack([torch.tensor(s, dtype=torch.float32) for s in normal_sequences])
elif isinstance(normal_sequences, np.ndarray):
normal_tensor = torch.tensor(normal_sequences, dtype=torch.float32)
else:
normal_tensor = normal_sequences
self.model.fit_latent_distribution(normal_tensor)
self.normal_latent_mean = self.model.latent_mean
self.normal_latent_cov_inv = self.model.latent_cov_inv
return
self.model.eval()
latents = []
if isinstance(normal_sequences, list):
normal_tensor = torch.stack([torch.tensor(s, dtype=torch.float32) for s in normal_sequences])
elif isinstance(normal_sequences, np.ndarray):
normal_tensor = torch.tensor(normal_sequences, dtype=torch.float32)
else:
normal_tensor = normal_sequences
if normal_tensor.ndim == 2:
normal_tensor = normal_tensor.unsqueeze(0)
with torch.no_grad():
normal_tensor = normal_tensor.to(self.device)
_, mu, _ = self.model(normal_tensor)
latents = mu.cpu().numpy()
self.normal_latent_mean = np.mean(latents, axis=0)
cov = np.cov(latents.T)
if cov.ndim == 0:
cov = np.array([[cov]])
cov_reg = cov + np.eye(cov.shape[0]) * 1e-6
self.normal_latent_cov_inv = np.linalg.inv(cov_reg)
def compute_anomaly_score(self, sequence: Union[np.ndarray, Dict[str, np.ndarray], torch.Tensor]) -> float:
"""
Computes Dual-Path Anomaly Score:
Score = alpha * MSE_recon + (1 - alpha) * D_mahalanobis
"""
if isinstance(self.model, EnhancedSequenceVAE):
if isinstance(sequence, dict):
tensor_in = self._format_input(sequence)
res = self.model.compute_anomaly_score(tensor_in)
else:
res = self.model.compute_anomaly_score(sequence)
return float(res.get("combined_score", res.get("recon_error", 0.0)))
self.model.eval()
tensor_in = self._format_input(sequence)
with torch.no_grad():
recon_x, mu, _ = self.model(tensor_in)
mse = float(F.mse_loss(recon_x, tensor_in).item())
# Path 2: Latent Mahalanobis Distance
if self.normal_latent_mean is not None and self.normal_latent_cov_inv is not None:
mu_np = mu.cpu().numpy()[0]
diff = mu_np - self.normal_latent_mean
mahal_dist = float(np.sqrt(np.clip(diff @ self.normal_latent_cov_inv @ diff.T, 0.0, None)))
else:
mahal_dist = float(torch.norm(mu).item())
combined_score = float((self.alpha * mse) + ((1.0 - self.alpha) * mahal_dist))
return combined_score
def fit_calibration(self, normal_windows: Union[np.ndarray, torch.Tensor], percentile: float = 99.0):
"""Fits latent distribution and Sigmoid distance calibrator on nominal track validation sequences."""
self.fit_latent_distribution(normal_windows)
scores = []
if isinstance(normal_windows, torch.Tensor):
arr = normal_windows.cpu().numpy()
else:
arr = np.asarray(normal_windows)
for seq in arr:
scores.append(self.compute_anomaly_score(seq))
self.calibrator.fit(scores, percentile=percentile)
return self.calibrator.threshold
def predict(self, geometry_window: Union[np.ndarray, Dict[str, np.ndarray], torch.Tensor]) -> CalibratedSignal:
"""
Runs the 1D-CNN Dual-Path VAE on the 20m geometry window.
Returns a calibrated CalibratedSignal indicating structural novelty.
"""
if isinstance(self.model, EnhancedSequenceVAE):
if isinstance(geometry_window, dict):
tensor_in = self._format_input(geometry_window)
pred = self.model.predict(tensor_in)
else:
pred = self.model.predict(geometry_window)
raw_score = float(pred["raw_score"])
calibrated_prob = float(pred["calibrated_prob"])
is_anomaly = bool(pred["is_anomaly"])
thresh = float(pred.get("threshold", self.threshold))
else:
raw_score = self.compute_anomaly_score(geometry_window)
calibrated_prob = float(self.calibrator.scale(raw_score))
is_anomaly = bool(calibrated_prob >= self.threshold)
thresh = self.threshold
explanation = {
"combined_anomaly_score": round(raw_score, 4),
"reconstruction_error": round(raw_score, 4),
"calibrated_prob": round(calibrated_prob, 4),
"p99_threshold": round(self.calibrator.threshold, 4),
"anomaly_detected": is_anomaly,
}
return CalibratedSignal(
name="sequence_vae_anomaly",
stream_name="geometry_vae",
model_version="0.1.0",
signal_type=SignalType.GEOMETRY_NOVEL,
value=calibrated_prob,
raw_score=raw_score,
calibrated_prob=calibrated_prob,
threshold=thresh,
fired=is_anomaly,
is_anomaly=is_anomaly,
predicted_class=DefectClass.GEOMETRY_ANOMALY if is_anomaly else DefectClass.NORMAL,
bbox=None,
explanation=explanation,
metadata={
"model_name": "sequence_vae_dual_path",
"latent_dim": self.model.latent_dim,
"combined_score": raw_score,
"calibrator_threshold": self.calibrator.threshold,
},
)
|