File size: 20,261 Bytes
bd4ab66 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 | """
cell_p_class_probe.py β geometric structure probe for P-Class batteries
Loads P-rank09 (h64_V32_D3_dp0_nx0_adam, MSE 0.028, CV 0.03) and asks
what its 32 row vectors in 3D space actually look like.
Four hypothesis tests:
1. RANK STRUCTURE β SVD on the 32Γ3 row matrix M.
- Polynomial basis: rank β€ 2 (Vandermonde collapses)
- Trig basis: rank = 2 or 3 with specific singular value ratio
- Cluster: rank 3, all SVs comparable
- Collapsed: rank 1, one dominant SV
2. PARAMETRIC ORDERING β Try ordering rows by their first coordinate
(or first principal axis projection). If rows form a smooth curve
when ordered, we're seeing a parametric structure (polynomial,
trig, etc). If they're scattered with no order, it's clusters.
Metric: smoothness of consecutive Ξ when sorted along PC1.
3. POLYNOMIAL FIT TEST β Fit a Vandermonde matrix to the ordered rows.
If RΒ² > 0.95 with cubic, polynomial hypothesis confirmed.
Try [1, x, xΒ²], [1, x, xΒ², xΒ³], [1, sin(x), cos(x)].
4. CLUSTER COUNT β k-means with k = 2..8 on the 32 rows. If silhouette
score is high at small k, it's clustered. If silhouette is low for
all k, the rows are spread continuously (consistent with parametric).
Outputs:
- Console verdict for each hypothesis
- /content/phaseQ_reports/p_rank09_probe.png β 4-panel diagnostic plot
- /content/phaseQ_reports/p_rank09_probe.json β all numerical results
"""
import json
import math
from pathlib import Path
import numpy as np
import torch
import torch.nn.functional as F
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D # noqa
from sklearn.cluster import KMeans
from sklearn.metrics import silhouette_score
CKPT_DIR = Path("/content/phaseQ_reports")
RANK09_CKPT = CKPT_DIR / "Q_rank09_h64_V32_D3_dp0_nx0_adam" / "epoch_1_checkpoint.pt"
OUTPUT_PLOT = CKPT_DIR / "p_rank09_probe.png"
OUTPUT_JSON = CKPT_DIR / "p_rank09_probe.json"
def load_rank09():
"""Reconstruct P-rank09 model and load its trained weights."""
cfgs = get_phaseQ_configs()
rank09_cfg = next(c for c in cfgs if 'rank09' in c['variant'])
cfg = build_run_config(rank09_cfg)
overrides = rank09_cfg['overrides']
model = PatchSVAE_F_Ablation(
matrix_v=cfg.matrix_v, D=cfg.D, patch_size=cfg.patch_size,
hidden=cfg.hidden, depth=cfg.depth,
n_cross_layers=cfg.n_cross_layers, n_heads=cfg.n_heads,
max_alpha=overrides.get('max_alpha', cfg.max_alpha),
alpha_init=cfg.alpha_init,
activation=overrides.get('activation', 'gelu'),
row_norm=overrides.get('row_norm', 'sphere'),
svd_mode=overrides.get('svd', 'fp64'),
linear_readout=overrides.get('linear_readout', False),
match_params=overrides.get('match_params', True),
init_scheme=overrides.get('init', 'orthogonal'),
)
ckpt = torch.load(RANK09_CKPT, map_location='cpu', weights_only=False)
# Trainer saves model weights under 'model_state'; the older
# 'model_state_dict' / 'state_dict' fallbacks are kept for compatibility.
state_dict = (
ckpt.get('model_state')
or ckpt.get('model_state_dict')
or ckpt.get('state_dict')
or ckpt
)
model.load_state_dict(state_dict)
model.eval()
return model, cfg
def collect_rows(model, cfg, n_batches=8, batch_size=64):
"""Run gaussian noise through encoder, collect M rows from one canonical
patch position to get a stable [n_samples, V, D] tensor of row matrices."""
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = model.to(device)
ds = OmegaNoiseDataset(
size=n_batches * batch_size,
img_size=cfg.img_size,
allowed_types=[0]) # gaussian
loader = torch.utils.data.DataLoader(
ds, batch_size=batch_size, shuffle=False)
all_M = [] # collect M from patch 0 of every sample
with torch.no_grad():
for imgs, _ in loader:
imgs = imgs.to(device)
out = model(imgs)
# M shape: [B, N_patches, V, D]
M_patch0 = out['svd']['M'][:, 0] # [B, V, D]
all_M.append(M_patch0.cpu())
return torch.cat(all_M, dim=0) # [n_samples, V, D]
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Hypothesis tests
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def test_rank_structure(M_avg):
"""Test 1: SVD on the canonical row matrix.
M_avg: averaged 32Γ3 row matrix. SVD gives 3 singular values.
Predictions:
Polynomial Vandermonde: top-1 SV dominates, rankβ1-2
Trig basis: balanced top-2 SVs, small 3rd
Sphere uniform (H2): ~equal SVs, full rank
Cluster: depends on cluster geometry
"""
U, S, Vt = np.linalg.svd(M_avg, full_matrices=False)
S_norm = S / S.sum()
erank = math.exp(-(S_norm * np.log(S_norm + 1e-12)).sum())
return {
'singular_values': S.tolist(),
'normalized_SV': S_norm.tolist(),
'effective_rank': erank,
'top1_share': S_norm[0],
'top2_share': S_norm[:2].sum(),
'verdict': (
'rank-1 (collapsed/aligned)' if S_norm[0] > 0.85 else
'rank-2 (planar β could be polynomial or trig)' if S_norm[:2].sum() > 0.92 else
'rank-3 (full, balanced)' if S_norm.std() < 0.05 else
'rank-3 (full, imbalanced)'
),
}
def test_parametric_ordering(M_avg):
"""Test 2: Project rows onto first principal axis, sort, check smoothness.
If rows lie on a smooth parametric curve (polynomial, trig), sorting
by PC1 projection should produce a smooth sequence. Smoothness =
1 / variance of consecutive Ξ in PC2/PC3 coords (after sort).
"""
U, S, Vt = np.linalg.svd(M_avg, full_matrices=False)
# Project rows onto principal axes
proj = M_avg @ Vt.T # [V, 3]
# Sort by PC1
sort_idx = np.argsort(proj[:, 0])
sorted_proj = proj[sort_idx]
# Ξ between consecutive sorted rows in PC2, PC3
deltas_pc2 = np.diff(sorted_proj[:, 1])
deltas_pc3 = np.diff(sorted_proj[:, 2])
# If smooth curve, Ξ should be small relative to overall PC2/PC3 spread
range_pc2 = sorted_proj[:, 1].max() - sorted_proj[:, 1].min()
range_pc3 = sorted_proj[:, 2].max() - sorted_proj[:, 2].min()
smoothness_pc2 = 1.0 - (np.abs(deltas_pc2).mean() / (range_pc2 + 1e-8))
smoothness_pc3 = 1.0 - (np.abs(deltas_pc3).mean() / (range_pc3 + 1e-8))
return {
'sort_order': sort_idx.tolist(),
'smoothness_pc2': float(smoothness_pc2),
'smoothness_pc3': float(smoothness_pc3),
'pc1_range': float(proj[:, 0].max() - proj[:, 0].min()),
'pc2_range': float(range_pc2),
'pc3_range': float(range_pc3),
'verdict': (
'smooth parametric curve' if min(smoothness_pc2, smoothness_pc3) > 0.85 else
'partial structure' if min(smoothness_pc2, smoothness_pc3) > 0.5 else
'scattered (cluster-like)'
),
}
def test_polynomial_fit(M_avg):
"""Test 3: Try polynomial bases of various orders.
Order rows by PC1 projection. Fit each PC2/PC3 coordinate as a function
of PC1. Polynomial degrees 1, 2, 3, 4. Best-fit RΒ² tells us the order.
Also tries [1, sin(x), cos(x)] for trigonometric basis.
"""
U, S, Vt = np.linalg.svd(M_avg, full_matrices=False)
proj = M_avg @ Vt.T
sort_idx = np.argsort(proj[:, 0])
x = proj[sort_idx, 0]
y2 = proj[sort_idx, 1]
y3 = proj[sort_idx, 2]
# Normalize x to [-1, 1] for stable polyfit
x_norm = 2 * (x - x.min()) / (x.max() - x.min() + 1e-8) - 1
def r2(y_true, y_pred):
ss_res = ((y_true - y_pred) ** 2).sum()
ss_tot = ((y_true - y_true.mean()) ** 2).sum()
return 1 - ss_res / (ss_tot + 1e-12)
poly_results = {}
for deg in [1, 2, 3, 4]:
coef2 = np.polyfit(x_norm, y2, deg)
coef3 = np.polyfit(x_norm, y3, deg)
pred2 = np.polyval(coef2, x_norm)
pred3 = np.polyval(coef3, x_norm)
poly_results[f'degree_{deg}'] = {
'r2_pc2': float(r2(y2, pred2)),
'r2_pc3': float(r2(y3, pred3)),
}
# Trigonometric fit: y = a + bΒ·sin(Οx) + cΒ·cos(Οx) + dΒ·sin(2Οx) + eΒ·cos(2Οx)
def trig_basis(x):
return np.column_stack([
np.ones_like(x),
np.sin(np.pi * x), np.cos(np.pi * x),
np.sin(2 * np.pi * x), np.cos(2 * np.pi * x),
])
B = trig_basis(x_norm)
coef2_t, _, _, _ = np.linalg.lstsq(B, y2, rcond=None)
coef3_t, _, _, _ = np.linalg.lstsq(B, y3, rcond=None)
trig_r2_pc2 = r2(y2, B @ coef2_t)
trig_r2_pc3 = r2(y3, B @ coef3_t)
# Pick the best fit
best_poly_deg = max([1, 2, 3, 4],
key=lambda d: poly_results[f'degree_{d}']['r2_pc2'])
best_poly_r2 = poly_results[f'degree_{best_poly_deg}']['r2_pc2']
return {
'polynomial': poly_results,
'trigonometric': {
'r2_pc2': float(trig_r2_pc2),
'r2_pc3': float(trig_r2_pc3),
'coefs_pc2': coef2_t.tolist(),
},
'best_poly_degree': best_poly_deg,
'best_poly_r2': float(best_poly_r2),
'verdict': (
f'polynomial degree {best_poly_deg} (RΒ²={best_poly_r2:.3f})'
if best_poly_r2 > 0.95 else
f'trigonometric (RΒ²={trig_r2_pc2:.3f})'
if trig_r2_pc2 > 0.95 else
f'no clean parametric fit (best poly RΒ²={best_poly_r2:.3f}, '
f'trig RΒ²={trig_r2_pc2:.3f})'
),
}
def test_cluster_structure(M_avg):
"""Test 4: k-means + silhouette across k = 2..8.
High silhouette at small k β genuine clusters. Low silhouette across
all k β continuous spread (consistent with parametric structure).
"""
results = {}
best_k = None
best_score = -1
for k in range(2, min(9, M_avg.shape[0])):
km = KMeans(n_clusters=k, n_init=10, random_state=42)
labels = km.fit_predict(M_avg)
if len(set(labels)) < 2:
continue
score = silhouette_score(M_avg, labels)
results[f'k={k}'] = {
'silhouette': float(score),
'inertia': float(km.inertia_),
}
if score > best_score:
best_score = score
best_k = k
return {
'per_k': results,
'best_k': best_k,
'best_silhouette': float(best_score),
'verdict': (
f'strong clusters (k={best_k}, silhouette={best_score:.3f})'
if best_score > 0.5 else
f'weak clusters (k={best_k}, silhouette={best_score:.3f})'
if best_score > 0.25 else
f'no clear clusters (best silhouette={best_score:.3f}) β '
f'consistent with continuous structure'
),
}
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Plotting
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def plot_diagnostic(M_avg, all_M, results, output_path):
"""4-panel diagnostic plot."""
fig = plt.figure(figsize=(16, 12))
# Panel 1: 3D scatter of the canonical 32 rows
ax1 = fig.add_subplot(2, 2, 1, projection='3d')
U, S, Vt = np.linalg.svd(M_avg, full_matrices=False)
proj = M_avg @ Vt.T
sort_idx = np.argsort(proj[:, 0])
colors = plt.cm.viridis(np.linspace(0, 1, len(sort_idx)))
for i, idx in enumerate(sort_idx):
ax1.scatter(M_avg[idx, 0], M_avg[idx, 1], M_avg[idx, 2],
c=[colors[i]], s=80, edgecolors='black', linewidths=0.5)
ax1.set_xlabel('D1')
ax1.set_ylabel('D2')
ax1.set_zlabel('D3')
ax1.set_title(f'P-rank09 row matrix M (V=32, D=3)\n'
f'colored by PC1 sort order\n'
f'effective rank: {results["rank"]["effective_rank"]:.2f}')
# Panel 2: Singular value spectrum
ax2 = fig.add_subplot(2, 2, 2)
SVs = np.array(results['rank']['singular_values'])
ax2.bar(['SV1', 'SV2', 'SV3'], SVs, color=['red', 'orange', 'yellow'])
ax2.set_ylabel('Singular value')
ax2.set_title(f'Singular values of M\n'
f'top1 share: {results["rank"]["top1_share"]:.2%}\n'
f'verdict: {results["rank"]["verdict"]}')
for i, sv in enumerate(SVs):
ax2.text(i, sv, f'{sv:.3f}', ha='center', va='bottom')
# Panel 3: PC2 and PC3 vs PC1 (parametric curve test)
ax3 = fig.add_subplot(2, 2, 3)
x = proj[sort_idx, 0]
y2 = proj[sort_idx, 1]
y3 = proj[sort_idx, 2]
ax3.plot(x, y2, 'o-', color='blue', label='PC2 vs PC1', markersize=6)
ax3.plot(x, y3, 's-', color='green', label='PC3 vs PC1', markersize=6)
ax3.set_xlabel('PC1 projection')
ax3.set_ylabel('PC2 / PC3 projection')
ax3.set_title(f'Parametric ordering test\n'
f'smoothness PC2: {results["parametric"]["smoothness_pc2"]:.3f}, '
f'PC3: {results["parametric"]["smoothness_pc3"]:.3f}\n'
f'verdict: {results["parametric"]["verdict"]}')
ax3.legend()
ax3.grid(alpha=0.3)
# Panel 4: Cluster silhouette across k
ax4 = fig.add_subplot(2, 2, 4)
ks = []
sils = []
for k_str, r in results['cluster']['per_k'].items():
ks.append(int(k_str.split('=')[1]))
sils.append(r['silhouette'])
ax4.plot(ks, sils, 'o-', color='purple', markersize=8)
ax4.axhline(0.5, color='red', linestyle='--', alpha=0.5,
label='strong cluster threshold')
ax4.axhline(0.25, color='orange', linestyle='--', alpha=0.5,
label='weak cluster threshold')
ax4.set_xlabel('k (number of clusters)')
ax4.set_ylabel('silhouette score')
ax4.set_title(f'Cluster structure test\n'
f'best k={results["cluster"]["best_k"]}, '
f'silhouette={results["cluster"]["best_silhouette"]:.3f}\n'
f'verdict: {results["cluster"]["verdict"]}')
ax4.legend(fontsize=8)
ax4.grid(alpha=0.3)
plt.tight_layout()
plt.savefig(output_path, dpi=120, bbox_inches='tight')
plt.show()
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Main
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def main():
print("Loading P-rank09 model...")
model, cfg = load_rank09()
print(f" Architecture: V={cfg.matrix_v}, D={cfg.D}, "
f"patch_size={cfg.patch_size}, hidden={cfg.hidden}")
n_params = sum(p.numel() for p in model.parameters())
print(f" Parameters: {n_params:,}")
print("\nCollecting M rows from gaussian inputs...")
all_M = collect_rows(model, cfg, n_batches=8, batch_size=64)
print(f" Collected {all_M.shape[0]} samples of M [V={all_M.shape[1]}, "
f"D={all_M.shape[2]}]")
# Average M over samples to get the canonical row matrix
M_avg = all_M.mean(dim=0).numpy()
M_std = all_M.std(dim=0).numpy()
print(f" M_avg shape: {M_avg.shape}")
print(f" Per-row variability (mean βΟββ across rows): "
f"{np.linalg.norm(M_std, axis=1).mean():.4f}")
print(f" Per-row mean magnitude (mean βΞΌββ): "
f"{np.linalg.norm(M_avg, axis=1).mean():.4f}")
# Sphere-norm verification
row_norms = np.linalg.norm(M_avg, axis=1)
print(f" Row norm range: [{row_norms.min():.4f}, {row_norms.max():.4f}]")
print(f" (sphere-normed rows should all have norm ~1.0)")
print("\n" + "β" * 70)
print("HYPOTHESIS TESTS")
print("β" * 70)
print("\n[1/4] Rank structure (SVD)...")
rank_results = test_rank_structure(M_avg)
print(f" Singular values: {[f'{s:.4f}' for s in rank_results['singular_values']]}")
print(f" Effective rank: {rank_results['effective_rank']:.2f}")
print(f" Top-1 share: {rank_results['top1_share']:.2%}")
print(f" VERDICT: {rank_results['verdict']}")
print("\n[2/4] Parametric ordering (PC1 sort + smoothness)...")
param_results = test_parametric_ordering(M_avg)
print(f" Smoothness PC2: {param_results['smoothness_pc2']:.3f}")
print(f" Smoothness PC3: {param_results['smoothness_pc3']:.3f}")
print(f" VERDICT: {param_results['verdict']}")
print("\n[3/4] Polynomial / trigonometric fit...")
fit_results = test_polynomial_fit(M_avg)
print(f" Polynomial fits (RΒ² for PC2):")
for deg in [1, 2, 3, 4]:
r2 = fit_results['polynomial'][f'degree_{deg}']['r2_pc2']
print(f" degree {deg}: RΒ² = {r2:.4f}")
print(f" Trigonometric fit (RΒ² for PC2): "
f"{fit_results['trigonometric']['r2_pc2']:.4f}")
print(f" VERDICT: {fit_results['verdict']}")
print("\n[4/4] Cluster structure (k-means silhouette)...")
cluster_results = test_cluster_structure(M_avg)
print(f" Per-k silhouette:")
for k_str, r in cluster_results['per_k'].items():
print(f" {k_str}: silhouette = {r['silhouette']:.3f}")
print(f" VERDICT: {cluster_results['verdict']}")
all_results = {
'config': {
'variant': 'P_rank09_h64_V32_D3_dp0_nx0_adam',
'V': cfg.matrix_v, 'D': cfg.D, 'params': n_params,
'gaussian_test_mse': 0.02782,
'observed_cv': 0.035,
},
'M_avg_shape': list(M_avg.shape),
'row_norms_mean': float(row_norms.mean()),
'row_norms_std': float(row_norms.std()),
'rank': rank_results,
'parametric': param_results,
'fit': fit_results,
'cluster': cluster_results,
}
print("\n" + "β" * 70)
print("OVERALL INTERPRETATION")
print("β" * 70)
print(f" Rank: {rank_results['verdict']}")
print(f" Parametric: {param_results['verdict']}")
print(f" Fit: {fit_results['verdict']}")
print(f" Clusters: {cluster_results['verdict']}")
# Composite verdict logic
is_polynomial = (
fit_results['best_poly_r2'] > 0.95 and
rank_results['effective_rank'] < 2.5
)
is_trig = (
fit_results['trigonometric']['r2_pc2'] > 0.95 and
not is_polynomial
)
is_clustered = cluster_results['best_silhouette'] > 0.5
is_collapsed = rank_results['top1_share'] > 0.85
print(f"\n Composite read:")
if is_polynomial:
deg = fit_results['best_poly_degree']
print(f" β POLYNOMIAL CONFIRMED (degree {deg}). "
f"P-Class naming validated.")
elif is_trig:
print(f" β TRIGONOMETRIC structure detected. "
f"P-Class might be better named F-Class (Fourier).")
elif is_collapsed:
print(f" β COLLAPSED β rows essentially 1-dimensional. "
f"Failed differentiation, not a useful battery.")
elif is_clustered:
k = cluster_results['best_k']
print(f" β CLUSTERED into {k} groups. "
f"P-Class might be better named K-Class "
f"(k-means / quantization).")
else:
print(f" β MIXED structure β not cleanly polynomial, trig, or "
f"clustered. Worth probing further with higher-order bases or "
f"deeper geometric analysis.")
with open(OUTPUT_JSON, 'w') as f:
json.dump(all_results, f, indent=2, default=str)
print(f"\n Results saved: {OUTPUT_JSON}")
plot_diagnostic(M_avg, all_M, all_results, OUTPUT_PLOT)
print(f" Plot saved: {OUTPUT_PLOT}")
return all_results
if __name__ == '__main__':
results = main() |