File size: 18,345 Bytes
d934656 | 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 | """
STER-GI Idea 3: Denoise-to-Sibling (去噪造兄弟)
================================================
- Train unconditional diffusion on ALL building property vectors (no labels)
- Generate "sibling" buildings via SDEdit (partial noise + denoise)
- Contrastive learning on (original, sibling) pairs
- Zero-shot evaluation on test pairs: encode → cosine sim → threshold → match
Baselines:
- real-only (raw property cosine): no training at all
- Idea 3 (denoise-to-sibling): diffusion + contrastive
"""
import os, sys, time, warnings, argparse
import numpy as np
import joblib, pickle as pkl
import torch, torch.nn as nn, torch.nn.functional as F
from torch.utils.data import DataLoader, TensorDataset
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import precision_score, recall_score, f1_score, average_precision_score
from collections import defaultdict
warnings.filterwarnings("ignore")
# ============================================================
# Config
# ============================================================
PROPERTY_NAMES = [
"bounding_box_width", "bounding_box_length", "area", "perimeter",
"perimeter_ind", "volume", "convex_hull_area", "convex_hull_volume",
"ave_centroid_distance", "height_diff", "num_floors", "axes_symmetry",
"compactness_2d", "compactness_3d", "density", "elongation",
"shape_ind", "hemisphericality", "fractality", "cubeness",
"circumference", "aligned_bounding_box_width", "aligned_bounding_box_length",
"aligned_bounding_box_height", "num_vertices"
]
CFG = {
'diffusion_steps': 1000,
'diffusion_hidden': 256,
'diffusion_layers': 4,
'diffusion_epochs': 500,
'diffusion_lr': 1e-3,
'diffusion_batch_size': 512,
# SDEdit: how much noise to add (0=none, 1000=full)
'sdedit_t0': 200,
'num_siblings': 3,
# Encoder
'encoder_hidden': 128,
'encoder_dim': 64,
'contrastive_epochs': 200,
'contrastive_lr': 1e-3,
'contrastive_temp': 0.07,
'contrastive_batch': 1024,
# Eval: threshold sweep
'eval_thresholds': [0.5, 0.6, 0.7, 0.75, 0.8, 0.85, 0.9, 0.92, 0.95, 0.97, 0.99],
'device': 'cuda' if torch.cuda.is_available() else 'cpu',
}
# ============================================================
# Data: Extract property vectors
# ============================================================
def extract_vectors(prop_dict, source, id_list=None):
"""Extract property matrix for given source ('cands'/'index') and optional id filter"""
first_prop = PROPERTY_NAMES[0]
all_ids = list(prop_dict[first_prop][source].keys()) if id_list is None else id_list
X = np.zeros((len(all_ids), len(PROPERTY_NAMES)), dtype=np.float32)
valid_ids = []
for i, bid in enumerate(all_ids):
vec = []
ok = True
for pname in PROPERTY_NAMES:
val = prop_dict[pname][source].get(bid, None)
if val is None or (isinstance(val, float) and np.isnan(val)):
val = 0.0
vec.append(float(val))
X[i] = vec
valid_ids.append(bid)
return X, valid_ids
def load_all_train_vectors(seed):
"""Load property vectors for ALL training buildings (both cands and index)"""
train_path = (f"data/property_dicts/Hague_allmodels_v1_train_matching_medium_"
f"neg_samples_num=2_vector_normalization=True_seed={seed}.joblib")
prop = joblib.load(train_path)
X_cands, ids_cands = extract_vectors(prop, 'cands')
X_index, ids_index = extract_vectors(prop, 'index')
X = np.concatenate([X_cands, X_index], axis=0)
all_ids = ids_cands + ids_index
print(f"Loaded {len(X)} train building vectors ({len(X_cands)} cands + {len(X_index)} index)")
return X, all_ids, prop
def load_test_pairs(seed):
"""Load test pairs with labels: list of (cand_id, index_id, label)"""
test_path = (f"data/property_dicts/Hague_allmodels_v1_test_matching_medium_"
f"neg_samples_num=2_vector_normalization=True_seed={seed}.joblib")
prop = joblib.load(test_path)
partition = pkl.load(open(f"data/dataset_partitions/Hague_seed{seed}.pkl", 'rb'))
test_pairs = partition['test']['matching']['negative_sampling']['medium'][2]
# Build cand/index ID -> vector mappings
cand_map = {}
for bid in prop[PROPERTY_NAMES[0]]['cands'].keys():
vec = []
for pname in PROPERTY_NAMES:
val = prop[pname]['cands'].get(bid, None)
if val is None or (isinstance(val, float) and np.isnan(val)):
val = 0.0
vec.append(float(val))
cand_map[bid] = np.array(vec, dtype=np.float32)
index_map = {}
for bid in prop[PROPERTY_NAMES[0]]['index'].keys():
vec = []
for pname in PROPERTY_NAMES:
val = prop[pname]['index'].get(bid, None)
if val is None or (isinstance(val, float) and np.isnan(val)):
val = 0.0
vec.append(float(val))
index_map[bid] = np.array(vec, dtype=np.float32)
cand_vecs, index_vecs, labels = [], [], []
for cid, iid in test_pairs:
if cid in cand_map and iid in index_map:
cand_vecs.append(cand_map[cid])
index_vecs.append(index_map[iid])
labels.append(1 if cid == iid else 0)
cand_vecs = np.array(cand_vecs, dtype=np.float32)
index_vecs = np.array(index_vecs, dtype=np.float32)
labels = np.array(labels, dtype=np.int32)
print(f"Test pairs: {len(labels)} ({labels.sum()} pos, {(1-labels).sum()} neg)")
return cand_vecs, index_vecs, labels
# ============================================================
# Diffusion Model (DDPM on property vectors)
# ============================================================
class MLPDiffusion(nn.Module):
def __init__(self, dim, hidden=256, layers=4):
super().__init__()
self.time_embed = nn.Sequential(
nn.Linear(1, hidden), nn.SiLU(), nn.Linear(hidden, hidden))
net = [nn.Linear(dim + hidden, hidden), nn.SiLU()]
for _ in range(layers - 1):
net += [nn.Linear(hidden, hidden), nn.SiLU()]
net.append(nn.Linear(hidden, dim))
self.net = nn.Sequential(*net)
def forward(self, x, t):
t_emb = self.time_embed(t.unsqueeze(-1).float())
return self.net(torch.cat([x, t_emb], dim=-1))
class DiffusionScheduler:
def __init__(self, steps=1000, beta_start=1e-4, beta_end=0.02):
self.steps = steps
self.betas = torch.linspace(beta_start, beta_end, steps)
self.alphas = 1 - self.betas
self.alpha_bars = torch.cumprod(self.alphas, dim=0)
def add_noise(self, x0, t):
ab = self.alpha_bars[t].view(-1, 1)
noise = torch.randn_like(x0)
return torch.sqrt(ab) * x0 + torch.sqrt(1 - ab) * noise, noise
@torch.no_grad()
def denoise_step(self, model, xt, t):
"""Single DDPM reverse step"""
a = self.alphas[t].view(-1, 1)
ab = self.alpha_bars[t].view(-1, 1)
b = self.betas[t].view(-1, 1)
eps = model(xt, t.float())
x0_hat = (xt - torch.sqrt(1 - ab) * eps) / torch.sqrt(a)
if t.min() == 0:
return x0_hat
ab_prev = self.alpha_bars[t - 1].view(-1, 1)
mean = (torch.sqrt(ab_prev) * b / (1 - ab) * x0_hat +
torch.sqrt(a) * (1 - ab_prev) / (1 - ab) * xt)
var = b * (1 - ab_prev) / (1 - ab)
return mean + torch.sqrt(var) * torch.randn_like(xt)
@torch.no_grad()
def sdedit(self, model, x0, t0, device):
"""Add noise to t0, then denoise back → sibling"""
n = x0.shape[0]
t0_t = torch.full((n,), t0, device=device, dtype=torch.long)
ab_t0 = self.alpha_bars[t0]
noise = torch.randn_like(x0)
xt = torch.sqrt(ab_t0) * x0 + torch.sqrt(1 - ab_t0) * noise
for t in range(t0, -1, -1):
tb = torch.full((n,), t, device=device, dtype=torch.long)
xt = self.denoise_step(model, xt, tb)
return xt
# ============================================================
# Contrastive Encoder
# ============================================================
class BuildingEncoder(nn.Module):
def __init__(self, dim, hidden=128, out_dim=64):
super().__init__()
self.net = nn.Sequential(
nn.Linear(dim, hidden), nn.BatchNorm1d(hidden), nn.ReLU(),
nn.Linear(hidden, hidden), nn.BatchNorm1d(hidden), nn.ReLU(),
nn.Linear(hidden, out_dim))
def forward(self, x):
return F.normalize(self.net(x), dim=-1)
def info_nce_loss(embs, temp=0.07):
"""embs: [2B, D] where (0,1), (2,3)... are positives"""
n = embs.shape[0] // 2
sim = embs @ embs.T / temp
# Mask self
sim = sim.masked_fill(torch.eye(2 * n, device=embs.device, dtype=torch.bool), -1e9)
# Labels: for row i, positive is i^1 (flip last bit)
labels = torch.arange(2 * n, device=embs.device)
labels = labels ^ 1 # (0->1, 1->0, 2->3, 3->2, ...)
return F.cross_entropy(sim, labels)
# ============================================================
# Training
# ============================================================
def train_diffusion(X, device):
print("\n" + "=" * 50)
print("Stage 1: Train Diffusion on Building Vectors")
print("=" * 50)
dim = X.shape[1]
model = MLPDiffusion(dim, CFG['diffusion_hidden'], CFG['diffusion_layers']).to(device)
sched = DiffusionScheduler(CFG['diffusion_steps'])
sched.betas = sched.betas.to(device)
sched.alphas = sched.alphas.to(device)
sched.alpha_bars = sched.alpha_bars.to(device)
opt = torch.optim.Adam(model.parameters(), lr=CFG['diffusion_lr'])
ds = TensorDataset(torch.FloatTensor(X))
dl = DataLoader(ds, batch_size=CFG['diffusion_batch_size'], shuffle=True)
model.train()
for ep in range(CFG['diffusion_epochs']):
total = 0
for (xb,) in dl:
xb = xb.to(device)
bs = xb.shape[0]
t = torch.randint(0, CFG['diffusion_steps'], (bs,), device=device)
xt, noise = sched.add_noise(xb, t)
loss = F.mse_loss(model(xt, t.float()), noise)
opt.zero_grad()
loss.backward()
opt.step()
total += loss.item() * bs
if (ep + 1) % 100 == 0:
print(f" Epoch {ep+1}/{CFG['diffusion_epochs']} | Loss: {total/len(ds):.6f}")
print(f" Done. Final loss: {total/len(ds):.6f}")
return model, sched
def generate_siblings(model, sched, X, device):
print("\n" + "=" * 50)
print(f"Stage 2: Generate Siblings (t0={CFG['sdedit_t0']})")
print("=" * 50)
model.eval()
origs, sibs = [], []
bs = CFG['diffusion_batch_size']
for i in range(0, len(X), bs):
xb = torch.FloatTensor(X[i:i+bs]).to(device)
for _ in range(CFG['num_siblings']):
sib = sched.sdedit(model, xb, CFG['sdedit_t0'], device)
origs.append(xb.cpu().numpy())
sibs.append(sib.cpu().numpy())
origs = np.concatenate(origs, axis=0)
sibs = np.concatenate(sibs, axis=0)
l2 = np.mean(np.linalg.norm(origs - sibs, axis=1))
print(f" Generated {len(origs)} pairs | Mean L2 diff: {l2:.4f}")
return origs, sibs
def train_encoder(origs, sibs, device):
print("\n" + "=" * 50)
print("Stage 3: Contrastive Encoder Training")
print("=" * 50)
dim = origs.shape[1]
encoder = BuildingEncoder(dim, CFG['encoder_hidden'], CFG['encoder_dim']).to(device)
opt = torch.optim.Adam(encoder.parameters(), lr=CFG['contrastive_lr'])
# Interleave: [orig1, sib1, orig2, sib2, ...]
n = len(origs)
data = np.zeros((n * 2, dim), dtype=np.float32)
data[0::2] = origs
data[1::2] = sibs
ds = TensorDataset(torch.FloatTensor(data))
dl = DataLoader(ds, batch_size=CFG['contrastive_batch'], shuffle=True)
encoder.train()
for ep in range(CFG['contrastive_epochs']):
total = 0
for (xb,) in dl:
xb = xb.to(device)
emb = encoder(xb)
loss = info_nce_loss(emb, CFG['contrastive_temp'])
opt.zero_grad()
loss.backward()
opt.step()
total += loss.item() * xb.shape[0]
if (ep + 1) % 50 == 0:
print(f" Epoch {ep+1}/{CFG['contrastive_epochs']} | Loss: {total/len(ds):.4f}")
print(f" Done. Final loss: {total/len(ds):.4f}")
return encoder
# ============================================================
# Evaluation
# ============================================================
@torch.no_grad()
def eval_zero_shot(encoder, cand_vecs, index_vecs, labels, device, tag="Model"):
"""Zero-shot pair classification via cosine similarity"""
encoder.eval()
ct = torch.FloatTensor(cand_vecs).to(device)
it = torch.FloatTensor(index_vecs).to(device)
ce = encoder(ct).cpu().numpy()
ie = encoder(it).cpu().numpy()
# Cosine sim for each pair
sims = np.sum(ce * ie, axis=1) # [N]
# Sweep thresholds
best_f1, best_thresh, best_res = 0, 0.5, None
for th in CFG['eval_thresholds']:
pred = (sims >= th).astype(np.int32)
p = precision_score(labels, pred, zero_division=0)
r = recall_score(labels, pred, zero_division=0)
f = f1_score(labels, pred, zero_division=0)
if f > best_f1:
best_f1, best_thresh, best_res = f, th, (p, r, f)
print(f"\n {tag} (best threshold={best_thresh:.2f}):")
print(f" Precision: {best_res[0]:.4f}")
print(f" Recall: {best_res[1]:.4f}")
print(f" F1: {best_res[2]:.4f}")
# Also AP score
ap = average_precision_score(labels, sims)
print(f" AvgPrecision: {ap:.4f}")
return {'precision': best_res[0], 'recall': best_res[1], 'f1': best_res[2],
'ap': ap, 'threshold': best_thresh}
def eval_real_only(cand_vecs, index_vecs, labels):
"""Baseline: raw property cosine similarity, no training"""
# Normalize
cn = cand_vecs / (np.linalg.norm(cand_vecs, axis=1, keepdims=True) + 1e-8)
i_n = index_vecs / (np.linalg.norm(index_vecs, axis=1, keepdims=True) + 1e-8)
sims = np.sum(cn * i_n, axis=1)
best_f1, best_thresh, best_res = 0, 0.5, None
for th in CFG['eval_thresholds']:
pred = (sims >= th).astype(np.int32)
p = precision_score(labels, pred, zero_division=0)
r = recall_score(labels, pred, zero_division=0)
f = f1_score(labels, pred, zero_division=0)
if f > best_f1:
best_f1, best_thresh, best_res = f, th, (p, r, f)
ap = average_precision_score(labels, sims)
print(f"\n Baseline Real-Only (best th={best_thresh:.2f}):")
print(f" P={best_res[0]:.4f} R={best_res[1]:.4f} F1={best_res[2]:.4f} AP={ap:.4f}")
return {'precision': best_res[0], 'recall': best_res[1], 'f1': best_res[2],
'ap': ap, 'threshold': best_thresh}
# ============================================================
# Main
# ============================================================
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--seed', type=int, default=1)
parser.add_argument('--t0', type=int, default=200, help='SDEdit noise level')
parser.add_argument('--skip_diff', action='store_true')
parser.add_argument('--skip_enc', action='store_true')
args = parser.parse_args()
CFG['sdedit_t0'] = args.t0
device = CFG['device']
seed = args.seed
print(f"Device: {device} | t0: {args.t0} | Seed: {seed}")
# ---- Load data ----
X_train, train_ids, train_prop = load_all_train_vectors(seed)
cand_vecs, idx_vecs, labels = load_test_pairs(seed)
# Normalize (fit on train, apply to test)
scaler = StandardScaler()
X_train_s = scaler.fit_transform(X_train)
cand_s = scaler.transform(cand_vecs)
idx_s = scaler.transform(idx_vecs)
# ---- Stage 1+2: Diffusion → Siblings ----
diff_path = f"saved_model_files/diff_idea3_s{seed}_t{args.t0}.pt"
sib_path = f"saved_model_files/siblings_idea3_s{seed}_t{args.t0}.npz"
if args.skip_diff and os.path.exists(diff_path):
print(f"Loading cached diffusion model: {diff_path}")
ck = torch.load(diff_path, map_location=device)
model = MLPDiffusion(X_train_s.shape[1], CFG['diffusion_hidden'],
CFG['diffusion_layers']).to(device)
model.load_state_dict(ck['model'])
sched = DiffusionScheduler(CFG['diffusion_steps'])
sched.betas = sched.betas.to(device)
sched.alphas = sched.alphas.to(device)
sched.alpha_bars = sched.alpha_bars.to(device)
else:
model, sched = train_diffusion(X_train_s, device)
torch.save({'model': model.state_dict()}, diff_path)
if os.path.exists(sib_path):
print(f"Loading cached siblings: {sib_path}")
data = np.load(sib_path)
origs, sibs = data['origs'], data['sibs']
else:
origs, sibs = generate_siblings(model, sched, X_train_s, device)
np.savez_compressed(sib_path, origs=origs, sibs=sibs)
# ---- Stage 3: Contrastive Encoder ----
enc_path = f"saved_model_files/enc_idea3_s{seed}_t{args.t0}.pt"
if args.skip_enc and os.path.exists(enc_path):
print(f"Loading cached encoder: {enc_path}")
ck = torch.load(enc_path, map_location=device)
encoder = BuildingEncoder(X_train_s.shape[1], CFG['encoder_hidden'],
CFG['encoder_dim']).to(device)
encoder.load_state_dict(ck['encoder'])
else:
encoder = train_encoder(origs, sibs, device)
torch.save({'encoder': encoder.state_dict()}, enc_path)
# ---- Evaluation ----
print("\n" + "=" * 50)
print("RESULTS")
print("=" * 50)
r_base = eval_real_only(cand_s, idx_s, labels)
r_idea3 = eval_zero_shot(encoder, cand_s, idx_s, labels, device,
f"Idea 3 (t0={args.t0})")
print("\n" + "=" * 50)
print(f"SUMMARY (seed={seed}, t0={args.t0})")
print("=" * 50)
print(f" Baseline (real-only): F1={r_base['f1']:.4f} AP={r_base['ap']:.4f}")
print(f" Idea 3 (denoise-sib): F1={r_idea3['f1']:.4f} AP={r_idea3['ap']:.4f}")
print(f" ΔF1: {r_idea3['f1'] - r_base['f1']:.4f} ΔAP: {r_idea3['ap'] - r_base['ap']:.4f}")
return r_base, r_idea3
if __name__ == '__main__':
main()
|