File size: 15,746 Bytes
9c44d9a | 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 | """
STER-GI Idea 2: Identity-Realization Disentanglement (身份-实现解耦)
====================================================================
World Model approach:
- VAE disentangles z_id (what building) from z_style (how it's captured)
- Multi-view data: same building appears in both cand and index sets
- Generate new views: freeze z_id, sample z_style → decode
- Contrastive learning on (original, generated) pairs
- Zero-shot: encode test buildings, cosine similarity → match
"""
import numpy as np, joblib, pickle as pkl, torch, torch.nn as nn, torch.nn.functional as F, time, os, argparse, sys
from torch.utils.data import DataLoader, TensorDataset
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import precision_score, recall_score, f1_score
PROPS = ["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"]
# ===== Data =====
def get_vecs(prop_dict, source):
ids = list(prop_dict[PROPS[0]][source].keys())
X = np.zeros((len(ids), len(PROPS)), dtype=np.float32)
for i, bid in enumerate(ids):
for j, pn in enumerate(PROPS):
v = prop_dict[pn][source].get(bid, None)
X[i,j] = float(v) if v is not None and not (isinstance(v,float) and np.isnan(v)) else 0.0
return X, ids
def load_all(seed):
"""Load all buildings + multi-view pairs for disentanglement"""
tp = f"data/property_dicts/Hague_allmodels_v1_train_matching_medium_neg_samples_num=2_vector_normalization=True_seed={seed}.joblib"
ep = f"data/property_dicts/Hague_allmodels_v1_test_matching_medium_neg_samples_num=2_vector_normalization=True_seed={seed}.joblib"
trp = joblib.load(tp); epd = joblib.load(ep)
Xtc, id_tc = get_vecs(trp, 'cands'); Xti, id_ti = get_vecs(trp, 'index')
Xec, id_ec = get_vecs(epd, 'cands'); Xei, id_ei = get_vecs(epd, 'index')
# Multi-view pairs: same building ID → two views (cand, index)
# Build ID -> vector mapping for both sources
cand_all, idx_all = {}, {}
for sp, ids in [(trp,id_tc),(epd,id_ec)]:
for bid in ids:
if bid not in cand_all:
cand_all[bid] = np.array([float(sp[pn]['cands'].get(bid,0) or 0) for pn in PROPS], dtype=np.float32)
for sp, ids in [(trp,id_ti),(epd,id_ei)]:
for bid in ids:
if bid not in idx_all:
idx_all[bid] = np.array([float(sp[pn]['index'].get(bid,0) or 0) for pn in PROPS], dtype=np.float32)
# Multi-view pairs: buildings present in BOTH cand and index
common = sorted(set(cand_all.keys()) & set(idx_all.keys()))
views1 = np.array([cand_all[bid] for bid in common], dtype=np.float32)
views2 = np.array([idx_all[bid] for bid in common], dtype=np.float32)
print(f"Multi-view pairs: {len(common)} buildings with cand+index views", flush=True)
# All buildings for reconstruction pre-training
X_all = np.concatenate([Xtc, Xti, Xec, Xei], axis=0)
# Eval pairs (same as Idea 3)
part = pkl.load(open(f"data/dataset_partitions/Hague_seed{seed}.pkl", 'rb'))
all_pairs = list(part['train']['negative_sampling']['medium'][2]) + list(part['test']['matching']['negative_sampling']['medium'][2])
cv, iv, lbs = [], [], []
for cid, iid in all_pairs:
if cid in cand_all and iid in idx_all:
cv.append(cand_all[cid]); iv.append(idx_all[iid]); lbs.append(1 if cid == iid else 0)
cv, iv, lbs = np.array(cv,dtype=np.float32), np.array(iv,dtype=np.float32), np.array(lbs,dtype=np.int32)
print(f"Eval pairs: {len(lbs)} ({lbs.sum()} pos, {(1-lbs).sum()} neg)", flush=True)
return X_all, views1, views2, cv, iv, lbs
# ===== World Model: Disentangled VAE =====
class DisentangledVAE(nn.Module):
"""Encoder → (z_id, z_style); Decoder → reconstruction"""
def __init__(self, d, z_id_dim=32, z_style_dim=16, h=256):
super().__init__()
# Encoder
self.enc = nn.Sequential(
nn.Linear(d, h), nn.BatchNorm1d(h), nn.ReLU(),
nn.Linear(h, h), nn.BatchNorm1d(h), nn.ReLU())
self.mu_id = nn.Linear(h, z_id_dim)
self.logvar_id = nn.Linear(h, z_id_dim)
self.mu_style = nn.Linear(h, z_style_dim)
self.logvar_style = nn.Linear(h, z_style_dim)
# Decoder
self.dec = nn.Sequential(
nn.Linear(z_id_dim + z_style_dim, h), nn.BatchNorm1d(h), nn.ReLU(),
nn.Linear(h, h), nn.BatchNorm1d(h), nn.ReLU(),
nn.Linear(h, d))
# Discriminator: predict building ID from z_style (for adversarial disentanglement)
self.disc = nn.Sequential(
nn.Linear(z_style_dim, 64), nn.ReLU(),
nn.Linear(64, 32), nn.ReLU(),
nn.Linear(32, 1)) # Binary: same or different building?
def encode(self, x):
h = self.enc(x)
return self.mu_id(h), self.logvar_id(h), self.mu_style(h), self.logvar_style(h)
def reparameterize(self, mu, logvar):
std = torch.exp(0.5 * logvar)
eps = torch.randn_like(std)
return mu + eps * std
def forward(self, x):
mu_id, lv_id, mu_style, lv_style = self.encode(x)
z_id = self.reparameterize(mu_id, lv_id)
z_style = self.reparameterize(mu_style, lv_style)
recon = self.dec(torch.cat([z_id, z_style], dim=-1))
return recon, mu_id, lv_id, mu_style, lv_style, z_id, z_style
@torch.no_grad()
def generate(self, x, device):
"""Generate a new view: extract z_id, sample new z_style, decode"""
mu_id, _, _, _ = self.encode(x.to(device))
z_style = torch.randn(x.shape[0], self.mu_style.out_features).to(device) * 0.5
return self.dec(torch.cat([mu_id, z_style], dim=-1))
def vae_loss(recon, x, mu_id, lv_id, mu_style, lv_style, z_id, z_style, disc,
v1, v2, mu_id1, mu_id2, z_style1, z_style2, beta=0.1):
"""VAE loss: reconstruction + KL + identity consistency + adversarial"""
# Reconstruction
recon_loss = F.mse_loss(recon, x, reduction='sum') / x.shape[0]
# KL divergence
kl_id = -0.5 * torch.sum(1 + lv_id - mu_id.pow(2) - lv_id.exp(), dim=-1).mean()
kl_style = -0.5 * torch.sum(1 + lv_style - mu_style.pow(2) - lv_style.exp(), dim=-1).mean()
# Identity consistency: same building (v1, v2) → same z_id
id_consistency = F.mse_loss(mu_id1, mu_id2)
# Adversarial: discriminator should NOT be able to tell if two z_styles are from same building
# Positive: z_style1 and z_style2 from same building → disc should NOT distinguish
# Negative: z_style from different buildings → disc should NOT distinguish either (goal: style is ID-independent)
# We train disc to distinguish, and encoder to fool disc
# Simplified: minimize cosine similarity between z_styles of same building
style_disentangle = F.cosine_similarity(z_style1, z_style2, dim=-1).abs().mean()
total = recon_loss + beta * (kl_id + kl_style) + 0.5 * id_consistency + 0.1 * style_disentangle
return total, {'recon': recon_loss.item(), 'kl_id': kl_id.item(), 'kl_style': kl_style.item(),
'id_cons': id_consistency.item(), 'style_dis': style_disentangle.item()}
# ===== Contrastive Encoder =====
class Encoder(nn.Module):
def __init__(self,d,h=128,o=64):
super().__init__()
self.net=nn.Sequential(nn.Linear(d,h),nn.BatchNorm1d(h),nn.ReLU(),
nn.Linear(h,h),nn.BatchNorm1d(h),nn.ReLU(),nn.Linear(h,o))
def forward(self,x):
z = self.net(x)
return z / (torch.norm(z, dim=-1, keepdim=True).clamp(min=1e-8))
def infonce(emb,temp=0.07):
n=emb.shape[0]//2; sim=emb@emb.T/temp
sim=sim.masked_fill(torch.eye(2*n,device=emb.device,dtype=torch.bool),-1e9)
return F.cross_entropy(sim,torch.arange(2*n,device=emb.device)^1)
# ===== Eval =====
def eval_pairs(encoder, cv, iv, lbs, dev, tag=""):
encoder.eval()
with torch.no_grad():
ce = encoder(torch.FloatTensor(cv).to(dev)).cpu().numpy()
ie = encoder(torch.FloatTensor(iv).to(dev)).cpu().numpy()
sims = np.sum(ce * ie, axis=1)
best_f1, best_th = 0, 0
for th in np.arange(0.3, 1.0, 0.02):
pred = (sims >= th).astype(np.int32)
f = f1_score(lbs, pred, zero_division=0)
if f > best_f1: best_f1, best_th = f, th
p = precision_score(lbs, (sims>=best_th).astype(np.int32), zero_division=0)
r = recall_score(lbs, (sims>=best_th).astype(np.int32), zero_division=0)
print(f" {tag}: P={p:.4f} R={r:.4f} F1={best_f1:.4f} (th={best_th:.2f})", flush=True)
return best_f1
def baseline_raw(cv, iv, lbs):
cn = cv/(np.linalg.norm(cv,axis=1,keepdims=True)+1e-8)
i_n = iv/(np.linalg.norm(iv,axis=1,keepdims=True)+1e-8)
sims = np.sum(cn * i_n, axis=1)
best_f1, best_th = 0, 0
for th in np.arange(0.3, 1.0, 0.02):
pred = (sims >= th).astype(np.int32)
f = f1_score(lbs, pred, zero_division=0)
if f > best_f1: best_f1, best_th = f, th
p = precision_score(lbs, (sims>=best_th).astype(np.int32), zero_division=0)
r = recall_score(lbs, (sims>=best_th).astype(np.int32), zero_division=0)
print(f" Baseline Raw: P={p:.4f} R={r:.4f} F1={best_f1:.4f} (th={best_th:.2f})", flush=True)
return best_f1
# ===== Main =====
def main():
a = argparse.ArgumentParser()
a.add_argument('--seed',type=int,default=1); a.add_argument('--z_id',type=int,default=32)
a.add_argument('--z_style',type=int,default=16); a.add_argument('--vae_ep',type=int,default=200)
a.add_argument('--enc_ep',type=int,default=100)
a.add_argument('--skip_vae',action='store_true'); a.add_argument('--skip_enc',action='store_true')
args = a.parse_args()
dev = 'cpu'
print(f"Device: {dev} | z_id={args.z_id} z_style={args.z_style} | Seed: {args.seed}", flush=True)
X_all, v1, v2, cv, iv, lbs = load_all(args.seed)
sc = StandardScaler(); Xs = sc.fit_transform(X_all)
v1s = sc.transform(v1); v2s = sc.transform(v2)
cv_s, iv_s = sc.transform(cv), sc.transform(iv)
print("\n=== Baseline ===", flush=True)
b_raw = baseline_raw(cv_s, iv_s, lbs)
# Stage 1: Train Disentangled VAE
vae_path = f"saved_model_files/vae_i2_s{args.seed}.pt"
gen_path = f"saved_model_files/gen_i2_s{args.seed}.npz"
if args.skip_vae and os.path.exists(vae_path):
print(f"Loading cached VAE: {vae_path}", flush=True)
ck = torch.load(vae_path, map_location=dev)
vae = DisentangledVAE(Xs.shape[1], args.z_id, args.z_style).to(dev)
vae.load_state_dict(ck['vae'])
else:
print(f"\n=== Training Disentangled VAE ({args.vae_ep} epochs) ===", flush=True)
vae = DisentangledVAE(Xs.shape[1], args.z_id, args.z_style).to(dev)
opt = torch.optim.Adam(vae.parameters(), lr=1e-3)
ds1 = TensorDataset(torch.FloatTensor(v1s), torch.FloatTensor(v2s))
ds2 = TensorDataset(torch.FloatTensor(Xs))
dl1 = DataLoader(ds1, batch_size=256, shuffle=True)
dl2 = DataLoader(ds2, batch_size=256, shuffle=True)
vae.train(); t0_t = time.time()
for ep in range(args.vae_ep):
tot_loss, tot_comp = 0, {}
for (x1, x2) in dl1:
x1, x2 = x1.to(dev), x2.to(dev)
# Encode both views
mu_id1, lv_id1, mu_s1, lv_s1 = vae.encode(x1)
mu_id2, lv_id2, mu_s2, lv_s2 = vae.encode(x2)
z_id1 = vae.reparameterize(mu_id1, lv_id1)
z_s1 = vae.reparameterize(mu_s1, lv_s1)
z_id2 = vae.reparameterize(mu_id2, lv_id2)
z_s2 = vae.reparameterize(mu_s2, lv_s2)
recon1 = vae.dec(torch.cat([z_id1, z_s1], -1))
recon2 = vae.dec(torch.cat([z_id2, z_s2], -1))
loss1, comp1 = vae_loss(recon1, x1, mu_id1, lv_id1, mu_s1, lv_s1, z_id1, z_s1, vae.disc,
x1, x2, mu_id1, mu_id2, z_s1, z_s2)
loss2, comp2 = vae_loss(recon2, x2, mu_id2, lv_id2, mu_s2, lv_s2, z_id2, z_s2, vae.disc,
x2, x1, mu_id2, mu_id1, z_s2, z_s1)
loss = loss1 + loss2
opt.zero_grad(); loss.backward(); opt.step()
tot_loss += loss.item()
for k in comp1: tot_comp[k] = tot_comp.get(k,0) + comp1[k] + comp2[k]
if (ep+1)%20==0:
n_b = max(1, len(dl1))
print(f" VAE ep {ep+1}/{args.vae_ep}: loss={tot_loss/n_b:.4f} "
f"recon={tot_comp.get('recon',0)/n_b:.4f} "
f"kl_id={tot_comp.get('kl_id',0)/n_b:.4f} "
f"id_cons={tot_comp.get('id_cons',0)/n_b:.4f} "
f"t={time.time()-t0_t:.0f}s", flush=True)
print(f" Done: loss={tot_loss/max(1,len(dl1)):.4f}", flush=True)
torch.save({'vae':vae.state_dict()}, vae_path)
# Stage 2: Generate new views
if os.path.exists(gen_path):
d = np.load(gen_path); gens = d['gens']
print(f"Loaded cached generated views: {len(gens)}", flush=True)
else:
print("\n=== Generating New Views ===", flush=True)
vae.eval(); gens = []
with torch.no_grad():
for i in range(0, len(Xs), 256):
xb = torch.FloatTensor(Xs[i:i+256]).to(dev)
for _ in range(2):
gens.append(vae.generate(xb, dev).cpu().numpy())
gens = np.concatenate(gens)
print(f" Generated {len(gens)} views", flush=True)
np.savez_compressed(gen_path, gens=gens)
# Stage 3: Contrastive Encoder
enc_path = f"saved_model_files/enc_i2_s{args.seed}.pt"
if args.skip_enc and os.path.exists(enc_path):
print(f"Loading cached encoder: {enc_path}", flush=True)
ck = torch.load(enc_path, map_location=dev)
encoder = Encoder(Xs.shape[1]).to(dev); encoder.load_state_dict(ck['e'])
else:
print(f"\n=== Contrastive Encoder ({args.enc_ep} epochs) ===", flush=True)
encoder = Encoder(Xs.shape[1]).to(dev)
opt = torch.optim.Adam(encoder.parameters(), lr=1e-4)
n_g = len(gens); idx = np.random.permutation(min(n_g, len(Xs)))
data = np.zeros((len(idx)*2, Xs.shape[1]), dtype=np.float32)
data[0::2] = Xs[idx]; data[1::2] = gens[idx]
ds = TensorDataset(torch.FloatTensor(data)); dl = DataLoader(ds, batch_size=1024, shuffle=False)
encoder.train(); t0_t = time.time()
for ep in range(args.enc_ep):
tot = 0
for (xb,) in dl:
xb = xb.to(dev); emb = encoder(xb); loss = infonce(emb, 0.07)
opt.zero_grad(); loss.backward()
torch.nn.utils.clip_grad_norm_(encoder.parameters(), 1.0)
opt.step(); tot += loss.item()*xb.shape[0]
if (ep+1)%10==0: print(f" Enc ep {ep+1}/{args.enc_ep}: loss={tot/len(ds):.4f} t={time.time()-t0_t:.0f}s", flush=True)
print(f" Done: loss={tot/len(ds):.4f}", flush=True)
torch.save({'e':encoder.state_dict()}, enc_path)
# Results
print("\n=== RESULTS ===", flush=True)
f1_i2 = eval_pairs(encoder, cv_s, iv_s, lbs, dev, f"Idea2 (WM disentangle)")
print(f"\n Baseline (raw): F1={b_raw:.4f}")
print(f" Idea2 (WM): F1={f1_i2:.4f}")
print(f" Supervised XGBoost: F1=0.982")
print(f" Δ over baseline: {f1_i2-b_raw:+.4f}", flush=True)
if __name__=='__main__': main()
|