File size: 12,404 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 | """
STER-GI Idea 3: Denoise-to-Sibling
- Train: diffusion + contrastive on ALL building vectors (NO labels)
- Eval: ALL train+test pairs, zero-shot cosine similarity → F1
"""
import numpy as np, joblib, pickle as pkl, torch, torch.nn as nn, torch.nn.functional as F, time, os, argparse
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"]
CFG = {'diff_steps':1000,'diff_hidden':256,'diff_layers':4,'diff_epochs':100,'diff_lr':1e-3,'diff_bs':512,
't0':200,'n_sib':2,'enc_hidden':128,'enc_dim':64,'ctr_epochs':100,'ctr_lr':1e-3,'ctr_temp':0.07,'ctr_bs':1024}
# ---------- 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_data(seed):
"""Load ALL building vectors (train+test) and ALL pairs (train+test)"""
# Train buildings
tp = f"data/property_dicts/Hague_allmodels_v1_train_matching_medium_neg_samples_num=2_vector_normalization=True_seed={seed}.joblib"
trp = joblib.load(tp)
Xtc, id_tc = get_vecs(trp, 'cands')
Xti, id_ti = get_vecs(trp, 'index')
X_all = np.concatenate([Xtc, Xti], axis=0)
print(f"Train buildings: {len(Xtc)} cands + {len(Xti)} index = {len(X_all)}", flush=True)
# Test buildings
ep = f"data/property_dicts/Hague_allmodels_v1_test_matching_medium_neg_samples_num=2_vector_normalization=True_seed={seed}.joblib"
epd = joblib.load(ep)
Xec, id_ec = get_vecs(epd, 'cands')
Xei, id_ei = get_vecs(epd, 'index')
# Combine ALL building vectors (for diffusion training)
X_all = np.concatenate([X_all, Xec, Xei], axis=0)
print(f"All buildings for diffusion: {len(X_all)}", flush=True)
# ALL pairs (train+test) for evaluation
part = pkl.load(open(f"data/dataset_partitions/Hague_seed{seed}.pkl", 'rb'))
train_pairs = part['train']['negative_sampling']['medium'][2]
test_pairs = part['test']['matching']['negative_sampling']['medium'][2]
all_pairs = list(train_pairs) + list(test_pairs)
# Build lookup for all buildings
cand_map = {}
for src_prop, src_ids in [(trp,id_tc), (epd,id_ec)]:
for bid in src_ids:
if bid not in cand_map:
v = [float(src_prop[pn]['cands'].get(bid,0) or 0) for pn in PROPS]
cand_map[bid] = np.array(v, dtype=np.float32)
index_map = {}
for src_prop, src_ids in [(trp,id_ti), (epd,id_ei)]:
for bid in src_ids:
if bid not in index_map:
v = [float(src_prop[pn]['index'].get(bid,0) or 0) for pn in PROPS]
index_map[bid] = np.array(v, dtype=np.float32)
# Build pair vectors + labels
cv, iv, lbs = [], [], []
for cid, iid in all_pairs:
if cid in cand_map and iid in index_map:
cv.append(cand_map[cid])
iv.append(index_map[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"All eval pairs: {len(lbs)} ({lbs.sum()} pos, {(1-lbs).sum()} neg)", flush=True)
return X_all, cv, iv, lbs
# ---------- Diffusion ----------
class DiffMLP(nn.Module):
def __init__(self,d,h=256,L=4):
super().__init__()
self.te = nn.Sequential(nn.Linear(1,h),nn.SiLU(),nn.Linear(h,h))
net = [nn.Linear(d+h,h),nn.SiLU()]
for _ in range(L-1): net += [nn.Linear(h,h),nn.SiLU()]
net.append(nn.Linear(h,d))
self.net = nn.Sequential(*net)
def forward(self,x,t):
return self.net(torch.cat([x,self.te(t.unsqueeze(-1).float())],-1))
class DiffSched:
def __init__(self,S=1000):
self.S=S; self.b=torch.linspace(1e-4,0.02,S); self.a=1-self.b; self.ab=torch.cumprod(self.a,0)
def noise(self,x0,t):
ab=self.ab[t].view(-1,1); eps=torch.randn_like(x0)
return torch.sqrt(ab)*x0+torch.sqrt(1-ab)*eps, eps
@torch.no_grad()
def step(self,m,xt,t):
a=self.a[t].view(-1,1); ab=self.ab[t].view(-1,1); b_=self.b[t].view(-1,1)
e=m(xt,t.float()); x0h=(xt-torch.sqrt(1-ab)*e)/torch.sqrt(a)
if t.min()==0: return x0h
abp=self.ab[t-1].view(-1,1)
mu=torch.sqrt(abp)*b_/(1-ab)*x0h+torch.sqrt(a)*(1-abp)/(1-ab)*xt
return mu+torch.sqrt(b_*(1-abp)/(1-ab))*torch.randn_like(xt)
@torch.no_grad()
def sdedit(self,m,x0,t0,dev):
n=x0.shape[0]; abt=self.ab[t0]
xt=torch.sqrt(abt)*x0+torch.sqrt(1-abt)*torch.randn_like(x0)
for t in range(t0,-1,-1):
xt=self.step(m,xt,torch.full((n,),t,device=dev,dtype=torch.long))
return xt
# ---------- 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): return F.normalize(self.net(x),-1)
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():
p = argparse.ArgumentParser()
p.add_argument('--seed',type=int,default=1)
p.add_argument('--t0',type=int,default=200)
p.add_argument('--skip_diff',action='store_true')
p.add_argument('--skip_enc',action='store_true')
a = p.parse_args()
CFG['t0']=a.t0
dev = 'cuda' if torch.cuda.is_available() else 'cpu'
print(f"Device: {dev} | t0: {a.t0} | Seed: {a.seed}", flush=True)
# Load
X_all, cv, iv, lbs = load_all_data(a.seed)
sc = StandardScaler(); Xs = sc.fit_transform(X_all)
cv_s, iv_s = sc.transform(cv), sc.transform(iv)
# Baseline
print("\n=== Zero-Shot Baseline ===")
b_raw = baseline_raw(cv_s, iv_s, lbs)
# Diffusion
diff_path = f"saved_model_files/diff_i3_s{a.seed}_t{a.t0}.pt"
sib_path = f"saved_model_files/sib_i3_s{a.seed}_t{a.t0}.npz"
if a.skip_diff and os.path.exists(diff_path):
ck = torch.load(diff_path, map_location=dev)
model = DiffMLP(Xs.shape[1], CFG['diff_hidden'], CFG['diff_layers']).to(dev)
model.load_state_dict(ck['m'])
sched = DiffSched(CFG['diff_steps'])
for attr in ['b','a','ab']: setattr(sched, attr, getattr(sched, attr).to(dev))
else:
print("\n=== Training Diffusion ===")
model = DiffMLP(Xs.shape[1], CFG['diff_hidden'], CFG['diff_layers']).to(dev)
sched = DiffSched(CFG['diff_steps'])
for attr in ['b','a','ab']: setattr(sched, attr, getattr(sched, attr).to(dev))
opt = torch.optim.Adam(model.parameters(), lr=CFG['diff_lr'])
ds = TensorDataset(torch.FloatTensor(Xs)); dl = DataLoader(ds, batch_size=CFG['diff_bs'], shuffle=True)
model.train()
for ep in range(CFG['diff_epochs']):
tot = 0
for (xb,) in dl:
xb = xb.to(dev); bs = xb.shape[0]
t = torch.randint(0, CFG['diff_steps'], (bs,), device=dev)
xt, noise = sched.noise(xb, t)
loss = F.mse_loss(model(xt, t.float()), noise)
opt.zero_grad(); loss.backward(); opt.step(); tot += loss.item()*bs
if (ep+1)%10==0: print(f" Diff ep {ep+1}/{CFG['diff_epochs']}: loss={tot/len(ds):.6f}", flush=True)
print(f" Done: loss={tot/len(ds):.6f}", flush=True)
torch.save({'m':model.state_dict()}, diff_path)
# Siblings
if os.path.exists(sib_path):
d = np.load(sib_path); origs, sibs = d['o'], d['s']
else:
print("\n=== Generating Siblings ===")
model.eval(); origs, sibs = [], []
with torch.no_grad():
for i in range(0, len(Xs), CFG['diff_bs']):
xb = torch.FloatTensor(Xs[i:i+CFG['diff_bs']]).to(dev)
for _ in range(CFG['n_sib']):
sib = sched.sdedit(model, xb, CFG['t0'], dev)
origs.append(xb.cpu().numpy()); sibs.append(sib.cpu().numpy())
origs = np.concatenate(origs); sibs = np.concatenate(sibs)
print(f" {len(origs)} pairs, mean L2 diff: {np.mean(np.linalg.norm(origs-sibs,axis=1)):.4f}", flush=True)
np.savez_compressed(sib_path, o=origs, s=sibs)
# Encoder
enc_path = f"saved_model_files/enc_i3_s{a.seed}_t{a.t0}.pt"
if a.skip_enc and os.path.exists(enc_path):
ck = torch.load(enc_path, map_location=dev)
encoder = Encoder(Xs.shape[1], CFG['enc_hidden'], CFG['enc_dim']).to(dev)
encoder.load_state_dict(ck['e'])
else:
print("\n=== Training Encoder ===")
encoder = Encoder(Xs.shape[1], CFG['enc_hidden'], CFG['enc_dim']).to(dev)
opt = torch.optim.Adam(encoder.parameters(), lr=CFG['ctr_lr'])
n = len(origs)
# Shuffle pairs (not individual samples) to keep (orig, sib) adjacent
pair_idx = np.random.permutation(n)
origs_shuffled = origs[pair_idx]
sibs_shuffled = sibs[pair_idx]
data = np.zeros((n*2, Xs.shape[1]), dtype=np.float32)
data[0::2]=origs_shuffled; data[1::2]=sibs_shuffled
ds = TensorDataset(torch.FloatTensor(data)); dl = DataLoader(ds, batch_size=CFG['ctr_bs'], shuffle=False)
encoder.train()
for ep in range(CFG['ctr_epochs']):
tot = 0
for (xb,) in dl:
xb = xb.to(dev); emb = encoder(xb); loss = infonce(emb, CFG['ctr_temp'])
opt.zero_grad(); loss.backward(); opt.step(); tot += loss.item()*xb.shape[0]
if (ep+1)%10==0: print(f" Enc ep {ep+1}/{CFG['ctr_epochs']}: loss={tot/len(ds):.4f}", flush=True)
print(f" Done: loss={tot/len(ds):.4f}", flush=True)
torch.save({'e':encoder.state_dict()}, enc_path)
# Eval
print("\n=== Results ===")
eval_pairs(encoder, cv_s, iv_s, lbs, dev, f"Idea3 (t0={a.t0})")
print(f"\n Baseline (raw props): F1={b_raw:.4f}", flush=True)
print(f" Supervised XGBoost ref: F1=0.982", flush=True)
# Patched
if __name__=='__main__': main()
sibs_shuffled = sibs[pair_idx]
in()
|