STER / code /ster_zeroshot.py
eduzrh's picture
Full autonomous run: 8 proposals with real results (few-shot WM wins K<=10; zero-shot via multi-LoD 0.67; noisy/cross-LoD solved by baseline) + all code, results, derived data
99f65bd verified
Raw
History Blame Contribute Delete
7.79 kB
"""
STER — Zero-shot Geospatial ER (GPU).
TRUE zero-shot: train ONLY on free 3DBAG multi-LoD self-supervision (same building's
LoD1.2 vs LoD2.2). NO Hague labels are ever used — even the decision threshold is
calibrated on a held-out multi-LoD split. Then transfer to the Hague CityModel<->3DBAG
test set (hard negatives). Tests whether transformation/identity structure learned on
the LoD gap transfers to the (unseen) source gap.
Methods:
base_raw_transfer_bagging : Bagging on multi-LoD ratio pairs (hard negs) -> Hague
base_raw_cos : cosine of raw std property vecs, thr from multi-LoD
wm_zeroshot_cos : InfoNCE identity encoder cosine, thr from multi-LoD (WM)
flow_zeroshot : Conditional Flow Matching residual score, thr from ML (4B)
"""
import os, json, numpy as np, torch, torch.nn as nn, torch.nn.functional as F
from sklearn.ensemble import BaggingClassifier
from sklearn.preprocessing import StandardScaler
from sklearn.neighbors import NearestNeighbors
from sklearn.metrics import f1_score, precision_score, recall_score
NPZ = "/root/ster/data/ster_wm_vectors.npz"
OUT = "/root/ster/exp/ster_zeroshot_results.json"
DEV = "cuda" if torch.cuda.is_available() else "cpu"
torch.manual_seed(1); np.random.seed(1)
z = np.load(NPZ, allow_pickle=True)
lod12, lod22 = z["lod12_X"], z["lod22_X"]
candX, indexX = z["cand_X"], z["index_X"]
teP, teY = z["test_pairs"], z["test_y"]
N = lod12.shape[0]; DIM = lod12.shape[1]
print(f"multiLoD={lod12.shape} Hague test={teP.shape} pos={teY.sum()} dev={DEV}", flush=True)
scaler = StandardScaler().fit(np.vstack([lod12, lod22, candX, indexX]))
def T(x): return torch.tensor(scaler.transform(x), dtype=torch.float32, device=DEV)
L12t, L22t = T(lod12), T(lod22)
CANDt, INDEXt = T(candX), T(indexX)
L12s, L22s = scaler.transform(lod12), scaler.transform(lod22)
CANDs, INDEXs = scaler.transform(candX), scaler.transform(indexX)
# split multi-LoD into train / calib (for thresholds) — no Hague labels used
idx = np.random.permutation(N); ntr = int(0.8 * N)
mtr, mcal = idx[:ntr], idx[ntr:]
# ---- build multi-LoD ratio pairs with HARD negatives (top-k NN among LoD22) ----
nn = NearestNeighbors(n_neighbors=6).fit(L22s)
_, knn = nn.kneighbors(L12s) # for each LoD12, nearest LoD22 buildings
def ratio(a, b):
with np.errstate(divide='ignore', invalid='ignore'):
r = np.where(b != 0, a / b, 1000.0)
return np.clip(np.round(r, 3), None, 1000.0)
def make_ml_pairs(ids):
Xr, Y = [], []
for i in ids:
Xr.append(ratio(L12s[i], L22s[i])); Y.append(1)
for j in knn[i][1:3]:
if j != i:
Xr.append(ratio(L12s[i], L22s[j])); Y.append(0)
return np.array(Xr), np.array(Y)
Xr_tr, Yr_tr = make_ml_pairs(mtr)
# Hague test raw ratio features
Xr_te = ratio(candX[teP[:, 0]], indexX[teP[:, 1]])
# calibration pairs (multi-LoD held out): scores for pos and hard-neg
def ml_calib_scores(score_fn):
pos, neg = [], []
for i in mcal:
pos.append(score_fn(i, i, kind='ml'))
for j in knn[i][1:3]:
if j != i:
neg.append(score_fn(i, j, kind='ml'))
return np.array(pos), np.array(neg)
def thr_from(pos, neg):
return float(np.median([np.quantile(pos, 0.2), np.quantile(neg, 0.8)]))
report = {}
# 1) raw transfer Bagging
clf = BaggingClassifier(n_estimators=100, random_state=1).fit(Xr_tr, Yr_tr)
pred = clf.predict(Xr_te)
report["base_raw_transfer_bagging"] = dict(
f1=round(f1_score(teY, pred, zero_division=0), 4),
precision=round(precision_score(teY, pred, zero_division=0), 4),
recall=round(recall_score(teY, pred, zero_division=0), 4))
# 2) raw cosine
def raw_cos(a, b): return float(np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b) + 1e-8))
pcos = np.array([raw_cos(L12s[i], L22s[i]) for i in mcal])
ncos = np.array([raw_cos(L12s[i], L22s[j]) for i in mcal for j in knn[i][1:3] if j != i])
thr = thr_from(pcos, ncos)
cos_te = np.array([raw_cos(CANDs[c], INDEXs[d]) for c, d in teP])
report["base_raw_cos"] = dict(thr=round(thr, 3),
f1=round(f1_score(teY, (cos_te >= thr).astype(int), zero_division=0), 4))
# 3) WM InfoNCE encoder
class Enc(nn.Module):
def __init__(s, din, d=32):
super().__init__(); s.net = nn.Sequential(nn.Linear(din, 64), nn.GELU(), nn.Linear(64, 64), nn.GELU(), nn.Linear(64, d))
def forward(s, x): return F.normalize(s.net(x), dim=-1)
def info_nce(za, zb, tau=0.1):
lg = za @ zb.t() / tau; lab = torch.arange(za.size(0), device=za.device)
return 0.5 * (F.cross_entropy(lg, lab) + F.cross_entropy(lg.t(), lab))
enc = Enc(DIM).to(DEV); opt = torch.optim.Adam(enc.parameters(), 1e-3, weight_decay=1e-5)
A, B = L12t[mtr], L22t[mtr]
for ep in range(500):
p = torch.randperm(A.size(0), device=DEV)
for i in range(0, A.size(0), 256):
b = p[i:i+256]
if b.numel() < 8: continue
loss = info_nce(enc(A[b]), enc(B[b])); opt.zero_grad(); loss.backward(); opt.step()
enc.eval()
with torch.no_grad():
e12 = enc(L12t).cpu().numpy(); e22 = enc(L22t).cpu().numpy()
ecand = enc(CANDt).cpu().numpy(); eindex = enc(INDEXt).cpu().numpy()
pe = np.array([np.dot(e12[i], e22[i]) for i in mcal])
ne = np.array([np.dot(e12[i], e22[j]) for i in mcal for j in knn[i][1:3] if j != i])
thrw = thr_from(pe, ne)
wcos_te = np.sum(ecand[teP[:, 0]] * eindex[teP[:, 1]], axis=1)
report["wm_zeroshot_cos"] = dict(thr=round(thrw, 3),
f1=round(f1_score(teY, (wcos_te >= thrw).astype(int), zero_division=0), 4),
precision=round(precision_score(teY, (wcos_te >= thrw).astype(int), zero_division=0), 4),
recall=round(recall_score(teY, (wcos_te >= thrw).astype(int), zero_division=0), 4))
# 4) Conditional Flow Matching (4B)
class VNet(nn.Module):
def __init__(s, din):
super().__init__(); s.net = nn.Sequential(nn.Linear(din + 1, 128), nn.GELU(),
nn.Linear(128, 128), nn.GELU(), nn.Linear(128, din))
def forward(s, x, t): return s.net(torch.cat([x, t], -1))
vnet = VNet(DIM).to(DEV); optf = torch.optim.Adam(vnet.parameters(), 1e-3, weight_decay=1e-5)
x0, x1 = L12t[mtr], L22t[mtr]
for ep in range(600):
p = torch.randperm(x0.size(0), device=DEV)
for i in range(0, x0.size(0), 256):
b = p[i:i+256]
if b.numel() < 8: continue
a0, a1 = x0[b], x1[b]
t = torch.rand(a0.size(0), 1, device=DEV)
xt = (1 - t) * a0 + t * a1
loss = F.mse_loss(vnet(xt, t), a1 - a0)
optf.zero_grad(); loss.backward(); optf.step()
vnet.eval()
@torch.no_grad()
def flow_resid(a, b):
a = torch.tensor(a, dtype=torch.float32, device=DEV); b = torch.tensor(b, dtype=torch.float32, device=DEV)
if a.dim() == 1: a = a[None]; b = b[None]
t = torch.full((a.size(0), 1), 0.5, device=DEV)
xt = 0.5 * (a + b)
v = vnet(xt, t)
return (-(v - (b - a)).pow(2).mean(1)).cpu().numpy() # higher = more consistent
pf = flow_resid(L12s[mcal], L22s[mcal])
nf = np.concatenate([flow_resid(L12s[[i]], L22s[[j]]) for i in mcal for j in knn[i][1:3] if j != i])
thrf = thr_from(pf, nf)
flow_te = flow_resid(CANDs[teP[:, 0]], INDEXs[teP[:, 1]])
report["flow_zeroshot"] = dict(thr=round(float(thrf), 3),
f1=round(f1_score(teY, (flow_te >= thrf).astype(int), zero_division=0), 4),
precision=round(precision_score(teY, (flow_te >= thrf).astype(int), zero_division=0), 4),
recall=round(recall_score(teY, (flow_te >= thrf).astype(int), zero_division=0), 4))
print("\n=== ZERO-SHOT (train on multi-LoD, 0 Hague labels) ===", flush=True)
for k, s in sorted(report.items(), key=lambda kv: -kv[1]['f1']):
print(f" {k:28s} F1={s['f1']:.4f}", flush=True)
os.makedirs(os.path.dirname(OUT), exist_ok=True)
json.dump(report, open(OUT, "w"), indent=2)
print("Saved ->", OUT, flush=True)