STER / code /README.md
eduzrh's picture
Few-shot optimization closeout: prototype+combo heads; WM gain confirmed K<=5; honest K>=10 tree recovery; docs+logs updated
44673a5 verified
|
Raw
History Blame Contribute Delete
1.59 kB

STER code

Scripts run against the official 3dSAGER repo (BarGenossar/3dSAGER) checkout.

file purpose
ster_repro.py Reproduce official matching pipeline (Bagging F1=0.946 on Hague-small).
ster_diagnostic.py Headroom map: few-shot / long-tail / discrepancy strata.
ster_fewshot.py Few-shot baseline landscape across 8 classifiers (bar-setting).
crawl_3dbag_multilod.py Crawl 3DBAG multi-LoD (LoD1.2/1.3/2.2) via api.3dbag.nl.
ster_crosslod.py Cross-LoD matching + zero-shot difficulty (run after crawl).
ster_export_vectors.py Compute per-building 25-D vectors (multi-LoD + Hague) -> data/ster_wm_vectors.npz.
ster_wm.py STER-WM: InfoNCE self-supervised identity encoder + few-shot fine-tune + hybrid (GPU).
ster_fusion.py Rank-fusion of WM cosine + raw Bagging across K (GPU).
ster_proto.py Few-shot prototype/metric head on WM relation vectors (GPU).
ster_combo.py Sample-efficient linear head on [raw-ratio ⊕ WM-relation] + fusion (GPU).
ster_zeroshot.py Zero-shot: train only on multi-LoD (0 target labels), transfer to Hague (GPU).
ster_noisy.py Cross-LoD hard-negative matching (noisy regime) (GPU).
ster_flow_v2.py Flow-Matching zero-shot (scale-invariant direction cosine); confirmed non-transfer (GPU).
clip.py Stub so import clip succeeds (ViT baseline unused in matching).

Dependencies: numpy, pandas, scikit-learn, scipy, xgboost, shapely, pyproj, faiss-cpu, torch. Pod (CPU) computes properties; GPU (torch 2.8+cu128) trains WM/Flow.