| """Fixed-rule ensembles of methods already in the prediction store. |
| |
| Rule, fixed before any ensemble was scored (from the literature shortlist, item 6): |
| log p = w0 * mean(log p of zero-shot members) + (1 - w0) * mean(log p of shot-using members) |
| w0 = K / (n + K), K = 10 (the zero-shot anchor fades as the labelled sample grows) |
| No weights are fitted, so no Amsterdam label is used beyond what the members used. |
| |
| python tools/nnlab/ensemble.py NAME --zs zs_xgb_dann --shot lgbm_upw,tab_dann_rz_joint [--k 10] |
| writes method ens_NAME. |
| """ |
| import argparse, sys |
| from pathlib import Path |
| sys.path.insert(0, str(Path(__file__).resolve().parent)) |
| import lab |
| import numpy as np |
|
|
| ap = argparse.ArgumentParser() |
| ap.add_argument("name") |
| ap.add_argument("--zs", default="") |
| ap.add_argument("--shot", default="") |
| ap.add_argument("--k", type=float, default=10.0) |
| a = ap.parse_args() |
| zs = [m for m in a.zs.split(",") if m] |
| sh = [m for m in a.shot.split(",") if m] |
| assert zs or sh |
| name = "ens_" + a.name |
| logm = lambda ms, n, r: np.mean([np.log(np.clip(lab.load_pred(m, n, r), 1e-6, 1)) for m in ms], 0) |
| done = 0 |
| for n, rep in lab.draws(): |
| if any(not lab.have_pred(m, n, rep) for m in sh) or any(not (lab.have_pred(m, n, rep) or lab.have_pred(m, 0, 0)) for m in zs): |
| continue |
| w0 = (a.k / (n + a.k) if sh else 1.0) if zs else 0.0 |
| L = (w0 * logm(zs, n, rep) if zs else 0) + ((1 - w0) * logm(sh, n, rep) if sh else 0) |
| lab.save_pred(name, n, rep, lab.softmax(L), meta=dict(zs=zs, shot=sh, k=a.k, rule="w0=k/(n+k), mean of log-probs")) |
| done += 1 |
| print(f"{name}: {done} draws written") |
|
|
