"""Sanity test for lgbm_feda: train with the SHOT LABELS SHUFFLED. If the score stays high, the gain does not come from the shots' labels and something is wrong.""" import os, sys os.environ.setdefault("NNLAB_THREADS", "4") sys.path.insert(0, "tools/nnlab") import numpy as np, pandas as pd import lab, lgbm_var as V W = lab.W d = lab.load(official60=False) feats = W.feature_names() mad = pd.read_parquet(W.DATA / "madrid_features.parquet", columns=feats) ams = pd.read_parquet(W.DATA / "amsterdam_features.parquet", columns=feats) Xm, Xa = mad[feats].to_numpy(float), ams[feats].to_numpy(float) Xm, Xa = W.robustz(Xm, W.robustz_fit(Xm)), W.robustz(Xa, W.robustz_fit(Xa)) ctx = {"top": V.top_gain_idx()} params = {**W.LGBM_PARAMS, **W.RETRAIN_OVERRIDES, "n_jobs": lab.THREADS} ya = d["_ya"] # scoring only for n in (5, 200): shot = lab.shots_for(d, n, 0) ys = lab.shot_labels(d, shot) keep = np.setdiff1d(np.arange(len(ya)), shot) ref = np.load(lab.pred_path("lgbm_feda", n, 0)).astype(float) f_ref = W.macro_f1(ya[keep], W.decide(ref[keep], d["prior"])) rng = np.random.default_rng(123 + n) P = V.fit_predict("lgbm_feda", Xm, d["ym"], Xa, shot, rng.permutation(ys), ctx, params) f_shuf = W.macro_f1(ya[keep], W.decide(P[keep], d["prior"])) Pt = V.fit_predict("lgbm_feda", Xm, d["ym"], Xa, shot, ys, ctx, params) print(f"n={n}: stored true-label F1 {f_ref:.4f} | refit true-label F1 {W.macro_f1(ya[keep], W.decide(Pt[keep], d['prior'])):.4f} " f"(max abs diff vs stored {np.abs(Pt-ref).max():.2e}) | SHUFFLED-label F1 {f_shuf:.4f}", flush=True)