File size: 14,629 Bytes
63ccb7e | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 | """Stacked out-of-fold (OOF) ensemble with nested cross-validation.
The base models are trained on exactly the same five folds (``--split_seed``
fixed), so every sample has one out-of-fold prediction per model. A per-target
linear blend is then fitted *inside* the loop: for fold ``k`` the weights are
estimated on the other four folds and applied to fold ``k``. No sample ever
contributes to the weights that predict it, so the reported score is an honest
out-of-sample estimate of the stacked model.
Why stacking: the architectures fail in different places (the physics-informed
Mamba is best at the absolute stress level, the attention hybrids at the curve
shape). A fixed average cannot exploit that; a fitted blend can.
Usage::
python -m pimamba.stack --prefix best --models hymba gdn llama pimamba \
--seeds 0 1 2 --out stack_best
"""
from __future__ import annotations
import argparse
import json
import re
from pathlib import Path
from typing import Dict, Iterable, List, Sequence
import numpy as np
import pandas as pd
from .data import TARGETS, TARGET_KEYS
from .metrics import PRIMARY, _r2, regression_metrics
from .pooled import RUNS, TABLES
TAG_RE = re.compile(r"^(?P<prefix>[a-z0-9]+)_(?P<model>.+)_f(?P<fold>\d+)_s(?P<seed>\d+)$")
# ---------------------------------------------------------------------------
# data assembly
# ---------------------------------------------------------------------------
def collect(prefix: str, models: Sequence[str], seeds: Sequence[int],
n_folds: int = 5, pmap: Dict[str, str] | None = None,
smap: Dict[str, Sequence[int]] | None = None
) -> Dict[str, Dict[int, np.ndarray]]:
"""Read ``predictions.npz`` for every (model, fold, seed).
Returns ``{model: {sid: (pred, true, mask)}}`` where each array is
``(T, 13)``. Seeds are averaged *after* inverse normalisation (the stored
predictions are already in physical units), which is the same operation as
averaging the sampled curves.
"""
pmap = pmap or {}
smap = smap or {}
acc: Dict[str, Dict[int, List[np.ndarray]]] = {m: {} for m in models}
for m in models:
pref = pmap.get(m, prefix)
for f in range(n_folds):
for s in smap.get(m, seeds):
d = RUNS / f"{pref}_{m}_f{f}_s{s}" / "predictions.npz"
if not d.exists():
continue
z = np.load(d, allow_pickle=True)
P, Y, M = z["pred"], z["true"], z["mask"]
sids = np.asarray(z["sids"]).astype(int)
for i, sid in enumerate(sids):
slot = acc[m].setdefault(int(sid), {"p": [], "y": None, "m": None})
slot["p"].append(P[i])
slot["y"] = Y[i]
slot["m"] = M[i]
out: Dict[str, Dict[int, tuple]] = {}
for m, per in acc.items():
out[m] = {sid: (np.mean(v["p"], 0), v["y"], v["m"]) for sid, v in per.items()}
return out
def fold_of(prefix: str, models: Sequence[str], seeds: Sequence[int],
n_folds: int = 5, pmap: Dict[str, str] | None = None,
smap: Dict[str, Sequence[int]] | None = None) -> Dict[int, int]:
"""Recover the fold membership of every sample from the run folders."""
pmap = pmap or {}
smap = smap or {}
fmap: Dict[int, int] = {}
m = models[0]
pref = pmap.get(m, prefix)
for f in range(n_folds):
d = RUNS / f"{pref}_{m}_f{f}_s{smap.get(m, seeds)[0]}" / "predictions.npz"
if not d.exists():
continue
z = np.load(d, allow_pickle=True)
for sid in np.asarray(z["sids"]).astype(int):
fmap[int(sid)] = f
return fmap
# ---------------------------------------------------------------------------
# blending
# ---------------------------------------------------------------------------
def _fit_weights(X: np.ndarray, y: np.ndarray, ridge: float = 1e-3,
nonneg: bool = True) -> tuple:
"""Ridge regression with optional non-negativity on the model weights.
Returns ``(w, b)`` such that ``pred = X @ w + b``. Centring removes the
intercept from the penalty; the solution is obtained with a small
projected-gradient loop (``nonneg``) or a closed form.
"""
mu_x, mu_y = X.mean(0), float(y.mean())
Xc, yc = X - mu_x, y - mu_y
A = Xc.T @ Xc + ridge * np.eye(X.shape[1])
b_vec = Xc.T @ yc
if not nonneg:
w = np.linalg.solve(A, b_vec)
else:
w = np.clip(np.linalg.solve(A, b_vec), 0, None)
# projected gradient refinement (a few hundred cheap iterations)
step = 1.0 / (np.linalg.eigvalsh(A).max() + 1e-9)
for _ in range(400):
w = np.clip(w - step * (A @ w - b_vec), 0, None)
b = mu_y - float(mu_x @ w)
return w, b
def nested_stack(data: Dict[str, Dict[int, tuple]], fmap: Dict[int, int],
models: Sequence[str], n_folds: int = 5,
ridge: float = 1e-3, nonneg: bool = True):
"""Fit per-target blends in a leave-one-fold-out loop.
Returns ``(P, Y, M, S, weights)`` with stacked predictions in the same
(sample, time, target) layout as a single model, plus the per-fold weight
matrix for inspection.
"""
sids = sorted(fmap, key=lambda s: (fmap[s], s))
keep = [s for s in sids if all(s in data[m] for m in models)]
S = np.asarray(keep, dtype=int)
shp = data[models[0]][keep[0]][0].shape
P = np.zeros((len(keep),) + shp, np.float32)
Y = np.zeros_like(P)
M = np.zeros_like(P)
weights = []
fold_arr = np.array([fmap[s] for s in keep])
for f in range(n_folds):
te = fold_arr == f
tr = ~te
for i, s in enumerate(keep):
Y[i] = data[models[0]][s][1]
M[i] = data[models[0]][s][2]
for j in range(shp[-1]):
Xtr, ytr = [], []
for i in np.where(tr)[0]:
msk = M[i, :, j] > 0
if not msk.any():
continue
Xtr.append(np.stack([data[m][keep[i]][0][msk, j] for m in models], -1))
ytr.append(Y[i, msk, j])
if not Xtr:
continue
Xc = np.concatenate(Xtr, 0)
yc = np.concatenate(ytr, 0)
w, b = _fit_weights(Xc, yc, ridge=ridge, nonneg=nonneg)
for i in np.where(te)[0]:
msk = M[i, :, j] > 0
if not msk.any():
continue
Xte = np.stack([data[m][keep[i]][0][msk, j] for m in models], -1)
P[i, msk, j] = Xte @ w + b
weights.append({"fold": f, "target": TARGET_KEYS[j],
**{f"w_{m}": float(wi) for m, wi in zip(models, w)},
"b": float(b)})
return P, Y, M, S, pd.DataFrame(weights)
def mean_ensemble(data: Dict[str, Dict[int, tuple]], fmap: Dict[int, int],
models: Sequence[str]):
sids = sorted(fmap, key=lambda s: (fmap[s], s))
keep = [s for s in sids if all(s in data[m] for m in models)]
S = np.asarray(keep, dtype=int)
P = np.mean([np.stack([data[m][s][0] for s in keep]) for m in models], 0)
Y = np.stack([data[models[0]][s][1] for s in keep])
M = np.stack([data[models[0]][s][2] for s in keep])
return P, Y, M, S
def greedy_blend(data: Dict[str, Dict[int, tuple]], fmap: Dict[int, int],
models: Sequence[str], n_folds: int = 5, iters: int = 15):
"""Caruana-style forward selection with replacement, fitted per fold.
For every target we start from an empty blend and repeatedly add the model
that most reduces the squared error on the *training* folds; the final
prediction on the held-out fold is the average of the selected models
(a model can be picked several times, which gives it a larger weight).
Selection uses only training folds, so the score stays out-of-sample.
Why this instead of ridge: with 26 training trajectories the models are
strongly correlated, and unconstrained least squares produces unstable,
sometimes negative weights. Greedy selection is a coarse but much more
stable procedure on this sample size.
"""
sids = sorted(fmap, key=lambda s: (fmap[s], s))
keep = [s for s in sids if all(s in data[m] for m in models)]
S = np.asarray(keep, dtype=int)
shp = data[models[0]][keep[0]][0].shape
P = np.zeros((len(keep),) + shp, np.float32)
Y = np.zeros_like(P)
M = np.zeros_like(P)
fold_arr = np.array([fmap[s] for s in keep])
for i, s in enumerate(keep):
Y[i] = data[models[0]][s][1]
M[i] = data[models[0]][s][2]
picks = []
for f in range(n_folds):
te = np.where(fold_arr == f)[0]
tr = np.where(fold_arr != f)[0]
for j in range(shp[-1]):
trn = [i for i in tr if (M[i, :, j] > 0).sum() >= 4]
if not trn:
continue
ytr = np.concatenate([Y[i, M[i, :, j] > 0, j] for i in trn])
Xtr = {m: np.concatenate([data[m][keep[i]][0][M[i, :, j] > 0, j]
for i in trn]) for m in models}
sel: List[str] = []
cur = np.zeros_like(ytr)
for _ in range(iters):
best = None
for m in models:
trial = (cur * len(sel) + Xtr[m]) / (len(sel) + 1)
err = float(np.mean((trial - ytr) ** 2))
if best is None or err < best[0]:
best = (err, m, trial)
sel.append(best[1])
cur = best[2]
cnt = {m: sel.count(m) / float(len(sel)) for m in models}
picks.append({"fold": f, "target": TARGET_KEYS[j], **cnt})
for i in te:
msk = M[i, :, j] > 0
if not msk.any():
continue
P[i, msk, j] = sum(cnt[m] * data[m][keep[i]][0][msk, j]
for m in models if cnt[m] > 0)
return P, Y, M, S, pd.DataFrame(picks)
# ---------------------------------------------------------------------------
# scoring
# ---------------------------------------------------------------------------
def score(P: np.ndarray, Y: np.ndarray, M: np.ndarray, S: np.ndarray) -> Dict:
pt = regression_metrics(Y.reshape(-1, len(TARGETS)),
P.reshape(-1, len(TARGETS)),
M.reshape(-1, len(TARGETS)))
row = {k: (round(pt[k]["r2"], 3) if np.isfinite(pt[k]["r2"]) else np.nan)
for k in TARGET_KEYS}
prim = [pt[k]["r2"] for k in PRIMARY if np.isfinite(pt[k]["r2"])]
row["primary_mean"] = round(float(np.mean(prim)), 3) if prim else np.nan
j = TARGET_KEYS.index("sigma")
mm = M[:, :, j] > 0
yv, pv = Y[:, :, j][mm], P[:, :, j][mm]
row["sigma_r2"] = round(_r2(yv, pv), 3)
ys, ps = [], []
for i in range(len(S)):
msk = M[i, :, j] > 0
if msk.sum() < 4:
continue
ys.append(Y[i, msk, j] - Y[i, msk, j].mean())
ps.append(P[i, msk, j] - P[i, msk, j].mean())
row["shape_r2"] = round(_r2(np.concatenate(ys), np.concatenate(ps)), 3)
row["n_samples"] = int(len(S))
return row
def main():
ap = argparse.ArgumentParser("pimamba.stack")
ap.add_argument("--prefix", default="best")
ap.add_argument("--models", nargs="+",
default=["hymba", "gdn", "llama", "pimamba"])
ap.add_argument("--seeds", nargs="+", type=int, default=[0])
ap.add_argument("--n_folds", type=int, default=5)
ap.add_argument("--ridge", type=float, default=1e-3)
ap.add_argument("--allow-negative", action="store_true")
ap.add_argument("--method", default="greedy",
choices=["greedy", "ridge", "mean"],
help="greedy = forward selection with replacement "
"(stable on small samples); ridge = per-target "
"non-negative least squares; mean = equal weights")
ap.add_argument("--iters", type=int, default=15)
ap.add_argument("--out", default="stack")
ap.add_argument("--prefixes", nargs="*", default=[],
help="per-model run prefix, e.g. mlp=main pimamba=main")
ap.add_argument("--seeds_map", nargs="*", default=[],
help="per-model seeds, e.g. mlp=0 pimamba=0,1,2")
a = ap.parse_args()
pmap = {}
for kv in a.prefixes:
k, v = kv.split("=")
pmap[k] = v
smap = {}
for kv in a.seeds_map:
k, v = kv.split("=")
smap[k] = [int(x) for x in v.split(",")]
data = collect(a.prefix, a.models, a.seeds, a.n_folds, pmap, smap)
fmap = fold_of(a.prefix, a.models, a.seeds, a.n_folds, pmap, smap)
rows = []
for m in a.models:
if not data.get(m):
continue
sids = sorted(data[m])
P = np.stack([data[m][s][0] for s in sids])
Y = np.stack([data[m][s][1] for s in sids])
M = np.stack([data[m][s][2] for s in sids])
rows.append({"model": m, **score(P, Y, M, np.asarray(sids))})
P, Y, M, S = mean_ensemble(data, fmap, a.models)
rows.append({"model": "mean-ensemble", **score(P, Y, M, S)})
if a.method == "mean":
Ps, Ys, Ms, Ss, W = P, Y, M, S, None
elif a.method == "greedy":
Ps, Ys, Ms, Ss, W = greedy_blend(data, fmap, a.models, a.n_folds,
iters=a.iters)
else:
Ps, Ys, Ms, Ss, W = nested_stack(data, fmap, a.models, a.n_folds,
ridge=a.ridge,
nonneg=not a.allow_negative)
rows.append({"model": f"stacked-{a.method}", **score(Ps, Ys, Ms, Ss)})
df = pd.DataFrame(rows)
df.to_csv(TABLES / f"tab_{a.out}.csv", index=False)
if W is not None:
W.to_csv(TABLES / f"tab_{a.out}_weights.csv", index=False)
print(df.to_string(index=False))
print("\nmean weight per model (greedy: selection frequency):")
wcols = [c for c in W.columns if c not in ("fold", "target")]
print(W[wcols].mean().round(3).to_string())
else:
print(df.to_string(index=False))
np.savez_compressed(TABLES / f"{a.out}_pred.npz", pred=Ps, true=Ys,
mask=Ms, sids=Ss)
(TABLES / f"{a.out}_meta.json").write_text(json.dumps(
{"models": a.models, "seeds": a.seeds, "prefixes": pmap,
"seeds_map": smap, "method": a.method, "ridge": a.ridge,
"nonneg": not a.allow_negative}, ensure_ascii=False, indent=2))
if __name__ == "__main__":
main() |