File size: 10,919 Bytes
239a1be 773c429 63ccb7e 239a1be 773c429 239a1be 773c429 63ccb7e 773c429 239a1be 773c429 63ccb7e 239a1be 63ccb7e 239a1be 63ccb7e 239a1be 63ccb7e 239a1be 63ccb7e 239a1be 63ccb7e 239a1be 63ccb7e 773c429 63ccb7e 773c429 63ccb7e 239a1be 63ccb7e 239a1be | 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 | """Experiment orchestrator: main comparison / cross-alloy / ablations / data efficiency.
All runs use a fixed epoch budget with a full OneCycle anneal and no internal
validation split (``use_val=False``): with 33 samples a 15% hold-out is only
~4 trajectories, which makes early stopping pure noise (see report §5.1).
"""
from __future__ import annotations
import argparse
import json
import time
from pathlib import Path
from typing import Dict, List
import pandas as pd
from .data import PROC
from .train import Cfg, run
ROOT = Path(__file__).resolve().parents[2]
RUNS = ROOT / "results" / "runs"
SUMMARY = ROOT / "results" / "summary.csv"
MODELS = ["mlp", "gru", "lstm", "transformer", "mamba", "mamba2", "pimamba"]
MODELS_NEXT = ["llama", "xlstm", "gdn", "hymba", "koopman", "node", "rssm", "flow"]
# performance-first group: larger model, longer schedule, a fixed fold split so
# seeds differ only in initialisation (required for seed averaging / stacking).
MODELS_BEST = ["hymba", "gdn", "llama", "pimamba"]
ABLATIONS = [
# name, cfg overrides
("pimamba", {}),
("pimamba_nocouple", {"coupling": False}),
("pimamba_phys", {"w_mono": 1.0, "w_init": 1.0, "w_smooth": 0.1}),
("pimamba_nogeom", {"drop_features": ("geom",)}),
("pimamba_noflags", {"drop_features": ("geom_flags",)}),
("pimamba_noensemble", {"drop_features": ("ensemble",)}),
("pimamba_noinit", {"drop_features": ("init_rho", "init_phase")}),
("pimamba_nonorm", {"norm_mode": "none"}),
]
def _row(res: Dict, suite: str, variant: str = "", fold: int = -1,
seed: int = 0) -> Dict:
cfg = res["config"]
pt = res["per_target"]
sig = pt.get("sigma", {})
dec = res.get("decomp", {}) or {}
row = {
"suite": suite,
"variant": variant or cfg["model"],
"model": cfg["model"],
"tag": Path(res.get("tag", "")).name or "",
"fold": fold if fold >= 0 else cfg.get("fold", -1),
"seed": seed,
"holdout": cfg.get("holdout", ""),
"n_train": res["n_train"],
"n_test": res["n_test"],
"n_params": res["n_params"],
"epochs": res["epochs_run"],
"wall_time_s": res["wall_time_s"],
"macro_r2": res["summary"]["macro_r2"],
"macro_r2_primary": res["summary"].get("macro_r2_primary"),
"core_r2": res["summary"].get("core_r2"),
"median_r2": res["summary"]["median_r2"],
"macro_r2_log": res["summary"]["macro_r2_log"],
"macro_nmae": res["summary"]["macro_nmae"],
"sigma_r2": sig.get("r2"),
"sigma_mae": sig.get("mae"),
"sigma_nmae": sig.get("nmae"),
"sigma_r2_log": sig.get("r2_log"),
"sigma_shape_r2": dec.get("shape_r2"),
"sigma_level_r2": dec.get("level_oracle_r2"),
"sigma_level_corr": dec.get("level_corr"),
"sigma_level_mae": dec.get("level_mae_GPa"),
"xi_FCC_r2": pt.get("xi_FCC", {}).get("r2"),
"rho_total_r2": pt.get("rho_total", {}).get("r2"),
"rho_total_r2log": pt.get("rho_total", {}).get("r2_log"),
"twin_r2": pt.get("twin", {}).get("r2"),
"peak_rel_err": res["curve"].get("peak_rel_err_mean"),
"curve_rmse": res["curve"].get("curve_rmse_mean"),
"sig@0.05_rel_err": res["curve"].get("sig@0.05_rel_err_mean"),
}
return row
def _maybe_run(cfg: Cfg, suite: str, variant: str, tag: str,
rows: List[Dict], force: bool = False):
out = RUNS / tag / "metrics.json"
if out.exists() and not force:
r = json.loads(out.read_text())
# a cached smoke run (fewer epochs than requested) must not satisfy the
# request -- otherwise a 2-epoch debugging run silently blocks the real
# 300-epoch experiment.
if int(r.get("epochs_run", 0)) >= int(cfg.epochs):
rows.append(_row(r, suite, variant, cfg.fold, cfg.seed))
print(f" cached {tag}")
return
print(f" stale {tag}: {r.get('epochs_run')} < {cfg.epochs} epochs, rerun")
cfg.tag = tag
res = run(cfg, verbose=True)
res["tag"] = tag
rows.append(_row(res, suite, variant, cfg.fold, cfg.seed))
def suite_main(seeds, folds, models, force, epochs, lr=3e-4):
rows = []
for m in models:
for f in folds:
for s in seeds:
cfg = Cfg(model=m, seed=s, split="kfold", fold=f, n_folds=5,
epochs=epochs, d_model=64, n_layers=2, use_val=False, lr=lr)
_maybe_run(cfg, "main", m, f"main_{m}_f{f}_s{s}", rows, force)
return rows
def suite_loao(seeds, models, force, epochs, lr=3e-4):
meta = pd.read_csv(PROC / "samples_meta.csv")
alloys = sorted(meta.alloy.unique())
rows = []
for m in models:
for al in alloys:
for s in seeds:
cfg = Cfg(model=m, seed=s, split="loao", holdout=al,
epochs=epochs, d_model=64, n_layers=2, use_val=False, lr=lr)
safe = "".join(ch if ch.isalnum() else "_" for ch in al)[:40]
_maybe_run(cfg, "loao", m, f"loao_{m}_{safe}_s{s}", rows, force)
return rows
def suite_ablation(seeds, folds, force, epochs, lr=3e-4):
rows = []
for name, over in ABLATIONS:
for f in folds:
for s in seeds:
cfg = Cfg(model="pimamba", seed=s, split="kfold", fold=f,
n_folds=5, epochs=epochs, d_model=64, n_layers=2,
use_val=False, lr=lr, **over)
_maybe_run(cfg, "ablation", name, f"abl_{name}_f{f}_s{s}", rows, force)
return rows
def suite_dataeff(fractions, folds, models, force, epochs, lr=3e-4):
rows = []
for m in models:
for fr in fractions:
for f in folds:
cfg = Cfg(model=m, seed=0, split="kfold", fold=f, n_folds=5,
epochs=epochs, d_model=64, n_layers=2, train_frac=fr,
use_val=False, lr=lr)
_maybe_run(cfg, "dataeff", f"{m}@{fr:.2f}",
f"de_{m}_{int(fr*100):03d}_f{f}", rows, force)
return rows
def suite_next(seeds, folds, models, force, epochs, lr=3e-4):
"""Next-generation backbones (2024-2026), same protocol as ``suite_main``."""
rows = []
for m in models:
for f in folds:
for s in seeds:
cfg = Cfg(model=m, seed=s, split="kfold", fold=f, n_folds=5,
epochs=epochs, d_model=64, n_layers=2, use_val=False,
lr=lr)
_maybe_run(cfg, "next", m, f"next_{m}_f{f}_s{s}", rows, force)
return rows
def suite_best(seeds, folds, models, force, epochs, d_model=128, n_layers=4,
lr=3e-4, split_seed=0):
"""Performance-first group: no equal-budget constraint.
``split_seed`` is pinned so every seed and architecture sees the same five
folds; otherwise the OOF predictions cannot be averaged or stacked.
"""
rows = []
for m in models:
for f in folds:
for s in seeds:
cfg = Cfg(model=m, seed=s, split="kfold", fold=f, n_folds=5,
epochs=epochs, d_model=d_model, n_layers=n_layers,
use_val=False, lr=lr, split_seed=split_seed)
_maybe_run(cfg, "best", m, f"best_{m}_f{f}_s{s}", rows, force)
return rows
def suite_v2(seeds, folds, force, epochs, lr=3e-4):
"""PI-Mamba v2: Taylor-form coupling and the curve-level peak loss."""
variants = [
("pimamba2", {"model": "pimamba2"}),
("pimamba2_peak", {"model": "pimamba2", "w_peak": 0.5}),
("pimamba_peak", {"model": "pimamba", "w_peak": 0.5}),
("pimamba2_nocouple", {"model": "pimamba2", "coupling": False}),
]
rows = []
for name, over in variants:
for f in folds:
for s in seeds:
cfg = Cfg(seed=s, split="kfold", fold=f, n_folds=5, epochs=epochs,
d_model=64, n_layers=2, use_val=False, lr=lr, **over)
_maybe_run(cfg, "v2", name, f"v2_{name}_f{f}_s{s}", rows, force)
return rows
def append_summary(rows: List[Dict]):
if not rows:
return
df = pd.DataFrame(rows)
if SUMMARY.exists():
old = pd.read_csv(SUMMARY)
df = pd.concat([old, df], ignore_index=True)
df = df.drop_duplicates(subset=["tag"], keep="last")
SUMMARY.parent.mkdir(parents=True, exist_ok=True)
df.to_csv(SUMMARY, index=False)
print(f"\nsummary -> {SUMMARY} ({len(df)} rows)")
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--suite", required=True,
choices=["main", "loao", "ablation", "dataeff", "next",
"v2", "best", "all"])
ap.add_argument("--models", nargs="*", default=MODELS)
ap.add_argument("--seeds", nargs="*", type=int, default=[0])
ap.add_argument("--folds", nargs="*", type=int, default=[0, 1, 2, 3, 4])
ap.add_argument("--fractions", nargs="*", type=float,
default=[0.25, 0.5, 0.75, 1.0])
ap.add_argument("--epochs", type=int, default=None)
ap.add_argument("--lr", type=float, default=3e-4)
ap.add_argument("--d_model", type=int, default=None)
ap.add_argument("--n_layers", type=int, default=None)
ap.add_argument("--split_seed", type=int, default=0)
ap.add_argument("--force", action="store_true")
a = ap.parse_args()
rows: List[Dict] = []
ep = 300 if a.epochs is None else a.epochs # equal-budget suites
ep_best = 600 if a.epochs is None else a.epochs # performance-first suite
if a.suite in ("main", "all"):
print("== main ==")
rows += suite_main(a.seeds, a.folds, a.models, a.force, ep, a.lr)
if a.suite in ("loao", "all"):
print("== loao ==")
rows += suite_loao(a.seeds, ["pimamba", "mamba", "gru", "mlp"], a.force, ep, a.lr)
if a.suite in ("ablation", "all"):
print("== ablation ==")
rows += suite_ablation(a.seeds, a.folds, a.force, ep, a.lr)
if a.suite in ("next", "all"):
print("== next ==")
ms = [m for m in a.models if m in MODELS_NEXT] or MODELS_NEXT
rows += suite_next(a.seeds, a.folds, ms, a.force, ep, a.lr)
if a.suite in ("v2", "all"):
print("== v2 ==")
rows += suite_v2(a.seeds, a.folds, a.force, ep, a.lr)
if a.suite in ("best", "all"):
print("== best ==")
ms = [m for m in a.models if m in MODELS_BEST] or MODELS_BEST
rows += suite_best(a.seeds, a.folds, ms, a.force, ep_best,
a.d_model or 128, a.n_layers or 4, a.lr, a.split_seed)
if a.suite in ("dataeff", "all"):
print("== dataeff ==")
rows += suite_dataeff(a.fractions, a.folds,
["pimamba", "mamba", "gru", "mlp"], a.force, ep, a.lr)
append_summary(rows)
if __name__ == "__main__":
main() |