Datasets:

eduzrh
/

training-model-material

"""Generate the result tables / figures / RESULTS.md from ``results/``.
Run after the experiment suites have finished::
    python -m pimamba.report
"""
from __future__ import annotations

import json
import os
from pathlib import Path
from typing import Dict, List

import numpy as np
import pandas as pd

from . import analysis as A

ROOT = Path(os.environ.get("PIMAMBA_ROOT")
            or Path(__file__).resolve().parents[2])
RUNS = ROOT / "results" / "runs"
TABLES = ROOT / "results" / "tables"
FIGS = ROOT / "results" / "figures"

MAIN_ORDER = ["mlp", "gru", "lstm", "transformer", "mamba", "mamba2", "pimamba"]
COLS = ["sigma_r2", "sigma_shape_r2", "sigma_level_corr", "macro_r2_primary",
        "rho_total_r2", "twin_r2", "xi_FCC_r2", "peak_rel_err"]


def _pm(df: pd.DataFrame, key: str, col: str, nd: int = 3) -> str:
    g = df[df[key[0]] == key[1]][col]
    if not len(g):
        return "n/a"
    return A.fmt_pm(float(g.mean()), float(g.std()) if len(g) > 1 else float("nan"), nd)


def _maybe_csv(path: Path) -> "pd.DataFrame | None":
    if path.exists():
        return pd.read_csv(path)
    return None


def table_main(df: pd.DataFrame) -> pd.DataFrame:
    """Per-fold mean+-sd *and* pooled out-of-fold metrics side by side."""
    sub = df[df.suite == "main"].copy()
    ps = _maybe_csv(TABLES / "tab_pooled_sigma.csv")
    pt = _maybe_csv(TABLES / "tab_pooled_per_target.csv")
    if ps is not None:
        ps = ps.set_index("model")
    if pt is not None:
        pt = pt.set_index("model")
    rows = []
    for m in MAIN_ORDER:
        g = sub[sub.model == m]
        if not len(g):
            continue
        r = {"model": A.MODEL_LABEL[m], "n": len(g)}
        r["应力 R²(逐折)"] = A.fmt_pm(float(g.sigma_r2.mean()), float(g.sigma_r2.std()))
        r["形状 R²(逐折)"] = A.fmt_pm(float(g.sigma_shape_r2.mean()),
                                     float(g.sigma_shape_r2.std()))
        if ps is not None and m in ps.index:
            r["池化应力 R²"] = f"{float(ps.loc[m, 'sigma_r2']):.3f}"
            r["池化形状 R²"] = f"{float(ps.loc[m, 'shape_r2']):.3f}"
        if pt is not None and m in pt.index:
            r["池化主指标 R²"] = f"{float(pt.loc[m, 'primary_mean']):.3f}"
        r["水平相关"] = A.fmt_pm(float(g.sigma_level_corr.mean()),
                               float(g.sigma_level_corr.std()))
        # rare targets (twin: 6/33 labelled) have a meaningless per-fold R2 --
        # a fold can contain a single labelled sample.  Report pooled OOF.
        if pt is not None and m in pt.index:
            r["池化总位错 R²"] = f"{float(pt.loc[m, 'rho_total']):.3f}"
            r["池化孪晶 R²"] = f"{float(pt.loc[m, 'twin']):.3f}"
            r["池化FCC R²"] = f"{float(pt.loc[m, 'xi_FCC']):.3f}"
        else:
            r["总位错 R²"] = A.fmt_pm(float(g.rho_total_r2.mean()),
                                   float(g.rho_total_r2.std()))
        r["峰值相对误差"] = f"{g.peak_rel_err.mean():.3f}"
        r["曲线 RMSE/GPa"] = A.fmt_pm(float(g.curve_rmse.mean()), float(g.curve_rmse.std()))
        r["逐折宏平均 R²"] = A.fmt_pm(float(g.macro_r2_primary.mean()),
                                   float(g.macro_r2_primary.std()))
        r["参数量"] = f"{int(g.n_params.mean()):,}"
        rows.append(r)
    return pd.DataFrame(rows)


NEXT_ORDER = ["llama", "xlstm", "gdn", "hymba", "koopman", "node", "rssm",
              "flow", "pimamba2"]


def table_next(df: pd.DataFrame) -> pd.DataFrame:
    """Next-generation backbones: per-fold + pooled OOF, sorted by R2."""
    sub = df[df.suite == "next"].copy()
    if not len(sub):
        return pd.DataFrame()
    ps = _maybe_csv(TABLES / "tab_pooled_next_sigma.csv")
    pt = _maybe_csv(TABLES / "tab_pooled_next_per_target.csv")
    if ps is not None:
        ps = ps.set_index("model")
    if pt is not None:
        pt = pt.set_index("model")
    rows = []
    order = sub.groupby("model").sigma_r2.mean().sort_values(ascending=False).index
    for m in order:
        g = sub[sub.model == m]
        r = {"模型": A.MODEL_LABEL.get(m, m), "n": len(g),
             "应力 R²(逐折)": A.fmt_pm(float(g.sigma_r2.mean()),
                                    float(g.sigma_r2.std())),
             "形状 R²(逐折)": A.fmt_pm(float(g.sigma_shape_r2.mean()),
                                    float(g.sigma_shape_r2.std()))}
        if ps is not None and m in ps.index:
            r["池化应力"] = f"{float(ps.loc[m, 'sigma_r2']):.3f}"
            r["池化形状"] = f"{float(ps.loc[m, 'shape_r2']):.3f}"
        if pt is not None and m in pt.index:
            r["池化主指标"] = f"{float(pt.loc[m, 'primary_mean']):.3f}"
        r["峰值误差"] = f"{g.peak_rel_err.mean():.3f}"
        r["参数量"] = f"{int(g.n_params.mean()):,}"
        r["耗时/run(s)"] = f"{g.wall_time_s.mean():.0f}"
        rows.append(r)
    return pd.DataFrame(rows)


def table_v2(df: pd.DataFrame) -> pd.DataFrame:
    """PI-Mamba v2 variants (Taylor coupling + peak loss)."""
    sub = df[df.suite == "v2"].copy()
    if not len(sub):
        return pd.DataFrame()
    ps = _maybe_csv(TABLES / "tab_pooled_v2_sigma.csv")
    pt = _maybe_csv(TABLES / "tab_pooled_v2_per_target.csv")
    if ps is not None:
        ps = ps.set_index("model")
    if pt is not None:
        pt = pt.set_index("model")
    rows = []
    # reference = the main-suite PI-Mamba, so every variant is comparable to the
    # baseline it is meant to improve (the v2 suite has no "pimamba" row).
    bm = df[(df.suite == "main") & (df.model == "pimamba")].sigma_r2
    base = float(bm.mean()) if len(bm) else float(sub.sigma_r2.max())
    order = sub.groupby("variant").sigma_r2.mean().sort_values(ascending=False).index
    for v in order:
        g = sub[sub.variant == v]
        r = {"变体": v, "n": len(g),
             "应力 R²(逐折)": A.fmt_pm(float(g.sigma_r2.mean()),
                                    float(g.sigma_r2.std())),
             "形状 R²(逐折)": f"{g.sigma_shape_r2.mean():.3f}"}
        if ps is not None and v in ps.index:
            r["池化应力"] = f"{float(ps.loc[v, 'sigma_r2']):.3f}"
            r["池化形状"] = f"{float(ps.loc[v, 'shape_r2']):.3f}"
        if pt is not None and v in pt.index:
            r["池化主指标"] = f"{float(pt.loc[v, 'primary_mean']):.3f}"
        r["峰值误差"] = f"{g.peak_rel_err.mean():.3f}"
        r["ΔR²(vs PI-Mamba)"] = f"{g.sigma_r2.mean() - base:+.3f}"
        rows.append(r)
    return pd.DataFrame(rows)


def table_loao(df: pd.DataFrame) -> pd.DataFrame:
    sub = df[df.suite == "loao"].copy()
    if not len(sub):
        return pd.DataFrame()
    rows = []
    for al, g in sub.groupby("holdout"):
        r = {"留出合金体系": al, "n": len(g)}
        for m in ["mlp", "gru", "mamba", "pimamba"]:
            gg = g[g.model == m]
            r[A.MODEL_LABEL[m]] = (f"{gg.sigma_r2.mean():.3f}"
                                   if len(gg) else "n/a")
        rows.append(r)
    rows.append({"留出合金体系": "**平均**", "n": len(sub),
                 **{A.MODEL_LABEL[m]: f"{sub[sub.model == m].sigma_r2.mean():.3f}"
                    for m in ["mlp", "gru", "mamba", "pimamba"]}})
    return pd.DataFrame(rows)


def table_ablation(df: pd.DataFrame) -> pd.DataFrame:
    sub = df[df.suite == "ablation"].copy()
    if not len(sub):
        return pd.DataFrame()
    rows = []
    base = sub[sub.variant == "pimamba"].sigma_r2.mean()
    order = sub.groupby("variant").sigma_r2.mean().sort_values(ascending=False).index
    for v in order:
        g = sub[sub.variant == v]
        rows.append({
            "变体": A.MODEL_LABEL.get(v, v), "n": len(g),
            "应力 R²": A.fmt_pm(float(g.sigma_r2.mean()), float(g.sigma_r2.std())),
            "形状 R²": A.fmt_pm(float(g.sigma_shape_r2.mean()),
                                float(g.sigma_shape_r2.std())),
            "总位错 R²": f"{g.rho_total_r2.mean():.3f}",
            "ΔR²(应力)": f"{g.sigma_r2.mean() - base:+.3f}",
        })
    return pd.DataFrame(rows)


def table_dataeff(df: pd.DataFrame) -> pd.DataFrame:
    sub = df[df.suite == "dataeff"].copy()
    if not len(sub):
        return pd.DataFrame()
    def _frac(v):
        f = float(str(v).split("@")[-1])
        return f / 100.0 if f > 1.0 else f

    sub["frac"] = sub.variant.map(_frac)
    rows = []
    for m, g in sub.groupby("model"):
        r = {"模型": A.MODEL_LABEL.get(m, m)}
        for fr, gg in sorted(g.groupby("frac")):
            r[f"{int(fr*100)}%"] = f"{gg.sigma_r2.mean():.3f}"
        r["n_train@100%"] = int(g[g.frac == 1.0].n_train.mean()) if len(g[g.frac == 1.0]) else np.nan
        rows.append(r)
    return pd.DataFrame(rows)


def table_per_target(tag: str) -> pd.DataFrame:
    m = json.loads((RUNS / tag / "metrics.json").read_text())
    rows = []
    for k, v in m["per_target"].items():
        if v["n"] == 0:
            continue
        rows.append({"目标": k, "组": v["group"], "n": v["n"],
                     "R²": round(v["r2"], 3) if np.isfinite(v["r2"]) else "nan",
                     "R²(log)": round(v["r2_log"], 3) if np.isfinite(v.get("r2_log", np.nan)) else "–",
                     "nMAE": round(v["nmae"], 3) if np.isfinite(v["nmae"]) else "nan",
                     "σ_y": f"{v['y_std']:.3g}", "单位": v["unit"]})
    return pd.DataFrame(rows)


def rebuild_summary():
    """Reconstruct ``results/summary.csv`` from the individual run folders.
    Safety net: the orchestrator only flushes the summary at the end of a suite,
    so a crash would otherwise lose the table.
    """
    from .run_experiments import _row
    rows = []
    for d in sorted(RUNS.iterdir()):
        mp = d / "metrics.json"
        if not mp.exists():
            continue
        res = json.loads(mp.read_text())
        res["tag"] = d.name
        n = d.name
        if n.startswith("main_"):
            suite, var = "main", res["config"]["model"]
        elif n.startswith("abl_"):
            suite, var = "ablation", "_".join(n.split("_")[1:-2])
        elif n.startswith("loao_"):
            suite, var = "loao", res["config"]["model"]
        elif n.startswith("de_"):
            frac = int(n.split("_")[2]) / 100.0
            suite, var = "dataeff", f'{res["config"]["model"]}@{frac:.2f}'
        elif n.startswith("next_"):
            suite, var = "next", res["config"]["model"]
        elif n.startswith("v2_"):
            suite, var = "v2", "_".join(n.split("_")[1:-2])
        else:
            continue
        try:
            rows.append(_row(res, suite, var, res["config"].get("fold", -1),
                             res["config"].get("seed", 0)))
        except Exception as e:  # noqa: BLE001
            print("skip", n, e)
    if rows:
        df = pd.DataFrame(rows).drop_duplicates(subset=["tag"], keep="last")
        (ROOT / "results" / "summary.csv").parent.mkdir(parents=True, exist_ok=True)
        df.to_csv(ROOT / "results" / "summary.csv", index=False)
        print(f"rebuilt summary.csv: {len(df)} rows")


def main():
    import sys
    if "--rebuild" in sys.argv:
        rebuild_summary()
    TABLES.mkdir(parents=True, exist_ok=True)
    FIGS.mkdir(parents=True, exist_ok=True)
    df = A.load_summary()
    out: Dict[str, str] = {}

    tm = table_main(df)
    if len(tm):
        tm.to_csv(TABLES / "tab_main.csv", index=False)
        out["main"] = A.markdown_table(tm, list(tm.columns))
        A.fig_model_comparison(df)

    tn = table_next(df)
    if len(tn):
        tn.to_csv(TABLES / "tab_next.csv", index=False)
        out["next"] = A.markdown_table(tn, list(tn.columns))
        A.fig_next(df)

    tv = table_v2(df)
    if len(tv):
        tv.to_csv(TABLES / "tab_v2.csv", index=False)
        out["v2"] = A.markdown_table(tv, list(tv.columns))

    tl = table_loao(df)
    if len(tl):
        tl.to_csv(TABLES / "tab_loao.csv", index=False)
        out["loao"] = A.markdown_table(tl, list(tl.columns))

    ta = table_ablation(df)
    if len(ta):
        ta.to_csv(TABLES / "tab_ablation.csv", index=False)
        out["ablation"] = A.markdown_table(ta, list(ta.columns))
        A.fig_ablation(df)

    td = table_dataeff(df)
    if len(td):
        td.to_csv(TABLES / "tab_dataeff.csv", index=False)
        out["dataeff"] = A.markdown_table(td, list(td.columns))
        A.fig_dataeff(df)

    best = df[(df.suite == "main") & (df.model == "pimamba")].sort_values(
        "sigma_r2", ascending=False)
    if len(best):
        tag = str(best.iloc[0]["tag"])
        if (RUNS / tag / "predictions.npz").exists():
            A.fig_curves(tag)
            out["curve_tag"] = tag
        tp = table_per_target(tag)
        tp.to_csv(TABLES / "tab_per_target.csv", index=False)
        out["per_target"] = A.markdown_table(tp, list(tp.columns))
        out["per_target_tag"] = tag

    lines = ["# 自动生成的结果汇总", "",
             "> 口径说明：`(逐折)` 为 5 折 × 2 种子的 mean ± sd；`池化` 为把 5 折测试预测\n"
             "> 拼成 33 个样本后统一计算的 out-of-fold 指标（更稳定，报告 §6 主用）。\n"
             "> `逐折宏平均 R²` 仍包含 `rho_perfect` / `rho_112_111` 等退化目标，\n"
             "> 数值大幅为负属正常，不可与池化主指标直接比较。", ""]
    for k, title in [("main", "主对比"), ("next", "新一代架构（2024–2026）"),
                     ("v2", "PI-Mamba v2 改进"), ("loao", "跨合金（LOAO）"),
                     ("ablation", "消融"), ("dataeff", "数据效率")]:
        if k in out:
            lines += [f"## {title}", "", out[k], ""]
    if "per_target" in out:
        lines += [f"## 逐目标指标（{out['per_target_tag']}）", "", out["per_target"], ""]
    (ROOT / "results" / "RESULTS.md").write_text("\n".join(lines), encoding="utf-8")
    print("\n".join(lines))


if __name__ == "__main__":
    main()

	"""Generate the result tables / figures / RESULTS.md from ``results/``.

	Run after the experiment suites have finished::

	python -m pimamba.report
	"""
	from __future__ import annotations

	import json
	import os
	from pathlib import Path
	from typing import Dict, List

	import numpy as np
	import pandas as pd

	from . import analysis as A

	ROOT = Path(os.environ.get("PIMAMBA_ROOT")
	or Path(__file__).resolve().parents[2])
	RUNS = ROOT / "results" / "runs"
	TABLES = ROOT / "results" / "tables"
	FIGS = ROOT / "results" / "figures"

	MAIN_ORDER = ["mlp", "gru", "lstm", "transformer", "mamba", "mamba2", "pimamba"]
	COLS = ["sigma_r2", "sigma_shape_r2", "sigma_level_corr", "macro_r2_primary",
	"rho_total_r2", "twin_r2", "xi_FCC_r2", "peak_rel_err"]


	def _pm(df: pd.DataFrame, key: str, col: str, nd: int = 3) -> str:
	g = df[df[key[0]] == key[1]][col]
	if not len(g):
	return "n/a"
	return A.fmt_pm(float(g.mean()), float(g.std()) if len(g) > 1 else float("nan"), nd)


	def _maybe_csv(path: Path) -> "pd.DataFrame \| None":
	if path.exists():
	return pd.read_csv(path)
	return None


	def table_main(df: pd.DataFrame) -> pd.DataFrame:
	"""Per-fold mean+-sd and pooled out-of-fold metrics side by side."""
	sub = df[df.suite == "main"].copy()
	ps = _maybe_csv(TABLES / "tab_pooled_sigma.csv")
	pt = _maybe_csv(TABLES / "tab_pooled_per_target.csv")
	if ps is not None:
	ps = ps.set_index("model")
	if pt is not None:
	pt = pt.set_index("model")
	rows = []
	for m in MAIN_ORDER:
	g = sub[sub.model == m]
	if not len(g):
	continue
	r = {"model": A.MODEL_LABEL[m], "n": len(g)}
	r["应力 R²(逐折)"] = A.fmt_pm(float(g.sigma_r2.mean()), float(g.sigma_r2.std()))
	r["形状 R²(逐折)"] = A.fmt_pm(float(g.sigma_shape_r2.mean()),
	float(g.sigma_shape_r2.std()))
	if ps is not None and m in ps.index:
	r["池化应力 R²"] = f"{float(ps.loc[m, 'sigma_r2']):.3f}"
	r["池化形状 R²"] = f"{float(ps.loc[m, 'shape_r2']):.3f}"
	if pt is not None and m in pt.index:
	r["池化主指标 R²"] = f"{float(pt.loc[m, 'primary_mean']):.3f}"
	r["水平相关"] = A.fmt_pm(float(g.sigma_level_corr.mean()),
	float(g.sigma_level_corr.std()))
	# rare targets (twin: 6/33 labelled) have a meaningless per-fold R2 --
	# a fold can contain a single labelled sample. Report pooled OOF.
	if pt is not None and m in pt.index:
	r["池化总位错 R²"] = f"{float(pt.loc[m, 'rho_total']):.3f}"
	r["池化孪晶 R²"] = f"{float(pt.loc[m, 'twin']):.3f}"
	r["池化FCC R²"] = f"{float(pt.loc[m, 'xi_FCC']):.3f}"
	else:
	r["总位错 R²"] = A.fmt_pm(float(g.rho_total_r2.mean()),
	float(g.rho_total_r2.std()))
	r["峰值相对误差"] = f"{g.peak_rel_err.mean():.3f}"
	r["曲线 RMSE/GPa"] = A.fmt_pm(float(g.curve_rmse.mean()), float(g.curve_rmse.std()))
	r["逐折宏平均 R²"] = A.fmt_pm(float(g.macro_r2_primary.mean()),
	float(g.macro_r2_primary.std()))
	r["参数量"] = f"{int(g.n_params.mean()):,}"
	rows.append(r)
	return pd.DataFrame(rows)


	NEXT_ORDER = ["llama", "xlstm", "gdn", "hymba", "koopman", "node", "rssm",
	"flow", "pimamba2"]


	def table_next(df: pd.DataFrame) -> pd.DataFrame:
	"""Next-generation backbones: per-fold + pooled OOF, sorted by R2."""
	sub = df[df.suite == "next"].copy()
	if not len(sub):
	return pd.DataFrame()
	ps = _maybe_csv(TABLES / "tab_pooled_next_sigma.csv")
	pt = _maybe_csv(TABLES / "tab_pooled_next_per_target.csv")
	if ps is not None:
	ps = ps.set_index("model")
	if pt is not None:
	pt = pt.set_index("model")
	rows = []
	order = sub.groupby("model").sigma_r2.mean().sort_values(ascending=False).index
	for m in order:
	g = sub[sub.model == m]
	r = {"模型": A.MODEL_LABEL.get(m, m), "n": len(g),
	"应力 R²(逐折)": A.fmt_pm(float(g.sigma_r2.mean()),
	float(g.sigma_r2.std())),
	"形状 R²(逐折)": A.fmt_pm(float(g.sigma_shape_r2.mean()),
	float(g.sigma_shape_r2.std()))}
	if ps is not None and m in ps.index:
	r["池化应力"] = f"{float(ps.loc[m, 'sigma_r2']):.3f}"
	r["池化形状"] = f"{float(ps.loc[m, 'shape_r2']):.3f}"
	if pt is not None and m in pt.index:
	r["池化主指标"] = f"{float(pt.loc[m, 'primary_mean']):.3f}"
	r["峰值误差"] = f"{g.peak_rel_err.mean():.3f}"
	r["参数量"] = f"{int(g.n_params.mean()):,}"
	r["耗时/run(s)"] = f"{g.wall_time_s.mean():.0f}"
	rows.append(r)
	return pd.DataFrame(rows)


	def table_v2(df: pd.DataFrame) -> pd.DataFrame:
	"""PI-Mamba v2 variants (Taylor coupling + peak loss)."""
	sub = df[df.suite == "v2"].copy()
	if not len(sub):
	return pd.DataFrame()
	ps = _maybe_csv(TABLES / "tab_pooled_v2_sigma.csv")
	pt = _maybe_csv(TABLES / "tab_pooled_v2_per_target.csv")
	if ps is not None:
	ps = ps.set_index("model")
	if pt is not None:
	pt = pt.set_index("model")
	rows = []
	# reference = the main-suite PI-Mamba, so every variant is comparable to the
	# baseline it is meant to improve (the v2 suite has no "pimamba" row).
	bm = df[(df.suite == "main") & (df.model == "pimamba")].sigma_r2
	base = float(bm.mean()) if len(bm) else float(sub.sigma_r2.max())
	order = sub.groupby("variant").sigma_r2.mean().sort_values(ascending=False).index
	for v in order:
	g = sub[sub.variant == v]
	r = {"变体": v, "n": len(g),
	"应力 R²(逐折)": A.fmt_pm(float(g.sigma_r2.mean()),
	float(g.sigma_r2.std())),
	"形状 R²(逐折)": f"{g.sigma_shape_r2.mean():.3f}"}
	if ps is not None and v in ps.index:
	r["池化应力"] = f"{float(ps.loc[v, 'sigma_r2']):.3f}"
	r["池化形状"] = f"{float(ps.loc[v, 'shape_r2']):.3f}"
	if pt is not None and v in pt.index:
	r["池化主指标"] = f"{float(pt.loc[v, 'primary_mean']):.3f}"
	r["峰值误差"] = f"{g.peak_rel_err.mean():.3f}"
	r["ΔR²(vs PI-Mamba)"] = f"{g.sigma_r2.mean() - base:+.3f}"
	rows.append(r)
	return pd.DataFrame(rows)


	def table_loao(df: pd.DataFrame) -> pd.DataFrame:
	sub = df[df.suite == "loao"].copy()
	if not len(sub):
	return pd.DataFrame()
	rows = []
	for al, g in sub.groupby("holdout"):
	r = {"留出合金体系": al, "n": len(g)}
	for m in ["mlp", "gru", "mamba", "pimamba"]:
	gg = g[g.model == m]
	r[A.MODEL_LABEL[m]] = (f"{gg.sigma_r2.mean():.3f}"
	if len(gg) else "n/a")
	rows.append(r)
	rows.append({"留出合金体系": "平均", "n": len(sub),
	**{A.MODEL_LABEL[m]: f"{sub[sub.model == m].sigma_r2.mean():.3f}"
	for m in ["mlp", "gru", "mamba", "pimamba"]}})
	return pd.DataFrame(rows)


	def table_ablation(df: pd.DataFrame) -> pd.DataFrame:
	sub = df[df.suite == "ablation"].copy()
	if not len(sub):
	return pd.DataFrame()
	rows = []
	base = sub[sub.variant == "pimamba"].sigma_r2.mean()
	order = sub.groupby("variant").sigma_r2.mean().sort_values(ascending=False).index
	for v in order:
	g = sub[sub.variant == v]
	rows.append({
	"变体": A.MODEL_LABEL.get(v, v), "n": len(g),
	"应力 R²": A.fmt_pm(float(g.sigma_r2.mean()), float(g.sigma_r2.std())),
	"形状 R²": A.fmt_pm(float(g.sigma_shape_r2.mean()),
	float(g.sigma_shape_r2.std())),
	"总位错 R²": f"{g.rho_total_r2.mean():.3f}",
	"ΔR²(应力)": f"{g.sigma_r2.mean() - base:+.3f}",
	})
	return pd.DataFrame(rows)


	def table_dataeff(df: pd.DataFrame) -> pd.DataFrame:
	sub = df[df.suite == "dataeff"].copy()
	if not len(sub):
	return pd.DataFrame()
	def _frac(v):
	f = float(str(v).split("@")[-1])
	return f / 100.0 if f > 1.0 else f

	sub["frac"] = sub.variant.map(_frac)
	rows = []
	for m, g in sub.groupby("model"):
	r = {"模型": A.MODEL_LABEL.get(m, m)}
	for fr, gg in sorted(g.groupby("frac")):
	r[f"{int(fr*100)}%"] = f"{gg.sigma_r2.mean():.3f}"
	r["n_train@100%"] = int(g[g.frac == 1.0].n_train.mean()) if len(g[g.frac == 1.0]) else np.nan
	rows.append(r)
	return pd.DataFrame(rows)


	def table_per_target(tag: str) -> pd.DataFrame:
	m = json.loads((RUNS / tag / "metrics.json").read_text())
	rows = []
	for k, v in m["per_target"].items():
	if v["n"] == 0:
	continue
	rows.append({"目标": k, "组": v["group"], "n": v["n"],
	"R²": round(v["r2"], 3) if np.isfinite(v["r2"]) else "nan",
	"R²(log)": round(v["r2_log"], 3) if np.isfinite(v.get("r2_log", np.nan)) else "–",
	"nMAE": round(v["nmae"], 3) if np.isfinite(v["nmae"]) else "nan",
	"σ_y": f"{v['y_std']:.3g}", "单位": v["unit"]})
	return pd.DataFrame(rows)


	def rebuild_summary():
	"""Reconstruct ``results/summary.csv`` from the individual run folders.

	Safety net: the orchestrator only flushes the summary at the end of a suite,
	so a crash would otherwise lose the table.
	"""
	from .run_experiments import _row
	rows = []
	for d in sorted(RUNS.iterdir()):
	mp = d / "metrics.json"
	if not mp.exists():
	continue
	res = json.loads(mp.read_text())
	res["tag"] = d.name
	n = d.name
	if n.startswith("main_"):
	suite, var = "main", res["config"]["model"]
	elif n.startswith("abl_"):
	suite, var = "ablation", "_".join(n.split("_")[1:-2])
	elif n.startswith("loao_"):
	suite, var = "loao", res["config"]["model"]
	elif n.startswith("de_"):
	frac = int(n.split("_")[2]) / 100.0
	suite, var = "dataeff", f'{res["config"]["model"]}@{frac:.2f}'
	elif n.startswith("next_"):
	suite, var = "next", res["config"]["model"]
	elif n.startswith("v2_"):
	suite, var = "v2", "_".join(n.split("_")[1:-2])
	else:
	continue
	try:
	rows.append(_row(res, suite, var, res["config"].get("fold", -1),
	res["config"].get("seed", 0)))
	except Exception as e: # noqa: BLE001
	print("skip", n, e)
	if rows:
	df = pd.DataFrame(rows).drop_duplicates(subset=["tag"], keep="last")
	(ROOT / "results" / "summary.csv").parent.mkdir(parents=True, exist_ok=True)
	df.to_csv(ROOT / "results" / "summary.csv", index=False)
	print(f"rebuilt summary.csv: {len(df)} rows")


	def main():
	import sys
	if "--rebuild" in sys.argv:
	rebuild_summary()
	TABLES.mkdir(parents=True, exist_ok=True)
	FIGS.mkdir(parents=True, exist_ok=True)
	df = A.load_summary()
	out: Dict[str, str] = {}

	tm = table_main(df)
	if len(tm):
	tm.to_csv(TABLES / "tab_main.csv", index=False)
	out["main"] = A.markdown_table(tm, list(tm.columns))
	A.fig_model_comparison(df)

	tn = table_next(df)
	if len(tn):
	tn.to_csv(TABLES / "tab_next.csv", index=False)
	out["next"] = A.markdown_table(tn, list(tn.columns))
	A.fig_next(df)

	tv = table_v2(df)
	if len(tv):
	tv.to_csv(TABLES / "tab_v2.csv", index=False)
	out["v2"] = A.markdown_table(tv, list(tv.columns))

	tl = table_loao(df)
	if len(tl):
	tl.to_csv(TABLES / "tab_loao.csv", index=False)
	out["loao"] = A.markdown_table(tl, list(tl.columns))

	ta = table_ablation(df)
	if len(ta):
	ta.to_csv(TABLES / "tab_ablation.csv", index=False)
	out["ablation"] = A.markdown_table(ta, list(ta.columns))
	A.fig_ablation(df)

	td = table_dataeff(df)
	if len(td):
	td.to_csv(TABLES / "tab_dataeff.csv", index=False)
	out["dataeff"] = A.markdown_table(td, list(td.columns))
	A.fig_dataeff(df)

	best = df[(df.suite == "main") & (df.model == "pimamba")].sort_values(
	"sigma_r2", ascending=False)
	if len(best):
	tag = str(best.iloc[0]["tag"])
	if (RUNS / tag / "predictions.npz").exists():
	A.fig_curves(tag)
	out["curve_tag"] = tag
	tp = table_per_target(tag)
	tp.to_csv(TABLES / "tab_per_target.csv", index=False)
	out["per_target"] = A.markdown_table(tp, list(tp.columns))
	out["per_target_tag"] = tag

	lines = ["# 自动生成的结果汇总", "",
	"> 口径说明：`(逐折)` 为 5 折 × 2 种子的 mean ± sd；`池化` 为把 5 折测试预测\n"
	"> 拼成 33 个样本后统一计算的 out-of-fold 指标（更稳定，报告 §6 主用）。\n"
	"> `逐折宏平均 R²` 仍包含 `rho_perfect` / `rho_112_111` 等退化目标，\n"
	"> 数值大幅为负属正常，不可与池化主指标直接比较。", ""]
	for k, title in [("main", "主对比"), ("next", "新一代架构（2024–2026）"),
	("v2", "PI-Mamba v2 改进"), ("loao", "跨合金（LOAO）"),
	("ablation", "消融"), ("dataeff", "数据效率")]:
	if k in out:
	lines += [f"## {title}", "", out[k], ""]
	if "per_target" in out:
	lines += [f"## 逐目标指标（{out['per_target_tag']}）", "", out["per_target"], ""]
	(ROOT / "results" / "RESULTS.md").write_text("\n".join(lines), encoding="utf-8")
	print("\n".join(lines))


	if __name__ == "__main__":
	main()