nmWTAI-Platform/ML/nmWTAI-ML/scripts/replay_pso_trace_screening.py

"""回放 PSO 轨迹并评估代理模型筛选效果。

脚本读取已有 PSO trace 中的候选参数和真实求解器目标值，重新用正演代理模型打分，
比较两套目标函数的相关性与 Top-K/保留比例效果，帮助判断代理是否能减少真实求解器
调用量而不明显损失最优候选。
"""

from __future__ import annotations

import argparse
import csv
import json
import math
import sys
from collections import defaultdict
from pathlib import Path

ROOT = Path(__file__).resolve().parents[1]
sys.path.append(str(ROOT))

import joblib
import matplotlib.pyplot as plt
import numpy as np

from scripts.compare_single_case import build_schedule_vector, load_model
from scripts.validate_autofit_local_ranking import infer_curve_layout, predict_surrogate_curve
from src.common.config import Config
from src.common.experiment_paths import config_for_stage, model_checkpoint_for_tag, normalize_tag, processed_path_for_tag
from src.data.curve_processing import clean_curve_for_dataset, is_valid_curve, resample_curve_to_features
from src.data.params import Params, Schedule
from src.evaluation.autofit_objective import dual_log_objective


PARAM_COLUMNS = ["k", "skin", "wellboreC", "phi", "h", "Cf"]
DEFAULT_KEEP_FRACS = [0.5, 0.6, 0.7]


def parse_args() -> argparse.Namespace:
    """解析 PSO trace、代理评分表和 keep_ratio，用于回放筛选策略。"""
    parser = argparse.ArgumentParser(
        description="Replay a C++ full-solver PSO trace with the forward surrogate as a screening model."
    )
    parser.add_argument("--trace-csv", type=str, default=None, help="Path to pso_baseline_trace_*.csv")
    parser.add_argument("--trace-meta", type=str, default=None, help="Path to matching pso_baseline_trace_*.meta.json")
    parser.add_argument("--tag", type=str, default="family_random_mixed_50k_logparam")
    parser.add_argument("--stage", type=str, default="family_random")
    parser.add_argument("--processed", type=str, default=None)
    parser.add_argument("--model", type=str, default=None)
    parser.add_argument("--config", type=str, default=None)
    parser.add_argument("--output-dir", type=str, default=None)
    parser.add_argument("--keep-fracs", type=str, default="0.5,0.6,0.7")
    return parser.parse_args()


def latest_trace_pair(temp_dir: Path) -> tuple[Path, Path]:
    """在 PSO 结果目录中找到最新的轨迹文件和候选文件组合。"""
    csv_paths = sorted(temp_dir.glob("pso_baseline_trace_*.csv"), key=lambda p: p.stat().st_mtime, reverse=True)
    for csv_path in csv_paths:
        meta_path = csv_path.with_suffix(".meta.json")
        if meta_path.exists():
            return csv_path, meta_path
    raise FileNotFoundError(f"No pso_baseline_trace_*.csv + .meta.json pair found in {temp_dir}")


def resolve_paths(args: argparse.Namespace) -> tuple[Path, Path, Path, Path, Path, Path]:
    """解析 PSO 回放所需的轨迹文件、候选文件、配置文件和输出目录。"""
    tag = normalize_tag(args.tag)
    if args.trace_csv is None:
        trace_csv, trace_meta = latest_trace_pair(ROOT / "data" / "temp")
    else:
        trace_csv = Path(args.trace_csv)
        trace_meta = Path(args.trace_meta) if args.trace_meta is not None else trace_csv.with_suffix(".meta.json")

    config_path = Path(args.config) if args.config is not None else config_for_stage(args.stage)
    if config_path is None:
        raise ValueError(f"Cannot resolve config for stage={args.stage!r}; pass --config explicitly")

    processed_path = Path(args.processed) if args.processed is not None else processed_path_for_tag(tag)
    model_path = Path(args.model) if args.model is not None else model_checkpoint_for_tag(tag, use_schedule=True)

    run_id = trace_csv.stem.replace("pso_baseline_trace_", "")
    output_dir = Path(args.output_dir) if args.output_dir is not None else Path("results") / f"pso_trace_replay_{run_id}"

    return (
        trace_csv.resolve(),
        trace_meta.resolve(),
        config_path.resolve(),
        processed_path.resolve(),
        model_path.resolve(),
        output_dir.resolve(),
    )


def read_trace_csv(path: Path) -> list[dict]:
    """读取 PSO 轨迹 CSV，并保留每代候选的代理分数和真实求解器分数。"""
    with path.open("r", newline="", encoding="utf-8-sig") as f:
        return list(csv.DictReader(f))


def to_float(value: str | None, default: float = float("nan")) -> float:
    """将表格字段安全转为 float；无法解析时返回 NaN。"""
    if value is None or value == "":
        return default
    try:
        return float(value)
    except ValueError:
        return default


def is_valid_solver_row(row: dict) -> bool:
    """判断轨迹行是否包含可用的真实求解器目标值。"""
    obj = to_float(row.get("solver_objective"))
    return row.get("phase") == "particle_solver" and row.get("solver_success") == "1" and math.isfinite(obj) and obj < 1e9


def build_schedule_from_meta(meta: dict) -> Schedule:
    """从 CSV 元数据字段还原 Schedule 对象。"""
    target = meta["target"]
    flow_points = target.get("flow_points") or []
    time_q: list[float] = []
    q: list[float] = []
    for item in flow_points:
        dt = float(item["x"])
        rate = float(item["y"])
        if dt <= 0 and not time_q:
            continue
        time_q.append(dt)
        q.append(rate)
    return Schedule(sectionIndex=int(target["section_index"]), timeQ=time_q, q=q)


def build_target_curve_from_meta(cfg: Config, processed: dict, meta: dict) -> np.ndarray:
    """从 CSV 元数据字段还原目标双对数曲线向量。"""
    target = meta["target"]["target_loglog"]
    t = np.asarray(target["time"], dtype=np.float64)
    p = np.asarray(target["pressure"], dtype=np.float64)
    d = np.asarray(target["derivative"], dtype=np.float64)
    cleaned = clean_curve_for_dataset(cfg, t, p, d)
    if cleaned is None:
        raise RuntimeError("Target loglog curve from meta cannot be cleaned for replay")
    t_clean, p_clean, d_clean = cleaned
    valid, reason = is_valid_curve(cfg, t_clean, p_clean, d_clean)
    if not valid:
        raise RuntimeError(f"Target loglog curve from meta is invalid: {reason}")

    curve = resample_curve_to_features(cfg, t_clean, p_clean, d_clean)
    curve_dim = int(processed["meta"]["curve_dim"])
    if curve.size != curve_dim:
        raise RuntimeError(f"Target curve dim mismatch after resample: {curve.size} != {curve_dim}")
    return curve.astype(np.float32)


def params_from_row(row: dict) -> Params:
    """从 CSV 行读取物理参数字段并构造 Params 对象。"""
    values = {name: to_float(row.get(name)) for name in PARAM_COLUMNS}
    return Params(
        k=values["k"],
        skin=values["skin"],
        wellboreC=values["wellboreC"],
        phi=values["phi"],
        h=values["h"],
        Cf=values["Cf"],
    )


def rank_positions(values: np.ndarray) -> np.ndarray:
    """把目标函数值转换成排名位置，用于比较代理排序和真实排序。"""
    order = np.argsort(values)
    ranks = np.empty_like(order)
    ranks[order] = np.arange(values.size)
    return ranks


def corr_pearson(a: np.ndarray, b: np.ndarray) -> float:
    """计算 Pearson 相关系数，用于评估代理分数与真实分数的线性相关。"""
    if a.size < 2:
        return float("nan")
    if np.std(a) <= 1e-12 or np.std(b) <= 1e-12:
        return float("nan")
    return float(np.corrcoef(a, b)[0, 1])


def corr_spearman(a: np.ndarray, b: np.ndarray) -> float:
    """计算 Spearman 秩相关，用于评估候选排序是否一致。"""
    if a.size < 2:
        return float("nan")
    return corr_pearson(rank_positions(a).astype(np.float64), rank_positions(b).astype(np.float64))


def summarize_generation(rows: list[dict], keep_fracs: list[float]) -> tuple[dict, list[dict]]:
    """按 PSO 代数汇总代理筛选保留真实优质候选的效果。"""
    valid = [r for r in rows if r["solver_success"] == "1" and math.isfinite(float(r["solver_objective"])) and float(r["solver_objective"]) < 1e9]
    if len(valid) < 2:
        return {}, []

    solver_obj = np.asarray([float(r["solver_objective"]) for r in valid], dtype=np.float64)
    surrogate_obj = np.asarray([float(r["surrogate_objective"]) for r in valid], dtype=np.float64)
    solver_rank = rank_positions(solver_obj)
    surrogate_rank = rank_positions(surrogate_obj)

    # 给每个粒子同时标注真实排名和代理排名，便于后续定位“代理看好但真实较差”的粒子。
    for i, row in enumerate(valid):
        row["solver_rank"] = int(solver_rank[i])
        row["surrogate_rank"] = int(surrogate_rank[i])
        row["rank_gap"] = int(surrogate_rank[i] - solver_rank[i])

    best_solver_idx = int(np.argmin(solver_obj))
    best_surrogate_idx = int(np.argmin(surrogate_obj))
    top_n = min(10, len(valid))
    top_solver = set(np.argsort(solver_obj)[:top_n].tolist())
    top_surrogate = set(np.argsort(surrogate_obj)[:top_n].tolist())

    summary = {
        "generation": int(valid[0]["generation"]),
        "n_particles": len(rows),
        "n_valid_solver": len(valid),
        "pearson_objective": corr_pearson(solver_obj, surrogate_obj),
        "spearman_objective": corr_spearman(solver_obj, surrogate_obj),
        "top10_overlap": int(len(top_solver & top_surrogate)),
        "best_solver_particle_id": int(valid[best_solver_idx]["particle_id"]),
        "best_solver_surrogate_rank": int(surrogate_rank[best_solver_idx]),
        "best_surrogate_particle_id": int(valid[best_surrogate_idx]["particle_id"]),
        "best_surrogate_solver_rank": int(solver_rank[best_surrogate_idx]),
        "solver_best_objective": float(solver_obj[best_solver_idx]),
        "solver_objective_at_surrogate_best": float(solver_obj[best_surrogate_idx]),
        "surrogate_top1_regret": float(solver_obj[best_surrogate_idx] - solver_obj[best_solver_idx]),
    }

    screening_rows: list[dict] = []
    all_surrogate = np.asarray([float(r["surrogate_objective"]) for r in rows], dtype=np.float64)
    valid_by_particle = {int(r["particle_id"]): r for r in valid}
    for keep_frac in keep_fracs:
        keep_n = max(1, int(math.ceil(len(rows) * float(keep_frac))))
        kept_all_positions = np.argsort(all_surrogate)[:keep_n]
        kept_particle_ids = {int(rows[i]["particle_id"]) for i in kept_all_positions}
        kept_valid_indices = [i for i, r in enumerate(valid) if int(r["particle_id"]) in kept_particle_ids]
        kept_solver_obj = solver_obj[kept_valid_indices] if kept_valid_indices else np.asarray([], dtype=np.float64)

        # 模拟“先用代理筛选，再让真实求解器评估保留下来的粒子”的效果。
        best_missed = int(int(valid[best_solver_idx]["particle_id"]) not in kept_particle_ids)
        top_recall = float(len({int(valid[i]["particle_id"]) for i in top_solver} & kept_particle_ids) / max(top_n, 1))
        if kept_solver_obj.size:
            screening_regret = float(np.min(kept_solver_obj) - solver_obj[best_solver_idx])
        else:
            screening_regret = float("inf")

        screening_rows.append(
            {
                "generation": summary["generation"],
                "keep_frac": float(keep_frac),
                "keep_n": int(keep_n),
                "solver_top10_recall": top_recall,
                "best_missed": best_missed,
                "screening_regret": screening_regret,
                "kept_valid_solver": int(len(kept_valid_indices)),
                "kept_failed_or_invalid": int(sum(1 for i in kept_all_positions if int(rows[i]["particle_id"]) not in valid_by_particle)),
            }
        )

    return summary, screening_rows


def write_csv(path: Path, rows: list[dict]) -> None:
    """按字段名写出 CSV 明细或汇总结果。"""
    if not rows:
        return
    fieldnames: list[str] = []
    for row in rows:
        for key in row.keys():
            if key not in fieldnames:
                fieldnames.append(key)
    with path.open("w", newline="", encoding="utf-8-sig") as f:
        writer = csv.DictWriter(f, fieldnames=fieldnames)
        writer.writeheader()
        writer.writerows(rows)


def plot_objective_scatter(rows: list[dict], path: Path) -> None:
    """绘制代理目标与真实目标的散点图，直观看排序筛选偏差。"""
    valid = [r for r in rows if r.get("solver_success") == "1" and float(r.get("solver_objective", "inf")) < 1e9]
    if len(valid) < 2:
        return
    solver = np.asarray([float(r["solver_objective"]) for r in valid], dtype=np.float64)
    surrogate = np.asarray([float(r["surrogate_objective"]) for r in valid], dtype=np.float64)
    gen = np.asarray([int(r["generation"]) for r in valid], dtype=np.int32)

    plt.figure(figsize=(6, 5), dpi=150)
    sc = plt.scatter(solver, surrogate, c=gen, cmap="viridis", alpha=0.8)
    lo = min(float(np.min(solver)), float(np.min(surrogate)))
    hi = max(float(np.max(solver)), float(np.max(surrogate)))
    plt.plot([lo, hi], [lo, hi], "k--", linewidth=1)
    plt.xlabel("C++ solver objective")
    plt.ylabel("Surrogate objective to target")
    plt.title("PSO trace replay objective ranking")
    plt.colorbar(sc, label="generation")
    plt.tight_layout()
    plt.savefig(path)
    plt.close()


def main() -> None:
    """在既有 PSO 轨迹上离线模拟代理筛选，评估真实最优粒子是否被保留。"""
    args = parse_args()
    trace_csv, trace_meta, config_path, processed_path, model_path, output_dir = resolve_paths(args)
    keep_fracs = [float(x) for x in args.keep_fracs.split(",") if x.strip()]

    cfg = Config(config_path)
    processed = joblib.load(processed_path)
    model, use_schedule, device = load_model(model_path)
    curve_layout = infer_curve_layout(processed["meta"], int(processed["meta"]["curve_dim"]))

    meta = json.loads(trace_meta.read_text(encoding="utf-8"))
    rows = read_trace_csv(trace_csv)
    schedule = build_schedule_from_meta(meta)
    target_curve = build_target_curve_from_meta(cfg, processed, meta)

    output_dir.mkdir(parents=True, exist_ok=True)

    replay_rows: list[dict] = []
    for row in rows:
        if row.get("phase") != "particle_solver":
            continue
        # 回放时不重跑真实求解器，只用保存的粒子参数重新计算代理目标，和 trace 中真实目标对比。
        params = params_from_row(row)
        params.schedule = schedule
        pred_curve = predict_surrogate_curve(
            processed=processed,
            model=model,
            device=device,
            use_schedule=use_schedule,
            params=params,
            schedule=schedule,
            cfg=cfg,
        )
        sur_obj = dual_log_objective(target_curve, pred_curve, curve_layout)
        replay_row = dict(row)
        replay_row["surrogate_objective"] = sur_obj["dual_log_objective"]
        replay_row["surrogate_p_obj"] = sur_obj["log_pressure_objective"]
        replay_row["surrogate_d_obj"] = sur_obj["log_derivative_objective"]
        replay_rows.append(replay_row)

    by_generation: dict[int, list[dict]] = defaultdict(list)
    for row in replay_rows:
        by_generation[int(row["generation"])].append(row)

    # 按 PSO 代分组统计，因为早期和后期粒子分布差异很大，不能只看全局平均。
    generation_summaries: list[dict] = []
    screening_rows: list[dict] = []
    for generation in sorted(by_generation):
        summary, gen_screening_rows = summarize_generation(by_generation[generation], keep_fracs)
        if summary:
            generation_summaries.append(summary)
            screening_rows.extend(gen_screening_rows)

    write_csv(output_dir / "candidate_objectives.csv", replay_rows)
    write_csv(output_dir / "generation_summary.csv", generation_summaries)
    write_csv(output_dir / "screening_summary.csv", screening_rows)
    plot_objective_scatter(replay_rows, output_dir / "objective_scatter.png")

    summary = {
        "trace_csv": str(trace_csv),
        "trace_meta": str(trace_meta),
        "processed_path": str(processed_path),
        "model_path": str(model_path),
        "run_id": meta.get("run_id"),
        "n_particle_rows": len(replay_rows),
        "n_generations": len(generation_summaries),
        "target_points_raw": len(meta["target"]["target_loglog"]["time"]),
        "target_curve_dim_resampled": int(target_curve.size),
        "use_schedule": bool(use_schedule),
        "keep_fracs": keep_fracs,
        "generation_summary_mean": {},
        "screening_mean": {},
    }
    if generation_summaries:
        # generation_summary_mean 给出跨代平均排序质量，用来和不同代理模型版本横向比较。
        for key in [
            "pearson_objective",
            "spearman_objective",
            "top10_overlap",
            "best_solver_surrogate_rank",
            "best_surrogate_solver_rank",
            "surrogate_top1_regret",
        ]:
            values = np.asarray([float(r[key]) for r in generation_summaries], dtype=np.float64)
            summary["generation_summary_mean"][key] = float(np.nanmean(values))

    for keep_frac in keep_fracs:
        rows_k = [r for r in screening_rows if abs(float(r["keep_frac"]) - keep_frac) < 1e-12]
        if not rows_k:
            continue
        # screening_mean 直接回答：保留某个比例时，真实好粒子是否被代理筛掉。
        summary["screening_mean"][str(keep_frac)] = {
            "solver_top10_recall": float(np.mean([float(r["solver_top10_recall"]) for r in rows_k])),
            "best_missed_ratio": float(np.mean([float(r["best_missed"]) for r in rows_k])),
            "screening_regret": float(np.mean([float(r["screening_regret"]) for r in rows_k])),
            "kept_failed_or_invalid": float(np.mean([float(r["kept_failed_or_invalid"]) for r in rows_k])),
        }

    (output_dir / "summary.json").write_text(json.dumps(summary, indent=2, ensure_ascii=False), encoding="utf-8")

    print(f"Trace replay output: {output_dir}")
    print(f"particle_rows={summary['n_particle_rows']}, generations={summary['n_generations']}")
    if summary["generation_summary_mean"]:
        print(
            "mean Spearman="
            f"{summary['generation_summary_mean']['spearman_objective']:.4f}, "
            f"mean best_solver_surrogate_rank={summary['generation_summary_mean']['best_solver_surrogate_rank']:.2f}"
        )
    for keep_frac, metrics in summary["screening_mean"].items():
        print(
            f"keep={keep_frac}: recall={metrics['solver_top10_recall']:.4f}, "
            f"best_missed={metrics['best_missed_ratio']:.4f}, "
            f"regret={metrics['screening_regret']:.6f}"
        )


if __name__ == "__main__":
    main()