"""
Hyperparameter search table (reviewer priority #6 supporting evidence).

The reviewer questions whether all 11 models received equally disciplined
hyperparameter optimization, or whether some were left at library defaults
while others (implicitly LightGBM) were tuned. train_classifier.py's
_build_candidate_families() already runs a per-family holdout search for
every ML model family (see _select_best_candidates), and
training_results.json records each model's selected_params and
validation_auc. This script turns that into a single audit table showing,
per model: how many candidates were tried, what the search space was, and
which configuration was selected — so the answer is verifiable rather than
asserted.

Usage:
    python -m app.training.hyperparameter_table
"""

from __future__ import annotations

import csv
import json
from pathlib import Path

MODELS_DIR = Path(__file__).resolve().parents[2] / "models"
TABLES_DIR = Path(__file__).resolve().parents[3] / "docs/academic/paper/real_tables"

# Number of hand-specified candidates per family, taken directly from
# _build_candidate_families() in train_classifier.py.
_N_CANDIDATES = {
    "Logistic Regression": 4,   # C in (0.25, 0.5, 1.0, 2.0)
    "Random Forest": 3,
    "Gradient Boosting": 3,
    "SVM (RBF)": 4,
    "MLP Neural Network": 3,
    "XGBoost": 3,
    "LightGBM": 3,
}

_SEARCH_SPACE_SUMMARY = {
    "Logistic Regression": "C in {0.25, 0.5, 1.0, 2.0}; class_weight=balanced fixed",
    "Random Forest": "n_estimators in {300,450,500}; max_depth in {12,18,None}; "
                      "max_features in {sqrt,sqrt,log2}",
    "Gradient Boosting": "n_estimators in {180,200,260}; max_depth in {2,3,4}; "
                          "learning_rate in {0.04,0.05,0.07}",
    "SVM (RBF)": "C in {1.0,3.0,6.0,10.0}; gamma in {scale,scale,0.02,0.05}",
    "MLP Neural Network": "hidden_layer_sizes in {(128,64),(192,96,32),(256,128)}; "
                           "alpha in {5e-4,1e-3,2e-3}",
    "XGBoost": "n_estimators in {240,300,500}; max_depth in {3,4,5}; "
               "learning_rate in {0.03,0.05,0.06}",
    "LightGBM": "n_estimators in {220,300,500}; max_depth in {-1,6,8}; "
                "num_leaves in {18,24,31}; learning_rate in {0.03,0.05,0.07}",
}

# DL models: architecture is fixed per model (no per-family candidate search
# like the ML side), but all four share identical training hyperparameters
# (optimizer, LR, epochs, early stopping) — recorded here for the same
# fairness audit.
_DL_SHARED_TRAINING = {
    "optimizer": "AdamW",
    "learning_rate": "1e-3",
    "weight_decay": "1e-4",
    "lr_scheduler": "ReduceLROnPlateau(mode=max, factor=0.5, patience=5)",
    "loss": "BCEWithLogitsLoss(pos_weight=n_neg/n_pos)",
    "epochs_max": "100",
    "early_stopping_patience": "10",
    "batch_size": "64",
    "seed": "42 (+fold index)",
}


def run() -> None:
    with open(MODELS_DIR / "training_results.json", "r", encoding="utf-8") as f:
        results = json.load(f)

    rows = []
    for name in _N_CANDIDATES:
        entry = results.get(name, {})
        rows.append({
            "model": name,
            "n_candidates_evaluated": _N_CANDIDATES[name],
            "selection_method": "stratified holdout (80/20), best validation AUC wins",
            "validation_auc": entry.get("validation_auc"),
            "selected_params": json.dumps(entry.get("selected_params", {}), ensure_ascii=False),
            "selection_time_sec": entry.get("selection_time_sec"),
            "search_space": _SEARCH_SPACE_SUMMARY[name],
        })

    TABLES_DIR.mkdir(parents=True, exist_ok=True)
    ml_path = TABLES_DIR / "hyperparameter_search_table.csv"
    with open(ml_path, "w", newline="", encoding="utf-8") as f:
        writer = csv.DictWriter(f, fieldnames=list(rows[0].keys()))
        writer.writeheader()
        writer.writerows(rows)

    print(f"ML hyperparameter search table written: {ml_path}")
    for r in rows:
        print(f"  {r['model']:22s} candidates={r['n_candidates_evaluated']} "
              f"val_auc={r['validation_auc']} time={r['selection_time_sec']}s")

    dl_path = TABLES_DIR / "dl_training_hyperparameters.csv"
    with open(dl_path, "w", newline="", encoding="utf-8") as f:
        writer = csv.writer(f)
        writer.writerow(["hyperparameter", "value"])
        writer.writerow(["note", "shared across all 4 DL architectures (Deep MLP, 1D-CNN, "
                                  "Residual MLP, Attention MLP) — architecture differs, "
                                  "training recipe does not"])
        for k, v in _DL_SHARED_TRAINING.items():
            writer.writerow([k, v])

    print(f"\nDL shared training hyperparameters written: {dl_path}")
    print(
        "\nNOTE: DL models do not go through a per-family candidate search "
        "like the ML models — each of the 4 DL architectures is trained "
        "once with an identical, fixed training recipe (see dl_training_"
        "hyperparameters.csv). This is architecturally different fairness "
        "(same recipe, different architecture) vs. the ML side (same "
        "architecture family, tuned hyperparameters) — both are internally "
        "consistent, but the paper should state this distinction explicitly "
        "rather than implying all 11 models went through identical tuning."
    )


if __name__ == "__main__":
    run()