examples/example_classification.py

"""
Example 2 — Classification: Keyhole vs Conduction Mode

Binary classifier predicting melting regime from four process parameters.

Labels are derived automatically from melt-pool depth (zmax − zmin) in
monitor/position-bounds_melt.dat: experiments whose steady-state depth
exceeds the dataset median are labeled Keyhole (1), the rest Conduction (0).
Rows with sentinel values (|val| > 1e30) are tagged "Initial Emptiness" and
excluded from depth computation.  The dataset is then balanced via random
undersampling of the majority class before any ML step.

Three models evaluated via leave-one-out cross-validation:
  1. Logistic Regression
  2. Random Forest
  3. SVM (RBF kernel)

Set N_SUBSET to a small number (e.g. 30) so a reviewer can run this quickly.
Set N_SUBSET = None to use the full balanced dataset.

Outputs saved to runs/classification_<timestamp>/:
  classification_diagnostics.png   — LOO confusion matrices + feature relevance
  run.log                          — full training log

This is a proof-of-concept, not a benchmark.
"""

import logging
import random
import sys
from datetime import datetime
from pathlib import Path

import numpy as np
from sklearn.base import clone
from sklearn.ensemble import RandomForestClassifier
from sklearn.inspection import permutation_importance
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import ConfusionMatrixDisplay
from sklearn.model_selection import LeaveOneOut
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.svm import SVC
import matplotlib.pyplot as plt

# ------------------------------------------------------------------
# Config  ← edit DATA_DIRS to point at your data directories
# ------------------------------------------------------------------

DATA_DIRS = [
    Path(__file__).parent.parent / "rnl" / "final_data_processed",
    Path(__file__).parent.parent / "rnl" / "lrz_data_new_format",
]
OUT_ROOT  = Path(__file__).parent.parent / "runs"

N_SUBSET    = 30      # reviewer-friendly subset size (None = full dataset)
N_STABLE    = 50      # last N valid timesteps for steady-state depth estimate
RANDOM_SEED = 42

INPUT_PARAMS = ["laser_power", "scan_speed", "laser_spot_size", "substrate_temp"]
MODEL_NAMES  = ["LogReg", "RandomForest", "SVM-RBF"]

# ------------------------------------------------------------------
# Logger
# ------------------------------------------------------------------

class _ColorFormatter(logging.Formatter):
    _COLORS = {logging.DEBUG: "\033[37m", logging.INFO: "\033[32m",
               logging.WARNING: "\033[33m", logging.ERROR: "\033[31m"}
    _RESET = "\033[0m"; _BOLD = "\033[1m"

    def format(self, record):
        color = self._COLORS.get(record.levelno, self._RESET)
        t = self.formatTime(record, "%H:%M:%S")
        return f"{self._BOLD}{t}{self._RESET}  {color}{record.levelname:<8}{self._RESET}  {record.getMessage()}"


run_id  = datetime.now().strftime("%Y%m%d_%H%M%S")
out_dir = OUT_ROOT / f"classification_{run_id}"
out_dir.mkdir(parents=True, exist_ok=True)

_log = logging.getLogger("clf")
_log.setLevel(logging.DEBUG)
_h = logging.StreamHandler(sys.stdout); _h.setFormatter(_ColorFormatter()); _log.addHandler(_h)
_f = logging.FileHandler(out_dir / "run.log")
_f.setFormatter(logging.Formatter("%(asctime)s  %(levelname)-8s  %(message)s", datefmt="%H:%M:%S"))
_log.addHandler(_f)

_log.info("=" * 60)
_log.info(f"Run ID  : {run_id}")
_log.info(f"Results : {out_dir}")
_log.info("=" * 60)

# ------------------------------------------------------------------
# 1. Load experiments and auto-label
# ------------------------------------------------------------------

def load_params(sim_dir: Path) -> dict | None:
    """Read process parameters from parameters.json."""
    pjson = sim_dir / "parameters.json"
    if not pjson.exists():
        return None
    try:
        raw = __import__("json").loads(pjson.read_text())
        return {
            "laser_power":           float(raw["laser_power"]["value"]),
            "scan_speed":            float(raw["scan_speed_x"]["value"]),
            "laser_spot_size":       float(raw["laser_spot_size"]["value"]),
            "substrate_temperature": float(raw["substrate_temperature"]["value"]),
        }
    except Exception:
        return None


def melt_depth(sim_dir: Path, n_stable: int) -> float | None:
    """Mean melt-pool depth (zmax − zmin) over the last n_stable valid rows.
    Rows where any value |v| > 1e30 are Initial Emptiness sentinels → excluded.
    """
    dat = sim_dir / "monitor" / "position-bounds_melt.dat"
    if not dat.exists():
        return None
    try:
        b = np.loadtxt(dat, delimiter=",")
        if b.ndim == 1:
            b = b.reshape(1, -1)
        b = b[~np.any(np.abs(b) > 1e30, axis=1)]   # drop Initial Emptiness rows
        if len(b) < 10:
            return None
        depth = (b[-n_stable:, 5] - b[-n_stable:, 4]).mean()   # zmax - zmin
        return depth if depth > 0 else None
    except Exception:
        return None


all_sims = []
for data_dir in DATA_DIRS:
    sims = sorted(data_dir.iterdir()) if data_dir.is_dir() else []
    _log.info(f"Scanning {data_dir} → {len(sims)} dirs")
    all_sims.extend(sims)
_log.info(f"Total simulation directories: {len(all_sims)}")

X_list, depth_list, names = [], [], []
skipped = 0
for sim_dir in all_sims:
    if not sim_dir.is_dir():
        continue
    params = load_params(sim_dir)
    if params is None:
        skipped += 1; continue
    d = melt_depth(sim_dir, N_STABLE)
    if d is None:
        skipped += 1; continue
    X_list.append(list(params.values()))
    depth_list.append(d)
    names.append(sim_dir.name)

_log.info(f"Valid experiments: {len(X_list)}  (skipped {skipped})")

X_all      = np.array(X_list)
depths_all = np.array(depth_list)
median_depth = np.median(depths_all)
y_all = (depths_all > median_depth).astype(int)

_log.info(f"Depth threshold (median): {median_depth*1e6:.2f} µm")
_log.info(f"Keyhole: {y_all.sum()}   Conduction: {(y_all==0).sum()}")

# ------------------------------------------------------------------
# 2. Balance classes (undersample majority) then optionally subset
# ------------------------------------------------------------------

rng = random.Random(RANDOM_SEED)

kh_idx = np.where(y_all == 1)[0].tolist()
cd_idx = np.where(y_all == 0)[0].tolist()
n_bal  = min(len(kh_idx), len(cd_idx))

rng.shuffle(kh_idx); rng.shuffle(cd_idx)
balanced_idx = sorted(kh_idx[:n_bal] + cd_idx[:n_bal])

X_bal = X_all[balanced_idx]
y_bal = y_all[balanced_idx]
_log.info(f"After balancing: {len(y_bal)} experiments  ({n_bal} Keyhole + {n_bal} Conduction)")

if N_SUBSET is not None and N_SUBSET < len(y_bal):
    # Stratified subsample: N_SUBSET/2 from each class
    n_each = N_SUBSET // 2
    kh_sub = [i for i in balanced_idx if y_all[i] == 1][:n_each]
    cd_sub = [i for i in balanced_idx if y_all[i] == 0][:n_each]
    sub_idx = sorted(kh_sub + cd_sub)
    X = X_all[sub_idx]
    y = y_all[sub_idx]
    _log.info(f"Reviewer subset: {len(y)} experiments  ({n_each} Keyhole + {n_each} Conduction)")
else:
    X, y = X_bal, y_bal
    _log.info("Using full balanced dataset")

# ------------------------------------------------------------------
# 3. Model definitions
# ------------------------------------------------------------------

def make_models():
    return [
        make_pipeline(StandardScaler(), LogisticRegression(max_iter=1000)),
        make_pipeline(StandardScaler(), RandomForestClassifier(n_estimators=200, random_state=RANDOM_SEED)),
        make_pipeline(StandardScaler(), SVC(kernel="rbf", probability=True, random_state=RANDOM_SEED)),
    ]

# ------------------------------------------------------------------
# 4. LOO cross-validation
# ------------------------------------------------------------------

splits  = list(LeaveOneOut().split(X))
n_folds = len(splits)
y_preds = {}

for name, base_pipe in zip(MODEL_NAMES, make_models()):
    preds   = np.empty(n_folds, dtype=int)
    correct = 0
    _log.info(f"[{name}] LOO CV ({n_folds} folds)")

    for fold, (train_idx, test_idx) in enumerate(splits):
        pipe = clone(base_pipe)
        pipe.fit(X[train_idx], y[train_idx])
        preds[test_idx] = pipe.predict(X[test_idx])
        correct += int(preds[test_idx[0]] == y[test_idx[0]])
        if (fold + 1) % 10 == 0 or fold == n_folds - 1:
            _log.info(f"  fold {fold+1:3d}/{n_folds}  running acc = {correct/(fold+1):.3f}")

    _log.info(f"[{name}] Final LOO accuracy: {(preds==y).mean():.3f}")
    y_preds[name] = preds

# ------------------------------------------------------------------
# 5. Fit on full subset for importance plots
# ------------------------------------------------------------------

_log.info("Fitting on full subset for feature importance ...")
fitted = {}
for name, pipe in zip(MODEL_NAMES, make_models()):
    pipe.fit(X, y)
    fitted[name] = pipe


def feature_importances(name, pipe):
    if name == "LogReg":
        return pipe.named_steps["logisticregression"].coef_[0]
    if name == "RandomForest":
        return pipe.named_steps["randomforestclassifier"].feature_importances_
    res = permutation_importance(pipe, X, y, n_repeats=30, random_state=0, scoring="accuracy")
    imp = res.importances_mean
    return imp / (np.abs(imp).max() or 1)

# ------------------------------------------------------------------
# 6. Plots — 2 rows × 3 columns
# ------------------------------------------------------------------

_log.info("Generating plots ...")
fig, axes = plt.subplots(2, 3, figsize=(13, 8))

for col, name in enumerate(MODEL_NAMES):
    acc = (y_preds[name] == y).mean()

    ConfusionMatrixDisplay.from_predictions(
        y, y_preds[name],
        display_labels=["Conduction", "Keyhole"],
        cmap="Blues", ax=axes[0, col], colorbar=False,
    )
    axes[0, col].set_title(f"{name}  (LOO acc={acc:.3f})")

    imp = feature_importances(name, fitted[name])
    colors = ["#e06c75" if v < 0 else "#61afef" for v in imp]
    axes[1, col].barh(INPUT_PARAMS, imp, color=colors)
    axes[1, col].axvline(0, color="k", lw=0.6)
    xlabel = ("Standardised coef. (+ → Keyhole)" if name == "LogReg" else
              "Mean decrease in impurity"          if name == "RandomForest" else
              "Permutation importance (norm.)")
    axes[1, col].set_xlabel(xlabel)
    axes[1, col].set_title(f"{name} — input relevance")

subset_note = f"N={len(y)} (reviewer subset)" if N_SUBSET else f"N={len(y)} (full balanced)"
plt.suptitle(
    f"LOO confusion matrices and input relevance — Keyhole vs Conduction  [{subset_note}]",
    y=1.01,
)
plt.tight_layout()

plot_path = out_dir / "classification_diagnostics.png"
plt.savefig(plot_path, dpi=150, bbox_inches="tight")
_log.info(f"Plot saved → {plot_path}")
plt.show()
_log.info("Done.")