engine

tit.opt.ex.engine

Single-class exhaustive search engine.

Replaces: runner.py (5 classes), roi_utils.py (4 functions). Direct SimNIBS usage — ROI metrics are computed inline.

ExSearchEngine

ExSearchEngine(leadfield_hdf: str, roi_file: str | tuple[str, int] | list[str | tuple[str, int]], roi_name: str, logger: Logger)

Exhaustive TI electrode search engine.

Owns the full pipeline: leadfield loading, ROI resolution and the simulation loop.

Source code in tit/opt/ex/engine.py

def __init__(
    self,
    leadfield_hdf: str,
    roi_file: str | tuple[str, int] | list[str | tuple[str, int]],
    roi_name: str,
    logger: logging.Logger,
):
    self.leadfield_hdf = leadfield_hdf
    self.roi_file = roi_file
    self.roi_name = roi_name
    self.logger = logger

    self.leadfield = None
    self.mesh = None
    self.idx_lf = None
    self.roi_coords = None
    self.roi_centers = None
    self.roi_indices = None
    self.roi_volumes = None
    self.gm_indices = None
    self.gm_volumes = None
    self._eval_subset = None

initialize

initialize(roi_radius: float = 3.0) -> None

Load leadfield, resolve the ROI (CSV/mask/atlas), find ROI + GM elements.

Source code in tit/opt/ex/engine.py

def initialize(self, roi_radius: float = 3.0) -> None:
    """Load leadfield, resolve the ROI (CSV/mask/atlas), find ROI + GM elements."""
    self._load_leadfield()
    self._load_roi_coordinates()
    self._find_roi_elements(roi_radius)
    self._find_gm_elements()
    self._eval_subset = None

compute_ti_fields

compute_ti_fields(e1_plus: str, e1_minus: str, e2_plus: str, e2_minus: str, current_ratios: list[tuple[float, float]]) -> list[dict[str, float]]

ROI metrics of one electrode montage at every current split.

The field is linear in the injected current, so each channel's unit-current field is gathered once and scaled per split -- the same current * (lf[a] - lf[b]) arithmetic TI.get_field performs, restricted to the elements the metrics read.

Source code in tit/opt/ex/engine.py

def compute_ti_fields(
    self,
    e1_plus: str,
    e1_minus: str,
    e2_plus: str,
    e2_minus: str,
    current_ratios: list[tuple[float, float]],
) -> list[dict[str, float]]:
    """ROI metrics of one electrode montage at every current split.

    The field is linear in the injected current, so each channel's
    unit-current field is gathered once and scaled per split -- the
    same ``current * (lf[a] - lf[b])`` arithmetic ``TI.get_field``
    performs, restricted to the elements the metrics read.
    """
    lf = self.leadfield
    idx = self.idx_lf
    subset, roi_pos, gm_pos = self._evaluation_subset()

    unit1 = TI.get_field([e1_plus, e1_minus, 1.0], lf, idx)[subset]
    unit2 = TI.get_field([e2_plus, e2_minus, 1.0], lf, idx)[subset]

    results = []
    for current_ch1_mA, current_ch2_mA in current_ratios:
        ef1 = (current_ch1_mA / 1000) * unit1
        ef2 = (current_ch2_mA / 1000) * unit2
        ti_max = envelope_max([ef1, ef2])
        data = self._roi_metrics(ti_max[roi_pos], ti_max[gm_pos])
        peak = hf_peak([ef1, ef2])
        data.update(
            self._carrier_metrics(
                peak[roi_pos], peak[gm_pos], data[f"{self.roi_name}_Focality"]
            )
        )
        data["current_ch1_mA"] = current_ch1_mA
        data["current_ch2_mA"] = current_ch2_mA
        results.append(data)
    return results

run

run(e1_plus: list[str], e1_minus: list[str], e2_plus: list[str], e2_minus: list[str], current_ratios: list[tuple[float, float]], all_combinations: bool, output_dir: str, n_jobs: int = 1, symmetry_mirror_map: dict[str, str] | None = None, symmetry_pairing: str = 'within_pairs') -> dict[str, dict[str, float]]

Run the full simulation loop. Returns {mesh_key: metrics}.

Candidates are evaluated in enumeration order, one electrode montage (all its current splits) per task, on n_jobs forked workers (n_jobs < 1: the job's CPU budget, see :func:tit.opt.ex.parallel.resolve_n_jobs; 1: in-process). A symmetry_mirror_map (bucket mode) restricts the enumeration to left/right mirrored montages (see :mod:tit.opt.ex.symmetry).

Signals are left to the caller: Ctrl-C raises KeyboardInterrupt and a job cancel's SIGTERM ends the process (tit/jobs/runner.py), so an interrupted search returns nothing and its run folder never gets the files of a finished run.

Source code in tit/opt/ex/engine.py

def run(
    self,
    e1_plus: list[str],
    e1_minus: list[str],
    e2_plus: list[str],
    e2_minus: list[str],
    current_ratios: list[tuple[float, float]],
    all_combinations: bool,
    output_dir: str,
    n_jobs: int = 1,
    symmetry_mirror_map: dict[str, str] | None = None,
    symmetry_pairing: str = "within_pairs",
) -> dict[str, dict[str, float]]:
    """Run the full simulation loop. Returns {mesh_key: metrics}.

    Candidates are evaluated in enumeration order, one electrode
    montage (all its current splits) per task, on ``n_jobs`` forked
    workers (``n_jobs < 1``: the job's CPU budget, see
    :func:`tit.opt.ex.parallel.resolve_n_jobs`; ``1``: in-process).
    A *symmetry_mirror_map* (bucket mode) restricts the enumeration to
    left/right mirrored montages (see :mod:`tit.opt.ex.symmetry`).

    Signals are left to the caller: Ctrl-C raises ``KeyboardInterrupt``
    and a job cancel's SIGTERM ends the process (``tit/jobs/runner.py``),
    so an interrupted search returns nothing and its run folder never
    gets the files of a finished run.
    """
    total = count_combinations(
        e1_plus,
        e1_minus,
        e2_plus,
        e2_minus,
        current_ratios,
        all_combinations,
        symmetry_mirror_map,
        symmetry_pairing,
    )
    n_jobs = resolve_n_jobs(n_jobs)
    self._log_config_summary(
        current_ratios,
        all_combinations,
        total,
        n_jobs,
        symmetry_pairing if symmetry_mirror_map is not None else None,
    )

    results: dict[str, dict[str, float]] = {}
    start_time = time.time()
    ratios = list(current_ratios)

    montages = list(
        _electrode_combinations(
            e1_plus,
            e1_minus,
            e2_plus,
            e2_minus,
            all_combinations,
            symmetry_mirror_map,
            symmetry_pairing,
            both_channel_orders=not ratios_mirror_symmetric(ratios),
        )
    )
    evaluations = evaluate_ordered(
        self,
        "compute_ti_fields",
        ((*montage, ratios) for montage in montages),
        n_jobs,
        n_tasks=len(montages),
    )

    i = 0
    for (ep1, em1, ep2, em2), montage_results in zip(montages, evaluations):
        montage_start = time.time()
        for (ch1, ch2), data in zip(ratios, montage_results):
            i += 1
            name = (
                f"{ep1}_{em1}_and_{ep2}_{em2}"
                f"_I1-{format_current_mA(ch1)}mA_I2-{format_current_mA(ch2)}mA"
            )
            key = f"TI_field_{name}.msh"

            elapsed = time.time() - start_time
            rate = i / elapsed if elapsed > 0 else 0
            eta = (total - i) / rate if rate > 0 else 0

            self.logger.info(f"[{i}/{total}] {name}")
            self.logger.info(
                f"  {100 * i / total:.1f}% | {rate:.2f}/s | ETA {eta / 60:.1f}min"
            )

            data["electrodes"] = (ep1, em1, ep2, em2)
            results[key] = data
            self.logger.info(
                f"  {(time.time() - montage_start) / len(ratios):.2f}s | "
                f"TImax={data[f'{self.roi_name}_TImax_ROI']:.4f} "
                f"TImean={data[f'{self.roi_name}_TImean_ROI']:.4f} "
                f"Foc={data[f'{self.roi_name}_Focality']:.4f}"
            )

    if results:
        t = time.time() - start_time
        self.logger.info(f"\n{'=' * 60}")
        self.logger.info(
            f"Done: {len(results)}/{total} in {t / 60:.1f}min "
            f"({t / len(results):.2f}s each)"
        )
        self.logger.info(f"Output: {output_dir}")

    return results