"""Read frozen WristHARBench fold exports (train/valid/test NPZ) for the package benchmarks. Only numeric arrays are read (``allow_pickle=False``); the global label vocabulary is taken from the fold's ``export_manifest.json``. ``X`` is converted to float64 exactly as the paper runners did, and train+validation are concatenated in that order. """ from __future__ import annotations import json from pathlib import Path import numpy as np STREAMS = ( "adl_wrist_accel_v1", "capture24_wearable_activity_v1", "domino_watch_accel_v1", "gotov_wrist_accel_v1", "handy_wrist_accel_v1", "harmes_left_wrist_accel_v1", "harmes_right_wrist_accel_v1", "iuwds_wrist_accel_v1", "mhealth_right_lower_arm_v1", "paal_adl_wrist_accel_v1", "pamap2_hand_imu_v1", "wisdm_watch_accel_v1", ) def load_part(root, dataset, fold, part): folder = Path(root) / dataset / f"fold{int(fold)}" manifest = json.loads((folder / "export_manifest.json").read_text()) with np.load(folder / f"{part}.npz", allow_pickle=False) as z: out = {"X": np.asarray(z["X"], dtype=np.float64), "y": np.asarray(z["y"], dtype=np.int64), "subject": np.asarray(z["subject"]), "time_index": np.asarray(z["time_index"])} out["label_names"] = list(manifest["label_names"]) out["n_classes"] = int(manifest["n_classes"]) return out def concat(a, b): out = {k: np.concatenate([a[k], b[k]], axis=0) for k in ("X", "y", "subject", "time_index")} out["label_names"], out["n_classes"] = a["label_names"], a["n_classes"] return out def load_fold(root, dataset, fold): """Return (train, valid, train+valid, test) dictionaries.""" train, valid, test = (load_part(root, dataset, fold, p) for p in ("train", "valid", "test")) return train, valid, concat(train, valid), test def paper_scores(y_true, y_pred, n_classes): """The frozen four-metric definitions used by the release (``wisp_release.scoring``).""" from wisp_release.scoring import score s = score(np.asarray(y_true, dtype=np.int64), np.asarray(y_pred, dtype=np.int64), int(n_classes)) return {k: s[k] for k in ("macro_f1", "accuracy", "balanced_accuracy", "worst_class_f1")}