"""Modality registry: which file each stream lives in and which columns form the model input. The feature rules are the ones of the original benchmark pipeline, made explicit and strict: * ``emg`` 8 Myo channels ``emg_1..emg_8``. The Myo armband is worn on the RIGHT forearm by every volunteer. * ``imu`` 10 WitMotion sensors x 18 columns = 180 columns (WT0-WT7 on the fingers, WT8 right wrist, WT9 left wrist). A file with any other column count (volunteer v5: 72 columns, 4 sensors) does not provide the modality. * ``imu_wrist`` / ``imu_finger`` sensor subsets of ``imu`` (36 / 144 cols). * ``eyetrack`` the 24 Dikablis core gaze columns below. Recordings missing any of them (one-eye exports, e.g. v1s1 and v14s8) do not provide the modality. * ``eyetrack_gaze`` / ``eyetrack_ocular`` 2 scene-camera gaze columns / the 22 frame-independent ocular columns. * ``pressure`` 50 fingertip cells ``R1(g)..R25(g), L1(g)..L25(g)``. * ``mocap`` Qualisys skeleton TSV (``aligned_vXsY_s_Q.tsv``): every position/quaternion column, positions made hip-relative, plus the frame-to-frame velocity of the position columns. * ``mocap_markers`` the 100 Hz marker CSV ``aligned_mocap_100hz.csv`` (marker XYZ, ``Type`` columns dropped). * ``myo_quat`` Myo orientation quaternion (4 columns). Quality rules (applied to the selected feature matrix after non-numeric values and NaN/inf are mapped to 0): any ``|x| > 1e6`` is rejected as corrupted; for modalities other than ``emg`` and ``pressure`` (which are legitimately zero at rest) a zero fraction above 0.9 is rejected as dropout. """ from __future__ import annotations from dataclasses import dataclass, field from typing import Dict, Optional, Tuple #: Columns that are never features. SKIP_COLS = frozenset({"Frame", "Time", "time", "UTC"}) SKIP_COL_SUFFIXES = (" Type",) EMG_COLS: Tuple[str, ...] = tuple(f"emg_{i}" for i in range(1, 9)) PRESSURE_COLS: Tuple[str, ...] = tuple( [f"R{i}(g)" for i in range(1, 26)] + [f"L{i}(g)" for i in range(1, 26)]) MYO_QUAT_COLS: Tuple[str, ...] = tuple(f"quat_{i}" for i in range(1, 5)) IMU_N_SENSORS = 10 IMU_COLS_PER_SENSOR = 18 IMU_N_COLS = IMU_N_SENSORS * IMU_COLS_PER_SENSOR # 180 IMU_SENSOR_SUBSETS: Dict[str, Tuple[str, ...]] = { "imu_wrist": ("WT8_", "WT9_"), "imu_finger": tuple(f"WT{i}_" for i in range(8)), } EYETRACK_CORE_COLS: Tuple[str, ...] = ( "Dikablis Glasses 3_Eye Data_Original_Pupil X", "Dikablis Glasses 3_Eye Data_Original_Pupil Y", "Dikablis Glasses 3_Eye Data_Original_Left Eye_Pupil X", "Dikablis Glasses 3_Eye Data_Original_Left Eye_Pupil Y", "Dikablis Glasses 3_Eye Data_Original_Left Eye_Pupil Area", "Dikablis Glasses 3_Eye Data_Original_Left Eye_Pupil Height", "Dikablis Glasses 3_Eye Data_Original_Left Eye_Pupil Width", "Dikablis Glasses 3_Eye Data_Original_Left Eye_Fixations_Fixations", "Dikablis Glasses 3_Eye Data_Original_Left Eye_Fixations_Fixations Duration", "Dikablis Glasses 3_Eye Data_Original_Left Eye_Saccades_Saccades", "Dikablis Glasses 3_Eye Data_Original_Left Eye_Saccades_Saccades Duration", "Dikablis Glasses 3_Eye Data_Original_Left Eye_Saccades_Saccades Angle", "Dikablis Glasses 3_Eye Data_Original_Right Eye_Pupil X", "Dikablis Glasses 3_Eye Data_Original_Right Eye_Pupil Y", "Dikablis Glasses 3_Eye Data_Original_Right Eye_Pupil Area", "Dikablis Glasses 3_Eye Data_Original_Right Eye_Pupil Height", "Dikablis Glasses 3_Eye Data_Original_Right Eye_Pupil Width", "Dikablis Glasses 3_Eye Data_Original_Right Eye_Fixations_Fixations", "Dikablis Glasses 3_Eye Data_Original_Right Eye_Fixations_Fixations Duration", "Dikablis Glasses 3_Eye Data_Original_Right Eye_Saccades_Saccades", "Dikablis Glasses 3_Eye Data_Original_Right Eye_Saccades_Saccades Duration", "Dikablis Glasses 3_Eye Data_Original_Right Eye_Saccades_Saccades Angle", "Dikablis Glasses 3_Field Data_Scene Cam_Original_Gaze_Gaze X", "Dikablis Glasses 3_Field Data_Scene Cam_Original_Gaze_Gaze Y", ) EYETRACK_SUBSETS: Dict[str, Tuple[str, ...]] = { "eyetrack_gaze": ("Scene Cam_Original_Gaze",), # 2 columns "eyetrack_ocular": ("Eye Data_Original",), # 22 columns } #: Raw streams (one file each). ``mocap_skeleton`` is the TSV. STREAM_FILES: Dict[str, Optional[str]] = { "emg": "aligned_emg_100hz.csv", "imu": "aligned_imu_100hz.csv", "eyetrack": "aligned_eyetrack_100hz.csv", "pressure": "aligned_pressure_100hz.csv", "mocap_markers": "aligned_mocap_100hz.csv", "mocap_skeleton": None, # aligned_vXsY_s_Q.tsv in raw/ or at the recording root "myo_quat": "aligned_myo_quat_100hz.csv", "myo_pose": "aligned_myo_pose_100hz.csv", } #: Name of each stream in ``alignment_metadata.json`` (``aligned_modalities``). #: This is the authority for whether a file must exist: the skeleton TSV is #: ``mocap_sq``; ``mocap`` is the marker CSV. STREAM_ALIGNMENT_NAME: Dict[str, str] = { "emg": "emg", "imu": "imu", "eyetrack": "eyetrack", "pressure": "pressure", "mocap_markers": "mocap", "mocap_skeleton": "mocap_sq", "myo_quat": "myo_quat", "myo_pose": "myo_pose", } #: Name of a stream in metadata/recordings.csv ``aligned_modalities`` (that #: file has no entry for the skeleton TSV or the Myo orientation streams; #: its ``mocap`` is the marker CSV). STREAM_RECORDINGS_CSV_NAME: Dict[str, Optional[str]] = { "emg": "emg", "imu": "imu", "eyetrack": "eyetrack", "pressure": "pressure", "mocap_markers": "mocap", "mocap_skeleton": None, "myo_quat": None, "myo_pose": None, } @dataclass(frozen=True) class ModalitySpec: name: str stream: str # raw stream it is computed from expected_dim: Optional[int] # None: data dependent (mocap skeleton) zero_check: bool = True # apply the >90 % zeros dropout rule description: str = "" extra: Dict[str, object] = field(default_factory=dict) MODALITIES: Dict[str, ModalitySpec] = { "emg": ModalitySpec("emg", "emg", 8, zero_check=False, description="Myo sEMG, right forearm, 8 channels"), "imu": ModalitySpec("imu", "imu", IMU_N_COLS, description="10 IMUs x 18 columns (fingers WT0-7, wrists WT8/9)"), "imu_wrist": ModalitySpec("imu_wrist", "imu", 36, description="IMU WT8+WT9 only"), "imu_finger": ModalitySpec("imu_finger", "imu", 144, description="IMU WT0-WT7 only"), "eyetrack": ModalitySpec("eyetrack", "eyetrack", 24, description="24 Dikablis core columns"), "eyetrack_gaze": ModalitySpec("eyetrack_gaze", "eyetrack", 2, description="scene-camera gaze X/Y"), "eyetrack_ocular": ModalitySpec("eyetrack_ocular", "eyetrack", 22, description="pupil + fixation/saccade columns"), "pressure": ModalitySpec("pressure", "pressure", 50, zero_check=False, description="fingertip pressure, 25 cells per hand"), "mocap": ModalitySpec("mocap", "mocap_skeleton", None, description="skeleton TSV, hip-relative + position velocity"), "mocap_markers": ModalitySpec("mocap_markers", "mocap_markers", None, description="100 Hz marker CSV (XYZ)"), "myo_quat": ModalitySpec("myo_quat", "myo_quat", 4, description="Myo orientation"), } #: The five benchmark modalities in the paper's order. BENCHMARK_MODALITIES: Tuple[str, ...] = ("mocap", "emg", "eyetrack", "imu", "pressure") def get_spec(modality: str) -> ModalitySpec: try: return MODALITIES[modality] except KeyError: raise KeyError(f"unknown modality {modality!r}; known: {sorted(MODALITIES)}") from None def base_stream(modality: str) -> str: return get_spec(modality).stream