muniverse-dynamic-v2 / benchmark_spec.json
emgbenchanon's picture
Add benchmark_spec.json
4f76c35 verified
Raw
History Blame
6.56 kB
{
"name": "muniverse-dynamic-v2",
"version": "1.0",
"description": "Non-stationary dynamic EMG benchmark for motor unit decomposition. 336 recordings spanning 3 pool-size tiers (26-113 MUs), 5 subjects, 7 muscle types, 8 conditions (sinusoid/triangular angle × 2 difficulties × 2 SNR).",
"source_package": "muniverse",
"source_repo_url": "https://github.com/dfarinagroup/muniverse",
"generation_pipeline": [
"Step 1: Generate MUAP caches via NeuroMotion+BioMime container (one per subject × muscle at Flexion-Extension DOF, 130 angle steps)",
"Step 2: Analyze pairwise MUAP cosine similarity, select clean MU subsets via greedy filtering at given threshold",
"Step 3: Write clean caches (same format as original, fewer MUs)",
"Step 4: Generate recordings by running container on clean caches with specified conditions",
"Step 5: (bug fix) Regenerate triangular-angle recordings with symmetric triangle patch"
],
"muap_generation": {
"subjects": [
{"seed": 0, "fibre_density": 194.0, "muscle_motor_unit_counts": [206, 210, 181, 187, 141, 486, 154]},
{"seed": 1, "fibre_density": 187.0, "muscle_motor_unit_counts": [176, 176, 208, 167, 168, 340, 135]},
{"seed": 2, "fibre_density": 190.0, "muscle_motor_unit_counts": [186, 217, 164, 178, 153, 437, 192]},
{"seed": 3, "fibre_density": 174.0, "muscle_motor_unit_counts": [167, 224, 198, 186, 157, 420, 189]},
{"seed": 4, "fibre_density": 196.0, "muscle_motor_unit_counts": [194, 205, 198, 205, 187, 408, 163]}
],
"muscles": ["PL", "FDSI", "EDI", "ECU", "ECRB", "ECRL", "FCU_h", "FCU_u"],
"movement_dof": "Flexion-Extension",
"total_caches": 40,
"container_image": "pranavm19/muniverse-test:neuromotion",
"muscle_label_slot_notes": {
"FCU_u": "Replaces FCU slot in MuscleLabels list (custom)",
"FCU_h": "Replaces FCU slot in MuscleLabels list (custom)",
"standard_labels": ["ECRB", "ECRL", "ECU", "EDI", "PL", "FCU", "FDSI"]
}
},
"crosstalk_filtering": {
"method": "greedy_henneman_order",
"similarity": "cosine_similarity at mid-pose (step 65), flattened over channels × time",
"description": "Starting with MU 0 (smallest), add MU i if max(sim(i, j)) < threshold for all j already selected"
},
"combos": [
{"tier": "small", "subject_seed": 0, "muscle": "PL", "threshold": 0.85, "clean_N": 27},
{"tier": "small", "subject_seed": 2, "muscle": "PL", "threshold": 0.85, "clean_N": 26},
{"tier": "small", "subject_seed": 4, "muscle": "PL", "threshold": 0.85, "clean_N": 30},
{"tier": "small", "subject_seed": 4, "muscle": "FCU_h", "threshold": 0.85, "clean_N": 36},
{"tier": "small", "subject_seed": 3, "muscle": "ECRL", "threshold": 0.85, "clean_N": 39},
{"tier": "small", "subject_seed": 0, "muscle": "FDSI", "threshold": 0.85, "clean_N": 45},
{"tier": "medium", "subject_seed": 2, "muscle": "ECU", "threshold": 0.85, "clean_N": 47},
{"tier": "medium", "subject_seed": 3, "muscle": "FDSI", "threshold": 0.85, "clean_N": 49},
{"tier": "medium", "subject_seed": 2, "muscle": "FDSI", "threshold": 0.85, "clean_N": 53},
{"tier": "medium", "subject_seed": 3, "muscle": "ECU", "threshold": 0.85, "clean_N": 55},
{"tier": "medium", "subject_seed": 0, "muscle": "EDI", "threshold": 0.85, "clean_N": 56},
{"tier": "medium", "subject_seed": 3, "muscle": "EDI", "threshold": 0.85, "clean_N": 62},
{"tier": "medium", "subject_seed": 0, "muscle": "ECRB", "threshold": 0.85, "clean_N": 71},
{"tier": "medium", "subject_seed": 0, "muscle": "FCU_u", "threshold": 0.85, "clean_N": 74},
{"tier": "large", "subject_seed": 3, "muscle": "FCU_u", "threshold": 0.85, "clean_N": 77},
{"tier": "large", "subject_seed": 2, "muscle": "EDI", "threshold": 0.90, "clean_N": 80},
{"tier": "large", "subject_seed": 1, "muscle": "ECU", "threshold": 0.90, "clean_N": 84},
{"tier": "large", "subject_seed": 4, "muscle": "EDI", "threshold": 0.90, "clean_N": 93},
{"tier": "large", "subject_seed": 2, "muscle": "ECRB", "threshold": 0.90, "clean_N": 96},
{"tier": "large", "subject_seed": 3, "muscle": "FCU_u", "threshold": 0.90, "clean_N": 112},
{"tier": "large", "subject_seed": 1, "muscle": "FCU_u", "threshold": 0.90, "clean_N": 113}
],
"conditions": {
"SA05-hi": {"angle_profile": "Sinusoid", "af": 0.5, "sin_amplitude_deg": 32.5, "noise_level_db": 25},
"SA05-lo": {"angle_profile": "Sinusoid", "af": 0.5, "sin_amplitude_deg": 32.5, "noise_level_db": 20},
"SA10-hi": {"angle_profile": "Sinusoid", "af": 1.0, "sin_amplitude_deg": 65.0, "noise_level_db": 25},
"SA10-lo": {"angle_profile": "Sinusoid", "af": 1.0, "sin_amplitude_deg": 65.0, "noise_level_db": 20},
"TA05-hi": {"angle_profile": "Triangular", "af": 0.5, "sin_amplitude_deg": 32.5, "noise_level_db": 25},
"TA05-lo": {"angle_profile": "Triangular", "af": 0.5, "sin_amplitude_deg": 32.5, "noise_level_db": 20},
"TA10-hi": {"angle_profile": "Triangular", "af": 1.0, "sin_amplitude_deg": 65.0, "noise_level_db": 25},
"TA10-lo": {"angle_profile": "Triangular", "af": 1.0, "sin_amplitude_deg": 65.0, "noise_level_db": 20}
},
"fixed_recording_parameters": {
"effort_level_pct": 50,
"effort_profile": "Constant",
"sin_frequency_hz": 0.3,
"duration_s": 10,
"sampling_frequency_hz": 2048,
"target_angle_deg": 0,
"n_electrodes": 320,
"electrode_grid": "10 rows × 32 cols",
"inter_electrode_distance_mm": 8,
"filter": {"type": "Butterworth", "cutoff_hz": 800, "order": 4},
"noise_model": "amplitude-based additive Gaussian, std = std(emg) * 10^(-snr_db/20)"
},
"channel_configs": {
"ch070": {"n_cols": 7, "n_channels": 70, "description": "7 columns × 10 rows centered on max-RMS column"},
"ch320": {"n_cols": 32, "n_channels": 320, "description": "Full grid, no subsampling"}
},
"recording_count": {
"combos": 21,
"conditions_per_combo": 8,
"channel_configs": 2,
"total": 336
},
"data_format": "See BENCHMARK_DATA_FORMAT.md — npz with emg (T, M), spike_mu_{i} arrays, metadata scalars",
"known_issues_fixed": [
"Initial v2 generation had a bug in NeuroMotion's Triangular angle profile (TargetAngle=0, RampDuration=0 yielded flat 0° signal). Fixed via scripts/patches/neuromotion_triangular_symmetric.patch which replaces the Triangular branch with a symmetric triangle wave using SinAmplitude + SinFrequency (matching Sinusoid's range and period). 168 triangular recordings regenerated."
]
}